Source-linked AI summary
Modeling User Exposure in Recommendation
Dawen Liang, Laurent Charlin, James McInerney, David M. Blei
TL;DR
Implicit-feedback recommenders cannot tell whether an unclicked item was disliked or simply unseen. The paper introduces ExpoMF, which models exposure as a latent variable and incorporates exposure covariates; it outperforms existing matrix-factorization approaches across four datasets. Its future direction is more realistic modeling of dynamic exposure processes that may differ between training and testing.
Problem
Implicit feedback treats unclicked items as evidence, although many may be unobserved because users were not exposed to them.
Method
ExpoMF separately models user exposure and item clicking, inferring exposure from data and allowing exposure covariates such as content and location.
Results
ExpoMF matches or surpasses existing approaches across four domains, and adding exposure covariates further improves performance.
Takeaways & Limitations
Modeling exposure provides a flexible collaborative-filtering framework that theoretically justifies downweighting unclicked items and supports richer exposure models.
Takeaways & Limitations
Exposure processes can be dynamic and may differ between training and test time, motivating more realistic assumptions about user-item interactions.
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
Implicit-feedback recommendation is difficult because unclicked items may reflect lack of exposure rather than dislike. ExpoMF models exposure separately, supports exposure covariates, and matches or surpasses established methods across four domains.
- The recommendation problem: Implicit feedback is accessible but ambiguous because an unclicked item may indicate either dislike or that the user never saw it.Unlike explicit ratings, implicit data arise naturally from behavior but require unclicked items during preference inference.
- The recommendation problem: Existing methods reduce the influence of unclicked items through heuristic confidence adjustments or subsampling.Hu et al. lower confidence for unclicked items, while Rendle et al. subsample them at a lower rate.
- Exposure-aware modeling: ExpoMF separately models exposure and clicking, iteratively estimating user preferences and why unclicked items were unclicked.It naturally downweights unclicked items that the model expects the user would like because it infers they were not exposed.
- Exposure-aware modeling: Exposure can be conditioned on item content, search exposure, or user/item location, enabling domain-specific exposure models.The paper illustrates location-aware weighting by upweighting unvisited restaurants near a user and downweighting distant ones.
- Empirical evaluation: Across music, scientific papers, bookmarks, and venue check-ins, ExpoMF matches or surpasses the state-of-the-art method, while extra exposure information improves performance further.Content-informed ExpoMF also outperforms a document recommendation method designed around content and user clicks.
2. BACKGROUND
Collaborative filtering represents users and items through latent factors, while implicit-feedback methods must handle the uncertainty of unobserved interactions. WMF downweights all unobserved interactions uniformly, but its formulation is not a proper generative model.
- Matrix factorization: Matrix factorization infers latent user preferences and item attributes by factorizing a user-item click or consumption matrix.The generative view draws user and item factors, then generates observations from a distribution parameterized by their dot product.
- Matrix factorization: Gaussian matrix factorization models observations with user factors θu and item factors βi, using hyperparameters for regularization and noise.The identity matrix IK has dimension K.
- Implicit-feedback models: Weighted matrix factorization treats consumed pairs as one and unobserved pairs as zero, uniformly downweighting evidence from unobserved interactions.WMF is also known as one-class collaborative filtering.
- Implicit-feedback models: WMF is not a generative model because its confidence depends directly on the click being modeled.Adding an exposure latent variable yields a proper generative model.
3. EXPOSURE MATRIX FACTORIZATION
ExpoMF models user exposure as a latent variable distinct from consumption, allowing unobserved non-clicks to reflect either non-exposure or ignored items. The model supports exposure covariates, recovers WMF as a special case, and uses EM for inference and exposure-aware prediction.
- 3.1 Model Description: ExpoMF introduces a latent exposure matrix alongside the observed user-item click matrix.A positive click implies exposure, but a zero click leaves exposure unknown because the user may have seen and ignored the item or never encountered it.
- 3.1 Model Description: Conditional on exposure, user preferences and item attributes are modeled with matrix factorization.The exposure variable gates the matrix-factorization evidence: low exposure probability downweights the contribution of a user-item pair, while exposure equal to one recovers standard matrix factorization.
- 3.1 Model Description: WMF is a special case of ExpoMF obtained by fixing the exposure matrix using fixed exposure probabilities.Fixing all exposure entries to one instead recovers Gaussian probabilistic matrix factorization.
- 3.2 Hierarchical Modeling of Exposure: Exposure priors can be specified using item popularity or external covariates such as document topics and venue locations.The model conditions exposure probabilities on covariates and exposure parameters, while empirical performance can be sensitive to how these priors are chosen.
- 3.2 Hierarchical Modeling of Exposure: Exposure covariate parameters represent the topics or geographical points of interest to which a user is usually exposed.For documents, content representations are derived with natural-language processing; for venues, locations are represented through clustered geographic assignments.
- 3.3 Inference: EM alternates between estimating missing exposure variables and maximizing parameters, exploiting conditional conjugacy for analytical updates.For prediction, ExpoMF can use either a simple dot product or an expectation integrating over exposure uncertainty; the better choice depends on the exposure model.
4. RELATED WORK
The related work connects ExpoMF to causal inference, biased collaborative filtering, random-graph exposure models, spike-and-slab models, and versatile feature-based recommenders. These connections distinguish latent exposure modeling from alternative treatments of selection, features, and unobserved interactions.
- Causal inference: Latent exposure models address counterfactual recommendation questions that are difficult to answer from observational logs alone.Online studies can answer such questions but are typically expensive, motivating interest in observational approaches.
- Causal inference: ExpoMF draws on causal inference by separating exposure as an assignment mechanism from consumption as a potential outcome.Consumption implies exposure, but items seen without consumption remain unknown; the model is therefore described as a latent assignment model.
- Biased CF models: Biased collaborative-filtering models similarly separate a selection or exposure matrix from a matrix-factorization data model.The paper places this shared structure alongside its own exposure formulation while noting that the interpretations differ.
- Modeling exposure with random graphs: Exposure can also be modeled with a hidden consider graph, but that alternative stochastically samples a graph instead of directly inferring the exposure posterior.The paper characterizes direct posterior inference in ExpoMF as computationally more demanding than the approximation used by the random-graph approach.
- Exposure in other contexts: Zero-inflated Poisson regression uses a similar latent binary variable to explain structural zeros in count data.The paper presents this as a related use of latent exposure-like variables outside recommendation.
- Other connections: ExpoMF resembles a spike-and-slab model in which exposure variables form the spike and matrix factorization forms the slab.The related-work discussion also connects ExpoMF to versatile collaborative-filtering models that incorporate external user, item, or interaction information.
5. EMPIRICAL STUDY
Across four user-item domains, ExpoMF performs comparably better than WMF, while exposure covariates further improve recommendation performance. Posterior analyses show that inferred exposure reflects popularity and user-specific content or listening interests.
- Empirical evaluation: ExpoMF performs comparably better than state-of-the-art WMF across four datasets covering clicks, checkins, bookmarks, and listening behavior.The datasets are TPS, arXiv, Mendeley, and Gowalla.
- Incorporating Exposure Covariates: Exposure covariates further improve performance: location and content variants outperform ExpoMF with per-item µi, and Content ExpoMF outperforms a state-of-the-art document recommendation model.The study uses venue locations and document content as exposure information.
- Empirical evaluation: ExpoMF performs comparably better than WMF on most datasets, although performance differences are generally small.Higher NDCG@100 and MAP@100 can occur even when Recall@50 is lower, indicating more relevant top-ranked items.
- Studying Exposure MF: Posterior exposure estimates for unlistened songs vary by user: songs from a user’s favored bands receive lower exposure probabilities than the prior, unlike WMF’s uniform downweighting.For user A, the model identifies unlistened Radiohead and Interpol tracks as exposure outliers.
- Studying Exposure MF: Popularity strongly influences ExpoMF exposure estimates, while listening history can distinguish similarly popular songs as likely unexposed or disliked.For user B, the model treats an unplayed popular Florence and the Machine song as more likely unexposed but infers dislike for another popular song.
- Incorporating Exposure Covariates: Content ExpoMF uses article topics to infer exposure, increasing exposure for papers aligned with a user’s interests and decreasing it for mismatched interests.The model uses LDA-based document content as an exposure covariate.
6. CONCLUSION
ExpoMF incorporates user exposure into collaborative filtering, theoretically justifies downweighting unclicked items, and outperforms existing matrix-factorization approaches across four diverse datasets. The paper identifies more realistic, dynamic exposure modeling and online evaluation as important directions for future work.
- ExpoMF models user exposure to items and theoretically justifies existing approaches that downweight unclicked items.
- Its flexible exposure framework supports more elaborate models based on logistic regression and can also analyze explicit feedback.
- ExpoMF’s additional flexibility helps it outperform existing matrix-factorization approaches on four datasets from various domains.
- Exposure processes may be more dynamic when users browse new content or seek recommendations in unfamiliar settings, differing between training and test time.
- The authors seek more realistic exposure assumptions and online evaluation settings where user interactions and exposed items may be observed.