How Cross‑Domain Embedding Boosts New User Recommendations in Alibaba’s Ecosystem

1. Background

Alipay’s “new user acquisition” strategy uses the “Oasis” marketing algorithm platform to select promotional items and display them personalized on Alipay’s homepage, aiming to give users a good shopping first experience and guide them to install the Taobao app.

New users lack behavior data and key profile features, making personalized targeting difficult. However, many of these users are mature Alipay users, so their offline consumption categories can be known via full‑domain account linking. We propose a cross‑domain user representation method called Cross‑Domain Embedding (CSDE), which generates vectors in the Taobao domain from Alipay features. Experiments and online tests show that CSDE improves representation for new users and yields more accurate recommendations.

In this article, “active”, “new”, and “lost” users refer to classifications within the Taobao app.

2. Algorithm Overview

2.1 Basic Idea

Our goal is to generate a Taobao‑domain profile for users who have little activity in Taobao by leveraging their Alipay features. The model consists of two stages: Cross‑Domain Embedding Learning (CSDE) , which learns to map Alipay features to Taobao features, and Adaptive Prediction with CSDE , which incorporates the generated embeddings into a deep prediction model to improve conversion‑rate (CVR) estimation.

2.2 Notation

Alipay‑domain user features (e.g., investment ability, Alipay consumption frequency, offline preference categories).

Taobao‑domain user features (e.g., purchase power, user tags, recent click sequences).

i – item features such as category, seller, price range, gender attribute, historical sales.

y – label indicating whether the exposure leads to a purchase.

Full user dataset covering active, lost, and new users.

Taobao active‑user dataset with rich behavior.

Taobao new‑user dataset with no behavior.

2.3 CSDE Learning

The loss function combines three terms: Lt (Taobao classification loss), Ls (Similarity Metric loss) and Lc (CSDE classification loss). Similarity Metric loss can be either an L2 distance or an adversarial discriminator that treats generated Taobao vectors as “fake” and real Taobao vectors as “real”. Experiments show that the distance‑based metric works better for our scenario.

2.4 Adaptive Prediction with CSDE

We train on all users Da = {x_i, y} to predict CVR and evaluate on active, new, and lost user groups. The CSDE network is injected into the user‑embedding pipeline, and a domain‑attention mechanism balances the contributions of Alipay and CSDE embeddings, especially benefiting new users with missing Taobao features.

3. Experiments and Online Tests

Both CSDE learning and adaptive prediction use 7‑day Alipay mini‑program logs as training data to predict next‑day conversion.

3.1 AUC Comparison

Compared with a baseline model without CSDE, the CSDE‑enhanced model achieves higher AUC (see figure).

3.2 Online Test

Deploying the CSDE‑augmented model to new‑user buckets improves conversion rate over the baseline bucket (see figure).

3.3 Attention Analysis

For high‑activity users, both Alipay and CSDE attentions are low because Taobao features dominate. For low‑activity or new users, CSDE attention is significantly higher, indicating the generated Taobao profile is crucial.

4. Exploring GAN‑Based Cross‑Domain User Representation

4.1 Objective

We investigate whether a generative adversarial network (GAN) can translate user behavior sequences between Taobao (Domain A) and Alipay (Domain B) without paired data.

4.2 Implementation

A bidirectional Cycle‑GAN‑style architecture maps A→B and B→A using fully‑connected generators and discriminators, with Wasserstein GAN loss and gradient penalty. The generated embeddings are fed directly into the target‑domain prediction task.

4.3 Results

Visualization shows that the generated A‑by‑B distribution closely matches the original A distribution. In CVR prediction, removing A features drops AUC from 0.6494 to 0.6056; filling the gap with generated A‑by‑B raises AUC to 0.6293.

5. Conclusion

We propose a novel recommendation model for new users that leverages cross‑domain embeddings and domain attention to compensate for missing features, achieving significant CVR improvement in both offline experiments and online deployment.