The Embeddings That Came in From the Cold

Dense representations of words (“embeddings”) have been enjoying exceptional success in all sorts of NLP tasks. In recent years, leading eCommerce companies have been translating that success to the product space, introducing “product embeddings” to be used directly in recommendation systems, or indirectly for personalization.

In this talk, we first describe best practices to train product embeddings from browsing data, and to evaluate their quality. After giving some examples of the effectiveness of product spaces in user modelling, we show that the quality of the representation tends to be far lower for infrequent products. In the second part, we explain our recipe to successfully deal with “cold embeddings”, proposing a scalable method that does not require human supervision or changes in existing downstream applications.

Bio:

Jacopo Tagliabue

Educated in several acronyms across the globe (UNISR, SFI, MIT), Jacopo Tagliabue is now the Lead A.I. Scientist at Coveo, after the acquisition of his own A.I. startup Tooso in 2019. He works on theoretical and practical challenges in A.I., shipping Machine Learning models to hundreds of millions of devices every year. In previous lives, he completed a PhD, worked for a professional basketball team, and gave an academic talk on video games (among others improbable "achievements"​); his work has been often featured in the general press and presented internationally in business venues and scientific conferences (including RecSys, ACL, WWW and SIGIR).

Christine Yu

Christine Yu is a machine learning developer at Coveo who is fascinated by how AI can

understand the human brain and enrich eCommerce search experiences. Graduating from McGill University with a Master’s degree, she has long been a traveler in the ocean of data, computer vision, NLP, and machine learning; and she has always enjoyed seeing and taking part in these unprecedented and wonderful transformations these technologies have brought to our world. Now at Coveo, she's excited to take on new challenges of enabling AI to speak, look and think like humans.

References:
https://www.aclweb.org/anthology/2020.ecnlp-1.2/
https://arxiv.org/abs/2007.14906
https://dl.acm.org/doi/10.1145/3383313.3411477

Topics: embeddings, deep learning, one-shot learning.

This event will be held in English and will be streamed via the Google meet platform.

The event is streamed live on Google Meet and you will also be able to interact and ask questions (link will be added to the event description a couple of days before the start).

Agenda:

18:30 Data Science Milan community opening

18:40 Talk (50 minutes)

---

Would you be happy to host one of our events? Increase your popularity and be known from our community by hosting one of our interesting events. Contact us at datasciencemilan@gmail.com.

If you are interested in doing a lightning talk or want to propose an event, hackathon, panel discussion, roundtable, or any sort of initiative please submit your proposal at Call for speakers or get in touch with one of the members of the staff.

---

This event is supported by IAML (Italian Association of Machine Learning).