Graph-Based Recommendation Algorithms and Cold‑Start Solutions

The talk begins with an introduction to the exploration‑exploitation (E&E) problem in recommendation systems and the challenges of cold‑start for new users and new items.

1. Recommendation System Challenges on E&E – New users lack historical behavior, making accurate recommendations essential for retention; new items have no feedback and suffer from exposure scarcity, leading to long‑tail issues.

2. Classic Graph Model – Collaborative Filtering – Describes item‑based CF as a bipartite graph, computing item similarity to build a directed graph for fast neighbor retrieval, while noting CF’s popularity bias and internal loops.

3. Graph Embedding – Introduces graph embedding analogous to word2vec, using random walks to generate node sequences and skip‑gram training, preserving temporal order and improving coverage at the cost of slight relevance loss. Discusses homophily vs. structural equivalence and sampling strategies (BFS, DFS) with node2vec parameters p and q.

4. Alibaba Graph‑Embedding with Side Information (GES/EGES) – Shows how Alibaba builds a session‑based directed graph, adds side‑information edges (brand, category, etc.), and learns weighted embeddings to address item cold‑start.

5. Airbnb Word2Vec and Sampling – Describes session segmentation (30 min), cluster‑based negative sampling, and paid‑behavior‑aware context windows; cold‑start items are initialized by averaging embeddings of top‑3 similar known items.

6. Our Work – Mixed Graph‑Vector Model – Combines CF similarity graph with node2vec random walks and word2vec training to produce both an item similarity graph and item embeddings; uses a two‑tower DSSM where the user tower processes real‑time behavior features and the item tower uses pre‑computed embeddings.

The two‑tower architecture enables fast online user vector computation while retrieving item vectors from a lookup table, reducing latency.

7. Graph‑Restricted Vector Search – Limits recommendation search to nodes adjacent to user‑clicked items, improving efficiency and enabling exploration for users with few or no interactions.

Q&A Session – Addresses session data filtering, differences between user and item scene vectors, and handling of disconnected sub‑graphs in embedding.

In summary, the presentation covers theoretical foundations, industry implementations, and a hybrid approach that leverages both collaborative filtering and graph embedding to improve recommendation quality and address cold‑start problems.