Cross‑Domain Recommendation and Heterogeneous Mixed‑Feed Ranking Practices in 58 Community

58 Community is a unified content platform under 58.com that aggregates various services such as real‑estate, jobs, cars, and local life, aiming to connect users across multiple scenarios.

The recommendation scenarios span multiple entry points in the app, displaying feeds that contain news, posts, topics, videos, and other heterogeneous items.

Cross‑domain recommendation is introduced to leverage user and item overlap between the target domain (58 Community) and source domains (other 58 business lines) to improve cold‑start performance, overall relevance, and diversity.

Three practical problems are identified: (1) how to use data from other business lines to boost Community recommendations, (2) how to jointly improve PGC and UGC content, and (3) how to rank multiple heterogeneous entities in a single feed.

Solution 1: Cross‑domain collaborative filtering based on Factorization Machines, where user one‑hot, item one‑hot, and source‑domain behavior multi‑hot features are concatenated, and embeddings are trained and served via Faiss, yielding ~8% conversion lift.

Solution 2: Attribute‑mapping using statistical co‑occurrence or word2vec‑derived tag embeddings to map source‑domain categories to Community tags, enabling a simple recall pipeline that improves UV conversion by ~4%.

Solution 3: Multi‑view DSSM model that shares a common user tower across domains while keeping domain‑specific item towers, achieving ~2% UV conversion gain.

For heterogeneous mixed‑feed ranking, a unified scoring model aligns scores from different sub‑models, applies bucketed feature handling, and merges wide‑&‑deep components into a single model, resulting in +1% AUC and +4% CTR.

Diversity is enhanced using two algorithms: Maximal Marginal Relevance (MMR) with a custom distance hierarchy and Determinantal Point Process (DPP) based on the paper “Fast Greedy MAP Inference for Determinantal Point Process to Improve Recommendation Diversity,” both delivering higher CTR, UV‑CTR, and average view counts.

The overall conclusion emphasizes that cross‑domain recommendation is a strategic approach akin to multi‑task and transfer learning, that algorithmic choices must align with business characteristics, and that diversity should be treated as a means to improve core business metrics.