Abstract
Feed recommendation is currently the mainstream mode for many real-worldapplications (e.g., TikTok, Dianping), it is usually necessary to model andpredict user interests in multiple scenarios (domains) within and even outsidethe application. Multi-domain learning is a typical solution in this regard.While considerable efforts have been made in this regard, there are still twolong-standing challenges: (1) Accurately depicting the differences amongdomains using domain features is crucial for enhancing the performance of eachdomain. However, manually designing domain features and models for numerousdomains can be a laborious task. (2) Users typically have limited impressionsin only a few domains. Extracting features automatically from other domains andleveraging them to improve the predictive capabilities of each domain hasconsistently posed a challenging problem. In this paper, we propose anAutomatic Domain Feature Extraction and Personalized Integration (DFEI)framework for the large-scale multi-domain recommendation. The frameworkautomatically transforms the behavior of each individual user into anaggregation of all user behaviors within the domain, which serves as the domainfeatures. Unlike offline feature engineering methods, the extracted domainfeatures are higher-order representations and directly related to the targetlabel. Besides, by personalized integration of domain features from otherdomains for each user and the innovation in the training mode, the DFEIframework can yield more accurate conversion identification. Experimentalresults on both public and industrial datasets, consisting of over 20 domains,clearly demonstrate that the proposed framework achieves significantly betterperformance compared with SOTA baselines. Furthermore, we have released thesource code of the proposed framework at https://github.com/xidongbo/DFEI.