Multi‑Business Recommendation System for the Tongcheng App Home Page Waterfall Flow

The homepage waterfall‑flow recommendation on the Tongcheng travel app is a crucial module for attracting users and converting traffic, providing personalized hotel, scenic spot, vacation, and content suggestions based on the user's location or selected city.

The overall recommendation pipeline (Figure 2) consists of basic data support, city‑intent determination, resource recall, and ranking. Basic data include static and dynamic attributes of resources, user demographics, behavior sequences, and statistical features, both offline and real‑time.

City‑intent modeling scores candidate cities (recent searches, travel orders, residence, location, and manually selected city) and selects the top‑N cities for separate resource recall.

Resource recall allocates quotas per city based on intent scores and applies different strategies for hotels, scenic spots, vacations, and content, such as popular items, similar items, nearby locations, and dual‑tower vector recall, enabling parallel multi‑path recall across business lines.

The ranking model employs a multi‑task learning framework with three components: feature preprocessing, a PLE‑CGC multi‑business sub‑network, and an ESMM structure for joint CTR and CVR prediction.

Feature preprocessing embeds discrete features, buckets numeric features, and models user behavior sequences with self‑attention and target‑attention. Geohash encoding via GRU captures geographic similarity.

The multi‑business sub‑network separates each business line into its own expert network while sharing a common expert, allowing the main network to learn user preferences and each sub‑network to learn line‑specific characteristics; gates weight the shared and private experts.

The ESMM component addresses sample selection bias by modeling the entire exposure‑click‑conversion space, combining a CTR sub‑network and a CVR sub‑network to output the final exposure‑to‑conversion (CTCVR) probability for ranking.

Online A/B testing shows a 0.6 percentage‑point increase in CTR and a 0.7 percentage‑point increase in conversion rate compared with a simpler ESMM model without multi‑business sub‑networks (Figure 4).

Future work includes enhancing recall algorithms with multi‑view DSSM or MIND‑style models, optimizing additional objectives such as novelty and depth of user engagement, and addressing cold‑start challenges for new users.