Quality-Aware Neural Complementary Item Recommendation Yin Zhang , - PowerPoint PPT Presentation

Quality-Aware Neural Complementary Item Recommendation Yin Zhang , Haokai Lu, Wei Niu, James Caverlee Department of Computer Science and Engineering Texas A&M University, USA ACM RecSys’18 :: October 3rd, 2018

Item-to-Item Recommendation substitutes ? ?

Filters? Lens? Others? Bags? Complementary Item Recommendation: items that might be purchased together

Complementary Item Recommendation: Ground Truth Also-Bought Bought-Together Amazon item relationship dataset: McAuley, Julian, et al. "Image-based recommendations on styles and substitutes." SIGIR, 2015.

Complementary Item Recommendation: Challenges 1. How to define “complementary" distance? • Previous methods rely on a single source to detect item relationships— images [McAuley SIGIR 2015] or text [Wang WSDM 2018]. 2. How to balance quality vs. complementary relationship? 1 star Complementary 2 star 3 star relationship 4 star 5 star Item Quality 0 10 20 30 40 1 star Recommendation 2 star 3 star 4 star “Bella Ladies” 5 star hoodie 0 100 200 300 400 3. How to model complex interactions? • Potential non-linear relationships between items features and quality. Melville, Prenn, et, al. “Content-boosted collaborative filtering for improved recommendations.” AAAI, 2002 McAuley, Julian, et al. "Image-based recommendations on styles and substitutes." SIGIR , 2015. Wang, Zihan, et al. “A Path-constrained Framework for Discriminating Substitutable and Complementary Products in E-commerce.” WSDM. 2018.

Our Solution: ENCORE 2. Quality-Aware 3. Transform via 1. Detect Complementary Recommendation Neural Model Items Complement threshold ENCORE: Neural COmplementary item REcommendation

1. Detecting Complementary Items • Influence factors (a) (b) (c) (d) Visual Textual USB Battery USB Flash Mac Air Mac Pro Mac Pro Drive Charger Charger Complementary relationship between items is influenced by style (image) and function (text) and this influence varies by items. • Idea: Embed Style + Function • • Style-Based Complements: Functional Complements: ( ct ) ( I i , I j ) = || ( t i − t j ) T E T || 2 ( cm ) ( I i , I j ) = || ( m i − m j ) T E M || 2 2 d j | i 2 d j | i Word2Vec Image Feature Vector Learned Low-ranked Learned Low-ranked Embedding for text Embedding for image

2. Quality-Aware Recommendation • Complement relationship vs Item Quality (A)(B) Bella Ladies pants: the nearest complement items (C) Spandex pants “Bella Ladies” hoodie Users may not choose the nearest complementary items but the highest- quality complementary items. • Item Quality Estimation Item 2 Item 1 Posterior Distribution

3. Neural Complementary Item Recommendation • Textual Relationship Items Vary by Quality Visual Categories Complementary item recommendation is influenced by the complex interactions of item visual, textual and quality information. • ENCORE Framework Image Input Text Input User Ratings Image Input Text Input User Ratings Image Input Text Input Image Distance Text Distance Text Distance Image Distance Image Embedding Text Embedding Image Embedding Text Embedding Non-linear Layers Asymmetric Quality- Asymmetric Quality- aware Recommendation aware Recommendation Non-linear Layers

Experiments • How well does ENCORE perform versus baselines? • What impact do the design choices of ENCORE have? (images, textual information, Non-linearity) Dataset * : Six categories in Amazon (Electronics, Cell Phones & Accessories (C & A), Clothing, Books, Digital Music, and Movies) * McAuley, Julian, et al. "Image-based recommendations on styles and substitutes." Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval . ACM, 2015.

Experimental Setup: Baselines • LR A : Logistic Regression with Average Rating • LR B : Logistic Regression with Bayesian Rating • WNN: Weighted Nearest Neighbor Same Inputs • FNN: Feedforward Neural Network • LMT: Low-rank Mahalanobis Transform [McAuley SIGIR 2015] • Monomer [He ICDM 2016] • Variations of ENCORE (see paper) Metrics: Accuracy, Precision at top-k * McAuley, Julian, et al. "Image-based recommendations on styles and substitutes." SIGIR ACM, 2015. * He, Ruining, Charles Packer, and Julian McAuley. "Learning compatibility across categories for heterogeneous item recommendation." Data Mining (ICDM), 2016.

Experiments: Recommendation E ff ectiveness Also-Bought Bought-Together ENCORE outperforms state-of-the-art methods in accuracy, precision@5 and precision@10 for most situations, especially for Electronics and Clothing categories.

Experiments: Case Study Complementary Items Query Items Recommended by ENCORE (a) IdeaPad U430 (b) (c)

Conclusions and Future Work • Complementary relationships vary for di ff erent items. Items visual and textual information can help find complement items. • Users may not choose the nearest complementary items but the highest-quality ones. Modeling item rating distribution by Bayesian inference can improve the accuracy and precision for complementary recommendation. • Neural network structure in ENCORE provides improvement to the accuracy and precision of complement item recommendation • Future work: • Personalized complementary item recommendation. • E ff ectively model textual information to improve the quality of recommendation.

Quality-Aware Neural Complementary Item Recommendation Yin Zhang , Haokai Lu, Wei Niu, James Caverlee Department of Computer Science and Engineering Texas A&M University, USA Thank you ! ACM RecSys’18 :: October 3rd, 2018