Learning-driven priority scheduling for technical ability-constrained IT projects: A TabNet-based sequential framework with quantitative assessment

Abstract

IT projects subject to technical skill constraints are complex systems that require effective resource and skill management to achieve their objectives. Given the complexity and dynamics of these projects, traditional project planning methods, based on fixed priority rules, lack flexibility and adaptability. To overcome the drawbacks of conventional IT project planning techniques, a novel scheduling paradigm called TANPS (Technical Ability-aware Neural Priority Selection) is presented in this study. The proposed approach incorporates TabNet, a supervised deep learning model, into a sequential scheduling engine based on the Serial Schedule Generation Scheme (SSGS) instead of predefined heuristics or opaque metaheuristic frameworks. This integration allows decision criteria to be dynamically adjusted at every planning stage in response to real-time project indicators, such as agent availability, technical skill gaps, and activity dependency network complexity. TANPS strikes a balance between algorithmic efficiency and human-centered adaptation by learning from past scheduling data. The architecture captures the project's fine-grained structural features while avoiding the high computational expense of stochastic optimization or combinatorial rule selection. To demonstrate improvements, evaluate performance, and provide insights into the causal impact of the proposed approach relative to random variation, we conducted a comprehensive quantitative validation based on simulations and extensive statistical and econometric analyses. According to simulation data, the method performs noticeably better than traditional planners on several important criteria, such as workload balance, makespan minimization, and robustness in the face of technical resource limitations. The approach guarantees interpretability, generalizability, and simplicity of integration with enterprise planning tools, in addition to these empirical benefits. Thus, it establishes the foundation for intelligent scheduling assistants that can learn continuously, optimize multiple objectives, and adjust in real time to uncertainty.