Boosting 1688’s “Guess You Like” with Wide‑ResNet and Batch Normalization

Background

"Guess You Like" is a classic recommendation scenario on 1688. The mobile homepage displays about 230,000 exposures per day, with roughly 72% of users clicking and an average of eight clicks per user, making it a major traffic source. The goal is to predict the most interesting items for each user based on real‑time behavior data to increase click‑through and purchase rates.

The overall framework (illustrated in the image below) collects real‑time user data in ABFS, passes it to TPP, performs a 1,000‑item recall in BE (coarse ranking), sends the candidates to RTP for online scoring, and finally displays the top 600 items (fine ranking). Offline model tuning is done on the Porsche platform before deployment.

ABFS (Ali Basic Feature Server) provides unified real‑time feature processing and statistics such as click, collect, add‑to‑cart, and aggregated counts.

TPP (The Personalization Platform) integrates RTP, iGraph, BE and other services to streamline data flow and reduce development cost.

BE (Basic Engine) is a vector‑based recall engine that retrieves items from multiple indexes within milliseconds.

iGraph stores large‑scale user features, category preferences, and hot‑item lists for infrequent updates.

RTP (Real Time Prediction) uses Swift incremental model transmission and real‑time BUILD indexing to generate features and score each item in the request list.

Porsche online learning platform is a Blink‑based distributed streaming framework that handles log processing, feature calculation, and real‑time model updates, feeding the refreshed model to RTP.

Model Overview

The core models discussed are Wide&Deep, PNN, DeepFM, and a custom Wide&ResNet.

Wide&Deep combines a linear (LR) wide component that memorizes feature co‑occurrences with a deep neural network (DNN) that generalizes to unseen combinations. It is the current online model for 1688’s recommendation.

PNN (Product‑based Neural Network) adds an explicit product layer to perform second‑order feature crossing, using either inner or outer product, requiring all embeddings to share the same dimension.

DeepFM merges the ideas of Wide&Deep and PNN by replacing the wide LR with a Factorization Machine (FM) for second‑order interactions while sharing embeddings with the DNN. FM handles sparse features effectively, but still requires uniform embedding dimensions.

Wide&ResNet is a novel architecture that replaces the DNN part of Wide&Deep with a ResNet‑style network featuring skip connections. Adding Batch Normalization (BN) to this structure significantly improves stability and AUC, as shown in offline experiments.

Training Data

Training data consists of user interactions on the 1688 homepage “Guess You Like” module from the seven days preceding the target date. Exposures that received a click are labeled 1; exposures without a click are labeled 0.

The feature system includes various behavior windows (1/3/5/7/15/30 days) and distinguishes B‑type users (business buyers) from C‑type consumers. B‑type users have features such as “is Taobao seller” but lack demographic attributes like age or gender, and product features do not include brand information.

B‑type users may have a “Taobao seller” flag.

Unlike C‑type users, B‑type users lack age, gender, and household attributes.

Products are generally non‑branded wholesale items.

Experiment Results

Offline experiments on the Porsche platform show that the Wide&ResNet model with BN outperforms the baseline Wide&Deep model by more than 1 % AUC on both training and test sets. After three incremental training rounds, AUC improves by 5 %–6 %.

Loss curves indicate that the BN‑enhanced model stabilizes below 0.3, whereas the baseline Wide&Deep loss fluctuates more widely.

Tuning Experience

Practical tips include:

Use Batch Normalization in the ResNet blocks to improve convergence and stability.

When embedding dimensions differ greatly (e.g., gender vs. high‑cardinality IDs), consider group‑wise product layers as done by the Taobao search team.

Iterative incremental training on fresh data can yield additional AUC gains.

Conclusion

The proposed Wide&ResNet architecture, enhanced with Batch Normalization, significantly improves the performance of the 1688 “Guess You Like” recommendation system, demonstrating the value of deep model innovations in large‑scale e‑commerce scenarios.