Meituan Search Ads Team's Solution for KDD Cup 2020 Multimodalities Recall Track

The article describes Meituan's participation in the KDD Cup 2020 Multimodalities Recall track, where they secured third place. It outlines the competition background, data characteristics, and the two main challenges: distribution mismatch between training and test sets, and complex multimodal matching.

To address distribution inconsistency, the team designed a diversified negative sampling strategy combined with distillation learning, gradually moving from easy to hard samples and using validation‑set fine‑tuned models to generate soft labels for training.

For complex multimodal matching, they built a fine‑grained text‑image matching network employing gated fully‑connected layers for semantic mapping, bidirectional attention between word and region features, and a diversified fusion strategy (Kronecker product, concatenation, self‑attention). The final model ensembled multiple neural networks and tree models.

Experimental results show significant NDCG@5 improvements over a baseline model. The proposed techniques were later transferred to Meituan's search advertising creative‑selection system, where knowledge distillation and shared embeddings helped balance model performance and efficiency.