Deep Learning Practices for Personalized Recommendation in a Cultural Artifact Auction Platform

Business Background and Technical Challenges The platform "Weipaitang" is a vertical e‑commerce for cultural artifacts where items are auctioned online. Personalized recommendation is crucial for feed streams and shopping paths, but challenges include lack of structured category taxonomy, short item lifecycles, diverse user expertise, and the need for constrained nearest‑neighbor retrieval.

Overall Technical Solution The team built NLP pipelines (custom tokenization and NER) and image algorithms to generate thousands of item tags, forming the foundation for downstream recall models.

Item Embedding Learning A skip‑gram‑style model samples positive and negative pairs from user behavior sequences, followed by a DeepMF + item‑unit network that learns embeddings from multi‑dimensional attribute IDs, enabling end‑to‑end training even for short‑lived items.

Cross‑Domain Keyword Intent Embedding Using a dual‑tower DSSM, item embeddings from the pretrained item‑unit are transferred to search‑keyword space. Positive samples are derived from post‑search clicks and title‑item relations, while negatives are sampled and tuned, aligning keyword and item vectors.

User Multi‑Interest Embedding Learning The model addresses three problems: multi‑peak interests within a session, fusion of short‑term and long‑term behaviors, and maintaining distinct interest vectors. A Fusion Gate produces three interest vectors, and a Selection Gate selects the most relevant one for interaction with target items.

Metric Learning To keep the three interest vectors diverse, a metric‑learning loss minimizes pairwise inner products, complementing the cross‑entropy loss and preventing collapse to a single dominant interest.

Constrained Nearest‑Neighbor Retrieval with Faiss By normalizing item and user vectors and adding large offsets on specific dimensions for each category, the system can perform category‑constrained nearest‑neighbor search in a single Faiss index, improving recall diversity and click‑through rates by about 20% over traditional collaborative filtering.

Summary and Future Outlook The work demonstrates how mathematical and deep‑learning methods can solve niche e‑commerce recommendation problems. Future directions include building a cultural‑artifact knowledge graph, applying reinforcement learning for auction behavior, and developing multimodal CTR models.