Meta Domain Adaptation Approach for Multi-Domain Ranking
In a real industry recommendation system, the distribution of recommended domains is very redundant. Different domains may address the same problem, such as the Click-Through Rate (CTR) prediction, and may share the same features. Multi-domain learning aims to use one model to complete the same task...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10975759/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In a real industry recommendation system, the distribution of recommended domains is very redundant. Different domains may address the same problem, such as the Click-Through Rate (CTR) prediction, and may share the same features. Multi-domain learning aims to use one model to complete the same task of all domains, in order to improve the model effect and achieve savings in computing resources and human effort. However, different domains have different data distributions and insufficient training data. In constructing a multi-domain model encompassing several domains with small sample sizes, how to make the best use of the existing data of other domains and ensure the generalization performance of the model in the training process has become the key problem of the research. To meet the challenge, in this paper, we propose a Meta Domain Adaptation model for Multi-Domain ranking (MDAMD), specifically designed for CTR prediction across multiple domains. It employs the unsupervised domain adaptation method to ensure the global unity of training for model optimization among all domains. Maximum Mean Discrepancy (MMD) loss calculation is added after the gate unit to ensure adaptive training of the model in feature extraction. Meanwhile, MDAMD applies a meta-learning method to avoid overfitting in the domain adaptation stage. Results from the industrial and public datasets demonstrate the effectiveness of MDAMD. Moreover, our model has been deployed online and obtained the CTR and Gross Merchandise Volume (GMV) improvements. |
|---|---|
| ISSN: | 2169-3536 |