Fantastic Embeddings and How To Align Them

Paper: https://arxiv.org/abs/2007.14906
Video: https://www.youtube.com/watch?v=xjyMGHrruQ8
Speaker: https://federicobianchi.io

This paper addresses the challenge of leveraging multiple embedding spaces for multi-shop personalization, proving that zero-shot inference is possible by transferring shopping intent from one website to another without manual intervention. We detail a machine learning pipeline to train and optimize embeddings within shops first, and support the quantitative findings with additional qualitative insights. We then turn to the harder task of using learned embeddings across shops: if products from different shops live in the same vector space, user intent - as represented by regions in this space - can then be transferred in a zero-shot fashion across websites. We propose and benchmark unsupervised and supervised methods to "travel" between embedding spaces, each with its own assumptions on data quantity and quality. We show that zero-shot personalization is indeed possible at scale by testing the shared embedding space with two downstream tasks, event prediction and type-ahead suggestions. Finally, we curate a cross-shop anonymized embeddings dataset to foster an inclusive discussion of this important business scenario.

Transcript

1. Fantastic Embeddings and How to Align Them Zero-Shot Inference in a Multi-Shop Scenario Federico Bianchi*, Jacopo Tagliabue*, Bingqing Yu*, Luca Bigon and Ciro Greco Bocconi University Coveo Labs * Equal Contribution @fb_vinid 2020 SIGIR E-com Workshop
2. Research Team Federico Bianchi Post-doc @ Bocconi Jacopo Tagliabue Lead AI Scientist @ Coveo Bingqing Yu ML Engineer @ Coveo Luca Bigon Sr. Data Engineer @ Coveo Ciro Greco Director of A.I @ Coveo
3. Use Case • Shoppers are likely to browse in multiple related digital shops before purchase action; • E-commerce websites tend to have high bounce rates; • We need to personalize user search as quickly as possible with as little data as possible.
4. Fantastic Embeddings
5. Fantastic Embeddings: Session Representations WORD2VEC (HISTORICAL MIKOLOV PAPER HERE) PROD2VEC1 THE CAT IS ON SKU: U12_AB7 SKU: U12_AB8 SKU: U12_AB9 SKU: U12_AB1 1 Grbovic, M., Radosavljevic, V., Djuric, N., Bhamidipati, N., Savla, J., Bhagwan, V., & Sharp, D. (2015, August). E-commerce in your inbox: Product recommendations at scale. In KDD.
6. Fantastic Embeddings: Session Representations WORD2VEC PROD2VEC CAT 1 2 8 0 2 5 6 4 8 3 2 9 sim(cat, dog) = 0.8, vector similarity (e.g., cosine similarity) sim( , ) = 0.7, vector similarity (e.g., cosine similarity) SKU: U12_AB8
7. • Intent is represented as the products the user interacts with in a session. • Products are represented as a multi-dimensional vector space: products that are “closer” in the space are similar. • Sessions are paths in the product space: you can tell a lot about people if you know where they are going. • Building such a space can be done in a purely unsupervised manner. Fantastic Embeddings: Representing User Intent
8. Fantastic Embeddings: Many Shops Different shops may have different products and different ids for each element; Product embeddings of different shops are not comparable. Shop A Shop B SKU: U12_AB7 SKU: U12_AB8 SKU: G1244 SKU: G1588
9. Fantastic Embeddings: The Alignment Roadmap 1. Learn fantastic embeddings of products; 2. Align the embeddings from multiple shops; 3. Use your aligned fantastic embeddings.
10. How to learn Fantastic Embeddings? Colors represent sport activities. On the left good representations, on the right bad representations. How do we know when an embedding is fantastic?
11. How to learn Fantastic Embeddings? Check embedding quality using a next event prediction task. How well can predict ? average check for this product in the nearest neighbours of the space Different shops might require different prod2vec parameters for us to get good representation (check the paper for more details)
12. Aligning Fantastic Embeddings
13. Aligning Fantastic Embeddings: How To We want to align/learn a mapping from the Source Shop to the Target Shop. We have some metadata and we have users that have browsed the two shops during the same session.
14. Aligning Fantastic Embeddings: How To We want to align/learn a mapping from the Source Shop to the Target Shop Image-based Model (IM) (unsupervised): image based model. We use VGG-16 to extract relevant image features for the products, find similar images in the target shop (using KNN) and build a seed dictionary to bootstrap a self learning framework; SKU: U12_AB8 SKU: U12_AB9
15. Aligning Fantastic Embeddings: How To We want to align/learn a mapping from the Source Shop to the Target Shop User-based Model (UM) (supervised): users browsing both shops: learn mapping function using the last element seen on the source shop and the first element seen on the target shop. SKU: U12_AB7 SKU: U12_AB8 SKU: U12_AB9 SKU: U12_AB1 Shop A Shop B
16. Aligning Fantastic Embeddings: How To We want to align/learn a mapping from the Source Shop to the Target Shop User-based Model (UM) (supervised): users browsing both shops: learn mapping function using the last element seen on the source shop and the first element seen on the target shop. SKU: U12_AB7 SKU: U12_AB8 SKU: U12_AB9 SKU: U12_AB1 Shop A Shop B
17. Aligning Fantastic Embeddings: How To We want to align/learn a mapping from the Source Shop to the Target Shop Translation Model (TM) (supervised): users browsing both shops: given the sequence of elements seen on the source shop and the target shop, we use a deep recurrent network1 to translate the source sequence into the target sequence. SKU: U12_AB7 SKU: U12_AB8 SKU: U12_AB9 SKU: U12_AB1 Shop A Shop B Translate sequence
18. Aligning Fantastic Embeddings: Two Tasks Next event prediction across shops (from source to target shop) Personalized Type Ahead across shops (from source to target shop)
19. Data We leverage behavioral data collected over 12 months. Two shops (a source (A) and a target (B) shop) from two mid-size shops (revenues >10M and <100M) in the same vertical (sport apparel). Users have been on the two shops in the same session.
20. Aligning Fantastic Embeddings: Baselines PM (unsupervised): simple product popularity; NM (unsupervised): state-of-the-art model for multi-lingual word embedding alignment1; AM (semi-supervised): we learn a mapping function by mapping products in the source and the target shop that come from the same category. 1Mikel Artetxe, Gorka Labaka, and Eneko Agirre. 2018. A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. InACL. Association for Computational Linguistics, Melbourne, Australia, 789–798.
21. Next event prediction across shops Each model is evaluated on its ability to find the next item in the target shop; In detail, we check how well the models predict the first item of the new session or any of the items browsed by the user in the new session.
22. Next event prediction across shops Evaluation metric: NCDG@10
23. Personalized Type Ahead across shops We train a conditional language model over the target shop using the target space embeddings; Encoder-decoder architecture that encodes user intent using product representations, and then use an LSTM based language model to sort query completions by their probability (Word2Seq); At test time we use the aligned embeddings as input to the model.
24. Personalized Type Ahead across shops Seed = number of characters given to the language models to suggest competitions Evaluation metric: MRR@5
25. Conclusions • Cross-shop alignment is crucial. There’s the need to personalize the shopper experience as soon as possible; • Product embeddings can be evaluated with a simple procedure that ensure an effective representation; • Alignment procedures help in next event prediction and type ahead across-shop tasks.
26. Thank you!