Building and Applying Relationship Graphs at Beike Real Estate: Architecture, Embedding, and Recommendation
The talk explains how Beike Real Estate constructs a large‑scale relationship graph from billions of user, house, and agent interactions, quantifies edge strengths, builds homogeneous and heterogeneous sub‑graphs, derives graph capabilities such as node influence, embedding, similarity and relation prediction, and finally deploys these capabilities in multi‑degree queries, house‑similarity recommendations and B‑side agent assistance, achieving measurable CTR improvements.
Beike Real Estate accumulates massive behavioral data of users, houses, and agents, and uses relationship‑graph techniques to extract actionable insights for decision‑making.
The construction pipeline starts with a basic graph defining various behavior relations, followed by sub‑graph extraction (both homogeneous and heterogeneous) and quantitative edge‑strength calculation based on type weight, frequency, and recency.
Four layers of the graph architecture are described: basic graph, sub‑graphs, graph capabilities, and applications. Capabilities include multi‑hop queries, node influence scoring (degree centrality or PageRank), graph embedding (Node2vec for homogeneous graphs, Metapath2vec for heterogeneous graphs, and Side‑Info‑enhanced Node2vec), clustering, similarity measurement, and relation prediction.
Embedding models are evaluated qualitatively (semantic similarity) and quantitatively (offline scoring, manual labeling, and online A/B tests). Results show that side‑info‑enhanced embeddings improve similarity accuracy to 67% and increase click‑through rates by 3‑5% in recommendation slots.
Practical applications cover user‑house matching, house‑similar recommendations on detail pages and push notifications, multi‑degree queries for agents (e.g., “房客通”), and B‑side tools that help agents discover suitable houses for clients, leading to up to 20% higher adoption rates.
The presentation concludes with a summary of the graph‑based AI stack, its impact on recommendation and recommendation‑related metrics, and future directions for integrating more algorithms and richer side information.
DataFunTalk
Dedicated to sharing and discussing big data and AI technology applications, aiming to empower a million data scientists. Regularly hosts live tech talks and curates articles on big data, recommendation/search algorithms, advertising algorithms, NLP, intelligent risk control, autonomous driving, and machine learning/deep learning.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.