Adversarial Feature Translation for Multi-domain ...

1 Adversarial Feature Translation for Multi-domainRecommendationXiaobo Hao WeChat, TencentBeijing, Liu WeChat, TencentBeijing, Xie WeChat, TencentBeijing, GeWeChat, TencentBeijing, TangWeChat, TencentBeijing, ZhangWeChat, TencentBeijing, LinWeChat, TencentBeijing, super platforms such as Google and WeChat usuallyhave different recommendation scenarios to provide heterogeneousitems for users diverse demands. Multi-domain recommendation(MDR) is proposed to improve all recommendation domains simul-taneously, where the key point is to capture informative domain-specific features from all domains.

2 To address this problem, we pro-pose a novel Adversarial Feature Translation (AFT) model for MDR,which learns the Feature translations between different domainsunder a generative Adversarial network framework. Precisely, inthe Multi-domain generator, we propose a domain-specific maskedencoder to highlight inter-domain Feature interactions, and thenaggregate these features via a transformer and a domain-specificattention. In the Multi-domain discriminator, we explicitly modelthe relationships between item, domain and users general/domain-specific representations with a two-step Feature Translation inspiredby the knowledge representation learning.

3 In experiments, we eval-uate AFT on a public and an industrial MDR datasets and achievesignificant improvements. We also conduct an online evaluation ona real-world MDR system. We further give detailed ablation testsand model analyses to verify the effectiveness of different compo-nents. Currently, we have deployed AFT on WeChat Top source code is in CONCEPTS Information systems Recommender systems. All authors contributed equally to this research. Ruobing Xie is the correspondingauthor to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page.

4 Copyrights for components of this work owned by others than ACMmust be honored. Abstracting with credit is permitted. To copy otherwise, or republish,to post on servers or to redistribute to lists, requires prior specific permission and/or afee. Request permissions from 21, August 14 18, 2021, Virtual Event, Singapore 2021 Association for Computing ISBN 978-1-4503-8332-5/21/08.. $ system; Multi-domain recommendation; GANACM Reference Format:Xiaobo Hao, Yudan Liu, Ruobing Xie, Kaikai Ge, Linyao Tang, Xu Zhang,and Leyu Lin. 2021. Adversarial Feature Translation for Multi-domain Rec-ommendation.

5 InProceedings of the 27th ACM SIGKDD Conference on Knowl-edge Discovery and Data Mining (KDD 21), August 14 18, 2021, Virtual Event, , New York, NY, USA, 10 pages. INTRODUCTIONR ecommendation systems have been widely used in various real-world scenarios for users to get information and entertainment [14].In recent years, the Matthew effect has spawned some influential su-per platforms such as Google, Twitter and WeChat. These platformsoften have various recommendation scenarios to provide heteroge-neous items. For example, the WeChat platform contains multiplerecommendation functions on different domains including videos,articles and mini-programs.

6 User behaviors in different recommen-dation domains could be linked via users shared accounts after userapproval. These Multi-domain behaviors should be jointly consid-ered to learn both user s general and domain-specific preferences,which are beneficial for recommendation in all recommendation (MDR) aims to improve all rec-ommendation domains simultaneously [17]. The key challenge inMDR is how to capture informative domain-specific features fromall domains for the target domain. An intuitive method is to jointlycombine Multi-domain behaviors as input features, while it caresless about the inter-domain Feature interactions.

7 Recently, [10,32]jointly consider multiple user behaviors with attention and graphconvolutional networks. Multi-task learning (MTL) is also an effec-tive method to jointly make full use of Multi-domain informationin different domains, where the recommendation in each domain istreated as a task [15,16]. A good MDR algorithm should learn theimplicit relationships between heterogeneous features and user be-haviors in different domains. However, MDR severely suffers fromits inherent sparsity issues, which derive from (1) the sparsity ofADS Track PaperKDD 21, August 14 18, 2021, Virtual Event, Singapore2964user-item pair-wise click behaviors, and (2)

8 The sparsity of featureinteractions within and between different masked encoderMulti-domain GeneratorTransformer Feature extractorDomain-specific aggregationItem/domain Feature extractorMulti-domain Discriminator1st Feature translation2nd Feature translationTop-k fakeclicked items in multiple domainsReal clickeditems in multiple domainsFakeclicksRealclicksMulti-domain items videoarticleAPPAPPG enerate diverse challenging fake items to confuse the discriminatorDistinguish real/fake clicked items via Multi-domain informationFigure 1: The overall framework of AFT in address these issues and improve all recommendation do-mains, we propose a novelAdversarial Feature Translation (AFT)framework for MDR.

9 As shown in Fig. 1, AFT highlights useful fea-ture interactions between different domains via a domain-specificmasked encoder and a two-step Feature Translation under a gen-erative Adversarial network (GAN) framework. Specifically, in thegenerator, we first design a domain-specific masked encoder, whichmerely masks behaviors in the target domain. It forces AFT to learnfrom informative features in other domains. Next, a transformerextractor and a domain-specific aggregation are conducted to learnuser aggregated preferences for the target domain. These domain-specific preferences are then used to generate the fake clicked itemswhich compete with the real clicked items in the discriminator.

10 Areinforcement learning (RL) method is conducted to update thegenerator. In the discriminator, inspired by the triplet-based train-ing in knowledge representation learning (KRL) [5], we learn item,domain and users general/domain-specific representations in thesame semantic space via two Feature translations. These translationsexplicitly model the interpretable relations among(item, domain,user)and(user general preference, domain information, user domain-specific preference). Precisely, we first combine user preferences onitems and domains to form the user representation, which reflectsthe user s general preferences in all domains.