• Have any questions?
  • info.zbook.org@gmail.com

Time To Shop For Valentine’s Day: Shopping Occasions And .

6m ago
26 Views
0 Downloads
7.18 MB
26 Pages
Last View : 3d ago
Last Download : n/a
Upload by : Bria Koontz
Share:
Transcription

Time to Shop for Valentine’s Day:Shopping Occasions and SequentialRecommendation in E-commerceJianling Wang, Raphael Louca*, Diane Hu*, Caitlin Cellier*,James Caverlee and Liangjie Hong*Texas A&M University* Etsy Inc.

Recommender Systems usually: Capture users' intrinsic preference from their long-term behaviorpatterns.Infer users’ current needs by emphasizing recent actions.

In E-commerceHowever, intrinsic user behavior may be shifted by occasions, such asbirthdays, anniversaries, or gifting celebrations (Valentine's Day orMother's Day).

In E-commerceOften, these occasion-based purchases deviatepreferences and are not related to recent actions.fromlong-term

Idea: Let’s incorporate occasion signals! Recommend more time or season-aware candidates, which mayalleviate the cold-start problem. Reduce the noise in modeling users' intrinsic preferences. Recommend relevant items to the user for upcoming reoccurringoccasions.

Global Occasions Happen at the same time for a large number of users. Encourage or lead to similar shopping decisions for crowds of users. Examples: Christmas, Valentine’s Day,

Personal Occasions Happen at different times for different users. Occur in a periodic and repeated pattern for a specific user. Examples: Birthdays (for you or your friends) and anniversaries.

To Illustrate: Temporal Shopping TrendsUsers' shopping preferences are dynamic and can reflect reoccurringoccasions (festivals, holidays, seasonal activities).

To Illustrate: From Personal PerspectiveThere are occasions that may re-occur within a certain period (e.g.annually or monthly) over the course of multiple years and trigger relevantpurchase. traceable patterns in personal occasion signals.Time Gap between Purchasesfor Wedding and Anniversarywithin a year. More than 50% ofpurchases for anniversary arenear the date of weddingpurchase within a time windowless than 30 days.

Our Proposal:Occasion-Aware Recommendation (OAR)Intrinsic Preference Personal Occasion Signals Global Occasion Signals

Occasion-Aware Recommendation (OAR) Model the repeated personal occasion signals with attention layers. Model the global occasion signals by memorizing the temporal trendsof shopping behaviors. With a gating component, we balance the global and local effects ofdifferent occasions. Experiments on Etsy and Amazon.

Attention Mechanisms Query Key-Value Pairs Scoring Function - Calculating similarity/relationship between a Query& a KeyValuesOutputsk1k2k3 q1 q2 q3v1v2v3o latexit sha1 base64 "KHLFA9dJh8UXBjlmeG4gSSqYavE " 0HW/ncLa St4Px7cxvP3FtRKwecJJwP4KhEqFggFbq9kAmI MMmnpV5qeAJsDEPetVRBxI2fzU e0jOrDGgYa1sK6Vz9PZFBZMwkCmxnBDgyy95M/M/rphhe 5lQSYpcscWiMJUUYzr7nw6E5gzlxBJgWthbKRuBBoY2pZINwVt ag8 K8Ox L1oKTzxyTP3A fwBTE5FH /latexit latexit sha1 base64 "vpeHugvKEkgkR/0j0altVfMTXNo " 0HW/ncLa Jllbyb QxlWwI3vLLq6RVq3oX1dr9ZaV k8dRhBM4hXPw4ArqcAcNaAKDBJ7hFd6c1Hlx3p2PRWvByWeO4Q czx9BnZHS /latexit latexit sha1 base64 "jAR3sPEfMoURVaK7hqHc6/mT8KA " lsJ 3SzSbsTsQS je8eFDEq3/Gm//GbZuDtj4YeLw3w8y8IJHCoOt O4W19Y3NreJ2aWd3b/ at7nCTQi9hQiVBwhlbyfYQnDMJsPO3X WKW3XnoKvEy0mF5Gj2y1/ eXc Fq0FJ585Jn/gfP4ARKWR1A /latexit latexit sha1 base64 "BWkoXlOkYejmq6A3DcUbfxqunms " AAAB8XicbVDLTgJBEOzFF tX3D9omCjRjNdZJCPd8qnhUiheR4GSt2LNaehL3vRH11O/ ci1EZG6w3HMuyEdKBEIRtFK9x3kT wSSfFDqJ4TFlIzrgbUsVDbnpprOLJ TEKn0SRNqWQjJTf0 kNDRmHPq2M6Q4NIveVPzPaycYXHZToeIEuWLzRUEiCUZk qiyMPR3AMp DBBVThBmpQBwYKnuEV3hzjvDjvzse8NedkM4fwB87nDx2HkTQ /latexit kL latexit sha1 base64 "XeR3m7IUyoUIUJ0yLNPrDRZNyDQ " AAAB83icbVA9SwNBEJ3zM8avqKXNYhCswl0UtAzaWFhEMB oRBmI0mvbteqexW3BnIMvFyUoYc9V7py /HLI24QiapMR3PTbCbUY2CST4p WFvJWxINWVoYyraELzFl5dJs1rxzivV fOYI/cD5/AGqJke0 /latexit Softmax latexit sha1 base64 "MYieos1SHUd3AQE445MyTQ0XbWo " 0HW/ncLa COxjfzvz2I2gjYvWAkwR6ERsqEQrO0Eq j/CEQZiNp/1av1xxq 4cdJV4OamQHI1 aQIii8WhamkGNNZAHQgNHCUE0sY18LeSvmIacbRxlSyIXjLL6 SVq3qXVRr95eV OyR84nz9DIZHT /latexit latexit sha1 base64 "iTEZbT2BAe/Lm 3dV9kMoHumgPI " 0HW/ncLa C2ZgOeddSRSNu/Gx 8pScWWVAwljbUkjm6u JjEbGTKLAdkYUR2bZm4n/ed0Uw2s/EypJkSu2WBSmkmBMZv TgdCcoZxYQpkW9lbCRlRThjalkg3BW355lbRqVe gtOPnMMf B8/gBUmJFI /latexit latexit sha1 base64 "HjnYOqDG 9mWsRnkxKGgT7/tGPs " lsN 3STTbuToQS jO8eFDEq7/Gm//GbZuDtj4YeLw3w8y8IJHCoOt O4W19Y3NreJ2aWd3b/ 75a/eQLE04jEyCcZ0PTdBPwONgkk B2RoAjs zNxP 8borhlZ 6 Rfn3flYtBacfOaY/IHz QNWHZFJ /latexit qL latexit sha1 base64 "kkjpS9t58amNksFaPDZ g H llA " MMp 1EUYwjTlvR6Gbqt56o0kyKBzNOaBjjQLA JYw7U39PZBhrPY4j2xmjGepFbyr 5wWp6V FGRNJaqgg80X9lLtGutP/3R5TlBg AGPlxCDW6hDg0gIOEZXuHNMc6L8 58zFtXnHzmCP7A fwBfBqRYg /latexit latexit sha1 base64 "LBrO2AG9JVSYbXOTEEaiIcFSFcM " 120i7dbMLupFhC/4YXD4p49c9489 /OxaF1z8pkT8gfO5w9SapHd /latexit latexit sha1 base64 "aZYsu15SPJ5E sMT1cDcJ3BeBvA " 120y7dbMLupFhC/4YXD4p49c9489 wbkQHSoSCUbSS7yN/wiDMxtNetVcquxV3DrJKvJyUIUe9V/ry zFLI66QSWpMx3MT7GZUo2CST4t woDs2yNxP/8zophjfdTKgkRa7YYlGYSoIxmQVA owCmcwQV4cA01uIc6NIBBAs/wCm9O6rw4787HonXNyWdO4A czx9T7pHe /latexit latexit sha1 base64 "22jywEjQYStxJ5VhI3BqGrGrjow " y2m3bpZhN2J8US je8eFDEq3/Gm//GbZuDtj4YeLw3w8y8IBFco 6yaKkSgQrBOM7 Z NSgRzNfvnLG8Q0jZhEKojWPddJ0M IQk4Fm5W8VLOE0DEZsp6hkkRM NSq0X6936WLYWrHzmFP7A vwBVXKR3w /latexit qKeysAttentionWeights QueriesvL latexit sha1 base64 "I0VC8gBrQq8yWNyo yFpgT4wz1Y " Fstpt26WYTdifFEvo3vHhQxKt/xpv/xm2bg7Y QQrbuuk6CfEYWcCjYtealmCaEjMmBdQyWJmPaz c1T 8wofTuMlSmJ9lz9PZGRSOtJFJjOiOBQL3sz8T vm2J47WdcJikySReLwlTYGNuzAOw ThBE7hHFy4gjrcQQOaQCGBZ3iFNyu1Xqx362PRWrDymWP4A vzB3tWkfg /latexit latexit sha1 base64 "MdMM0CCP6gkS6I9FmEsJNlSQy9E " 120y7dbMLuRCyh/8KLB0W8 m 8 KlV iSMtS2FZKb nshoZMw4CmxnRHFoFr2p gqe4RXeHOO8OO/Ox7x1xclnjuAPnM8fGn2RMg /latexit

Intrinsic Preference ModelingModel users' dynamic intrinsic preferences based on the correlationbetween the most recent purchase and the personal historic purchases. Query embedding of current item Keys, Values sequence of previous itemsShopping History IntrinsicPreferenceModelingItem EmbeddingPositionalEmbedding Self-AttentionRef: Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recommendation. In ICDM.QueryIntrinsic Preference

Personal Occasion ElicitationElicit personal occasion signals by tracing the user's previous shoppingbehavior in the neighboring days. Query timestamp of interest (Key, Value) timestamp embedding of previous purchase,corresponding item EmbeddingItem Embedding Shopping HistoryPersonal OccasionNextShopping

Global Occasion MemorizationMemorize the shopping behavior of the crowd under different globaloccasions. Query current timestamp embedding (Key, Value) (timestamp embedding, memory slot embedding)Global TableGlobalOccasionMemorizationMemory Slots Time KeyEmbedding Global OccasionQueryNextShopping

Gating LayerUtilize an attention (gating) layer to balance a user's intrinsic preferenceswith different occasion signals for personalization.Gating LayerGlobal OccasionPersonal Occasion QueryNextShoppingShopping HistoryPredictedPreferenceUser EmbeddingIntrinsic PreferenceGated Outputx

Experiment: DataEtsy: November 2006 to December 2018Amazon: May 1996 to July 2014Dataset# Users# Items# 1

Experiment: Metrics Normalized Discounted Cumulative Gain (NDCG@K) Hit Rate (HR@K) Mean Reciprocal Rank (MRR)

Experiment: Baselines Most Popular MF-BPR: Matrix Factorization with Bayesian Personalized Ranking Fossil: Fusing Similarity Models with Markov Chains GRU4Rec : Recurrent Neural Networks with Top-k Gains TCN: A Simple Convolutional Generative Network for Next ItemRecommendation HPMN: Lifelong Sequential Modeling with Hierarchical PeriodicMemory Network SARec: Self-attentive sequential recommendation

OAR vs BaselinesOAR achieves the best performance on both Etsy and AmazonEtsyAmazonModelNDCGHRNDCGHRMRRMRRK 5K 10K 5K 10K 5K 10K 5K 8930.2650GRU4Rec 2780.54330.30090.33850.40850.52510.2984OAR0.6078* 0.6415* 0.7425* 0.8462* 0.5847* 0.3200* 0.3580* 0.4301* 0.5476* 0.3165*e 2: Comparison of Di erent Models. indicates that the improvement of the best result is statistically signi cant cd with second best result with p 0.01.6% 8% improvement

470.71020.82780.54330Impact of eachOARcomponent?0.6078* 0.6415* 0.7425* 0.8462* 0.5847* 0Table occasion2: Comparisonof Di erentModels. indicatesthatthe improveThe globalmemorizationcomponentperformsbetterthanpared with second best result with p 0.01.recommending general most popular items.EtsyAmazonNDCG@5MRRNDCG@5MRRGlobal (G)0.18160.19530.22380.2294Intrinsic (I)0.56650.54330.30090.2984Personal (P)0.57910.55820.30690.3047I G0.58850.56420.30990.3063I P0.59160.56770.31360.3108Remove .3165Table 3: Ablation Test Results.Model

470.71020.82780.54330Impact of eachOARcomponent?0.6078* 0.6415* 0.7425* 0.8462* 0.5847* 0Table2: Comparisonof Di erentModels. indicatesthe improvePersonaloccasioncomponentworks betterthan theintrinsic thatpreferencepared with second best result with p 0.01.component.EtsyAmazonNDCG@5MRRNDCG@5MRRGlobal (G)0.18160.19530.22380.2294Intrinsic (I)0.56650.54330.30090.2984Personal (P)0.57910.55820.30690.3047I G0.58850.56420.30990.3063I P0.59160.56770.31360.3108Remove .3165Table 3: Ablation Test Results.Model

470.71020.82780.54330Impact of eachOARcomponent?0.6078* 0.6415* 0.7425* 0.8462* 0.5847* 0TableComparisonof simpleDi erentModels. indicatesthat theimproveReplacethe2:Gatinglayer withaddition Theperformancedropspared with second best result with p 0.01.EtsyAmazonNDCG@5MRRNDCG@5MRRGlobal (G)0.18160.19530.22380.2294Intrinsic (I)0.56650.54330.30090.2984Personal (P)0.57910.55820.30690.3047I G0.58850.56420.30990.3063I P0.59160.56770.31360.3108Remove .3165Table 3: Ablation Test Results.Model

Visualization of Occasion-driven ShoppingPreferences for Occasion-related items are dynamics in a year.

Summary Propose an occasion-aware sequential recommender Model the repeated personal occasion signals with attention layers,while modeling the global occasion signals by memorizing thetemporal trends of shopping behaviors Experiments on Etsy and Amazon show the positive impact ofincorporating occasion signals Next

Thank you!

There are occasions that may re-occur within a certain period (e.g. annually or monthly) over the course of multiple years and trigger relevant purchase. traceable patterns in personal occasion signals. Time Gap between Purchases for Wedding and Anniversary within a year.More than 50% of purchases for anniversary are near the date of wedding