POG: Personalized Outfit Generation forFashion Recommendation at Alibaba iFashion

Wen Chen, Pipei Huang, Jiaming Xu, Xin Guo, Cheng Guo, Fei Sun, Chao Li, Andreas Pfadler, Huan Zhao, Binqiang ZhaoAlibaba Group, Beijing, China. (Applied Data Science Track, Poster No.5, August 6)

Introduction

There exist two requirements for fashion out-fit recommendation: the Compatibility ofthe generated fashion outfits, and the Per-sonalization in the recommendation pro-cess.We attempt to build the bridge between fash-ion outfit generation and recommendation.For the Compatibility requirement, we pro-pose a Fashion Outfit Model (FOM) by set-ting up a masked item prediction task basedon the self-attention mechanism. For thePersonalization requirement, we proposea Personalized Outfit Generation (POG)model, which uses a Transformer architec-ture to model both signals from user pref-erence and outfit compatibility. This is thefirst study to generate personalized outfitsbased on users’ historical behaviors with anencoder-decoder framework in an industrial-scale application at Alibaba iFashion (Fig-ure 1).

Figure 1:Alibaba iFashion.

Dataset

We provide three datasets describing theoutfits, the items, and the user behaviorsseparately. To the best of our knowledge,our dataset is the largest publicly availabledataset of fashion items with rich informa-tion compared to existing datasets.

Table 1:Statistics of the datasets.

Dataset #Outfits #Users #ItemsOutfit data 1,013,136 - 583,464Item data - - 4,747,039User data 127,169 3,569,112 4,463,302

FOM

Masking items one at a time in the outfits, werequire the model to fill in the blank with thecorrect item according to the context. Sinceevery item in the outfit is masked to fuseits left and right context, the compatibilitybetween each item and all the other itemswithin the outfit can be learned from the self-attention mechanism.Given an outfit F = {f1, . . . , ft, . . . , fn},where ft is the t-th item, we use a particu-lar embedding [MASK] for the masked item.Non-masked items are represented by theirmulti-modal embeddings. We then representthe set of input embeddings as Fmask.

POG

Take the advantage of encoder-decoderstructure, we aim to translate an user’s his-torical behaviors to a personalized outfit.Let U denote the set of all users and F bethe set of all outfits. We use a sequence ofuser behaviors U = {u1, . . . , ui, . . . , um} tocharacterize an user, where ui are the clickeditems by the user. F = {f1, . . . , ft, . . . , fn}is the clicked outfit from the same user,where ft are the items in the outfit. At eachtime step, we predict the next outfit itemgiven previous outfit items and user’s clicksequence on items U .

Figure 2:The architecture of FOM. We mask the items in the outfit one at a time.

Figure 3:The architecture of POG, which is an encoder-decoder architecture with a Per network and a Gen network.

We minimize the following loss function:

LF = −1n

n∑mask=1

log Pr(fmask|Fmask; ΘF )

where ΘF denotes the model parameters,and Pr(·) is the probability of choosing thecorrect item conditioned on the non-maskeditems. The model architecture is shown inFigure 2.

For pair (U, F ), the objective function ofPOG can be written as:L(U,F ) = −1

n

n∑t=1

log Pr(ft+1|f1, . . . , ft, U ; Θ(U,F )

)

where Θ(U,F ) denotes the model parameters.Pr(·) is the probability of seeing ft+1conditioned on both previous outfit itemsand user clicked items. The modelarchitecture is shown in Figure 3.

Dida Platform

We develop a platform named Dida whichis able to generate personalized outfits auto-matically. Dida is widely used by more thanone million operators at Alibaba. About 6million personalized outfits are generated ev-eryday with high quality. So far, the outfitshave been recommended to more than 5.4million users.

Figure 4:Dida platform.

Experiment

Our model significantly outperforms otheralternative methods through outfit compat-ibility experiments, including pushing theFITB (Fill In The Blanks) benchmark to68.79% (5.98% relative improvement) andCP (Compatibility Prediction) benchmarkto 86.32% (25.81% relative improvement).Through extensive online experiments, weshow that POG clearly outperforms the CFmethod by 70% increase in CTR (Click-Through-Rate) metric. The online cases canbe found in Figure 5.

Figure 5:The online cases of POG.

Contact Information

•Email: chenyu.cw@alibaba-inc.com