Beyond activity counts: object-mediated participation in project-based learning is associated with outcomes at different levels of analysis

1 Introduction

1.1 Digital traces of collaborative project work

Contemporary professional practice requires well-developed collaboration skills, including the ability to work in distributed project teams (Chen et al., 2021). These skills are formed in purposefully organized learning formats that allow not only the outcome but the process of joint activity itself to be observed. A central format of this kind in higher education is team-based project-based learning (PBL), in which students jointly create a shared product. From the perspective of the activity-theoretical tradition, what matters for the formation of collaboration skills is precisely the fact of students working together on a shared educational object: activity that is not embedded in joint work does not develop these skills and may contribute nothing to the collective project outcome (Leont'ev, 1978; Rubtsov, 2024).

The digital transformation of higher education is changing the environments in which collaborative learning unfolds (Benavides et al., 2020). In current PBL models, students’ joint activity increasingly takes a distributed form, carried out in professional digital environments for software development and task management. Version control systems, wikis, and task management systems have become de facto organizational standards in distributed software development and engineering projects (Mockus et al., 2002; Wagstrom et al., 2010). The use of these platforms goes beyond technical support: they structure planning, the distribution of functions, the recording of decisions, and the alignment of actions—that is, they act as an organizational form of joint activity rather than merely a communication tool. This organizational function is of fundamental importance for higher education, and for engineering education in particular, because student project teams work in environments that approximate real professional practice (Patarakin and Vachkova, 2019; Cortés Ríos et al., 2019; Eraslan et al., 2020).

A key property of such environments is the stepwise recording of changes and versioning: participants’ actions leave observable digital traces—task creation and completion events, status changes, comments, version control records (commits), and document edits. As a result, collaborative project activity, previously accessible to instructors and researchers mainly through self-reports, peer assessment, observation, or the final product, becomes observable directly in the process of its performance (Mittal and Sureka, 2014; Saqr and Peeters, 2022; Feng et al., 2025). The transition from observability to pedagogical interpretation of digital traces as evidence of teamwork quality, however, is not automatic: a digital event records only the fact of an action in a system and carries no direct indication of its pedagogical meaning, its contribution to the shared product, or the quality of coordination (Wise and Shaffer, 2015; Winne, 2020; Gašević et al., 2015). The assessment of collaborative project activity still relies predominantly either on the final project outcome or on simple activity counts—numbers of actions, messages, tasks, or commits—which register the fact of activity but not the structure of collaborative participation. When joint activity is distributed across several platforms and unfolds over time, the interpretation problem becomes especially acute, creating the need for a theoretically grounded approach capable of reconstructing the structure of joint activity, the degree of students’ involvement in the shared objects of project work, and the dynamics of team interaction.

1.2 Related work

The development of learning analytics and collaboration analytics has created new opportunities for studying collaborative project activity from students’ digital traces. Unlike traditional observation and self-report methods, digital platforms enable researchers to record participants’ actual actions in the course of joint work and to reconstruct the structure of their interaction from the logs of educational and professional information systems (Martinez-Maldonado et al., 2021). Digital traces are becoming an important data source for studying collective activity in higher education, as they allow the analysis of both individual contributions and the patterns of interaction in solving shared tasks. The most common sources are learning management systems, collaborative software development platforms, and project management systems (Zamecnik et al., 2023; Kutuzov and Bogdanova, 2025). Below we review the main strands of research on digital traces of joint activity and the unresolved methodological problems associated with them.

Most studies of joint activity rely on the analysis of interaction networks between participants (social network analysis, SNA), reconstructed from communication—messages, comments, and other forms of direct information exchange (Kutuzov and Bogdanova, 2025; Saqr et al., 2018). The measures most frequently applied are degree, betweenness, closeness, and eigenvector centrality, which characterize students’ level of participation, their coordination functions, and their position in the interaction structure. The findings demonstrate an association between participants’ network position and educational outcomes: higher embeddedness in the interaction network is associated with success in group assignments and with the stability of team activity (Zamecnik et al., 2023; Saqr et al., 2020). This approach describes the structure of communication well, but it does not allow direct analysis of joint work on shared objects of activity.

Alongside static network analysis, a research strand has developed that focuses on the temporal dynamics of collaboration. The temporal approach takes into account not only the presence of interactions but also their sequence, intensity, and distribution over time (Saqr and Peeters, 2022; Knight et al., 2017). The analysis of temporal characteristics allows researchers to examine the stability of participation, changes in students’ roles, and the formation of group structures over the course of a project (Saqr et al., 2020; Reimann, 2009). Since project-based learning is inherently processual, reducing it to a static analysis of data aggregated over the whole period loses meaningful information about how the structure of joint activity formed and changed over time.

A further limitation is the underdevelopment of multi-platform approaches. When several tools are used simultaneously, the digital trace is distributed across environments, each of which captures its own layer of activity. Yet the integration of data from several environments is rare and requires data harmonization and participant identification (Jacobs and Castek, 2022). Even in studies where data are extracted from several environments, they are typically treated independently, and their integration amounts to a mechanical combination of separate metrics rather than the reconstruction of the entire structure of joint activity as a single network of interactions (Arthars et al., 2024). Moreover, many studies are based on single courses or limited samples, which makes it difficult to assess the robustness of the identified patterns (Zamecnik et al., 2023; Ouyang et al., 2023).

Despite the progress of network analysis of joint activity, the object-mediated mechanisms by which participants coordinate through shared digital artifacts remain understudied. The CSCL tradition and activity-theoretical research have long argued that collaboration is organized around shared objects rather than through direct communication alone (Stahl, 2006; Engeström, 2008), but this theoretical position has rarely been operationalized in trace-based analytics at scale. An open question is also whether there are effects that manifest differently at the individual and team levels of project activity: the same characteristics of digital participation may carry different substantive meaning depending on whether they are considered at the level of an individual student or of the team as a whole.

In sum, the current literature confirms the value of digital traces for studying collaborative learning activity while pointing to a set of unresolved problems: the development of multi-platform analytics; the extension of object-mediated models of collaboration, in which the object of joint activity is a digital artifact; the integration of temporal analysis of a participant’s structural position rather than of activity alone; and the differentiation of effects at the individual and team levels of project activity (Saqr and Peeters, 2022; Ouyang et al., 2023).

1.3 Research questions and contribution

The aim of this study is to identify the analytical potential of multi-platform digital traces for reconstructing students’ joint activity in project-based learning and to assess the association between the structure of students’ digital participation and educational outcomes. This shifts the analysis from individual activity to joint activity as a single system of interactions, in which the object of analysis is not students’ separate actions but the emerging structure of ties between team members, tasks, and digital artifacts. We posit that digital traces allow different facets of students’ participation in work on a shared product to be reconstructed: the configuration of traces on artifacts, network embeddedness, structural position, and the dynamics of participation over time. The final project grade is treated as an external educational indicator against which the identified characteristics of digital participation can be tested. The study is oriented not toward grade prediction as an end in itself, but toward analyzing which configurations of digital traces, network interaction, and temporal organization of participation correspond to different educational outcomes. A further motivation is practical. In the setting studied here, students and teachers already had access to a trace-based indicator—a weekly count of platform actions displayed on institutional dashboards—while no information about the structure or the development of participation was available to them. Whether such a count is an adequate proxy for participation in joint work is an empirical question, and it is the question this study addresses.

Based on the gaps identified above, the following research questions are addressed:

RQ1. How are the digital traces of collaborative project activity distributed across platforms, and what network structures and configurations do they form?

RQ2. How are object-mediated characteristics of individual participation—network embeddedness, isolation, and structural position—associated with educational outcomes, and what is their explanatory value compared with activity volume measures?

RQ3. Is the temporal organization of participation, including the dynamics of expansion of a student’s object-mediated position, associated with educational outcomes?

RQ4. Do the identified associations differ between the individual and team levels of project activity?

The study makes a threefold contribution. First, it proposes an object-mediated multi-platform model for integrating digital traces, which treats joint activity as a single network of ties between participants and shared artifacts rather than as a set of isolated per-platform indicators. Second, it separates the volume of digital activity from relational participation, showing that volume measures are associated with outcomes but that the structural configuration and dynamics of participation carry additional analytical information. Third, it identifies isolation within the observed trace network and the temporal expansion of a student’s object-mediated position as correlates of educational outcomes, with several associations manifesting differently at the individual and team levels. Practically, the approach yields a documented procedure for reconstructing students’ joint work in team projects and a vocabulary for describing participation patterns—isolation within the observed network, structural closure, absence of expansion—that activity counts cannot express, intended as prompts for formative enquiry rather than as an assessment instrument.

2 Materials and methods

2.1 Study design and context

The study follows a retrospective observational design based on digital traces of learning activity (Siemens, 2013; Winne, 2020). The empirical setting is the project-based learning model of the Moscow Institute of Electronics and Mathematics, HSE University (MIEM HSE), in which undergraduate and graduate students carry out team projects with mandatory documentation of their work in institutional digital environments. The unit of analysis is the student–project pair: the nature of participation of the same student may differ qualitatively across projects depending on team composition, task type, and the distribution of roles (Reimann, 2009). The observation period covers two academic years (2021/22 and 2022/23). All data were processed in de-identified form, with personal identifiers replaced by internal hash codes.

2.2 Data sources and integration

Four data sources were used: (1) final grade records of the project-based learning program; (2) digital logs of GitLab repositories¹; (3) digital logs of the Wekan task management system²; and (4) digital logs of the Trello task management system.³ Educational and digital data were matched using a user integration file that maps students’ institutional identifiers to their platform identifiers in each of the three digital environments.

2.3 Sample

The analytical sample was formed by a selection cascade (Figure 1). The grade records for the two academic years contained 1,800 student–project pairs in 554 projects; 16 pairs from staff-only records were removed (1,784 pairs in 549 projects). Because the object of analysis is team activity, 331 pairs in projects with fewer than three members according to the grade records were excluded (119 single-member and 212 two-member projects), leaving 1,453 pairs in 324 projects; one further pair without a final grade was removed (1,452 pairs). Of these, 838 pairs (57.7%) left no observable digital trace and 73 of the 324 projects had no trace-observable member, giving an analytical sample of 614 student–project pairs, 526 students and 251 projects. Observed and non-observed pairs do not differ in the final grade (δ = −0.018, p = 0.54) or in any grade component; observability is related to the examination board, the study programme, the academic year and team size, and the comparison is unchanged after adjustment for these covariates. Conclusions therefore apply to students whose participation was observable on the instrumented platforms.

Within the analytical sample, projects differ in the number of trace-observable members (n_obs, range 1–8; Mdn = 2, M = 2.45): 68 projects (27.1%) have a single observable member, 77 have two, and 106 have three or more. Because several network measures are undefined or degenerate at low observability, analyses use nested subsamples of increasing strictness; their definitions and sizes are given in Section 2.9 and Table 1.

Table 1. Analytic subsamples: Definitions, sizes and the measures computed on each.

Subsamples are nested as E ⊂ C ⊂ B ⊂ A and F, G ⊂ A; H1, H2 and D partition A. Counts are for the first-attempt outcome. Each measure is defined in Supplementary material.

The sample is highly heterogeneous in content, spanning web development, data analysis tools, computer vision, AR/VR, and hardware–software systems. This heterogeneity ensured that a variety of collaboration models were represented and allowed conclusions about the organization of joint activity that are independent of the narrow subject specifics of particular tasks.

2.4 Event-to-project attribution

Each digital event in the logs was attributed to a specific educational project only when four conditions were met simultaneously: (1) student—the event was performed by a user identified as a member of the project team; (2) platform space—the event occurred in a repository, board, or card container technically assigned to the project; (3) project time window—the event timestamp falls within the project period according to the grade records; and (4) team membership—at the time of the event, the student was a member of the project team according to the educational records.

A strict rule was applied: a single digital event of a student was attributed to at most one project. Where an event could potentially match several projects simultaneously (e.g., parallel participation with overlapping time windows), it was excluded, preventing artificial duplication of activity. Under this rule, 20 ambiguous event–student records were removed from the source log.

The final event log contains 14,289 digital events unambiguously attributed to a specific student and educational project. For each event, the student and project identifiers, platform, action type, timestamp, and digital artifact identifier were retained. The distribution of events across platforms and the structure of the object-mediated networks (numbers of edges and artifacts) are reported in Section 3.1.

2.5 Operationalization of digital traces

A digital artifact was defined as a discrete object of the project’s digital environment on which a student action can be recorded: a task card in Wekan or Trello, a board, a GitLab repository, an individual commit, or a file in a repository. Artifact type was determined by the technical nature of the object and the platform; for the purposes of network analysis, all artifact types were treated as homogeneous second-mode nodes in a bipartite network.

For each student–project pair, three groups of measures were computed. Activity volume measures characterize the intensity and persistence of the student’s own activity: the number of digital events, and the numbers of calendar days and of calendar weeks with at least one event. Configuration measures characterize the breadth of digital presence: the number of unique artifacts bearing the student’s traces, the mean number of actions per artifact, and the number of shared artifacts (those also bearing traces of at least one other team member).

The relative contribution measure (participation z-score) is the student’s total number of events standardized against the distribution of event counts across the members of their team (formula in the Supplementary material). Within-project standardization ensures cross-project comparability of a student’s relative standing; higher values correspond to a larger relative contribution. The z-score is undefined for the sole observable member of a project; such pairs are excluded from the corresponding correlation analyses, whereas in the regression models these values were set to 0 (the team mean), with a sensitivity analysis treating them as undefined (Section 2.9).

2.6 Object-mediated network construction

The methodological framework is the object-mediated approach, rooted in activity theory and developed in the computer-supported collaborative learning (CSCL) literature and the sociology of practice (Engeström, 2008; Stahl, 2006; Styhre, 2003): the coordination of team members is reconstructed through their joint work with shared digital artifacts rather than through direct communication alone. Joint engagement of several students with the same artifacts is interpreted as a manifestation of their involvement in shared activity (Patarakin and Vachkova, 2019).

For each project, a bipartite student–artifact network was constructed (an edge was recorded if the student performed at least one action on the artifact; edges are binary). The bipartite network was projected onto a unipartite object-mediated network of students: two students are joined by a tie if both left events on at least one common artifact. A tie therefore records observed co-use of an artifact, not an observed interaction; throughout the paper an edge denotes a student–artifact link in the bipartite network and a tie denotes a student–student link in the projection; the terms used for the trace data and the derived networks are defined in Supplementary Table S5. Following the logic of a weighted projection (Newman, 2001), the contribution of each artifact to a tie was set inversely proportional to the number of contributor pairs on that artifact, which downweights “large” artifacts and prevents inflation of projection density (Newman, 2001; Borgatti and Halgin, 2011). Formally, an artifact touched by k students contributes 1/C(k,2) to each of the C(k,2) ties it induces. Results are insensitive to this choice: under Newman’s 1/(k − 1) weighting and under an unweighted projection the tie-strength and structural measures retain their sign, magnitude and robustness status (Burt constraint ρ = −0.190, −0.209 and −0.224 respectively; within-project strength rank on the rescaled scale +0.187, +0.180 and +0.153). The projection formula is given in the Supplementary material. Four terms are used consistently and are not interchangeable: participation denotes a student’s presence and position in the observed trace network; joint work denotes activity on artifacts that also carry events of teammates; embeddedness denotes a student’s relative position among teammates within the same project; and contribution denotes what a student’s work changed inside a shared artifact, which is not measured in this study.

Based on the object-mediated network, the following measures were computed for each student–project pair: degree (number of partners)—the number of other team members with whom the student shares at least one edge in the projection, i.e., at least one shared artifact; tie strength—the sum of the weights of the student’s ties, reflecting the intensity of observed co-use of artifacts with teammates; isolation—a binary indicator of the absence of object-mediated ties in the project, that is, a situation in which the student left events but none of the artifacts they touched also carried events of another team member; and Burt constraint and effective size—measures of the structural non-redundancy of ties (Burt, 1992). Burt constraint is high when a student’s ties are concentrated in a dense subgroup of closely interconnected partners, and low when the partners are weakly connected to each other and span different parts of the team; effective size is the effective number of “independent” partners in the ego network. Low constraint and high effective size reflect the same structural openness of the ego network, so a negative association of constraint with the grade and a positive association of effective size are consistent. Operational formulas are provided in the Supplementary material.

Burt’s measures were computed in NetworkX 3.2⁴ on the weighted network. Because they degenerate to trivial values at degree = 1, their correlation analysis was performed on the subsample of projects with at least three trace-observable members and pairs with at least two object-mediated ties—324 student–project pairs in 90 projects. At the team level, the density of the object-mediated network was computed for each project under two denominators, reported as alternative measures rather than as a correct and an incorrect one: the number of trace-observable members, which reflects the observed structure, and the nominal team size from the grade records, which reflects the official team. The proportion of observable members in each team (median 0.667) is reported alongside both, and the association of density with the outcome is additionally estimated with that proportion partialled out.

2.7 Temporal analysis

For the analysis of dynamics, project time was divided into consecutive windows of fixed width counted from the first recorded event of the project rather than from the academic calendar, which ensures cross-project comparability of trajectories. A 7-day step is used in the primary specification; it reduces sensitivity to short-term fluctuations in activity and increases the stability of network measure estimates when the number of events per window is small. Widths of 5, 10 and 14 days are reported as sensitivity analyses.

For each window, a cumulative object-mediated network was built from all project events from the start of the project window through the end of that window inclusive. Windows are indexed by elapsed time from the start of the project window, and windows containing no events are retained rather than dropped, so that the index measures elapsed time rather than the order of active periods. The project window opens at the first event recorded in the project across all observable members, since official project start dates are not available in the source records, and closes either at the last recorded event of the project (activity anchor) or at the defence date taken from the grade sheet (deadline anchor); both anchors are reported for the shares of late activity.

To quantify the temporal trend of degree, the slope of its linear regression on the calendar-window index was computed, so that the coefficient is expressed in partners per week of elapsed time rather than per observed window. The slope is defined for pairs with at least three calendar windows in the project window (subsample F). Because cumulative degree is monotone non-decreasing and starts at zero for every pair by construction, a baseline degree cannot be controlled; observable team size and project duration are controlled instead. A flat slope is decomposed into persistently isolated students and students connected early and stable, since the two are otherwise conflated. A tie-formation timing model, in which the normalized time of each new tie is recorded, was estimated as an alternative specification (rules for handling gaps are given in the Supplementary material). The analysis of the slope was restricted to pairs with at least three calendar windows in the project window (subsample F: 534 pairs from 205 projects).

For each pair, the following additional temporal measures were computed: share_last_7 and share_last_14—the share of the student’s events falling within the last 7 and 14 days before the end of the project window, computed once under each closing anchor; temporal entropy—the normalized entropy of the distribution of the student’s events across 7-day windows, ranging from 0 (all activity concentrated in a single window) to 1 (activity distributed uniformly); and regularity—the share of the project’s 7-day windows in which the student performed at least one action. These measures are undefined when the project’s observed window is degenerate (all project events share a single timestamp); the corresponding analyses use 603 pairs from 240 projects.

2.8 Educational outcome

The external criterion was the student’s project grade on a 10-point scale, computed as 0.1 × supervisor score + 0.2 × documentation-upload score + 0.7 × examination board score. The documentation-upload component is itself an assessment of a trace event, so 20% of the composite is definitionally related to platform activity. To address this, the first-attempt composite is used as the primary outcome and two alternatives are reported throughout: the examination board component alone, and the maximum composite across assessment attempts as used in the earlier specification. All reported associations hold on the three outcomes; the decomposition reduces but does not remove the dependence, since the examination board could consult the institute’s activity dashboards. Those dashboards displayed a single quantity—the number of platform actions per week—and did not display any of the network, structural or temporal measures analysed here, which were computed only for this study. The residual dependence therefore concerns the volume measures; the network-position and temporal measures were not visible to students or to graders at any point. The retake flag was retained separately and used as a robustness control.

2.9 Statistical analysis

All computations were performed in Python 3.12⁵ using NumPy 1.26, SciPy 1.13, pandas 2.2, NetworkX 3.2, and statsmodels 0.14.⁶ The significance level was α = 0.05.

The association of each measure with the first-attempt grade was estimated with Spearman’s rank correlation. The non-parametric method was chosen because of non-normal distributions and the discrete nature of the dependent variable. Given the hierarchical structure of the data, with student–project pairs nested in teams (on average 2.4 observable members per project) and 80 students appearing in more than one project, the reported correlations are descriptive associations, and 95% confidence intervals were computed by a bootstrap that resamples whole projects (B = 2,500 resamples; fixed seed 42): whole projects, with all their students, were resampled with replacement, preserving the nested structure of the data. Because the data are dependent along two dimensions, intervals were additionally computed with multiway cluster-robust inference over projects and students (Cameron et al., 2011), which raised standard errors by 3.9% on average and changed no conclusion; a sensitivity check repeatedly drawing one project per student reproduced every estimate within ±0.02. The interval, rather than the p-value, is the criterion by which an association is described as robust; where the two criteria diverge, both are reported. The same cluster resampling was applied to the bootstrap intervals for Cliff’s δ; for project-level measures (network density), pair and cluster resampling coincide. Clustering is additionally addressed in the regression models through cluster-robust standard errors and within-project estimation. To control the false discovery rate under multiple testing of correlations, the Benjamini–Hochberg procedure was applied at q = 0.05 (Benjamini and Hochberg, 1995) within three families corresponding to the three research questions: volume and configuration measures (seven tests), network-position measures (five tests), and temporal measures (seven tests). Adjusted q-values are reported alongside raw p-values, and the composition of each family is stated in the note to the corresponding table. Because different measures are mathematically defined under different observability conditions, analyses use nested subsamples: A—all 614 pairs in 251 projects; B—projects with at least two observable members (546 pairs, 183 projects); C—at least three observable members (392 pairs, 106 projects); D—non-isolates (466 pairs, 158 projects); E—the Burt-valid subsample with at least three observable members and degree ≥ 2 (324 pairs, 90 projects); F—pairs for which the calendar-week slope is defined, that is with at least three calendar weeks in the project window (534 pairs, 205 projects); and G—projects with within-project variation in the grade (337 pairs, 105 projects). Two further groups are used in the analysis of isolation: H1 – sole trace-observable members (68 pairs) and H2—students isolated within the observed network (80 pairs); H1, H2 and D partition subsample A. Each measure is defined in Supplementary Table S1 with its formula, unit, range, level of analysis, required subsample and interpretation, and the subsample definitions with the metric-to-subsample mapping are given in Table 1. Descriptive statistics for grades and the main measures by academic year, observability group and dominant platform are reported in Supplementary Table S2.

Distributions of grades between groups (isolates vs. connected students; slope = 0 vs. slope > 0) were compared with the Mann–Whitney U test and Cliff’s δ effect size (Cliff, 1993) with 95% cluster-bootstrap intervals; two-sided tests are used throughout, including in the slope analysis, since no one-sided alternative was prespecified. Values of δ were interpreted following Cliff’s conventions (0.15, small, 0.33, medium, 0.47, large).

To separate the between-project and within-project components of the association of the slope and the share of late activity with the grade, a two-level decomposition was applied: at the between-project level, the Spearman correlation of project-mean values of the measure with project-mean grades; at the within-project level, the association of deviations of the measure and the grade from team means (Pearson correlation on deviations, and regression of demeaned variables with project fixed effects and cluster-robust standard errors; Cameron and Miller, 2015). The three estimates answer related but different questions and are reported side by side rather than interpreted as one effect. All three levels were estimated on a unified subsample—pairs for which the measure is defined, in projects with at least two observable members. To assess the association of a student’s object-mediated position with their relative standing within the team, within-project rank correlation was used: in each project, students were ranked by degree, by tie strength and by grade using mean ranks, and the ranks were rescaled to the unit interval as (r − 1)/(n − 1) before pooling. Rescaling is necessary because raw ranks from projects of different size are not comparable: pooling them allows project size to enter both variables, which inflates the association. Because ranks degenerate with a single observable member, the primary subsample comprises projects with at least three observable members, and the number of projects in which degree actually varies is reported alongside the pooled estimate.

To assess the joint and incremental explanatory value of the participation dimensions, hierarchical (block) regressions of the first-attempt grade were estimated. Predictors were entered in blocks: T0—control variables (study year, academic year, team size, project duration); T1—activity and coverage (number of active days, number of unique artifacts, participation z-score); T2—object-mediated network position (degree, isolation indicator); T3—temporal organization (share of events in the last 14 days, duration of the active period, temporal entropy, growth of degree). Variables are entered on their natural scales except for the binary academic-year and isolation indicators; standardized coefficients are reported alongside unstandardized ones. The significance of the increment in explained variance at each block was assessed with cluster-robust Wald tests rather than ordinary F-tests. Because the first-attempt grade has a pronounced between-project component (intraclass correlation ICC = 0.697), the model was estimated in two variants: Model A (pooled OLS on the analytic sample with cluster-robust standard errors at the project level) and Model B (within estimation with variables centered on project means, restricted to projects with within-project grade differentiation). Pairs with a complete set of predictors were used—459 of 614 pairs in 150 projects; the loss is driven by the participation z-score, which is undefined for sole observable members and for teams with zero within-team variance in event counts, and by the calendar-week slope, which requires at least three windows. As a sensitivity analysis, all regression models were re-estimated under a stricter missingness policy in which the participation z-score is undefined for projects with a single observable member and temporal measures are undefined for degenerate project windows (N = 459 pairs, 150 projects; within model: 291 pairs, 88 projects).

3 Results

3.1 Multi-platform structure of the trace corpus

The digital traces of collaborative project work are distinctly multi-platform in character and distributed across several environments, each capturing a different facet of teamwork. The distribution of the 14,289 events across platforms proved substantially uneven (Figure 2).

The largest volume of traces was recorded in Wekan—9,522 events (66.6%), followed by GitLab—4,521 (31.6%) and Trello—246 (1.7%). This distribution reflects the different roles of the platforms: GitLab captures work with software artifacts and the codebase (commits, file changes), whereas Wekan and Trello capture organizational and coordination activity (creating and assigning tasks, changing statuses). Observed project activity is not localized in a single environment: some participants left events predominantly in task management systems, others in code repositories, and a substantial share of activity emerged from the intersection of traces across several platforms—a pattern characteristic of engineering project-based learning, where organizational coordination and product work form interrelated but distinct layers. The 14,289 events involved 7,214 unique artifacts and formed 9,781 student–artifact edges; only 1,717 artifacts (23.8%) were shared. Sharedness, therefore, does not arise automatically from the presence of actions; it is formed through the intersection of traces on common objects.

3.2 Volume and configuration of digital traces

The count of digital actions does not fully characterize a student’s participation: two students with comparable activity volumes may differ fundamentally in the structure of their participation. Alongside volume, we therefore analyzed the configuration of a student’s trace—the distribution of that student’s events across artifacts and the extent to which those artifacts also carry events of other team members. At broadly similar activity volumes, the numbers of touched and shared artifacts differ substantially, reflecting different participation strategies—from predominantly individualized work (many actions with a minimum of shared artifacts) to pronounced engagement with common objects (Figure 3).

The relative contribution measure (participation z-score) complements this picture by capturing the student’s standing in the distribution of team activity. Together, these measures shift the analysis from the quantity of actions to the structure of digital participation. Correlation analysis of volume and configuration measures with the final grade quantifies this heterogeneity (Table 2).

Table 2. Association of the volume and configuration of a student’s digital trace with the first-attempt project grade.

Spearman’s ρ; q-values from Benjamini–Hochberg FDR correction applied within the family of volume and configuration measures (seven tests; see Section 2.9). 95% CIs from a bootstrap resampling whole projects, B = 2,500, fixed seed 42. The participation z-score is undefined for sole observable members and for teams with zero within-team variance in event counts (N = 518 of 546 pairs in projects with two or more observable members; 28 pairs have zero within-team standard deviation).

Four of the seven associations are robust under the interval criterion: the volume measures (total events, active days, unique artifacts, participation z-score) retain intervals excluding zero, whereas the three weakest—the number of shared artifacts, object-mediated tie strength and the share of a student’s artifacts that are shared—have intervals crossing zero and do not pass FDR correction either. The last of these is of separate interest: it varies independently of how much a student does (ρ = +0.019 with total events, p = 0.66) yet is not associated with the outcome, so the two dimensions are measurably distinct without one being a better profile than the other (Figure 3). The largest effect (ρ = +0.222 for the total number of events) is small by conventional benchmarks (Cohen, 1988), supporting the central claim: activity volume alone does not explain the outcome and must be complemented by an analysis of the structure and dynamics of participation. The associations replicate on the subsample of projects with at least two observable members (total events: ρ = +0.234; shared artifacts: ρ = +0.100). Values on subsample B are reported for comparability with the measures that require at least two observable members.

3.3 Object-mediated connectedness

We first assess how a student’s position in the team’s object-mediated network relates to their relative standing within the team, using within-project rank correlation. Ranks are computed separately in each project and rescaled to the unit interval as (r − 1)/(n − 1), so that projects of different size enter on a common scale; pooling raw ranks would allow project size to enter both variables and inflate the association (Section 2.9). On the strict subsample (392 pairs from 106 projects with at least three observable members), the rescaled rank by number of partners is positively associated with the rescaled rank by grade [ρ = +0.160; 95% CI (+0.056; +0.264); q = 0.002], as is the rescaled rank by tie strength [ρ = +0.185; 95% CI (+0.060; +0.310); q = 0.001]. Both associations are small; comparable estimates for the volume measures on the same subsample and the same scale are ρ = +0.280 for the number of artifacts and +0.237 for total events. Within-project ranks carry information only where degree varies among team members. In 84 of the 106 projects the observable members share an identical degree—at a median team size of three the projection is typically either complete or empty—and 18 projects vary in both degree and grade. On those 18 projects the pooled rescaled association is ρ = +0.385 [95% CI (+0.175; +0.596)], and the mean of the project-specific correlations is +0.392 [95% CI (+0.145; +0.597)], positive in 14 of 18 projects. Ties were resolved by mean ranks; 94.4% of observations lie in ties on degree, so the number of informative projects must be reported alongside the pooled estimate.

3.3.1 Isolation within the observed trace network

Of the 614 student–project pairs, 148 (24.1%) had no object-mediated tie despite leaving digital events; the remaining 466 (75.9%) shared at least one artifact with a teammate. These 148 divide into two groups that behave very differently (Figure 4). Sole observed team members (68 pairs) were the only trace-observable member of their project, so the absence of ties reflects instrumentation coverage rather than their work. Students isolated within the observed trace network (80 pairs) worked in projects with other observable members but shared no artifact with any of them.

Sole observed team members are indistinguishable from connected students [M = 6.78 vs. 6.91; δ = −0.036; 95% CI (−0.195; +0.120); p = 0.63], whereas students isolated within the observed network score markedly lower [M = 5.47; δ = −0.342; 95% CI (−0.535; −0.134); p < 0.001]. The distinction is diagnostic rather than cosmetic: only the second group carries information about participation, and the first is an artefact of coverage. An alternative reading is that these students specialised on a self-contained module rather than being less involved in shared artifact use. Their task profile is consistent with narrower work—a median of 3 task categories against 4 among connected students—but they are almost absent from coordination tasks (0.8% of their card events against 6.7%), which a legitimate division of labour would not predict. Specialisation and disengagement are not separable in these data, and the label refers to position in the observed trace network rather than to the substance of a contribution. Both estimates use project-clustered bootstrap intervals. The two groups also differ from each other directly [δ = −0.314; 95% CI (−0.520; −0.073); p = 0.001], confirming that the decomposition captures a distinction in observed participation rather than a purely technical one. Digital activity per se does not guarantee observable participation around shared team artifacts.

3.3.2 Structural position: brokerage and closure

The presence or absence of object-mediated ties reflects only the basic level of a student’s involvement in joint activity. Students differ not only in whether they have ties but also in how those ties are structurally organized. In some cases, interactions are concentrated within a limited circle of participants forming a dense local group. In others, the student links different parts of the project structure, acting as a broker between several strands of joint activity. Figure 5 presents typical structural positions of students in the object-mediated network.

Two measures from Burt’s structural holes framework were used to assess the structural organization of ties. Burt constraint captures the extent to which a student is confined to a narrow circle of closely interconnected partners: a high value means all of the student’s ties are concentrated in one dense subgroup whose members also work actively with each other; a low value means the partners are weakly connected among themselves. Effective size captures how many non-redundant partners the student is actually tied to—that is, how many of their ties lead to distinct, non-overlapping participants. The combination of low Burt constraint and high effective size indicates that the student’s work links different parts of the team rather than being confined to one dense subgroup. The analysis was performed on the subsample of students with at least two object-mediated ties in projects with at least three observable members (N = 324).

As Table 3 shows, Burt constraint showed a robust negative association with the grade [ρ = −0.190; 95% CI (−0.321; −0.059); q = 0.001], whereas the positive association for effective size was suggestive: it passes FDR correction but its interval crosses zero [ρ = +0.134; 95% CI (−0.022; +0.289); q = 0.016]. Higher grades were observed among students with a less closed structural position, whose object-mediated ties span different parts of the project network rather than concentrating within a limited circle. The same result is reproduced for the derived brokerage index (1 − Burt constraint; ρ = +0.190; q = 0.001). For display, panel D of Figure 5 groups students into terciles of Burt constraint (median constraint 0.79, 0.94 and 1.21), whose mean first-attempt grades are 7.44, 6.83 and 6.46; the estimate itself is computed on the continuous measure. The Burt constraint association survives adjustment for activity volume and ego-network size [partial ρ = −0.133; 95% CI (−0.240; −0.024)], so it is not reducible to activity volume, although it is not independent of ego-network size. On the broader subsample of all non-isolates (N = 466) the association retains sign and significance (ρ = −0.144; q = 0.003), whereas effective size weakens to borderline (ρ = +0.087; q = 0.061). The central conclusion is therefore formulated in terms of structural closure. Pedagogically, this means that the capacity to link different subgroups of the team (a brokerage position) is associated with higher outcomes and is not reducible to activity volume.

Table 3. Association of structural network characteristics with the first-attempt project grade (Burt-valid subsample E, N = 324 pairs, 90 projects).

Spearman’s ρ with BH-FDR correction applied within the family of network-position measures, which comprises five tests: degree, the rescaled within-project ranks of degree and of tie strength, reported in Section 3.3, and the two measures in this Table. 95% CIs from a bootstrap resampling whole projects and students (two-way), B = 2,500, fixed seed 42.

3.3.3 Team-level network density

In line with RQ4, beyond individual position (RQ2) we analyzed tie structure at the team level—the density of the object-mediated network, i.e., the share of actually observed ties between members through shared artifacts relative to the maximum possible (see Table 4).

Table 4. Association of object-mediated team network density with project outcomes.

Density is reported under both denominators. With trace-observable members in the denominator the association is positive; with the nominal team size it is not distinguishable from zero (ρ = +0.083 and +0.046, both n.s.). The proportion of observable members is itself unrelated to the outcome (ρ = −0.024, p = 0.74), and the association with observable-member density holds when that proportion is partialled out (ρ = +0.177, p = 0.017), so it is not an artefact of coverage. Neither denominator is treated as inherently correct: the first describes integration among the students whose work is visible, the second integration relative to the official roster. Project outcome is the team mean of the first-attempt grade; the association is of similar size for the median (ρ = +0.207) and strongest for the lowest grade in the team (ρ = +0.251).

Teams with denser observed artifact co-use tended to have higher project outcomes. The association is positive both for the sample with two or more observable members and for the strict subsample (n_obs ≥ 3), although in the latter the lower bound of the confidence interval virtually touches zero [(−0.000; +0.398)], so the team-level association should be regarded as moderate and less robust. The density of the object-mediated network characterizes the degree of observed integration through shared digital objects and is consistent with the view of members’ involvement in common work as a distinct factor of effectiveness.

3.4 Temporal organization of participation

3.4.1 Distribution of activity over time

The distribution of digital events over time is substantially uneven: for some students activity is spread across the entire project, while for others a large share of actions is concentrated in the final stages. Temporal measures are computed on calendar windows. The project window opens with the first event recorded in the project and closes either at the last recorded event (activity anchor) or at the defence date taken from the grade sheet (deadline anchor). Official project start dates are not available in the source records, so the opening anchor is later than the true start by an unknown margin; this is stated as a limitation. Empty calendar windows are retained and slopes are computed against elapsed time rather than event order. The two closing anchors answer different questions: teams set their own pace and some completed the work well before the examination session, so concentration relative to a team’s own completion is not the same as lateness relative to the institutional deadline. To assess temporal organisation we computed the number of active days, the duration of the active period, normalised temporal entropy, regularity, and the shares of events in the last 7 and 14 days under both anchors (Table 5).

Table 5. Association of temporal participation characteristics with the first-attempt project grade (family of seven tests).

3.4.2 Expansion of object-mediated position

Beyond the overall distribution of activity over time, we analysed how a student’s object-mediated position changed over the course of the project. For some students the number of object-mediated partners grew steadily; for others it stabilised early or never rose above zero. Because cumulative degree is monotone non-decreasing by construction, a flat trajectory conflates two different situations, and these are separated below. Examples of typical trajectories are shown in Figure 6.

For quantification we used the linear trend (slope) of cumulative object-mediated degree on calendar weeks, reflecting the rate at which a student’s structural position expands. The measure is defined for pairs with at least three calendar weeks in the project window (subsample F: 534 pairs in 205 projects). The association with the grade is positive [ρ = +0.182; 95% cluster-bootstrap CI (+0.062; +0.301); q < 0.001] and stable across window widths of 5, 10 and 14 days. It holds under a two-sided test with controls for observable team size and project duration [B = +3.68; 95% CI (+0.90; +6.46); p = 0.009]; baseline degree cannot be controlled, since cumulative degree starts at zero for every pair by construction. A tie-formation timing model, considered as an alternative specification, agrees in sign and magnitude—the mean normalised time at which ties appear is associated with the grade at ρ = +0.224—but is undefined for the 68 pairs that formed no tie, so the slope is retained as the primary measure. Table 6 characterises the trajectory groups.

Table 6. Outcome by rate of expansion of the object-mediated position (subsample F: N = 534 pairs, 205 projects).

Negative slopes do not occur by construction of the cumulative scheme. Retakes are too few for group comparison (24 of 534). The three expansion tertiles do not differ from one another, so the association is a threshold rather than a dose–response gradient.

Students who kept opening their circle of object-mediated ties over the project achieved, on average, higher grades than students whose position plateaued (M = 7.13 vs. 6.33; Figure 7). The association does not strengthen with the rate of expansion: each tertile of positive slope differs from the plateau group by a comparable margin (δ = +0.229, +0.217 and +0.215 respectively; panel B of Figure 7), and the tertiles do not differ among themselves. After controlling for total activity volume, number of artifacts, number of active days, and number of shared artifacts [partial rank correlation of slope with the grade on the unified decomposition subsample of 490 pairs from 161 projects: ρ = +0.154; 95% cluster-bootstrap CI (+0.026; +0.282); p = 0.001], which allows the expansion dynamics of the object-mediated position to be treated as a distinct characteristic of participation, not reducible to network position or activity volume.

3.4.3 Individual vs. team levels

The rate of expansion of the object-mediated position was examined separately at the team level and at the level of students within teams, on the unified decomposition subsample (490 pairs from 161 projects; Section 2.9). At the team level, no significant association with mean project outcomes was found (ρ = +0.142; p = 0.073): teams with faster average tie accumulation did not receive higher grades. At the within-project level, the picture differs—students who kept increasing their number of object-mediated partners over the project received higher grades than their teammates (r = +0.328; p < 0.001). The direction is preserved in the model with project fixed effects and cluster-robust errors (B = +8.49; SE = 1.53; p < 0.001).

The expansion of structural position therefore manifests predominantly at the individual level: across teams, the average rate of tie accumulation is unrelated to the outcome, whereas within a team students who keep opening their circle of shared objects receive higher grades. A flat trajectory separates into persistently isolated students (n = 113, M = 5.96) and students connected early and stable (n = 148, M = 6.61), against students expanding later (n = 273, M = 7.13); Kruskal–Wallis H = 23.54, p < 0.001. Among students who expanded at all, the rate of expansion does not differentiate outcomes (H = 0.27, p = 0.87), so the association is a threshold rather than a dose–response gradient.

3.4.4 A possible sign reversal across levels

Behind the bivariate negative association of the share of events in the last 14 days with the final grade (ρ = −0.134 under the activity anchor; Table 5) lie associations of opposite direction at different levels of analysis. As with the linear trend (slope), a level decomposition was performed for the share of activity in the last 14 days of the project (Table 7).

Table 7. Multilevel decomposition of the association between the share of late activity (activity anchor) and the first-attempt project grade (unified subsample: measure defined, ≥2 observable members).

Confidence intervals from a bootstrap resampling whole projects (B = 2,500, fixed seed 42); p-values from the rank test at the between level, the deviation correlation at the within level, and cluster-robust inference in the fixed-effects model. The three rows answer related but different questions: the first compares team means, the second compares students with their own team mean, and the third estimates the within-project association with project fixed effects. Leading zeros are shown throughout.

At the between-project level the direction is negative—teams with a larger share of late activity receive lower mean grades—and although the rank test is marginal (p = 0.064) the cluster-bootstrap interval excludes zero. At the within-project level the sign reverses (r = +0.129, p = 0.003), but the fixed-effects estimate with cluster-robust errors does not support it (B = +0.736, SE = 0.556, p = 0.185). We therefore describe this as a possible sign reversal across levels rather than an established instance of Simpson’s paradox: only the between-project side is supported by the cluster-aware procedure. Substantively, late concentration at the team level is consistent with postponement and weak coordination, whereas late activity of an individual student relative to teammates may reflect integration work at the assembly stage. The reversal is specific to this measure and to the activity anchor: under the deadline anchor the bivariate association is not robust [ρ = −0.067; 95% CI (−0.194; +0.060)], and for the expansion slope both levels carry the same sign, so no reversal arises. Temporal indicators are not interpretable without stating both the level of analysis and the anchor.

3.5 Hierarchical regression: incremental explanatory value

To test whether each dimension of participation—activity, network position, temporal organization—contributes independently beyond the others, hierarchical (block) regression models were estimated (Table 8; the procedure, block composition and control variables are described in Section 2.9). Because the grade has a pronounced between-project component (ICC = 0.697, so 30.3% of the total variance lies within projects), the model was estimated in two variants: Model A (pooled OLS with cluster-robust standard errors at the project level) and Model B (within estimation on projects with differentiated grades). The within-project R² of Model B and the within-project share of total variance have different denominators and are not compared.

Table 8. Hierarchical regression models: Incremental association of digital-trace measures with the first-attempt project grade.

Model A—pooled OLS with cluster-robust standard errors clustered on projects; Model B—within estimation on variables centred on project means, so the within-project R² is a share of within-project variance and is not comparable with the within-project share of total variance (ICC = 0.697). Block composition is described in Section 2.9. All increments are tested with cluster-robust Wald statistics, not ordinary F-tests. All VIF ΔR², χ², and p refer to the increment from adding each block; complete coefficients, standard errors, confidence intervals and standardized β for the final models of both panels are given in Supplementary Table S3. All VIF < 3.5; the correlation matrix of predictors is reported in Supplementary Table S4.

The results of sequential block entry are given in Table 8. In Model A, significant increments in explained variance come from the activity and coverage block (T1, ΔR² = 0.043, cluster-robust Wald p < 0.001) and from the object-mediated network position block (T2, ΔR² = 0.046, p = 0.020), whereas the temporal block adds nothing between projects (T3, ΔR² = 0.006, p = 0.619). In Model B the ordering differs: the temporal block carries a substantial increment within projects (ΔR² = 0.083, p < 0.001). The contribution of the network block in Model A is carried by the isolation indicator [B = −1.40; 95% CI (−2.45; −0.36); p = 0.009], whereas degree itself is not distinguishable from zero once isolation is in the model—consistent with a threshold rather than a graded association. All increments are tested with cluster-robust Wald statistics rather than ordinary F-tests. The key result lies in the contrast between the models: temporal organization has independent explanatory value only where grades are differentiated within projects (Model B), not on the aggregated sample (Model A), where the between-project component dominates—confirming the level-specific nature of the temporal associations (RQ4). Network position adds explained variance beyond activity volume (RQ2), and temporal organization contributes at the individual level but not between projects (RQ3).

The results are robust to the missingness policy and to the network specification. On the card-only network, which excludes container artifacts, all qualitative conclusions reproduce (degree ρ = +0.158 against +0.161; rescaled within-project degree rank +0.155 against +0.160; Burt constraint −0.193 against −0.190), and the temporal block remains non-significant in the pooled model and significant in the within model. Restricting the network to shared artifacts alone, that is removing artifacts touched by a single student, weakens the degree association (+0.082) and reverses the sign for shared artifacts (−0.062); this specification excludes every isolate by construction and is therefore reported for completeness rather than as an alternative primary specification. Excluding Trello leaves every estimate unchanged, since Trello artifacts are never shared and contribute no tie.

4 Discussion

4.1 From activity counts to object-mediated participation

The proposed approach advances learning analytics by moving the analysis of digital traces from the level of individual events to the level of object-mediated relations between participants and shared artifacts. In contrast to studies that reconstruct collaboration from communication interactions (Saqr et al., 2018; Zamecnik et al., 2023), the present approach operationalizes participation as observed co-use of shared objects of activity, measured as the intersection of students’ events on the same artifact, in line with the object-mediated view of collaboration developed in activity theory and CSCL research (Engeström, 2008; Stahl, 2006; Patarakin and Vachkova, 2019). The focus of analytics thus shifts from “how much a student did” to “how their work is embedded in the collective production of the shared result.”

The empirical results support this shift, though not by producing larger coefficients: all associations reported here are small, and the object-mediated measures are small alongside the volume measures (ρ = +0.222 for total events, and ρ = +0.160 for relative embeddedness once within-project ranks are placed on a common scale). What the object-mediated view adds is a distinction that volume cannot express at all. A student may be among the most active members of a team and still share no artifact with anyone in it: activity counts place such a student near the top, whereas the object-mediated network shows no observed artifact co-use with teammates. Isolation within the observed trace network emerges as a distinct risk pattern (ρ = −0.342) and is the only measure in this study that retains its sign and significance under pooled OLS, a project random-effects model and an ordinal specification alike. It is also specific rather than merely correlated: students who are the sole trace-observable member of their team, where the absence of ties reflects instrumentation coverage, are indistinguishable from connected students (δ = −0.036, p = 0.63). The network view therefore separates a diagnostic signal from an artefact of observability, which activity counts cannot do in principle. Among connected students the number of partners adds little, so the association is categorical rather than graded: a student may be active yet remain disconnected in the observed artifact co-use network, and such a position is associated with lower outcomes. Structural configuration carries further information among connected students: a less closed ego network is associated with higher grades (Burt constraint ρ = −0.190), and the association survives adjustment for activity volume and ego-network size (partial ρ = −0.133), extending findings on network position in collaborative learning (Saqr et al., 2020) to object-mediated networks. At the team level, the density of the object-mediated network is most closely related to the lowest grade in the team (ρ = +0.251) rather than to its mean, which is what an indicator intended for formative use should track. Temporal measures describe the development of participation and do so at the level where teachers act: between teams they add no incremental variance, whereas within a team they add a substantial share. Taken together, the three families of measures answer different questions rather than competing on explanatory power—volume describes how much a student did, object-mediated position distinguishes observed participation around shared artifacts from digital presence alone, and temporal organisation differentiates students inside a team as the project unfolds.

4.2 Inference under clustered educational data and a multi-platform pipeline

The methodological contribution of the study is a reproducible procedure that converts scattered event logs from several platforms into a single object-mediated network model. The procedure comprises strict attribution of events to projects, construction of the bipartite student–artifact network, its projection onto an object-mediated network of students, and the tracking of the dynamics of their positions over time. Strict event-to-project attribution is worth emphasizing: without it, parallel participation with overlapping time windows produces artificial duplication of activity, inflating both volume measures and network density.

Student–project pairs are nested in projects, and 80 students appear in more than one project, so uncertainty is dependent along two dimensions. We therefore use a single cluster-aware framework: estimates are reported with intervals from a bootstrap that resamples whole projects, and the interval rather than the p-value is the criterion by which an association is called robust; Benjamini–Hochberg correction is applied within three families corresponding to the three studies, whose composition is reported. Adding the student dimension through multiway cluster-robust inference raises standard errors by 3.9% on average and changes no conclusion, and repeatedly drawing one project per student reproduces every estimate within ±0.02: project clustering was already the binding constraint. This framework separates robust from fragile associations more sharply than significance testing alone—the number of shared artifacts, tie strength, effective size and team density on the strict subsample do not retain intervals excluding zero, and are reported as suggestive rather than established. Two specification choices proved consequential and are easy to overlook. Temporal windows must be defined on calendar time with empty intervals retained, since indexing only windows that contain events measures event order rather than elapsed time, and anchoring the final window on a student’s last observed event makes the anchor endogenous to the behavior measured. Within-project ranks from projects of different size are not comparable on their raw scale: pooling them allows project size to enter both variables and yields a substantially larger association than the one obtained after rescaling to a common interval. Both issues are invisible in the point estimates and are detectable only by examining the subsamples on which a measure is degenerate.

4.3 Participation patterns and formative enquiry

The interpretive logic of the results is summarized in Figure 8. It reflects a sequential transition from recording individual digital events to analyzing the structure of students’ participation in joint activity. At the first level, the digital trace is treated as a set of observable actions. At the second level, digital activity is analyzed as a quantitative characteristic of participation. At the third level, the object of analysis is the student’s digital participation in collective activity, expressed through their ties to shared artifacts and their structural position in the object-mediated network. Finally, at the fourth level, the contribution to the shared project product is assessed. The patterns identified in this study – isolation within the observed trace network, structural closure, absence of expansion of the circle of shared objects, and concentration of activity in the closing stages of the project – belong to the level of digital participation and can serve as prompts for formative enquiry into students’ joint work rather than as instruments of assessment.

The practical value of the approach lies in the diagnosis of these patterns, and the indicators are diagnostic prompts rather than scores. Students may be misclassified when they work offline, use tools outside the instrumented platforms, or occupy roles that leave few traces; in this corpus 57.7% of eligible pairs left no observable trace, and observability varied systematically with the examination board, the study programme and the academic year. Use of the indicators should therefore be limited to settings where platform coverage has been established, should be transparent to students, should give them a means of contesting an interpretation, and should be reviewed by a person before any intervention follows. The opposite directions of the late-activity association at the team and individual levels demonstrate that temporal indicators cannot be interpreted without specifying the level of analysis and the anchor used to close the project window. The measure that behaves most consistently—isolation within the observed trace network—identifies a group worth asking about, not a group to be graded down.

4.4 Limitations

The results are subject to several limitations. Observability is the principal constraint: 838 of 1,452 eligible pairs (57.7%) left no observable trace and 73 of 324 projects had no observable member, and observability varies with the examination board (25 to 77% across boards), the study programme, the academic year and team size, though not with the grade. Ties record observed co-use of artifacts rather than interaction: two students may work on the same card at different times and without contact, while joint work in meetings, messaging or local environments leaves no events here, and no external validation against task assignments, peer assessment or instructor observation was available; student role within the team is not recorded in the source data. The platforms contribute unequally—task-management cards carry 84.7% of all ties, a GitLab repository is a container artifact aggregating a median of 38 events into a single object, and Trello, at 1.7% of events, produced no shared artifact at all; results computed without container artifacts are reported as a sensitivity analysis and reproduce the findings, but that specification excludes students whose only traces are commits, and the exports provide no file-level granularity for GitLab, so the fourth conceptual level—what a contribution changed inside an artifact—could not be operationalised. Reactivity is confined to a single quantity: the weekly action count shown on the student dashboards was visible and incentivised, whereas the network, structural and temporal measures reported here were never displayed to students or graders. Assessment introduces two further constraints: the grade composite contains a component that assesses a trace event, addressed by outcome decomposition, and each project was examined by a single board, so project-level structural measures are not separable from differences in board strictness (boards account for R² = 0.20 of grade variance), whereas within-project ranks, computed inside a board, are unaffected. Within-project rank analyses are informative only where degree varies among team members: at a median team size of three the projection is often complete or empty, and 18 of 106 projects vary in both degree and grade. Official project start dates are unavailable, so the project window opens at the first observed event. Prior academic performance is not available in the institutional records released for this study, so the control block contains contextual covariates (team size, study year, academic year, project duration) rather than a prior-attainment baseline; the relative contribution of the blocks is therefore reported as a Shapley decomposition of the explained variance in addition to the sequential increments (network position 36.3% of the explained variance, controls 21.8%, temporal organisation 21.8%, activity and coverage 20.1%), which does not depend on the order of entry. Finally, the empirical base comprises data from a single institution (MIEM HSE) over two academic years, so generalization to other contexts requires verification, and the ordinal and mixed specifications agree with the pooled model on the isolation indicator but not on every predictor, which is a further reason to treat the smaller associations cautiously.

5 Conclusion

The study shows that digital traces of collaborative project activity lend themselves to meaningful pedagogical interpretation when the analysis moves from individual actions to the structural and temporal organisation of participation, and that the three families of measures answer different questions at different levels of analysis rather than competing on explanatory power.

• Object-mediated participation can be reconstructed from multi-platform traces at the level of individual students, teams and shared artifacts. The platforms contribute unequally: task-management cards carry 84.7% of all ties, a GitLab repository aggregates a median of 38 events into a single container artifact, and Trello produced no shared artifact at all. Integration requires strict attribution of events to projects and to artifact levels, without which participation is duplicated.
• Volume and configuration measures are associated with the grade, but all associations are small (ρ ≤ +0.222); statistical significance should not be read as practical importance, and full interpretation requires the network and temporal structure of participation.
• The clearest pattern is categorical rather than graded: students isolated within the observed trace network—digital presence without any shared artifact with an observable teammate—receive markedly lower grades (M = 5.47 vs. 6.91; δ = −0.342; p < 0.001). This indicator is the only measure that retains its sign and significance under pooled, mixed-effects and ordinal specifications alike, and it identifies a group that activity counts place among the active, whereas sole trace-observable members do not differ from connected students (δ = −0.036; p = 0.63). Trace-based indicators of isolation are interpretable only once platform coverage is taken into account.
• Among connected students, a less closed structural position is associated with higher grades (Burt constraint ρ = −0.190) and remains so after adjustment for activity volume and ego-network size. At the team level, integration through shared objects is most closely related to the lowest grade in the team (ρ = +0.251) rather than to its mean.
• Sustained, distributed activity is associated with higher grades, and continued expansion of a student’s circle of shared objects over the project is associated with higher outcomes (ρ = +0.182; M = 7.13 vs. 6.33 for a flat trajectory). The association is a threshold rather than a gradient: once expansion occurs, its rate does not differentiate outcomes (Kruskal–Wallis H = 0.27, p = 0.87).
• Temporal indicators are level-specific and anchor-dependent. The rate of expansion is associated with the grade within teams but not across projects; the share of late activity reverses sign between levels and loses its association when the project window closes at the institutional deadline rather than at a team’s own last event. Both the level of analysis and the anchor must be stated for these indicators to be interpretable.
• Observability is systematically related to the examination board, the study programme and the academic year, but not to the grade; conclusions apply to students whose participation was observable on the instrumented platforms. The approach supports the diagnosis of observable participation patterns in shared digital work and their use as a diagnostic vocabulary rather than a scoring instrument. Future research will require cross-institutional verification and content-level analysis of what a contribution changes inside a shared artifact.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. Generative AI was used in the preparation of this manuscript (translation from Russian and language editing). The authors reviewed and verified all content and take full responsibility for the final text.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fcomp.2026.1940743/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Minela Kerla, Consultant, Sarajevo, Bosnia and Herzegovina

Reviewed by: Abdelhafid Chadli, Ibn Khaldoun University, Algeria

Imron Fauzi, Universitas Islam Negeri Kiai Haji Achmad Siddiq Jember, Indonesia

Footnotes

https://gitlab.com/

https://wekan.fi/

https://trello.com/

https://networkx.org/documentation/stable/release/release_3.2.html

https://www.python.org/downloads/release/python-3120/

https://www.statsmodels.org/stable/gettingstarted.html