Construction and Application of an Interest Point Graph for Content Understanding in Information Feed Recommendation

The talk begins by outlining the basic paradigm of feed recommendation, highlighting two core challenges: inaccurate recommendations caused by loss of contextual information in user profiles, and the filter‑bubble problem arising from overly coarse tags such as categories or entities.

Traditional content‑understanding techniques—classification, entity extraction, keyword extraction, LDA topic modeling, and neural‑network models—are surveyed, noting their limitations in granularity, ambiguity, and interpretability.

To address these issues, a five‑layer interest point graph is introduced, comprising classification, entity, concept, topic, and event nodes, together with three relation types: hierarchical (up‑down), associative, and participation. The graph captures richer semantic information and user intent.

Interest‑point mining leverages massive UGC logs. Weak‑supervised methods include query‑title alignment (extracting common substrings) and pattern‑based bootstrapping. Coverage is improved with a unified GCTSP‑net framework: a bipartite query‑title graph for clustering, a query‑title interaction graph enriched with token adjacency and syntactic dependencies, followed by GCN‑based node embeddings and a classifier to identify interest‑point tokens. The ordering of tokens is solved as a traveling‑salesman problem to form complete interest points.

Relationship extraction uses co‑occurrence statistics for hierarchical links (category‑concept, concept‑entity, event‑topic) and supervised learning with automatically generated positive/negative samples. Associative relations are derived from document‑level and session‑level co‑occurrence, enhanced by vector similarity and triplet‑loss training.

The resulting graph contains millions of nodes (e.g., 46 M concepts, 1.98 M entities) and relationships, with >95 % accuracy across relation types.

In application, the graph supports a two‑stage pipeline: recall (hierarchical and semantic) followed by matching using a Match‑Pyramid interaction model combined with BOW similarity. Offline experiments show CTR improvements over traditional category tags and increased user retention, confirming that the graph better captures user consumption motives and mitigates information bubbles.

The Q&A section clarifies extraction methods for concept and topic layers, discusses the trade‑off between query‑based and article‑based mining, and explains term importance weighting (TF‑IDF) and event extraction via semantic recall and embedding similarity.

Overall, the interest point graph provides a more complete and inferential content‑understanding framework that enhances personalized recommendation performance.