Deep Interest Evolution Network for Click-Through Rate ...

1 deep Interest Evolution Network for Click-Through Rate PredictionGuorui Zhou*, Na Mou , Ying Fan, Qi Pi, Weijie Bian,Chang Zhou, Xiaoqiang ZhuandKun GaiAlibaba Inc, Beijing, China{ , , , , , , , rate (CTR) prediction, whose goal is to esti-mate the probability of a user clicking on the item, has be-come one of the core tasks in the advertising system. ForCTR prediction model, it is necessary to capture the latentuser Interest behind the user behavior data. Besides, consid-ering the changing of the external environment and the in-ternal cognition, user Interest evolves over time are several CTR prediction methods for Interest mod-eling, while most of them regard the representation of behav-ior as the Interest directly, and lack specially modeling forlatent Interest behind the concrete behavior.}

2 Moreover, littlework considers the changing trend of the Interest . In this pa-per, we propose a novel model, named deep Interest Evolu-tion Network (DIEN), for CTR prediction. Specifically, wedesign Interest extractor layer to capture temporal interestsfrom history behavior sequence. At this layer, we introducean auxiliary loss to supervise Interest extracting at each user interests are diverse, especially in the e-commercesystem, we propose Interest evolving layer to capture inter-est evolving process that is relative to the target item. At in-terest evolving layer, attention mechanism is embedded intothe sequential structure novelly, and the effects of relative in-terests are strengthened during Interest Evolution . In the ex-periments on both public and industrial datasets, DIEN sig-nificantly outperforms the state-of-the-art solutions.

3 Notably,DIEN has been deployed in the display advertisement systemof Taobao, and obtained improvement on per click (CPC) billing is one of the commonestbilling forms in the advertising system, where advertisers arecharged for each click on their advertisement. In CPC adver-tising system, the performance of Click-Through rate (CTR)prediction not only influences the final revenue of wholeplatform, but also impacts user experience and CTR prediction has drawn more and more atten-tion from the communities of academia and most non-searching e-commerce scenes, users do notexpress their current intention actively. Designing model to*Corresponding author is Guorui Zhou. This author is the one who did the really hardwork for Online Testing.

4 The source code is available 2019, Association for the Advancement of ArtificialIntelligence ( ). All rights user s interests as well as their dynamics is the keyto advance the performance of CTR prediction. Recently,many CTR models transform from traditional methodolo-gies (Friedman 2001; Rendle 2010) to deep CTR mod-els (Guo et al. 2017; Qu et al. 2016; Lian et al. 2018). Mostdeep CTR models focus on capturing interaction betweenfeatures from different fields and pay less attention to userinterest representation. deep Interest Network (DIN) (Zhouet al. 2018c) emphasizes that user interests are diverse, ituses attention based model to capture relative interests to tar-get item, and obtains adaptive Interest representation.

5 How-ever, most Interest models including DIN regard the behav-ior as the Interest directly, while latent Interest is hard tobe fully reflected by explicit behavior. Previous methods ne-glect to dig the true user Interest behind behavior. Moreover,user Interest keeps evolving, capturing the dynamic of inter-est is important for Interest on all these observations, we propose deep Inter-est Evolution Network (DIEN) to improve the performanceof CTR prediction. There are two key modules in DIEN,one is for extracting latent temporal interests from explicituser behaviors, and the other one is for modeling interestevolving process. Proper Interest representation is the foot-stone of Interest evolving model. At Interest extractor layer,DIEN chooses GRU (Chung et al.)

6 2014) to model the depen-dency between behaviors. Following the principle that inter-est leads to the consecutive behavior directly, we proposeauxiliary loss which uses the next behavior to supervise thelearning of current hidden state. We call these hidden stateswith extra supervision as Interest states. These extra super-vision information helps to capture more semantic meaningfor Interest representation and push hidden states of GRU torepresent interests effectively. Moreover, user interests arediverse, which leads to Interest drifting phenomenon: user sintentions can be very different in adjacent visitings, and onebehavior of a user may depend on the behavior that takeslong time ago. Each Interest has its own Evolution , the click actions of one user on different targetitems are effected by different parts of interests.

7 At inter-est evolving layer, we model the Interest evolving trajectorythat is relative to target item. Based on the Interest sequenceobtained from Interest extractor layer, we design GRU withattentional update gate (AUGRU). Using Interest state andtarget item to compute relevance, AUGRU strengthens rela- [ ] 16 Nov 2018tive interests influence on Interest Evolution , while weakensirrelative interests effect that results from Interest the introduction of attentional mechanism into updategate, AUGRU can lead to the specific Interest evolving pro-cesses for different target items. The main contributions ofDIEN are as following: We focus on Interest evolving phenomenon in e-commerce system, and propose a new structure of net-work to model Interest evolving process.

8 The model forinterest Evolution leads to more expressive Interest repre-sentation and more precise CTR prediction. Different from taking behaviors as interests directly, wespecially design Interest extractor layer. Pointing at theproblem that hidden state of GRU is less targeted for inter-est representation, we propose one auxiliary loss. Auxil-iary loss uses consecutive behavior to supervise the learn-ing of hidden state at each step. which makes hidden stateexpressive enough to represent latent Interest . We design Interest evolving layer novelly, where GPUwith attentional update gate (AUGRU) strengthens the ef-fect from relevant interests to target item and overcomesthe inference from Interest the experiments on both public and industrial datasets,DIEN significantly outperforms the state-of-the-art solu-tions.

9 It is notable that DIEN has been deployed in Taobaodisplay advertisement system and obtains significant im-provement under various WorkBy virtue of the strong ability of deep learning on featurerepresentation and combination, recent CTR models trans-form from traditional linear or nonlinear models (Fried-man 2001; Rendle 2010) to deep models. Most deep mod-els follow the structure of Embedding and Multi-ayer Per-ceptron (MLP) (Zhou et al. 2018c). Based on this basicparadigm, more and more models pay attention to the in-teraction between features: Both Wide & deep (Cheng et ) and deep FM (Guo et al. 2017) combine low-orderand high-order features to improve the power of expression;PNN (Qu et al.)

10 2016) proposes a product layer to captureinteractive patterns between interfield categories. However,these methods can not reflect the Interest behind data (Zhou et al. 2018c) introduces the mechanism of at-tention to activate the historical behaviors given targetitem locally, and captures the diversity characteristic of userinterests successfully. However, DIN is weak in capturingthe dependencies between sequential many application domains, user-item interactions canbe recorded over time. A number of recent studies showthat this information can be used to build richer individualuser models and discover additional behavioral patterns. Inrecommendation system, TDSSM (Song, Elkahky, and He2016) jointly optimizes long-term and short-term user inter-ests to improve the recommendation quality; DREAM (Yuet al.