Optimizing Vector Recall for Feizhu's Homepage "You May Like" Recommendation Feeds

Li Xi, a senior algorithm engineer from Alibaba's Feizhu recommendation team, shares recent optimizations made in the recall stage of Feizhu's homepage "You May Like" feeds.

The Feizhu homepage feeds mix various material types (products, hotels, POI, videos, etc.) and rely on a multi‑stage recommendation pipeline (recall, coarse ranking, fine ranking). Recall serves as the first funnel, quickly filtering massive candidates to a manageable set for downstream ranking.

Current industrial practice uses multi‑path recall to ensure diversity, including I2I recall, attribute recall, geographic recall, travel‑status recall, hot recall, cold‑start and volume‑guarantee strategies, with vector recall being a key focus.

Vector recall models the problem as nearest‑neighbor search in a low‑dimensional embedding space learned for users and items, which has been shown to improve diversity and generalization.

v1 – Dual‑Tower Model : A classic two‑tower architecture decouples user and item feature interactions. The user tower ingests basic user attributes, historical behavior sequences (processed with user‑attention), and context features to produce a user embedding. The item tower incorporates item attributes, statistical efficiency features, and sub‑material features (product, hotel, POI) via a gating mechanism, producing an item embedding. Training uses random negative sampling to mitigate sample‑selection bias between offline training and online exposure.

v2 – Enhanced‑Vector Dual‑Tower : Inspired by Meituan's approach, enhanced vectors are introduced on both towers to enable early feature interaction. The item‑side enhanced vector, representing aggregated information from all items a user has positively interacted with, yields measurable offline gains and online click‑through improvements, while the user‑side enhanced vector adds little benefit due to redundancy with existing behavior features.

v3 – Unbiased Recall Framework : To address bias in negative sampling, the team adopts an Inverse Propensity Weighting (IPW) scheme and multi‑task learning. The framework splits the candidate space into four mutually exclusive zones (clicked, exposed‑but‑not‑clicked, recalled‑but‑not‑exposed, never‑recalled) and assigns different weights to negatives from each zone. Two auxiliary tasks estimate the probabilities of transitioning from the full space to recall space and from recall space to exposure space, respectively. This work was presented as a short paper at SIGIR 2022.

v4 – State‑Aware Deep Recall : Recognizing that user travel state (seed, decision, pre‑travel, in‑travel) strongly influences preferences, the model predicts the next travel state and employs four expert networks, each specialized for a particular state. A state‑transition probability module processes historical click and state sequences, applies mean‑pooling and self‑attention, and outputs a probability distribution over the four states. State‑specific user embeddings are generated via feature‑level attention on user attributes and self‑attention on behavior sequences, then adaptively aggregated using the predicted state probabilities. The final similarity is computed by inner‑product of normalized user and item embeddings. This approach was presented at CIKM 2022.

Extensive offline experiments on Feizhu's internal data and public Taobao recommendation datasets show consistent superiority of each proposed version over baseline recall methods. Online A/B tests confirm stable click‑through rate gains when deploying the v3 and v4 models in Feizhu's "You May Like" sub‑scenes.

The presentation concludes with a thank‑you note to the audience.