Source-linked AI summary
Predicting Student Attrition in Competitive Programming: A Large-Scale Study Integrating Survey Insights and Global Behavioral Logs
Azuad Islam Ruhan, Golam Mostofa Naeem, Rakibul Islam Rafi, Sherin Afrin Mim, Nazira Bani Opi, Dewan Fahad Chowdhury, Md. Rejaul Korim Sadi
TL;DR
Competitive-programming attrition is understudied despite fragile participation and limited data-driven support for identifying disengaging students. The paper combines Codeforces behavioral logs with psychographic survey data to study and predict attrition. It reports behavioral and psychographic signals associated with attrition, benchmarks predictive models, and demonstrates a proof-of-concept early warning deployment.
Problem
Competitive-programming attrition is understudied, while mentors lack a data-driven mechanism to identify students drifting toward disengagement before they quit.
Method
The paper combines Codeforces behavioral logs with psychographic survey data in a dual-layer predictive framework and deploys a survey-trained Early Warning System.
Results
The Soft-Voting Ensemble achieved CV F1 = 0.737 on behavioral data, while Random Forest achieved CV F1 = 0.924 on survey data as a localized exploratory pilot.
Takeaways & Limitations
Behavioral and psychographic signals can support data-driven mentoring, with the proof-of-concept system identifying 4 high-risk active students.
Takeaways & Limitations
The survey is small and geographically limited, behavioral attrition labels use an inactivity proxy, and the survey model’s performance may overstate generalizability.
Abstract
from arXiv · showhide
Competitive programming (CP) offers computer science students an environment for developing algorithmic reasoning skills. However, sustained participation remains a challenge, as many students disengage after encountering skill plateaus or performance anxiety. While educational data mining (EDM) has studied dropout in MOOCs and academic courses, CP attrition remains understudied. This paper presents a dual-layer framework combining large-scale Codeforces activity logs (n=1,816) with a multi-institutional psychographic survey across 10 universities in Bangladesh (n=64). Analysis reveals that true attrition is preceded by an 83.71% reduction in contest participation and a 15.6% increase in struggle time. We identify a "Skill-Application Paradox": stopped students self-report higher mathematical confidence (3.88 vs. 3.41) and data structure understanding (3.57 vs. 3.09) than active peers, yet their independent practice and upsolving habits are significantly weaker (p < 0.001). Benchmark evaluations show that a Soft-Voting Ensemble achieves a 0.737 CV F1-score on behavioral logs, while Random Forest achieves 0.924 CV F1-score on psychographic data (an exploratory pilot). We deploy the survey-trained model as a proof-of-concept Early Warning System over 22 active students, identifying 4 at high risk. These findings show that behavioral and psychographic signals can support data-driven mentoring in computing programs.
1 INTRODUCTION
Competitive programming develops algorithmic and technical problem-solving skills, but sustained participation is fragile and CP-specific attrition remains insufficiently studied. This paper responds with a dual-layer predictive framework linking behavioral logs, psychographic data, and early intervention.
- Competitive programming provides feedback-rich practice for algorithmic reasoning and technical problem-solving under ambiguous, time-constrained conditions.
- Many students disengage after short practice periods, with self-efficacy fluctuations, perceived lack of progress, social isolation, and competing academic demands associated with this pattern.
- CP attrition remains underexamined because existing educational data-mining research mainly models dropout in formal institutional and MOOC settings.
- The study asks whether self-reported reasons, Codeforces behavioral proxies, and psychographic machine-learning models can explain and predict competitive-programming discontinuation.
- The paper introduces the Skill-Application Paradox, engineers the Intensity Ratio and Editorial Dependency Index, and combines Codeforces logs with survey data across Bangladeshi universities.
- A proof-of-concept Early Warning System applies survey-trained predictions to support proactive mentoring in university programming clubs.
2 RELATED WORK
Prior retention research has progressed from static demographic and performance features toward dynamic behavioral signals, while programming-judge and motivational studies add cognitive and psychosocial perspectives. However, competitive-programming attrition has not previously been examined by combining platform logs with psychographic survey data.
- Retention studies have shifted from demographic and first-year performance features toward dynamic behavioral measures, especially in MOOCs.Examples include submission patterns, login frequency, and forum engagement; Random Forest and gradient-boosted models are among the methods cited.
- Fine-grained programming-judge logs can proxy cognitive frustration and learning difficulty through errors, submission timing, and wrong-answer rates.Cognitive load theory provides a basis for interpreting high struggle times and repeated failures as possible overload markers rather than productive effort.
- Motivation, self-efficacy, and peer support are established persistence predictors, while leaderboards and volatile ratings can intensify performance anxiety and social comparison.South Asian academic research also highlights institutional mentoring gaps as relevant to persistence.
- No prior study had combined large-scale competitive-programming logs with psychographic surveys to study CP-specific attrition.Telemetry provides scale and objectivity but limited psychological context, whereas surveys provide depth but face recall bias and unverifiable behavioral claims.
3 METHODOLOGY
The study combines filtered Codeforces telemetry with a multi-institutional survey, engineers behavioral and psychographic features, and evaluates multiple classifiers using stratified cross-validation. Attrition labels distinguish true attrition from intentional exits, while leakage-prone or indirect features are excluded.
- Datasets and labeling: 6,927 Codeforces accounts were filtered to users with at least five rated contests and peak ratings below 2400.These criteria aimed to ensure meaningful competitive-programming engagement and focus on university-level learners.
- Datasets and labeling: 908 inactive users were labeled true attrition, while 416 higher-rated inactive users were classified as intentional exits and excluded.The final behavioral dataset contained 5,603 active users and 908 true attrition cases before balancing.
- Datasets and labeling: Random undersampling produced a balanced behavioral modeling dataset of 1,816 users, comprising 908 active and 908 true attrition cases.Sampling used seed = 42.
- Feature engineering: Twelve normalized behavioral features and nine survey psychographic features were constructed, including practice intensity, upsolving, struggle time, activity decay, self-rated skills, mentorship, and long breaks.Academic features such as semester and CGPA were excluded as indirect proxies not causally linked to attrition.
- Model architecture and training: Nine classifier architectures, including a soft-voting ensemble, were evaluated with class balancing, imputation, scaling, one-hot encoding, and stratified 5-fold cross-validation.F1-score was the primary metric because false negatives and false positives have asymmetric educational costs.
4 RESULTS
Behavioral logs show clear divergence between active and true attrition users, especially in contest participation, activity decay, and skill-related performance. Survey results reveal a Skill-Application Paradox, while classifier results provide exploratory evidence for identifying attrition risk.
- Behavioral Trends: 7.27 versus 1.19 contests per month represents an 83.71% reduction among true attrition users (p < 0.001).This was identified as the strongest single behavioral signal of impending disengagement.
- Behavioral Trends: 3,059 versus 2,646 minutes indicates a 15.6% increase in struggle time for true attrition users, but the difference was not statistically significant (p = 0.213).High within-group variance from outlier accounts contributed to the nonsignificant result.
- Behavioral Trends: True attrition users solved fewer problems per month and had lower math and data structure solve rates than active users, with problems solved showing a 42.7% lower value.The reported differences were 7.86 versus 13.70 problems per month, 0.469 versus 0.538 math solve rate, and 0.400 versus 0.480 data structure solve rate.
- Behavioral Trends: 45.01% of true attrition users versus 29.81% of active users experienced activity declines exceeding 50%, a 51% higher burnout incidence.The comparison concerns activity in the final three months.
- Behavioral Trends: Both groups predominantly never upsolved, while attrition users who did upsolve showed marginally higher rates, a pattern presented as the Editorial Dependency Hypothesis rather than a confirmed causal claim.The study could not directly observe editorial consultation from Codeforces logs.
- Survey Findings: The Skill-Application Paradox: Stopped students reported higher mean math confidence (3.88 versus 3.41) and data structure understanding (3.57 versus 3.09), despite significantly different weaker behavioral practice profiles.Upsolving habit, peer circle density, and long-break history all differed significantly between active and stopped students.
5 DISCUSSION
The two data layers converge on a coherent account of CP attrition: declining engagement and increased friction coincide with high theoretical confidence but irregular practice and limited peer support. The authors recommend targeted club interventions while cautioning that the evidence is observational, geographically limited, and not yet prospectively validated.
- Convergence Across Data Layers: Platform data shows that attrition is preceded by declining contest participation and increased problem-solving friction, while survey data identifies irregular practice and limited peer support.These signals describe the same attrition pattern from behavioral and psychographic perspectives.
- Convergence Across Data Layers: At-risk students may report strong theoretical understanding yet struggle to apply it consistently under contest conditions.The paper connects this mismatch to unexpected difficulty and recurring behavioral traces across both datasets.
- Practical Implications: Structured post-contest review, mentor-buddy systems, and monitoring after extended breaks are proposed as targeted club responses.The recommendations prioritize upsolving discipline and peer network density rather than simply increasing contests or study materials.
- Limitations: The survey evidence is geographically limited, while Codeforces attrition labels use inactivity thresholds that cannot distinguish dropout from temporary academic breaks.These scope constraints limit generalization across academic and cultural contexts and affect interpretation of behavioral labels.
- Limitations: Retrospective recall bias may inflate separation between active and stopped cohorts because stopped students reported past behaviors after disengaging.Current psychological states may systematically color recollections of earlier practice.
- Limitations: The Early Warning System remains a proof-of-concept because no longitudinal follow-up verifies whether four flagged students later disengaged or experienced performance decay.The authors identify 2–3 months of tracking as a priority for future validation and make no causal claims from the observational data.
6 CONCLUSION
The paper concludes that a dual-layer framework combining Codeforces logs and Bangladeshi university survey data can characterize CP attrition through behavioral and psychographic signals. Behavioral and survey models produced different benchmark leaders, while the Early Warning System remained an exploratory baseline requiring larger longitudinal validation.
- 6 CONCLUSION: The framework combines Codeforces behavioral logs (n = 1,816) with a psychographic survey from 10 Bangladeshi universities (n = 64 for predictive modeling).This dual-layer design links global platform telemetry with localized survey evidence.
- 6 CONCLUSION: 83.71% reduction in contest participation preceded true attrition, alongside consistent underperformance on skill-related metrics.The Skill-Application Paradox was validated through feature importance rankings and significant group-level tests for upsolving habit and peer circle density.
- 6 CONCLUSION: Soft-Voting Ensemble achieved CV F1 = 0.737 and Test Recall = 0.769 on behavioral data, while Random Forest achieved CV F1 = 0.924 on survey data as a localized exploratory pilot.The Intensity Ratio ranked as the third most predictive Codeforces feature.
- 6 CONCLUSION: The survey-trained Early Warning System identified 4 high-risk active students whose profiles matched the historical attrition signature.The flagged profiles combined strong academic performance, irregular upsolving, and absent mentorship.
- 6 CONCLUSION: The results are an exploratory baseline rather than a production-ready deployment, with future work requiring larger longitudinal cohorts and prospective EWS validation.Planned extensions include multiple universities and academic years, additional activity records, and partnership-based validation.