Exploration in Recommender Systems

Minmin Chen, Sr. Staff Research Scientist, Google

Recommender systems as the intermediary in the large ecosystem of users, items and creators, plays a critical role in its success in the long run. Industrial recommender systems, suffering from closed feedback loop effect, however tend to exploit popular choices, leaving a large majority of the item corpus and creator base under-discovered. In this talk, I will discuss our work in studying exploration, in particular what we referred to as item exploration that aims at discovering fresh and tail items/creators in recommender systems. The talk will shed light on the challenges in measuring the benefits of exploration, and bring new experiment designs to systematically quantify the value of exploration to the long term user experience and ecosystem health. I will discuss our work in building dedicated exploration stacks in an industrial recommendation platform, as well as making existing recommendation stacks more exploration friendly. I will go over some of the main technology bets we are investing to bring efficient exploration, leading to maximum long-term ecosystem growth with minimum short-term user cost.

Making Decisions in Sponsored Advertising

Ben Allison, Sr. Principal Applied Scientist, Amazon

Modern advertising systems are built on a combination of machine learning, distributed systems, statistics, and game theory. Recent work in recommender systems has illuminated the distinction between the prediction paradigm (which underlies much of machine learning) and the decision paradigm which accomodates feedback loops and partial observability. Advertising is further subject to strategic actors and marketplace-level dynamics. Our disciplines lack many of the formal tools to reason about large marketplaces that are intermediated by machine learning models.

In this talk I’ll give an introduction to our advertising domain at Amazon, before moving on to discuss recent work on decision making in ads: learning to bid, and learning to conduct auctions. In both cases we combine machine learning, counterfactual reasoning and optimized decision making with strategyproof-ness and incentive compatibility, highlighting the opportunities for exciting new work at the intersection of multiple disciplines.

Zero Shot Recommenders, Large Language Models and Prompt Engineering

Anoop Deoras, Sr. Manager of Applied Science and Engineering, Amazon

Abstract: Large scale deep learning recommender system models, which are used for predicting top-k items given a query (typically a user identifier), learn the user to item propensities from large amounts of training data containing many (user, item, time) tuples. Unlike the simpler bi-linear models such as Matrix Factorization, deep learning recommender system models almost always do not model users explicitly bur rather model them via the items they have consumed in the past, with or without temporal ordering information. This allows the model to extrapolate the preference matching on many unseen users as long as the items these unseen users have interacted with are seen during training. However, pure item identifier based models are often incapable in assigning faithful propensities to novel items not previously seen during training. This is the classical item cold start problem in recommender system. Item content features (e.g. item description, movie synopsis, product image, etc) become useful in such cold start scenarios as they become proxy for the item identifier. In this paper, we investigate a rather extreme cold start situation wherein the entire application is treated as cold start. Imagine two related businesses whose product inventory is similar in content type but not in identity. Further assume that one of them lacks sufficient user-item interactions limiting us from building a personalization model. Question we want to ask is whether we can obtain domain agnostic information from the data rich domain — source domain, and transfer it to the data poor domain — target domain ?

To this end, in this talk, we present Zero Shot Transfer Learning model that transfers such learned task-agnostic-knowledge from the source to the target domain. We conduct offline experiments on large scale real world datasets and present the results and findings. Time permitting, we will also talk about a novel concept we first introduced back in 2021 — Language Models as Recommender Systems, where we show how we can leverage large pre-trained language models and prompt engineering to generate recommendations in a zero shot setting.