Generalized Recommendation Solution for Transaction Scenarios

DeWu is an e‑commerce platform focused on trend culture. Its transaction recommendation module is responsible for traffic distribution and directly influences user experience, technical reputation, and brand image.

The rapid expansion of transaction scenarios (e.g., waterfall flow, gold‑position, category tabs) has led to many small‑scale scenes (female channel, luxury channel, brand channel) that still rely on basic, non‑personalized strategies. Maintaining dozens of data tables across ~30 modules makes upgrades difficult, prompting the need for a universal personalized recommendation algorithm.

2.1 Current Situation of Small Transaction Scenes

These scenes feature complex item hierarchies (products, brands, series, themes, tabs) but simple business logic, mainly recommending items based on positive user actions (click, collect, purchase) or related attributes via ICF inverted indexes. Maintenance is hard due to the large number of tables and manual effort.

2.2 Common Recall Algorithms

Two main recall families are considered:

i2i (item‑to‑item) Recall – offline similarity computation between items, storing top‑N similar items in an inverted index. Algorithms include ItemCF, static embeddings (Word2Vec, GloVe, FastText), dynamic embeddings (ELMo, GPT, BERT), and graph embeddings (DeepWalk, Node2vec, GCN, GraphSAGE). Advantages: simple, quick to launch. Disadvantage: a single inverted index cannot satisfy all small scenes.

u2i (user‑to‑item) Recall – offline user model training to infer item embeddings, followed by online ANN retrieval. Typical algorithms are YouTube‑style recall and DSSM. Advantages: flexible vector‑search service, can recommend items even without prior user interaction. Disadvantage: higher engineering complexity.

2.3 Determined General Recommendation Scheme

The chosen solution is a u2i model based on DSSM for universal recall, combined with a waterfall‑flow ranking model. The system can personalize recommendations for products, brands, categories, themes, and series, distinguish scenarios by scene‑id, and handle dimension‑specific recalls (category, series). A unified sampling logic enables easy extension to new scenes and future DSSM upgrades.

3 Algorithm Principles

DSSM (Dual‑Tower) maps queries and documents (or users and items) into a shared low‑dimensional space using DNNs, then computes relevance via cosine similarity. It originated from Microsoft’s 2013 paper on web search and is now a classic in recommendation.

4 Practical Dual‑Tower Model

The dual‑tower architecture splits user (including context) and item features into two separate towers, each producing an embedding. Similarity is calculated by inner product or cosine similarity.

4.2 Negative Sample Mining

Four strategies are discussed: in‑batch negatives, global random sampling, a mix of both, and a combination of global random sampling with exposure‑non‑click negatives. The implementation adopts the fourth mixed approach, adding category‑aware sampling and exposure‑based weighting (exposure^0.75) for negatives, while positives use click, collect, and purchase signals.

4.3 Embedding Norm

Applying L2‑norm to both user and item embeddings (i.e., using cosine similarity) yields better performance, so the model normalizes embeddings during training and inference.

4.4 Model Deployment and Monitoring

After training, the user tower is served online for real‑time embedding inference, while the item tower pre‑computes item embeddings stored in an ANN index. Monitoring covers model serving health, data volume fluctuations, and performance metrics (e.g., 3σ rule on AUC). Anomaly‑blocking mechanism halts downstream tasks when thresholds are breached.

5 Reflections

Deployed DSSM models have increased CTR by over 10% across small scenes. Future work includes expanding optimization objectives beyond CTR and designing separate DSSM models for different recommendation dimensions (e.g., brand, category).