Modeling User Exposure in Recommendation

Abstract

Collaborative filtering analyzes user preferences for items (e.g., books, movies, restaurants, academic papers) by exploiting the similarity patterns across users. In implicit feedback settings, all the items, including the ones that a user did not consume, are taken into consideration. But this assumption does not accord with the common sense understanding that users have a limited scope and awareness of items. For example, a user might not have heard of a certain paper, or might live too far away from a restaurant to experience it. In the language of causal analysis [9], the assignment mechanism (i.e., the items that a user is exposed to) is a latent variable that may change for various user/item combinations. In this paper, we propose a new probabilistic approach that directly incorporates user exposure to items into collaborative filtering. The exposure is modeled as a latent variable and the model infers its value from data. In doing so, we recover one of the most successful state-of-theart approaches as a special case of our model [8], and provide a plug-in method for conditioning exposure on various forms of exposure covariates (e.g., topics in text, venue locations). We show that our scalable inference algorithm outperforms existing benchmarks in four different domains both with and without exposure covariates.

Recommendations and privacy in the arXiv system: A simulation experiment using historical data

“…Liang, Altosaar, Charlin, and Blei (2016) simultaneously factorized the user-item interaction matrix and item--item term co-occurrence matrix. Liang, Charlin, McInerney, and Blei (2016) tried to estimate what they called "user exposure," by which they meant not "exposure of the user," but rather "exposure (of an item) to a user." In Charlin, Ranganath, McInerney, and Blei (2015), Dynamic Poisson Factorization was applied to arXiv usage data, to explore the evolution of a user's interests or of an article's audience over time.…”

Prediction of Sparse User-Item Consumption Rates with Zero-Inflated Poisson Regression

“…This observation has been made in application contexts as diverse as epidemiology, economics, and manufacturing (e.g., see [2]), but has seen relatively little application to the type of high-dimensional user-item consumption data that we investigate in this pape-exceptions are [9,16,18], which we discuss in more detail later in the paper.…”

“…Exposure Process: The exposure process describes whether or not a user i has been exposed to item j at time t. The concept of exposure captures the idea that for large item sets a typical user is likely to be unaware of (or unexposed to) most items in the "item vocabulary" (see also [9,16,18]), e.g., in music-listening many artists are unknown to many users. We define z t i j ∈ {0, 1} as an indicator variable to indicate if user i was exposed to item j at time t. We can model P(z t i j = 1) via a Bernoulli distribution with parameter π t i j , where the Bernoulli parameter will be a function of the past history of user i and item j.…”

“…Liang et al [16] proposed the framework of exposure matrix factorization (ExpoMF) which uses zero-inflation to explicitly account for exposure effects in matrix factorization of large binary user-item data sets. Liang et al found that ExpoMF systematically outperformed traditional PMF methods that did not account for exposure.…”

Learning User Embeddings from Emails

“…In the email domain, the mainstream approaches to representing users are based on bag-of-words or keywords features (Bekkerman et al, 2004;Dredze et al, 2008). Many previous efforts model users and their interactions in social networks or recommendation systems (Grover and Leskovec, 2016;Liang et al, 2016;Zhao et al, 2010). Emails, although can be viewed as a special kind of social platform, tend to generate interactions within a smaller group of participants, requiring a dense representation to help bridge the gap between even the farthest users.…”