Instant Interest Reinforcement and Extension for Taobao Detail Page Distribution

Taobao’s product detail page is one of the largest traffic modules in the mobile app, displaying billions of product details and serving as a crucial bridge between user decision making and platform flow distribution. Because the page handles massive traffic and external referrals, the platform designs several full‑network distribution scenarios within the detail page to enhance user experience and platform efficiency.

Background

In an information‑overloaded environment, recommendation systems act as the voice of massive content, responsible for quality control and user exposure. Modern recommenders deeply mine user behavior to capture real‑time interests, helping users quickly locate relevant items among billions of candidates.

Scenario Introduction

The full‑network distribution scenarios mainly include bottom‑of‑page information flow ("Neighbouring Good Products"), main‑image horizontal scroll, and add‑to‑cart pop‑up. These break the isolation of merchant private domains and improve cross‑domain distribution efficiency, while still protecting merchant interests by separating same‑store and cross‑store recommendation modules.

Technical Exploration

Algorithm Problem Definition – Instant Interest Reinforcement

When users enter a detail page, they exhibit a strong focus on the main product (trigger). Existing methods often replace the last behavior in a sequence with an inferred instant interest, which lacks the explicit intent provided by the trigger. This work strengthens the trigger information in various ways to better satisfy users’ immediate needs.

Recall

Inspired by deep‑learning‑based recall methods such as YouTube DNN, SDM, and User‑based Sequential Deep Match, we enrich the recall stage with trigger‑related features. By adding trigger‑filtered sequences (leaf and first‑level categories) to the original behavior sequence, we achieve offline HR improvements of +1.07% and +1.37%, and an online IPV lift of 1.1%.

Model – CIDM (Current Intention Reinforce Deep Match)

The CIDM model introduces four trigger‑aware components:

Trigger‑Layer: explicitly models the main product and fuses it with long‑ and short‑term user preferences.

Trigger‑Attention: replaces self‑attention with target‑attention centered on the trigger.

Trigger‑LSTM: injects trigger information into the LSTM cell and adds a trigger‑gate to bias memory toward the main product.

Trigger‑filter‑sequence: supplements the original sequence with trigger‑filtered category sequences.

Two variants of the trigger‑gate are explored: (1) feeding trigger as an additional input to the memory gate, and (2) adding a parallel instant‑interest gate. Both variants improve offline HR (+1.07% / +1.37%) and online IPV (+1.1%).

Variational Auto‑Encoder (VAE) for Sequence Denoising

To further exploit trigger information, we treat the sequence as a denoising problem and employ a VAE whose prior mean is set to the trigger embedding. The optimization objective combines KL divergence and reconstruction loss, encouraging the latent space to capture user intent while filtering out unrelated behaviors. This VAE‑based approach yields an offline HR gain of +2.23% (online not yet tested).

Effectiveness

Compared with the baseline SDM, CIDM improves online IPV by 4.69%.

Fine‑Ranking – DTIN (Deep Trigger‑based Interest Network)

DTIN builds on DIN by aligning trigger and candidate features, concatenating them into the wide part, and adding a dual‑attention mechanism: the first attention jointly processes trigger and candidate, while the second uses only the trigger as query. The element‑wise product of the two attention outputs serves as the sequence weighting vector. DTIN achieves online IPV +9.34% and offline AUC +4.6%.

Coarse‑Ranking – Tri‑tower (Triple‑tower Preparatory Ranking Framework)

To meet efficiency constraints in large‑scale recall, we extend the classic dual‑tower architecture with a third "trigger‑tower" that mirrors the item tower’s lightweight features. The three towers output logits that are combined before the final sigmoid. This lightweight model improves online IPV by 3.96% while maintaining fast inference.

Summary

Across recall, coarse‑ranking, and fine‑ranking, the common theme is to mine and reinforce the main product (trigger) information, thereby capturing users’ instant intent and balancing relevance with diversity. The proposed instant‑interest reinforcement pipeline consistently yields metric improvements and provides practical insights for industrial recommendation systems.

References

Tang, Jiaxi, et al. "Towards neural mixture recommender for long range dependent user sequences." WWW 2019.

Zhu, Yu, et al. "What to Do Next: Modeling User Behaviors by Time‑LSTM." IJCAI 2017.

Liang, Dawen, et al. "Variational autoencoders for collaborative filtering." WWW 2018.

Li, Xiaopeng, and James She. "Collaborative variational autoencoder for recommender systems." KDD 2017.

Zhao, Shengjia, Jiaming Song, and Stefano Ermon. "Towards deeper understanding of variational autoencoding models." arXiv 2017.

For further details or collaboration, feel free to contact the authors.