Impression-Aware Recommender Systems

TL;DR

Impression-Aware Recommender Systems reframes the learning problem to incorporate impressions—the items shown to users at a given time—alongside traditional interactions. It establishes a theoretical framework with a formal definition of events, user histories, and a prediction function, and articulates three recommendation phases: learning, predicting, and generating impressions. The paper introduces a three-way taxonomy (model-centric, data-centric, signal-centric) to classify impression-aware reco methods and systematically reviews 43 papers across heuristics, statistics, ML, DL, and RL, highlighting trends toward deep learning and reinforcement learning. It also surveys public and expired datasets, evaluates methodologies, and outlines open questions around impression signals, fatigue, biases, and debiasing, advocating for more public contextual datasets and standardized evaluation. Overall, the work positions impression-aware recommender systems as a distinct paradigm with strong potential to improve personalization by leveraging on-screen exposures and their signals, while outlining concrete research directions for model design, data usage, and evaluation practices.

Abstract

Novel data sources bring new opportunities to improve the quality of recommender systems and serve as a catalyst for the creation of new paradigms on personalized recommendations. Impressions are a novel data source containing the items shown to users on their screens. Past research focused on providing personalized recommendations using interactions, and occasionally using impressions when such a data source was available. Interest in impressions has increased due to their potential to provide more accurate recommendations. Despite this increased interest, research in recommender systems using impressions is still dispersed. Many works have distinct interpretations of impressions and use impressions in recommender systems in numerous different manners. To unify those interpretations into a single framework, we present a systematic literature review on recommender systems using impressions, focusing on three fundamental perspectives: recommendation models, datasets, and evaluation methodologies. We define a theoretical framework to delimit recommender systems using impressions and a novel paradigm for personalized recommendations, called impression-aware recommender systems. We propose a classification system for recommenders in this paradigm, which we use to categorize the recommendation models, datasets, and evaluation methodologies used in past research. Lastly, we identify open questions and future directions, highlighting missing aspects in the reviewed literature.

Figures (5)

• Figure 1: Categorization of items in a recommender system using impressions. Catalog are all items (solid line). Impressions are shown items (dashed line). Interactions are shown and interacted items (dotted line). Non-impressions are not shown items ($D = A - B$, i.e., between solid and dashed lines). Non-interactions are not interacted items ($E = A - C$, i.e., between solid and dotted lines).
• Figure 2: Number of papers reviewed in this work by their publication year (left) and venue (right).
• Figure 3: The three phases of any given for generating recommendations to a given user. (a) illustrates the first phase (learning), where the creates a prediction function ($f$) using the set of users profiles ($\mathcal{H}$). (b) illustrates the second phase (prediction), where the uses $f$ to predict the relevance score ($\Tilde{r}_{u\xspace, i\xspace}$) of any given user-item-profile triplet. (c) illustrates the third phase (recommendation), where the generates an impression (recommendation list) to a given user ($u$) by selecting their N-most relevant items based on their predicted relevance scores.
• Figure 4: Hierarchy of four learning paradigms, from top to bottom and left to right: cfside, impressionsbased, contextaware, and hybrid content-based collaborative filtering. As illustrated, the last three belong to cfside. is not equivalent to its sibling paradigms due to the theoretical and practical differences between them, differences that we analyze in \ref{['subsec:impressions-based-recsys:related-domains']}. Additionally, the diagram places our proposed taxonomies for impressionsbased, namely model-centric, data-centric, and signal-centric taxonomies; also presented and discussed in \ref{['subsec:impressions-based-recsys:classification']}.
• Figure 5: Types of inputs received by recommendation models from the reviewed literature according to the three categories within the data-centric taxonomy. (a) illustrates the first category (features), where recommendation models receive statistical features computed from impressions as part of their input. (b) illustrates the second category (learn), where recommendation models receive an impression; partially or fully, as part of their input. (c) illustrates the third category (sample), where recommendation models receive a vector containing sampled items from the catalog; each item from a possibly different part of the catalog, i.e., interacted, solely impressed, or non-impressed. The data-centric taxonomy allows for a recommendation model to belong to one or more categories within it.