Two-Stage Personalized Vehicle Ranking in Online Auctions Using Heterogeneous Graph Neural Networks and LambdaRank

Online vehicle auctions contain unique, short-lived items, requiring models to infer attribute preferences rather than memorize repeated consumption. We present a temporally controlled two-stage auction ranking system. A two-layer graph neural network (GNN) encodes a heterogeneous graph of user, vehicle, and property nodes connected by bid, purchase, favorite, and alert edges. Its embeddings are combined with behavioral, vehicle, compatibility, and auction context features in LightGBM LambdaRank. A pre-target graph snapshot limits leakage, and evaluation ranks positives against reconstructed as-of auction rosters. Across approximately 1.6 × 104 tasks, the full model achieved an aggregate NDCG@10 of 0.1043. For users represented in the LambdaRank training rows (ranking warm), NDCG@10 was 0.1093 versus 0.0747 for content affinity, 0.0177 for popularity rate, and 0.0122 for random ranking. On the interaction-derived candidate pool, GNN embeddings improved ranking warm NDCG@10 by 0.0098 when twelve handcrafted user–vehicle cross-features were removed. With the complete feature set, their matched five-seed mean contribution was +0.0008, with a confidence interval spanning zero. No stable improvement was observed for users absent from those rows (ranking cold). These results suggest graph representations are most useful when handcrafted compatibility signals are limited, while cold-start performance remains constrained and requires controlled online evaluation.