Boosting your recommender systems
Join this meet-up to hear stories of how Taboola, Agoda, and Toloka meet the challenges of recommendation systems.
Overview
Join our September meetup to hear stories of how Taboola, Agoda, and Toloka meet the challenges of recommendation systems.
Our guests will share practical solutions for tuning recommendation systems that will help you avoid pitfalls, adapt to unexpected market trends, and get the best recommendations possible.
This meetup will be relevant to ML engineers, ML researchers, data scientists, and anyone who interacts with AI. Content is geared toward an audience with an intermediate level of experience in ML or Data Science.
Agenda
* The time is indicated in Israeli time zone (GMT+3)
Doors open
Improving CVR with a CTR model
Small but important: How to grow a small segment within a big recommendation system
Boosting a recommendation system with human labeling
Break
Q&A session with experts
Improving CVR with a CTR model
Taboola's content recommendation system optimises for both click-through rate (CTR) and conversion rate (CVR). CTR can sometimes be misleading, so we use post-click dwell time on the advertiser site to measure whether the item matched the user interest.
In this talk we will focus on how to properly incorporate this post-click signal into the CTR model. Treating clicks with low dwell time differently, either by data manipulation or loss function manipulation, can introduce bias to the models. The problem becomes more severe because post-click dwell time signals are incomplete, and differ between advertisers. This can cause a disproportionate loss in scale for some advertisers in the system, along with an RPM loss for publishers.
To avoid introducing bias, we employ oversampling of positive events to replace low dwell time clicks with high engagement ones. Online tests reveal improvements in advertiser KPIs such as CVR and CPA without sacrificing scale or publishers' RPM.
Small but important: How to grow a small segment within a big recommendation system
The Covid-19 pandemic changed the tourism world profoundly. One of the evident trends was an increase in demand for apartments and homes over traditional hotels. As one of the world's leading online travel agencies, Agoda had to adapt to this change. Our ranking engine has long been optimised for hotels, which made it a challenge to start promoting new types of properties.
In this talk Yair will share how Agoda approached the problem of promoting a small but strategic segment within an existing framework that was not built specifically for it. We'll define the problem and relevant metrics and discuss some of the ideas we've tried and their tradeoffs.
Boosting a recommendation system with human labeling
Recommendation systems benefit both service providers and clients as they boost GMV and customer satisfaction. These systems are often based on implicit signals from logs and historical data gathered online, but a cutting-edge recommendation system needs explicit feedback as well.
In this talk we will look at how to take your recommendations to the next level by gathering targeted feedback about personalised recommendations based on previous search queries and clicks, complementary and substitute items, trending products, and more.
Q&A session with experts
Grab a drink and ask our guest experts everything you ever wanted to know about recommender systems! This is a great opportunity to discuss best practices and find out how other companies boost their recommendation solutions.
Don't miss out
tutorials, and webinars.
