Deep Learning for Search and Recommender Systems in Practice

Abstract

In this talk, we will go over the components of personalized search and recommender systems and demonstrate the applications of various deep learning techniques along the way.

Search and recommender systems are probably the most prevalent ML powered application across the industry. They share most of the components composition and provide a user a ranked list of items, while there is subtle difference that a search system typically acts passively with a clear user intention in terms of queries and a recommender system acts more proactively. Deep learning has been wildly successful in solving complex tasks such as image recognition, speech recognition, natural language processing and understanding, machine translation, etc. In the area of personalized recommender systems, deep learning has been showing tremendous impact in recent years. Search and recommender systems can be staged roughly in three phases: 1. User and query understanding, where a query or a user profile are processed so that the systems can use the processed information to 2. retrieve all the related items (high recall) and 3. rank the items by the order of the most relevance to the user’s intent (high precision). Each phase has its unique challenges but deep learning has been ubiquitously pushing beyond the limit. After walking through the talk, we hope the audience would gain some first-hand experience building a personalized search/recommender system using deep learning techniques.

Information

When: Wednesday, August 26 9:00AM-4:00PM

Slides
Python Notebooks (Use this Google Drive App to open notebooks directly in Google Colab)
Paper

Whova

Zoom

Tutorial Outline

1.1 Introduction to deep learning, search and recommender systems

System architectural overview
Major components of search and recommender systems and common approaches
Deep learning and and its applications in search and recommender systems

1.2 Understanding

(1) User and query understanding
• Query understanding infers the intent of a search engine user by extracting semantic meaning from the searcher’s keywords.
• User understanding provides personalization features for candidate retrieval and ranking.

(2) Hands-on session

•Train a query intent model.

•Train a query auto completion model.

1.3 Candidate Retrieval

(1) Candidate retrieval for a search system through indices

For a search system (search engine), the candidate selection is typically handled by a reverted index.

(2) Candidate selection for a recommender system
• A general recommender system usually has multiple sources of candidates.

(3) Hands-on session:
• Setup ElasticSearch with pre-populated indices;
• Train a deep KNN model that can be used for candidate selection of a recommender system.

1.4 Ranker and Re-ranker

(1) Learning to Rank (LTR) in search and recommender systems
• Different strategies applied to a ranking problem, including point-, pair- and list-wise ranking algorithms.

• Business rule-based mixer as a special kind of reranker.

(2) Hands-on session:

•Train a Generalized Deep Mixed Model(GDMix, an extension to GLMix with DeText)