[Paper Review] Modeling User Exposure in Recommendation

Dawen Liang, Laurent Charlin+2 more|arXiv (Cornell University)|Oct 23, 2015

Recommender Systems and Techniques30 references4 citations

TL;DR

This paper proposes Exposure MF (ExpoMF), a probabilistic collaborative filtering model that explicitly models user exposure to items as a latent variable, separating exposure from user preference. By inferring exposure from data and incorporating exposure covariates (e.g., content, location), ExpoMF improves recommendation accuracy over state-of-the-art methods, outperforming baselines on four diverse datasets with and without covariates.

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, 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-the-art approaches as a special case of our model, 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.

Motivation & Objective

• To address the limitation in implicit feedback recommendation where unclicked items are incorrectly assumed to be disliked, when they may simply not have been exposed.
• To formalize the exposure mechanism as a latent variable in a probabilistic model, enabling principled inference over user preferences.
• To develop a scalable inference algorithm that outperforms existing heuristic downweighting methods for unclicked items.
• To extend the model with exposure covariates (e.g., content, location) to improve exposure estimation and recommendation performance.
• To demonstrate the framework's effectiveness across diverse domains, including music, academic papers, bookmarks, and venue check-ins.

Proposed method

• ExpoMF models user exposure as a latent binary variable $a_{ui}$, representing whether user $u$ was exposed to item $i$, and jointly infers user preferences and exposure using a probabilistic matrix factorization framework.
• The model uses a logistic regression component to predict exposure: $\mathbb{E}[a_{ui}] = \sigma(\mathbf{\psi}_u^\top \mathbf{x}_i + \gamma_u)$, where $\mathbf{x}_i$ are exposure covariates (e.g., topic vectors, venue locations).
• Preference inference is performed by modeling user-item interactions as Bernoulli trials, with the probability of interaction depending on both user preference and exposure.
• The model employs an iterative expectation-maximization (EM)-style inference algorithm that alternates between estimating user preferences and exposure probabilities.
• For content-based exposure, the model uses LDA-derived topic vectors as covariates; for location-based exposure, it uses $K$-means clustering of venue locations to define exposure features.
• The framework allows for plug-in exposure modeling with various covariates, enabling flexible and interpretable exposure estimation.

Experimental results

Research questions

• RQ1 How can we model user exposure to items as a latent variable in collaborative filtering to improve recommendation accuracy?
• RQ2 Can separating exposure from preference estimation lead to better performance than heuristic downweighting of unclicked items?
• RQ3 How do exposure covariates such as item content or geographic location improve exposure estimation and downstream recommendation?
• RQ4 Does the proposed model generalize across diverse domains such as music, academic papers, bookmarks, and venue check-ins?
• RQ5 Can the model outperform state-of-the-art methods like weighted matrix factorization (WMF) and the confidence-based method of Hu et al.?

Key findings

• ExpoMF outperforms the state-of-the-art method of Hu et al. [8] across all four datasets studied: music, academic e-prints, bookmarks, and venue check-ins.
• On the Gowalla dataset, Location ExpoMF achieved a Recall@20 of 0.129, surpassing WMF (0.122) and ExpoMF with per-item $\mu_i$ (0.118).
• When using exposure covariates, Content ExpoMF and Location ExpoMF outperform the simpler ExpoMF model, demonstrating the benefit of incorporating external information.
• The model’s performance with exposure covariates is consistently better than baselines, including a simple FilterWMF that uses location to filter recommendations.
• The model provides a theoretically grounded alternative to heuristic downweighting, justifying such practices through latent variable modeling of exposure.
• Empirical results show that exposure modeling with covariates leads to better prediction of missing interactions, especially when exposure is not uniformly distributed.