Applying Graph Embedding and Vector Recall for Personalized Recommendation in a UGC Community

Background – The "Tǎngpíng" app, a lifestyle‑recording community, faces user cold‑start (sparse or no behavior) and content cold‑start (long‑tail items lacking exposure) during its growth phase.

Problem Statement – To attract new users and retain low‑activity users while ensuring long‑tail content gets recommendation opportunities.

Solution Overview – Augment the existing personalized and popular recall architecture with an additional vector‑recall path. The approach leverages two research directions: network representation learning (graph embeddings) and multimodal representation learning (image embeddings).

Graph Construction – Build a user‑item bipartite graph from 7‑day interaction sessions and an item‑item homogeneous graph derived from content similarity. Nodes include user attributes (age, gender, region) and item attributes (tags, circles, topics).

Embedding Models

Shallow models: item2vec, metapath2vec, userCluster2vec for initial item‑item and user‑item embeddings.

Deep GNN models: GraphSAGE (inductive learning with neighbor sampling) and GAT (attention‑based aggregation) to capture higher‑order topology and attribute information.

GraphSAGE Implementation – Pseudo‑code (shown as images) outlines neighbor sampling, two‑hop aggregation, concatenation with the node’s previous layer embedding, and a fully‑connected transformation to produce the k‑th layer representation.

GAT Implementation – Applies masked self‑attention on sampled neighbors to weight their contributions, improving robustness to noisy user actions.

Training Details – Use two‑hop aggregation, cross‑entropy loss, and offline experiments to balance performance and training cost.

Multimodal Image Embedding – Fine‑tune a ResNet backbone on the app’s category taxonomy, adding two fully‑connected layers for supervised tag classification. Image vectors from intermediate layers serve as an additional recall channel, mitigating content cold‑start.

Evaluation – Offline A/B tests show positive lifts in click‑through rate and exposure depth when incorporating attribute‑enhanced embeddings. Case studies demonstrate successful recall of niche content for both male and female users.

Conclusion & Outlook – The vector‑recall pipeline improves recall accuracy and diversity, helping new and long‑tail users. Future work includes incorporating richer user actions, sequential modeling, and richer multimodal signals (text, topics, circles).