Graph-Aware Recommendations with Neo4j GDS and Random Forest

Session track: Data Intelligence

Session time: 02:30 pm - 03:00 pm UTC November 12

Session description:

Recommending the right technology products to enterprise buyers is a classic cold-start and long-tail challenge. This session will show you how to build a production-grade technographics recommendation engine within Neo4j GDS. The speakers will walk you through a bipartite knowledge graph connecting ~49,000 company and product nodes via ~258,000 CONTEXT relationships. You will learn how FastRP embeddings (128 dimensions, undirected CONTEXT projection) capture co-adoption signals, how degree and PageRank mutations enrich node features, and how one-hot encodings tie structural graph signals to industry context. You will then see how graph-aware hard negative sampling — driven by constrained random walks (gds.randomWalk.stream, walkLength=5) over multi-hop co-purchase neighborhoods — produces realistic negatives that stress-test the model. A stratified resampling scheme corrects power-law skew, balancing the training set without destroying the graph structure. A Random Forest Regressor trained on element-wise FastRP embedding products achieves OOB AUPRC of 97.3% after Optuna TPE hyperparameter search (100 trials). You will learn how Platt re-calibration restores the true ~5.5% population positive rate from a 50% sampled prior, and how Hits@K evaluated over the full company×product score matrix quantifies real-world ranking quality. You will leave with a reusable blueprint: graph projection → GDS feature engineering → graph-aware sampling → resampling → RF scoring → calibration → Hits@K evaluation.