CF-RL-TOPSIS model boosts talent recommendation with interpretability

Özkan Canay introduces CF-RL-TOPSIS, a novel interpretable late-fusion model for skills-aware talent recommendation that balances behavioral transition patterns, trajectory-sensitive adaptation, and inspectable occupation-level criteria. The model combines three branches: a transition-aware collaborative filtering branch, a compact reinforcement-style occupation-family bandit, and an entropy-weighted TOPSIS branch built from six semantic proxies. Fusion coefficients are selected via validation and remain auditable. On the JobHop benchmark, the full hybrid achieves NDCG@5 = 0.3040 ± 0.0073, significantly outperforming repeat-last, item Markov, transition-aware CF, CF+TOPSIS, GRU4Rec, and SASRec (all p ≤ 0.0039 in paired Wilcoxon tests). The results demonstrate that the three branches reinforce each other in semantically rich, non-saturating talent-history regimes.

On the Karrierewege dataset, however, the hybrid remains competitive but does not significantly exceed the strongest Markov baseline, revealing a persistence-dominated setting. In this regime, the adaptive bandit branch appropriately shrinks to near-zero weight, while the collaborative backbone maintains performance. Proxy-sensitivity analysis, family-level deep Q-network checks, and runtime evaluations support this interpretation. A worked user-level example shows how branch scores, criterion weights, and rank shifts can be inspected for individual recommendations. The paper does not claim benchmark-agnostic superiority, but provides a reproducible account of when transparent late fusion adds value beyond simple continuation heuristics. This work contributes to more trustworthy and interpretable AI in HR tech.