LLM for Recommendations

RecSys IL has a new LinkedIn community! Click here to join >>

Join us at the upcoming RecSys IL meetup on March 4, where we'll explore how LLM are transforming review summarization, local content recommendations, and e-commerce personalization.

**Talks are in English.

Date: March 4, 17:30
Location: Booking.com office - Derech Menachem Begin 146, PWC Tower, 30th Floor, Tel-Aviv

Agenda:

17:30 - 18:00 Registration, Mingling, Snacks & Drinks
18:00 - 18:15 LLMs for Recommendations Overview (Bracha Shapira)
18:15 - 18:45 Abstractive Review Summarization Using a GenAI Approach (Eran Fainman)
18:45 - 19:15 Utilizing LLMs for localized Recommendations (Gali Katz)
19:15 - 19:45 Aspect affinity prediction using SLMs for e-commerce personalization (Yotam Eshel)

Details:

LLMs for Recommendations Overview
**Speaker:**Bracha Shapira
In this introductory talk, I will provide a short overview of the integration of LLMs into recommender systems. I will discuss the motivation behind this integration, highlighting how LLMs' advanced language understanding can enhance personalization and user engagement. Following this, I will address the challenges encountered, and will then examine the current state of research and implementation, focusing on models, training and evaluation. Finally, I will provide some insights into future directions, considering ongoing advancements in the field.

Talk 1

Title: Abstractive Review Summarization Using a GenAI Approach

Speaker: Eran Fainman (Senior Machine Learning Scientist @ Booking.com)

Bio:
Eran is a Senior Machine Learning Scientist in the Content Intelligence team at Booking.com, specializing in Generative AI and Natural Language Processing (NLP). He holds a Master’s degree in Software and Information Systems Engineering from Ben-Gurion University. Previously, he worked on search and personalization at Amazon’s Alexa Shopping Group. Currently, he develops text generation and summarization systems using Large Language Models (LLMs) and contributes to the AI Trip Planner, Booking.com’s chatbot.

Abstract:
User-generated reviews significantly impact travelers' decisions, especially when choosing accommodations. However, the vast amount of content can be overwhelming, making it difficult for travelers to find relevant information. To address this, Booking.com developed an automatic review summarization system, using GenAI capabilities to condense lengthy reviews into concise, user-friendly summaries. This talk will delve into the intricacies of the system, detailing the process of selecting and aggregating reviews at the accommodation level, and employing a Large Language Model (LLM) to generate the summaries. Additionally, we will share our learnings, experimental results, and future plans for the system.

Talk 2

Title: Utilizing LLMs for localized Recommendations

Speaker: Gali Katz (Senior Algorithm Engineer @ Taboola)

Bio:
Gali Katz is a senior algorithm engineer and public speaking advocate in Taboola's Algorithms group. She holds a PhD in Cognitive Science from Ben-Gurion University. Her recent work focuses on leveraging SOTA embeddings to serve related stories in Taboola-powered slots on digital websites and utilizing LLMs to deliver more localized content in Taboola’s personalized homepages.

Abstract:
Local news websites aim to engage audiences by delivering relevant local content. However, identifying the local context in news items can be challenging, especially when locations are mentioned implicitly. Traditional Named Entity Recognition (NER) models struggle to detect these implicit locations, making it difficult for personalized recommender systems like Taboola’s to serve truly local content.
In this talk, I'll present our approach to improving localization coverage, examining NERs, Knowledge Graphs, and ChatGPT along the way. I’ll demonstrate how we enhanced ChatGPT's prompts with Knowledge Graph information to better detect implicit locations. Finally, I’ll share insights into the engagement and system impact on our local news websites.

Talk 3

Title: Aspect affinity prediction using SLMs for e-commerce personalization

Speaker: Yotam Eshel (Senior Applied Researcher @ ebay)

Bio:
An applied researcher interested in ML, personalization, recommender systems and NLP. Have been working at eBay for the past 7+ years. Prior to that I did my master’s degree at the Technion under the supervision of Shaul Markovich working on applying deep learning to NLP.

Abstract:
We will present the problem of inferring buyer aspect-affinities in e-commerce: predicting which brands, sizes, colors, and other aspects are of interest to a user on our platform. Knowing these affinities about our users allows us to build a complex and nuanced user profile that can be leveraged to improve our personalization and recommendation systems. We will present a model using SLM (small language model) to solve the problem and evaluate it both intrinsically on the task accuracy itself, and extrinsically in a real-world recommendation scenario.