A dynamic recommendation algorithm for regional industry-adapted higher vocational courses integrating counterfactual causal inference and graph neural networks

Abstract

The mismatch between higher vocational curricula and the shifting skill demands of regional industries continues to constrain graduate employability, while existing recommendation engines remain trapped in correlational logic and selection-biased enrollment histories. This paper proposes CCIG-DRec, a dynamic course recommendation algorithm that couples heterogeneous graph neural networks with counterfactual causal inference. A time-evolving graph linking students, courses, occupational posts, skills, and industries is constructed from recruitment portals and open vocational catalogs across the Chengdu–Chongqing Economic Circle, and sliding-window snapshots are refreshed through relation-specific exponential decay whose half-lives are estimated from observed edge-persistence curves rather than assumed. A dual attention scheme operating at node and meta-path levels produces type-aware embeddings, after which a causal intervention layer reweights neighbor messages by clipped, self-normalized exposure propensities to neutralize industry popularity bias. Counterfactual student trajectories generated through abduction–action–prediction supply both training signals and interventional explanations for each recommended course. Against the strongest of five competitive baselines, each granted the same five-type graph and the same industry signal, CCIG-DRec raises Recall@10 by 12.7 percent and NDCG@10 by 11.8 percent on the regional corpus and narrows the counterfactual fairness gap by 30.4 percent, every margin significant at $$p$$ < 0.01 under a paired test over ten seeds. The industry adaptation margin behaves differently, and the paper says so plainly: it shrinks from 13.0 to 4.6 percent once every baseline receives an equivalent industry-fit regularizer, while a placement-based indicator that enters no training objective puts the surviving advantage at 5.1 percent. Results on the public MOOCCube benchmark are reported alongside the regional corpus. Cross-scenario tests across five industrial clusters confirm stable performance under varying demand structures, and a case study demonstrates auditable interventional justifications suitable for program-level curricular governance.