HEEF: A Hybrid Engineering Evaluation Framework for AI R&D Team Efficiency Prediction using Meta Kaggle

Accurately evaluating the efficiency of Artificial Intelligence (AI) research and development (R&D) teams is a critical yet challenging engineering problem, as traditional productivity metrics fail to capture the iterative, data-dependent, and non-linear nature of AI workflows. This paper proposes HEEF (Hybrid Engineering Evaluation Framework), a novel data-driven framework designed to predict and analyze AI R&D efficiency by leveraging the large-scale Meta Kaggle dataset. HEEF introduces a multi-dimensional feature engineering strategy that integrates individual capability baselines, R&D investment intensity, graph-based collaborative structures, and dynamic learning curves. To balance computational efficiency with modeling complexity, the framework employs a two-stage learning architecture: utilizing Node2vec for unsupervised representation of collaboration networks and LightGBM for robust supervised regression. Furthermore, SHAP-based interpretability is incorporated to transform "black-box" predictions into actionable diagnostic insights. Experimental results demonstrate that HEEF significantly outperforms conventional machine learning baselines, achieving a 12.7% improvement in NDCG@10 and an 18.2% reduction in RMSE. Beyond predictive accuracy, our analysis reveals critical industrial insights, such as the inverted U-shaped relationship between team size and efficiency and the superiority of adaptive experimentation rhythms. This work provides a systematic tool and empirical foundation for optimizing AI engineering resources in large-scale information systems.