Explainable Recommendation Algorithms at Alibaba Health: System Design, Feature Engineering, and Experimental Results

1. Recommendation Business Overview Alibaba Health operates self‑operated and industry stores with products ranging from regular goods to OTC and prescription drugs. Regulatory constraints prohibit promotional terms, so explanations rely on product attributes such as function, usage, and sales metrics.

2. Basic Data Preparation Features are extracted from product titles, detail images (via OCR), user comments, and drug instructions. Keywords are merged, de‑duplicated, and manually verified to build a standard tag dictionary. Feature encoding uses word2vec embeddings and defines positive samples (browse‑then‑purchase pairs) and negative samples (common browse or purchase pairs).

3. Explainable Recommendation Model The system combines a selection model (MLP with sparsemax) and a prediction model (deep FM‑style architecture with a deep network and a CNN‑based cross layer). The selection model highlights important feature distributions, while the prediction model predicts click probability using deep and cross features. Sparsemax provides sparse importance scores, improving interpretability.

4. Experimental Results Offline evaluation shows an AUC of 0.74. Online, the new algorithm improves predicted CTR by 9.13% and user‑level CTR by 3.4% when both the baseline CTR model and the new model exceed a threshold, with explanation texts displayed to users.

5. Q&A Highlights Discussed standard tag generation, use of LIME for explanations, linking of selection and prediction models, manual generation of explanation texts, and future directions such as template‑based slot filling and automated text generation.