Impression-aware recommender systems (IARS) are a type of recommenders that learn user preferences using their interactions and the recommendations (also known as impressions) shown to users. The community's interest in this type of recommenders has steadily increased in recent years. To aid in characterizing this type of recommenders, we propose a theoretical framework to define IARS and classify the recommenders present in the state-of-the-art. We start this work by defining core concepts related to this type of recommenders, such as impressions and user feedback. Based on this theoretical framework, we identify and define three properties and three taxonomies that characterize IARS. Lastly, we undergo a systematic literature review where we discover and select papers belonging to the state-of-the-art. Our review analyzes papers under the properties and taxonomies we propose; we highlight the most and least common properties and taxonomies used in the literature, their relations, and their evolution over time, among others.
tions or search systems also generate impressions.
In Pérez Maurera et al. [
Many types of RS exist in the literature, each tailored to specific tasks, data sources, or domains. In this work, 1. We define those recommender systems that we study one type of RS, which we term impression- leverage impressions to learn users’ preferences: aware recommender system (IARS). IARS learn user pref- impression-aware recommender system (IARS). erences by leveraging user interactions and impressions. 2. We identify diferent properties that IARS share. Interactions are the actions users perform over items of Furthermore, we propose several categories a recommender system, such as ratings, purchases, or within those identified properties. media watching. Instead, Impressions are the items the system recommends to users, i.e., the items presented 3. We propose a classification system for papers to the user on-screen. Impressions are not exclusive to describing IARS. Such taxonomies inspect each recommender systems; any system that selects a limited paper from diferent perspectives and provide a amount of items to show to the user can be considered a comprehensive paper overview. system that generates impressions, e.g., editorial selec4. We classify and discuss the current state-of-theart on IARS. In our discussion, we provide a comprehensive view of past works. work, we do not assume impressions represent either a need to be a RS. For instance, a search engine or an edipositive or negative type of user feedback; instead, we torial system also generate impressions, i.e., impressions identify which signals impressions may carry and classify may be the results of a search query or a selection of papers accordingly. items the editors of the system want to promote.
Several papers address topics related to the
characterization of IARS. For instance, Pérez Maurera et al. [6] Interaction An interaction is the implicit or explicit
provides an overview of IARS while describing five cate- action of the user over an item where this item is inside
gories of this type of recommenders. Compared to their an impression shown on-screen. Examples of implicit
paper, we integrate their proposed categories into a sin- interactions are clicks or purchases, while examples of
gle taxonomy and provide additional novel taxonomies explicit actions are ratings. Traditionally, implicit
interacthat analyze other aspects of IARS. Moreover, we analyze tions represent the users’ favorable preference, i.e., users
how papers can be classified among all the taxonomies, interact with items they consider align with their tastes.
present how these taxonomies evolve over time, and
discuss how they are associated to others. These discus- User Feedback User feedback refers to the implicit or
sions are not included in the extended version of this explicit preferences of the user to recommended items,
work Pérez Maurera et al. [
Other papers have pursued diferent directions for on their screen, it is assumed they scan the impression IARS, particularly evaluating this type of recommenders and decide to interact with certain items while others while considering the recent calls for sound evaluation remain without interactions. We catalog those items that strategies and better replicability measures in the rec- received an interaction as interacted impressions and ommender systems community. For instance, Zhao et al. those that did not as non-interacted impressions. [7] studied the user preference for those items in impressions which received no interactions. Their results Impression-Aware Recommender System (IARS) suggest that impressions are not strictly a type of neg- An IARS is a type of recommender system that leverages ative user feedback but are complex signals dependent impressions and interactions to learn users’ preferences on the context and time of the recommendation. Pérez toward items in the catalog. Inside a RS, one module Maurera et al. [8, 9] performed two evaluation studies called recommendation model is in charge of learning of some IARS from the literature. The results of their the user preferences. experiments suggest that using impressions increases the accuracy and beyond-accuracy metrics of some existing recommenders. Those papers emphasize the need Inputs The input of a recommendation model is, at for further experiments and the evaluation of various minimum, a user, an item identifier, and users’ profiles. recommenders in future works. This work difers from The users’ profiles are a collection of all the interactions those papers mainly in evaluating recommenders, as we the users performed with the system and the impressions do not perform an evaluation study of IARS. Instead, we the system showed to those users. Depending on the decover the topic of the evaluation of IARS by highlighting sign of a recommendation model, the input may change; the existing recommendation tasks and public datasets. commonly, the input may include additional information
Overall, this work aims to propose a theoretical frame- about users and items, e.g., user demographics or item work to define IARS and describe this type of recom- attributes. menders through their properties and taxonomies.
Output The output of any recommendation model is a real number called the predicted relevance score. Generally, this score tells the preference of a given user over an item in the catalog. Despite this typical case, the relevance score can also portray diferent meanings, e.g., it may reflect the expected click-through rate or the probability of purchase.
pression as a collection of items shown to the user. This section provides a broader definition of impressions and other subjects needed to formalize our taxonomies of this type of recommenders.
pressions that recommendation models compute and utilize to learn user preferences. For instance, one common Impression An impression is an ordered sequence of feature in the literature is the number of times a given potentially relevant items to the user whose contents are item has been impressed by a user. This feature is comarranged on the user’s screen. The source of an impres- monly called the number of impressions. sion, i.e., the entity that generates it, does not necessarily Computational Tasks Based on the definition of IARS, Recommendation Tasks The community in RS has recommendation models, their inputs, outputs, and fea- primarily focused on the task of top- recommendatures, the more evident computational task for a IARS tions, which generates a personalized collection of is computing the relevance score. However, the recom- items to a user called a recommendation list or impresmender computes more than the relevance score. De- sion. Another task, which has increased in popularity, is pending on the recommender’s design, it may compute called re-ranking. Under that task, the recommender several features or needs to interpret interactions and receives an impression holding items. Then the recomimpressions in specific manners. For instance, a recom- mender produces a permutation of such impression [12]. mender may be designed to predict the next song in a playlist; thus, it needs to order impressions and interac- Public Datasets To evaluate IARS, the community has tions by the date and time they occurred. access to thirteen datasets with impressions from diferent recommendation domains. The distribution of the 3.1. Related Recommenders datasets by recommendation domain is: three datasets in news [13, 14, 15, 16, 17], three in online advertiseBased on the definition of IARS shown in the previous ment [18, 19],1 three in media [20, 21], two in fashion [22], section, a recommender system is considered impression- and two in e-commerce [23, 24, 25]. aware when it learns user preferences using impres- The information in those datasets varies greatly. Some sions and interactions. Impressions are a data type that contain the entire impression list and the interactions complements interactions and do not pose any restric- each item received by the users, while others only hold tions or constraints to the underlying recommender. In the number of interactions or impressions of each userother words, a recommender of a diferent type may item pair. Pérez Maurera et al. [9] provide a thorough incorporate impressions without becoming exclusively description of three datasets. Another paper Pérez MauIARS. For instance, a sequential-aware recommender rera et al. [6] shortly describe other types of datasets learns user preferences from ordered sequences of in- in the literature, e.g., those not publicly available to the teractions [10]. Sequential-aware recommenders may community. also become impression-aware if they learn from the sequence of interactions and impressions i.e., they incorporate into their input sequence the impressions generated 4. Properties by the system.
Regarding seemingly similar types of recommenders, This section depicts the properties shared among IARS. Context-Aware Recommender Systems (CARS) seem In this context, a property is a characteristic of the RS comparable to IARS in their definition and may pose and how such characteristic is involved with impressions. as equivalent. A CARS learns user preferences using Notably, we study three properties: which type of impresinteractions and contextual features [11]. In this sense, sions IARS use, how IARS deem impressions in terms of impressions can be seen as contextual features of inter- the users’ preferences, and what kind of recommendation actions. However, these two types of recommenders are model uses an IARS. diferent because the inputs of the recommenders are diferent. Mainly, when producing a relevance score (as 4.1. Impressions Type defined in the previous section), part of the input of a CARS is the current context; IARS cannot receive an impression at this stage because it generates the impression after it computes all the relevance scores.
This section presents specific properties for the evaluation of IARS. In particular, we present the typical recommendation tasks for IARS and list the thirteen public datasets with impressions available for the recommender systems community. Other relevant aspects for the evaluation of IARS; for instance, evaluation methodologies and challenges, have already been discussed in other works [8, 6, 9]. • Contextual: the recommender has access to the user, a possibly interacted item, and the impression holding such an item. In other words, with a contextual impression, the recommender knows every impression shown to users and their feedback on every impressed item, i.e., the feedback indicates whether the user interacted with the impressed item. • Global: the recommender has access to users’ feedback on impressed items, i.e., interacted or non-interacted impressions, but does not have access to the contents of the impressions. In other words, the recommender knows whether a user interacted with an item but does not know to which impression such an item belongs.
In traditional recommender systems research, it is usually considered that interacted items convey positive signals of users’ preferences, i.e., users show their implicit acceptance of an item by interacting with it. Analogously, the non-interacted items are assumed to represent negative signals of user preferences. This scenario is the so-called the missing as negatives assumption [26].
As stated in Section 3, impressions may receive two types of user feedback: interacted impressions and noninteracted impressions. As commonly agreed by the community, we also assume that interacted impressions, i.e., interactions, represent positive signals. However, we do not assign a negative signal to non-interacted impressions beforehand as the literature in IARS does not converge on a single signal for non-interacted impressions. Moreover, we identify three signals to non-interacted impressions: • Neutral: meaning users do not have a positive or a negative preference for non-interacted impressions.
the module of an IARS that learns user preferences and computes the relevance score for any given user-item pair. Despite this, the design of the recommendation model depends on the recommendation task at hand, i.e., top- recommendations, or re-ranking of impressions. As such, we identify three types of recommenders: • End-to-end: a type of recommendation model used in top- recommendations. The model receives the user preferences and generates an impression containing elements. • Plug-in: a type of recommendation model used in top- recommendations. The model receives the user preferences and the predicted relevance scores created by another recommender. The model generates an impression containing elements by transforming those relevance scores. • Re-ranking: a type of recommendation model exclusively used in re-ranking tasks. This type of recommender not only receives the users’ proifles as input but also receives an impression. It generates a permutation of the input impression.
tions of this work: three taxonomies for IARS. This section thoroughly defines each taxonomy, indicating their similarities and diferences and their main categories. These taxonomies are diferent classification systems for recommenders in the literature, which we inspect under three perspectives: model-centric, data-centric, and signal-centric.
• Heuristics: contains all papers which describe a recommendation model using rules, associations, or ad-hoc approaches to model user preferences.
For example, Buchbinder et al. [27] does not recommend an item after a certain number of impressions. • Statistical: contains all papers which describe a recommendation model using probability or statistical models to model user preferences. For example, Zhang et al. [28] uses logistic regression. • Machine learning: contains all papers which describe a recommendation model using shallow machine learning approaches to model user preferences. For example, Liu et al. [29] uses gradient boosting decision trees. • Deep learning: contains all papers which describe a recommendation model using deep machine learning approaches to model user preferences. For example, Covington et al. [30] uses multilayer perceptrons. • Negative: meaning users dislike non-interacted impressions. Under the traditional missing as negatives assumption, non-interacted impressions are deemed as negative signals.
In the model-centric taxonomy, we analyze the recommendation model a specific paper uses, i.e., we examine • Positive: meaning users prefer non-interacted the design of the recommendation model and its learnimpressions; however, they decided not to inter- ing paradigm. Hence, this taxonomy does not classify act with them when shown. papers based on how they use impressions nor how they deem the user preferences to impressions. Instead, the taxonomy focuses on classifying the recommendation model of an IARS. Under this taxonomy, we propose five categories of IARS:
In the data-centric taxonomy, we analyze how a recommendation model uses impressions to learn users’ preferences. In other words, we analyze the input of the recommendation model. Under this taxonomy, we propose three categories of IARS: • Assume: contains all papers describing a recommendation model that assumes users’ specific preference toward non-interacted impressions.
For example, Xi et al. [
For example, Ma et al. [32] learn to classify items In this section, we classify relevant papers in the literainto two classes: interacted and non-interacted ture, which we deem as state-of-the-art, under the propimpressions. erties and taxonomies of IARS discussed in Section 4 and Section 5, respectively. Before the classification, we • Sample: contains all papers which describe a identify the state-of-the-art in IARS by conducting a sysrecommendation model where its input is a col- tematic literature exploration, describing the discovery lection (e.g., a set, a sequence, or a vector) con- process and selection criteria for papers. Overall, in two taining items sampled from interactions, impres- taxonomies, we find that works are distributed almost sions, or both. For example, Pérez Maurera et al. uniformly amongst the proposed categories, while in the [20] sample impressions as negative items and remaining taxonomy and the three properties, most painteractions as positive items when training a pers favor one category over the rest. We also find that BPR-optimized [33] matrix factorization recom- almost all taxonomies and properties show large statistimender. cally significant associations against others. In contrast, only one taxonomy does not have statistically significant associations with the rest of the properties and taxonomies. Table 1 shows the distribution of papers according to each taxonomy and property, and Table 2 shows the association between properties and taxonomies.
Unlike the previous taxonomy, in the data-centric taxonomy, one paper may belong to two categories simultaneously. This relaxed property is allowed to avoid the creation of categories that combine two of the previous categories. For example, Aharon et al. [34] propose a recommender that learns from impressions and computes frequency features; therefore, we classify the paper as learn and features under the data-centric taxonomy.
In this work, we consider a paper to belong to the state-ofthe-art of IARS when the paper meets these conditions:
ranking: https://www.scimagojr.com/journalrank.php?area=1700
The first condition excludes pre-prints. The second the query “recommender system AND (impression OR condition excludes posters, demos, extended abstracts, exposure OR slate OR past recommendation OR previworkshops, and short papers. The third condition ex- ous recommendation)”. This query matches the keyword cludes conference venues classified with a ranking of recommender system and keywords related to IARS. “B” or lower according to the CORE Rank 2021. It also excludes journals classified with a ranking of “Q2” or 6.2. Recommendation Models lower according to the Scimago 2021 ranking in computer science.2 Conversely, the third condition only accepts We identify a learning paradigm shift regarding the types “A” or “A*” conferences or “Q1” journals. Lastly, condi- of published recommenders. Notably, early works (before tion 4 ensures that discovered papers are relevant to the 2018) primarily published recommenders using heuristic, IARS community. The condition excludes papers with probabilistic, or shallow machine learning; moreover, keywords related to impressions without describing an only one paper uses reinforcement learning. We do not IARS. identify any recommendation model used to a greater
We searched through five popular academic search extent than others amongst reviewed papers on those
engines to discover, review, analyze, and select those pa- four groups.
pers deemed relevant to this work. We queried the ACM When analyzing more recent papers, i.e., those
pubDigital Library, IEEE Xplore, SpringerLink, ScienceDi- lished after 2018, we see that most papers describe
eirect, and Google Scholar academic search engines with ther recommenders using deep learning or
reinforcement learning. From those papers using deep learning
architectures, four papers [
the missing as negatives assumption, i.e., where noninteracted impressions are seen as negative signals of users’ preferences. 6.3. Data-Centric Taxonomy Despite many papers following the missing as negatives assumption, some papers in the literature do not Regarding the data-centric taxonomy, we see three major consider impressions as negative signals; instead, they trends: papers only learn using impressions, only use dissect the signals within non-interacted impressions. In features, or both. Moreover, only two papers sample particular, Zhao et al. [7] performed a user study where from impressions; those papers also use impressions as they surveyed participants to express their preference negative signals. toward non-interacted impressions. The results of such
Regarding features computed from impressions, the a study suggest that users are mostly unaware of all immost common feature is the number of impressions a pressed items, i.e., users only scan part of the impression given user received for a given item. Other features used list. Moreover, the paper shows that only 5.8 % of nonin a few papers include the average position of an item impressed items were disliked by users. The results of in an impression, the number of days between two im- those papers suggest that considering non-interacted impressions with the same item, average watch time of pressions as negative signals may be an oversimplification impressed items, among others. of users’ preferences.
On the signal-centric taxonomy, most papers assume In Section 3, we defined three types of recommenders: a specific signal for non-interacted impressions; from end-to-end, plug-in, and re-ranking. Their main difthose, the assumed signal is negative, i.e., the papers ference is how they generate an impression. Table 1 deem non-interacted impressions as negative items to shows that most papers, approximately 72 %, describe user preferences. end-to-end recommenders. Those papers describe a rec
Nine papers learn the user preferences for non- ommender that uses impressions and generates the final interacted impressions; seven assume non-interacted im- impression shown to the user without further processing pressions are neutral signals, while two assume negative by another recommender. signals. For the former, they allow the recommendation As shown in Table 2, the type of recommender has a models to learn such a preference. For the latter, they statistically significant association with the taxonomies design their recommenders to learn how much the user we propose: the model-centric, data-centric, and signaldislikes non-interacted impressions. centric taxonomies. Mainly, plug-in recommenders only appear in papers describing heuristic or statistical ap6.5. Types of Impressions proaches. On the data-centric taxonomy, re-ranking recommenders do not sample from impressions. Lastly, on In Section 4, we presented two types of impressions: the signal-centric taxonomy, plug-in recommenders exglobal and contextual. Their diference is that the former clusively learn the signal from non-interacted impresdoes not associate interactions and impressions, while sions, while the majority of re-ranking ones assume a the latter provides such association. specific signal.
Most papers, approximately a 66 % of them, use global impressions. Moreover, those papers use one particular type of global impressions: impressions as user-item- 7. Conclusions label triplets. Under this type of impressions, the recommender’s input is a triplet containing a user identifier, an item identifier, and a binary label that indicates whether the item is an interacted or non-interacted impression.
Despite most papers using a particular type of impressions, Table 2 shows no statistically significant associa
the definition, characterization, and classification of impression-aware recommender system, i.e., recommenders learning users’ preferences from impressions and interactions. In our first discussions, we provide the definitions of core concepts to IARS, such as impression, interactions, user feedback, among others. Also, we compare IARS to other types of recommenders, where we identified that IARS are a distinct type of recommenders despite their resemblance to others.
We present one of our main contributions to this work: the characterization of IARS in terms of their properties and taxonomies. For the former, we identify three properties that cover how recommenders consider user preferences to impressions, the kind of impressions they use, and how the recommender generates future impressions.
For the latter, we propose three classifications, covering diferent aspects of IARS, i.e., the learning paradigm of the recommender, how the recommender uses impressions, and whether the recommender assumes or learns users’ preference toward impressions.
Lastly, we classify papers belonging to the state-ofthe-art under our proposed properties and taxonomies.
We select relevant papers by applying specific selection criteria, focusing on papers published in high-level conferences and journals. In our classification, we discuss the general trends of the state-of-the-art under each property or taxonomy; at the same time, we indicate how they relate. Our study reveals a strong association between the signal of impressions and papers assuming or learning a specific signal in impressions. Furthermore, most papers assume a negative signal to non-interacted impressions, following the traditional missing as negatives assumption.
However, as indicated by Zhao et al. [7], this assumption may be erroneous. Some of our previous works [57, 8, 9] align to both assumptions, i.e., in some cases impressions represent negative signals, while in others it does not.
Such results call for future works that address this topic in particular. Information Retrieval, Madrid, Spain, July 11 - 15, 3290999. 2022, ACM, 2022, pp. 1185–1195. doi:10.1145/ [54] J. McInerney, B. Lacker, S. Hansen, K. Higley, 3477495.3532040. H. Bouchard, A. Gruson, R. Mehrotra, Explore, [48] Z. Chen, J. Wu, C. Li, J. Chen, R. Xiao, B. Zhao, Co- exploit, and explain: personalizing explainable rectraining disentangled domain adaptation network ommendations with bandits, in: S. Pera, M. D. for leveraging popularity bias in recommenders, Ekstrand, X. Amatriain, J. O’Donovan (Eds.), Proin: E. Amigó, P. Castells, J. Gonzalo, B. Carterette, ceedings of the 12th ACM Conference on RecJ. S. Culpepper, G. Kazai (Eds.), SIGIR ’22: The 45th ommender Systems, RecSys 2018, Vancouver, BC, International ACM SIGIR Conference on Research Canada, October 2-7, 2018, ACM, 2018, pp. 31–39. and Development in Information Retrieval, Madrid, doi:10.1145/3240323.3240354. Spain, July 11 - 15, 2022, ACM, 2022, pp. 60–69. [55] H. Cramér, Mathematical methods of statistics, doi:10.1145/3477495.3531952. Princeton university press, 1946. [49] Z. Zhao, L. Hong, L. Wei, J. Chen, A. Nath, S. An- [56] J. Cohen, Statistical power analysis for the behavdrews, A. Kumthekar, M. Sathiamoorthy, X. Yi, E. H. ioral sciences, 2 ed., Lawrence Erlbaum Associates, Chi, Recommending what video to watch next: a 1988. multitask ranking system, in: T. Bogers, A. Said, [57] F. B. Pérez Maurera, M. Ferrari Dacrema, P. Castells, P. Brusilovsky, D. Tikk (Eds.), Proceedings of the P. Cremonesi, Incorporating impressions to graph13th ACM Conference on Recommender Systems, based recommenders, in: LERI 2023: ProceedRecSys 2019, Copenhagen, Denmark, September ings of the 1st Workshop on Learning and Evaluat16-20, 2019, ACM, 2019, pp. 43–51. doi:10.1145/ ing Recommendations with Impressions, Singapore, 3298689.3346997. September 19, 2023, CEUR Workshop Proceedings, [50] Y. Ge, X. Zhao, L. Yu, S. Paul, D. Hu, C. Hsieh, CEUR-WS.org, 2023.
Y. Zhang, Toward pareto eficient fairness-utility trade-of in recommendation through reinforcement learning, in: K. S. Candan, H. Liu, L. Akoglu, X. L. Dong, J. Tang (Eds.), WSDM ’22: The Fifteenth ACM International Conference on Web Search and Data Mining, Virtual Event / Tempe, AZ, USA, February 21 - 25, 2022, ACM, 2022, pp. 316–324.
doi:10.1145/3488560.3498487. [51] R. Xie, S. Zhang, R. Wang, F. Xia, L. Lin, Explore, iflter and distill: Distilled reinforcement learning in recommendation, in: G. Demartini, G. Zuccon, J. S. Culpepper, Z. Huang, H. Tong (Eds.), CIKM ’21: The 30th ACM International Conference on Information and Knowledge Management, Virtual Event, Queensland, Australia, November 1 5, 2021, ACM, 2021, pp. 4243–4252. doi:10.1145/ 3459637.3481917. [52] C. Pei, X. Yang, Q. Cui, X. Lin, F. Sun, P. Jiang,
W. Ou, Y. Zhang, Value-aware recommendation based on reinforcement profit maximization, in: L. Liu, R. W. White, A. Mantrach, F. Silvestri, J. J.
McAuley, R. Baeza-Yates, L. Zia (Eds.), The World Wide Web Conference, WWW 2019, San Francisco, CA, USA, May 13-17, 2019, ACM, 2019, pp. 3123– 3129. doi:10.1145/3308558.3313404. [53] M. Chen, A. Beutel, P. Covington, S. Jain, F. Belletti, E. H. Chi, Top-k of-policy correction for a REINFORCE recommender system, in: J. S. Culpepper, A. Mofat, P. N. Bennett, K. Lerman (Eds.), Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining, WSDM 2019, Melbourne, VIC, Australia, February 11-15, 2019, ACM, 2019, pp. 456–464. doi:10.1145/3289600.