Remaining time prediction for IT project tasks based on temporal graph representations

Abstract

The subject of the study is models and approaches for predicting the time remaining until IT project tasks are completed, based on event logs and the network plan dependency graph in snapshot mode. The aim of the work is to improve the quality of remaining time prediction for IT project tasks in snapshot mode through a hybrid approach that combines temporal graph representations with regression models. Tasks: to formalize the remaining time as a target variable and define a leakage-free training protocol; to describe SPARK as a unified "tasks – events – dependencies" representation and reconcile dependency types with network plan constraints; to implement training example generation schemes in single- and multi-snapshot modes; to build a hybrid temporal graph network with gradient boosting and compare it with baselines; to analyze sensitivity to anomalously large durations, snapshot modes, and graph characteristics. Methods: time-based splitting and block cross-validation; event prefix aggregation; structural graph metrics; event encoding by a temporal graph neural network; prediction by gradient boosting; evaluation by mean absolute error and Spearman rank correlation. Results: on identical test observations (19,663 snapshots, 3,309 tasks), the hybrid model with the 64 most informative embedding dimensions selected achieved a task-level mean absolute error of 545.44 h against 585.90 h for gradient boosting on the same features without temporal embeddings, i.e., a reduction of 6.9%, and the lowest value among the compared models (XGBoost – 551.03 h; the advantage is confirmed by a paired bootstrap in 99.4% of iterations, while XGBoost remains better on typical error and rank correlation); an untrained-encoder control (572.88 h) shows that two-thirds of the gain comes from training the temporal memory ; the multi-snapshot mode raises the Spearman rank correlation from 0.314 to 0.380, whereas the single-snapshot mode is more accurate for a single early-stage estimate. Conclusions: the approach makes it possible to move from static duration estimates to regular updating of the network plan from actual events; the choice of model, snapshot mode, and metric must match the managerial objective.