Using Shapley Additive Explanations for supporting individual and group recommendation: survey and new perspectives

Using Shapley Additive Explanations for supporting individual and group recommendation: survey and new perspectives Raciel Yera, Bapi Dutta, and Luis Mart´ ınez Abstract Recommender systems are essential tools that assist users in navigating vast information spaces by filtering and ranking options based on their preferences and behaviors. While traditional recommender systems focus on individual users, group recommender systems extend this concept to scenarios where items are consumed collectively. These systems must aggregate diverse user preferences and resolve conflicts to generate recommendations that satisfy the group as a whole. Various aggregation techniques, including social choice-based methods, are employed to achieve this goal. Additionally, contextual factors, such as group size and decision-making styles, play a relevant role in shaping recommendations. Beyond enhancing accuracy, trustworthiness has become an important concern in recommender systems, aligning with European AI guidelines that emphasize transparency, accountability, and fairness. Explainability is central to fostering trust, enabling users to understand the rationale behind recommendations. This chapter explores post-hoc explainability approaches, which offer flexible explanations that can be integrated into various recommendation frameworks. Specifically, the chapter explores the use of SHapley Additive exPlanations for recommendation explanations. Despite some research efforts in this direction, a systematic analysis is lacking. This contribution aims to bridge that gap by examining SHAP’s applicability in recommender systems, offering through a PRISMA-based systematic review, insights into its potential for enhancing transparency and user trust. A general framework for explaining individual and group recommendation using Shapley Additive Explanations is provided. Furthermore, a case study is developed to show how the proposal can effectively accomplish its goal. Raciel Yera University of Ja´ en, Campus Las Lagunillas s/n, e-mail: [email protected], Bapi Dutta University of Ja´ en, Campus Las Lagunillas s/n, e-mail: [email protected], Luis Mart´ ınez University of Ja´ en, Campus Las Lagunillas s/n, e-mail: [email protected], 1 This is a preprint of the following chapter: Yera R., Dutta, B. & Martínez, L., Using Shapley Additive Explanations for supporting individual and group recommendation: survey and new perspectives, published in Decision-Making in Computational Intelligence-Based Systems, edited by Pedrycz, W., Rivera, G., Rodríguez, R.M., Ibarra-Martínez, S., 2026, Springer, reproduced with permission of Springer Nature Switzerland AG. The final authenticated version is available online at: http://dx.doi.org/10.1007/978-3-032-04056-5_1 2 Raciel Yera, Bapi Dutta, and Luis Mart´ ınez Key words: recommender system; Shapley additive explanations; group recommendation; survey; trustworthy artificial intelligence 1 Introduction Recommender systems are intelligent tools designed to assist users in discovering items or information of interest by filtering and ranking options based on their preferences or behavior [1]. These systems operate using various paradigms: content-based recommendation focuses on item attributes to match user preferences; collaborative recommendation leverages the preferences of similar users; knowledge-based recommendation relies on domain knowledge to suggest items based on explicit needs; demographic recommendation tailors suggestions based on user demographics; and hybrid recommendation combines multiple approaches to enhance accuracy. Recommender systems find applications across diverse domains, including e-commerce (personalized product suggestions), e-learning (customized learning resources), eservices (service recommendations), e-tourism (travel planning), and e-health (personalized health advice) [24]. By delivering relevant and timely suggestions, these systems improve user satisfaction, decision-making, and engagement in both personal and professional contexts. Being traditional recommender systems usually focused on individuals, group recommendation systems extend such concept by focusing on scenarios where items are consumed collectively by a group rather than an individual [11]. These systems are designed to be focused on the preferences, needs, and dynamics of multiple users to suggest items that satisfy the group as a whole. This is particularly relevant in domains where decisions are usually collaborative, such as selecting travel packages for families or friends, choosing movies or TV shows for group viewing, or planning group dining experiences. Group recommendation are centered on aggregating diverse preferences and resolving conflicts to find a solution acceptable to all members [10]. The main used techniques include preference aggregation and recommendation aggregation methods. Such aggregation is usually modeled through the use of social choice-based approaches such as additive, least misery, approval voting, or most pleasure aggregation [11]. Transversally, contextual factors, such as group size, relationships, and decision-making styles, also drive the generated recommendations [28]. Beyond accuracy, trustworthiness in recommender systems is essential for ensuring user acceptance, satisfaction, and long-term engagement, especially in light of the European AI guidelines emphasizing ethical and transparent AI [42]. These guidelines highlight the importance of accountability, fairness, and transparency in AI systems, pointing out recommender systems as AI-based tools, and the necessity of operating in a way that users can understand, trust, and rely on. The current contribution concerns explainability in recommender systems, which is the key concept managed by the research community for accomplishing the transparency goal [48]. It refers to the ability of the system to provide clear, understandable SHAP for Individual and Group Recommendations 3 reasons behind its suggestions, allowing users to comprehend why certain items are recommended. Zhang et al. [48] point out different types of explanations in recommender systems: 1. Intrinsic recommendation explanations, built as component of specific recommendation approaches, providing insights directly tied to its internal logic, such as highlighting shared features or remarkable directions that directly lead to the system’s output. 2. Post-hoc recommendation explanations, focused on explaining the output of the recommendation approach, instead of being focused on its inner logic. Particularly, the current contribution is centered on post-hoc explanations, which are flexible approaches that can be coupled with any recommendation framework and are focused on finding some common sense reasoning that would justify the delivery of some specific recommendation [48]. Due to their flexibility, they have attracted the attention of the research community over the last few years. Peake and Wang [27] present one of the first approaches focused on this goal, focused on mining local association rules representing the relationship between items, that are used for building the subsequent explanations of the recommended items. In parallel, Quijano-S´ anchez et al. [31] present a methodology for explaining group recommendations by exploiting social relationships between group members. Ariza et al. [2] recently presented a comparative analysis between current text-based explainable recommender systems. Furthermore, Yera and Mart´ ınez [46] have also developed a local rules-based approach that works over item features to build factual and counterfactual recommendations for individuals and groups. The development of explanation interfaces for supporting recommendations has been also a relevant research effort covered by several authors [37]. Graphical interfaces such as bar charts, connected graphs, or line charts; have been used for representing feature weights, feature contributions, neighborhood information etc, and therefore justify the recommendation. Over the last few years, visual explanation paradigms such as LIME or SHAP, have been also brought to the recommender systems scenario. The aim of the current book chapter is to explore the use of SHAP for supporting recommendation explanations. While some research efforts have been made in this direction, there remains an important gap in systematic analysis and contributions that would facilitate the structured and consistent application of this approach both in individual and group recommendation systems. Particularly, Section 2 introduces the necessary background, including individual and group recommendation and posthoc explanation in recommender systems. Section 3 presents the systematic review methodology on the use of SHAP for supporting recommender systems. Section 4 discusses a general framework for explaining individual and group recommendation using SHAP, being one of the first efforts on covering these research task in the group recommendation scenario. Section 5 develops a case study for applying the proposed framework in a real scenario. Section 6 concludes the contribution. 4 Raciel Yera, Bapi Dutta, and Luis Mart´ ınez 2 Background This section will be focused on providing a background on individual and group recommendations, which is necessary for introducing the subsequent stages of this contribution. Furthermore, a brief overview on post-hoc explanations in recommender systems is also provided. 2.1 Individual and group recommendation Two main paradigms have driven the development of individual recommendations since 90s [1]. These have been: 1) the content-based recommendation approach, focused on recommending items similar to those consumed in the past by the active user, and 2) the collaborative filtering approach, focused on recommending items that have been consumed in the past by other users with a rating behavior similar to the current one. Content-based recommendation methods, as described by Adomavicius and Tuzhilin [1], operate on the principle of “Give me more of what I’ve already enjoyed.” These systems are defined by three key components: a framework to represent the items that can be suggested, a mechanism to build a user profile capturing their preferences, and a method to match items against the user’s profile to decide which ones to recommend. The recommended items are typically characterized by a set of features or attributes, each with defined possible values. A user’s level of preference for certain items is assessed, and machine learning algorithms are employed to develop a profile based on those attribute values. This enables the system to identify, among the remaining items, those that offer the highest relevance or utility to the user. Consequently, the two central challenges in these systems are how the items 𝑠 are represented (𝐶𝑜𝑛𝑡𝑒𝑛𝑡(𝑠)) and how the user’s profile 𝑐is accurately learned (𝐶𝑜𝑛𝑡𝑒𝑛𝑡𝐵𝑎𝑠𝑒𝑑𝑃𝑟𝑜 𝑓 𝑖𝑙𝑒(𝑐)). The ultimate goal is to find the utility function for the user 𝑐, of item 𝑠: 𝑢(𝑐, 𝑠)=𝑠𝑐𝑜𝑟𝑒(𝐶𝑜𝑛𝑡𝑒𝑛𝑡𝐵𝑎𝑠𝑒𝑑𝑃𝑟𝑜 𝑓 𝑖𝑙𝑒(𝑐), 𝐶𝑜𝑛𝑡𝑒𝑛𝑡(𝑠)) (1) Collaborative filtering operates on the principle of leveraging collective preferences to recommend movies to a specific user. This method identifies, either explicitly or implicitly, patterns in users’ ratings that align with those of the current user and employs this shared information to generate recommendations. Collaborative filtering techniques can be broadly categorized into two main types: (1) memory-based approaches [13], which focus on directly identifying similar user groups, or “neighborhoods” for the active user and utilizing the preferences of these closest neighbors to make suggestions, and (2) model-based approaches [20], which involve creating intermediate models that encapsulate the preferences of the overall set of users, for leading to the recommendation generation. Summarizing, the goal of collaborative SHAP for Individual and Group Recommendations 5 filtering would be the calculation of the current rating 𝑟(𝑐, 𝑠), based on the aggregation of the ratings associated with the same item, provided by other users 𝑐′, similar to the current one 𝑐. 𝑟(𝑐, 𝑠)=𝑎𝑔𝑔𝑐′∈𝐶𝑟(𝑐′, 𝑠)(2) These methods have gained significant popularity in several recommendation scenarios, due to their ability to deliver precise suggestions by relying solely on user ratings, without requiring additional information. This distinguishes them from content-based methods, which depend on detailed item attributes to perform effectively. Beyond the recommendation generation for individuals, there are several contexts with items that are usually consumed by groups, such as travel packages, TV programs, restaurants, music, etc; and that therefore need different recommendation algorithms that suggest items that fit the necessity and preferences of a set of users, and not an individual one. Two alternative approaches drive the conception of a group recommender system: •Recommendation aggregation approach: Here each member of the group receives individual recommendations using a method tailored for the individual scenarios, and after that, such individual recommendations are aggregated to obtain the group recommendations. •Preference aggregation approach: In this case the aggregation is done over the individual preferences of the group members, to build a pseudo-user profile that represents the preferences of the group. The group recommendations are delivered to this pseudo-user profile, using an individual recommendation method. The aggregation of the user preferences or the generated recommendations has been usually done with the use of social choice-based aggregation approaches [11], such as the following ones: •Additive: It is focused on adding the preferences associated with each individual in the group, for composing the group preferences. 𝑟𝐺,𝑖 =∑︁ 𝑢∈𝐺 𝑟𝑢,𝑖 (3) •Least Misery: It is focused on identifying the preference associated with the group, as the minimum preference linked to any of their members. 𝑟𝐺,𝑖 =𝑀𝑖𝑛𝑢∈𝐺𝑟𝑢,𝑖 (4) •Approval Voting: It identifies the group preferences, as the amount of users whose individual preferences are over some predefined threshold. 𝑟𝐺,𝑖 =|𝑟:𝑟𝑢,𝑖 > 𝜃, 𝑢 ∈𝐺|(5) 6 Raciel Yera, Bapi Dutta, and Luis Mart´ ınez •Most Pleasure: It identifies the group preferences, as that associated with the member with the highest preference value. 𝑟𝐺,𝑖 =𝑀𝑎𝑥𝑢∈𝐺𝑟𝑢,𝑖 (6) Taking as base these foundations, in the last few years several authors have been focused on developing more sophisticated approaches. Dara et al. [10] conducted a comprehensive survey on the latest advancements in group recommender systems, covering various domains and analyzing different systems based on their aggregation strategies and user preference modeling. Yalcin et al. [41] introduced two innovative aggregation methods that combine additive utilitarian and approval voting approaches to highlight popular items agreed upon by group members. Similarly, Contreras et al. [8] proposed a collaborative framework grounded in the social dynamics within web-based conversational group recommender systems. Their model enables the system to implicitly identify different roles within the group, such as collaborative members and leaders. Lastly, Ismailoglu [18] applied the Expectation-Maximization (EM) algorithm to merge group members’ preferences, estimating group ratings and evaluating the expertise levels of individual members in group recommendation scenarios. 2.2 Post-hoc explanations in recommender systems This section will cover a brief overview of recent advances in post-hoc explanations in recommender systems. To accomplish this aim, the related works will be categorized into four main groups, recognizing that their objectives may be overlapped in some cases: 1) explanations supported by traditional data mining techniques, 2) explanations supported by alternative, simpler and transparent recommendation models, 3) explanations supported by generative models, usually with the aim of text generation, and 4) visual explanations supported by established model-agnostic machine learning approaches. Explanations supported by traditional data mining techniques: Peake and Wang [27] developed a representative approach belonging to this category, that use the discovery of association rule mining for matching preferred and recommended items. Specifically, given a set of recommended items, the approach is focused on discovering rules with the shape 𝐴−> 𝐵 where the consequent 𝐵of the rule is represented by recommended items, which recommendation would be justified by the preferences of the users over items in 𝐴, in case of it applies for the active user. This work exploits the concept of local explanations, by mining the rules just using data of the neighborhood of the current user. Across the recent years, this approach has been used for explaining recommendation in specific scenarios such as cooking recipes [42], as well as incorporated as key explanation approach in several recommender systems libraries [7]. SHAP for Individual and Group Recommendations 7 Furthermore, Yera and Martinez [46] has recently explored the use of rules that connect item features with preferred items, for performing factual and counterfactual explanations in individual and group recommendations. The explanation generation is supported by the path from the root to the leaves, of decision trees built taking as base the information of items similar to the recommended one. Herein the group recommendation explanations are built by aggregating the explanations generated for the individuals. Explanations supported by alternative, simpler and transparent recommendation models: Some previous works have explained black-box recommendation approaches through the use of simpler and transparent recommendation models, employed as a proxy for reaching the same output of the main black-box approach. This is the goal of the framework proposed by Schmaryahu et al. [34] (Figure 1), where initially the black-box recommendation algorithm delivers recommendations, which are attempted to be explained by several explanation algorithms, using additional data sources. The explanation algorithms are based on basic content-based and collaborative filtering approaches, such as item-item content-based, user-item content-based, item-item overview, item-item collaborative filtering, and user-user collaborative filtering. Fig. 1: General explanation approach based on simpler models In a different direction, Tran Trang et al. [38] and Barile et al. [5] have systematized the use of several social choice-based explainable aggregation approaches, such as additive utilitarian, least misery, approval voting, most pleasure, fairnessbased, and so on. Furthermore, they have shown that the generation of explanations focused on illustrating the internal reasoning behind these aggregation approaches could positively impact the perceived group satisfaction, consensus, and fairness. Additionally, the effect across several group configurations, such as divergent, minority, or coalitional, was investigated. Recently, Waterschoot et al. [40] evaluate the objective and subjective user understanding of social-choice based explainable aggregation approaches. Explanations supported by generative models, usually with the aim of text generation. With the rapid growth of LLMs-based models, some authors have used their benefits for generating text-based explanations of the output of recommender systems. This research line has the particularity of working over previous user reviews to generate textual explanations. Even though this could be a desired capability for 8 Raciel Yera, Bapi Dutta, and Luis Mart´ ınez most e-commerce environments, it may also imply a limited application in other recommendation scenarios. Ariza et al. [2] have recently presented an analysis and benchmarking of recent research focused on this aim. Herein, related works have been classified as text generation-based, text extraction-based, or hybrid. Generation-based models compose the explanations word by word in an autoregressive way, taking into account past user-item interactions. The primary shortcoming of these models is their lack of controllability in their generation process, which produces repetitive and shallow explanations. In this family of works, PETER [22] was one of the first explainable models that took advantage of a transformer-like architecture to accomplish multiple tasks such as rating prediction, context prediction, and explanation generation. Focused on a more specialized goal, SEQUER [3] extends PETER through the inclusion of user interaction sequences and explorations of different masking alternatives to improve the representation capabilities. More recently, POD [23] introduced a transformer-like model that uses a pre-trained multitask LLM, that is tuned to perform top 𝑁recommendations, next-item recommendation, and explanation generation. Extraction-based models build the explanation by selecting the top-𝑘explanation sentences for a candidate set that is composed of user-item review sentences. In this group of works, Pugoy and Kao [30] have applied unsupervised text clustering over review sentences encoded with a BERT-based approach, for creating user and item profiles, which can be used to generate rating predictions and their subsequent explanation. Furthermore, Wang et al. [39] construct a heterogeneous graph of useritem-attribute nodes taking as starting point the user and item reviews, which is used later to rank the candidate sentences and attributes presented in the graph, supported by a graph neural network solution. Overall, these works are limited by the diversity and completeness of the set of review sentences and then can be affected by the cold start problem when the set of reviews does not have sufficient coverage. Hybrid architectures have been also proposed, by following a retrieval-augmented generation viewpoint, in order to combine the generation-based and extraction-based approaches. Visual explanations supported by established model-agnostic machine learning approaches. In this group of works we will explore a set of works that have brought to the recommender systems environment, established model-agnostic approaches that have been systematically used for explaining the output of machine learning models, and that usually come with some kind of visual information. Formerly, Ribeiro et al. [32] proposed the creation of Local Interpretable Modelagnostic Explanation models (LIME), which leverage sparse linear models to approximate the behavior of a black-box classifier in the vicinity of a specific instance. These linear models can then identify the sample features that significantly influence the predicted outcome. Building on the LIME framework, N´ obrega and Marinho [25] developed LIME-RS, a variant tailored for recommender systems, designed to pinpoint the top-n item attributes that most effectively elucidate individual recommendations generated by a factorization machine. Additionally, Chanson et al. [6] introduced LIRE, an enhanced version of LIME for recommender systems, which SHAP for Individual and Group Recommendations 9 incorporates optimized sampling around the recommendation instance to construct a more accurate local surrogate model. Taking an alternative approach, SHapley Additive exPlanation values (SHAP) have also emerged as a post-hoc explanation framework for recommender systems [16]. This method emphasizes calculating the average marginal impact of each feature value on the model’s prediction across all possible permutations. Recently, Zhong and Negre [49] have used the SHAP framework as a tool for identifying possible candidates for composing counterfactual explanations. In contrast to text-based explanation approaches, these methods are able to generate explanations using just ratings and attribute values. However, in a different way in relation to the previous group of works based on traditional data mining techniques or simpler recommendation methods, they have been explored with a lesser extent. Particularly, we will be interested on SHAP values, which can be used both at the global and local level. Furthermore, in contrast to LIME-based approaches, they do not suffer from stability concerns [49]. The next section will be focused on discussing the direct antecedents on the use of SHAP for supporting explanations recommender systems at individual and group level. 3 Antecedents of the use of SHAP for supporting recommender systems This section is focused on identifying key works across the literature, focused on explaining the recommender system output based on SHAP. With this purpose in mind, the Preferred Reporting Items for Systematic reviews and Meta-Analyses methodology (PRISMA) will be driven. Additionally, the identified works will be screened in detail. 3.1 Systematic review methodology This section outlines the methodology used to perform a systematic review of the literature on food recommender systems for diabetic patients. The review was carried out according to the PRISMA 2020 guidelines (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) [26]. To achieve this, we followed the steps specified in the PRISMA 2020 checklist throughout the development of the survey. The main stages of this checklist are described in detail in this section, being those also used by previous works also in the domain in recommendation systems, and therefore generalized for systematic reviews in this context. These are: 1) eligibility criteria, 2) information sources and search strategy, 3) the studies selection, 4) data extraction and data items, 5) quality evaluation, and 6) the analysis of the obtained results. 16 Raciel Yera, Bapi Dutta, and Luis Mart´ ınez 3.2.4 SHAP-based exploration of recommender systems with longitudinal data in the finance service domain Diaz et al. [12] have recently explored the impact of longitudinal data in a recommendation environment, comparing traditional collaborative filtering approaches, multi-label classification algorithms, and a deep learning approach based on gated recurrent units. They are specifically focused on a dataset containing product ownership data for 35 distinct financial services and their recency, frequency, and monetary information. Demographic data is also available from 300000+ clients. Moreover, SHAP is deployed to better understand the output of the best performing models. With this purpose in mind, the products are grouped into different categories such as credits, insurance, savings, and services. The absolute SHAP values for the features belonging to each group are summed to find their global importance associated with each category. For each different category, different sets of relevant features were identified. For example, the recommendation of credit products was primarily influenced by features such as the recency and frequency of Savings products, as well as the frequency, monetary, and recency features for Insurance products (Figure 7). Fig. 7: Mean associated SHAP values for Credit products (Diaz et al. [12]) SHAP for Individual and Group Recommendations 17 4 Towards a general framework for explaining individual and group recommendations using Shapley Additive Explanations The previously analyzed works suggested the presence of a still incipient development towards the task of taking advantage of SHAP for explaining recommender system outputs. In the following section, we discuss a general framework that could be used for systematizing the research and development tasks towards this effort. Figure 8 presented this general framework, which would be composed of the steps discussed below, and that concerns individual and group recommendations. Fig. 8: Our framework for systematic SHAP-based explanations in recommender systems 4.1 Identify the item features and their associated output for each individual user, in the recommendation model. At first, it is necessary to define the information input that will be used for explanation generation. Based on the works analyzed in the previous section, there is a consensus of previous works on the use of item features for supporting the subsequent explanations generation. In this case, it is necessary to precise whether taking into account the nature of the feature (categorical or continuous), it would be necessary some kind of discretization procedures for transforming them in a way that could be managed by the traditional SHAP approaches. Furthermore, in the case of group recommendations, it is necessary to analyze the pros and cons of performing some kind of feature value aggregations at this stage (as done in content-based group recommendation approaches [29]), or leaving this process for a later stage and therefore avoiding information losing. It is also necessary to precise what would be the output of the model that will be explained later by SHAP. Zhong and Negre [49] consider a binary output, in which the value 1 indicates that the item would be in the recommendation list, and the value 0 indicates that it does not. Other works, such as Yin et al. [47], manage the probability of an item to be in the recommendation list. Finally, we stress that it would be interesting to explore alternative ways for representing items beyond the traditional features, such as the preferences of some 18 Raciel Yera, Bapi Dutta, and Luis Mart´ ınez key users in the dataset, or the preferences of certain groups in the case of group recommendation. 4.2 Calculate the feature contribution and ranking for each individual user. Once defined the shape of the item features and the output of the model, the main challenging task at this stage is the calculation of the feature contribution for each individual user. Here it is important to point out that in contrast to the scenario usually used for applying SHAP, which is composed of a set of instances characterized by features, and an output/class for each instance, in the recommender system scenario is necessary to also consider the presence of a user that would provided the output (rating), associated to the current instance (item). This fact leads us to consider the application of SHAP at two different levels: the global level considers an important part of all the available dataset, and the user level focuses on explaining just the feature contributions based on the active user preferences. The approach suggested by Yin et al. [47], focused on ”macroscopic” or ”microscopic” viewpoints of the SHAP explanations, could be an appropriate starting point for supporting this issue. Anyway, taking into account the sparsity of recommender systems datasets, it is important to analyze their feasibility, specifically for the user-level case. Another important design decision to be regarded here is the appropriateness of the specific SHAP implementation to be used, for the recommender systems environment. In practice, it is expected that most of the features that characterize items will be categorical [7, 28, 29]. Therefore, it would be necessary to exploit previous SHAP implementations tailored to categorical feature scenarios [21], and even the use of some classifier such as CatBoost [17], developed for such environment. 4.3 Generate the feature ranking at the group level. In addition to the use of SHAP at the global and individual levels, defining SHAPbased explanations to justify the recommendations delivered to a specific group of users in the context of group recommendation is a challenging task. To cover this goal, according to our viewpoint the first stage would be to define whether the aggregation of the group member information will be performed at the first stage of identification of features (e.g. the profiling stage), or at this current stage by obtaining the feature ranking based on the individual ranking linked to each of the group members’. In the latter case, some aggregation operators need to be formalized, which can be inspired by the proposed feature aggregation schemes for content-based group recommendation [14, 29]. SHAP for Individual and Group Recommendations 19 Finally, hybrid approaches focused on simultaneously presenting explanations at the individual and group levels could also be considered. 4.4 Generate the final explanations at individual and group level. In the last stage, generating the final explanations at both the individual and group levels is not a trivial task. Since the expected output of SHAP-based approaches is typically standardized (see Figures 3-7), some authors, such as Zhong and Negre [49], have pointed out that these outputs could be difficult for some users to interpret. Particularly, Zhong and Negre [49] used the output of SHAP to generate counterfactual explanations in the form of a set of < 𝑎𝑡𝑡𝑟𝑖𝑏𝑢𝑡𝑒, 𝑣𝑎𝑙𝑢𝑒 > pairs, similar to Yera and Mart´ ınez [46]. Subsequently, text-based handcrafted templates could be defined and populated with these pairs [15] to generate effective and plausible explanations. The previous four stages have referred several open problems to be cover towards systematizing the use of SHAP-based explanations in recommender systems. Beyond this, it is also important to highlight the necessity of boosting the development and use of robust evaluation framework for measuring the real contribution of such developments, both from the offline analysis and the user studies perspective. Offline analysis: The effectiveness of post-hoc explanations have been usually characterized by metrics that measure the fidelity of the explanation mechanism, i.e. the percentage of recommended items for which the proposed approach was able to find some explanation [42, 46]. Text-based explanation approaches, like those discussed in the background section, have also incorporated feature-oriented metrics, such as feature coverage rate (i.e. the percentage of features referred to in the generated explanations), that could be also important to consider in the SHAP environment. In a different direction, experts-mediated offline evaluation, such as the one developed by Guo et al. [16], would also be an interesting direction to explore. In this approach, the authors ask domain experts to identify the most relevant features, which are then contrasted with the results of SHAP. User studies: The development of user studies will answer important questions towards the appropriateness of SHAP-associated explanations in individual and group recommendations. In such a context, it is necessary to empirically determine the effect of such explanations in well-established aims regarding explanatory interfaces, such as transparency, trust, effectiveness, efficiency, and satisfaction [36]. Furthermore, for the case of group recommendation, tailored criteria such as perceived consensus, perceived fairness and group satisfaction [5], need to be covered. Also, Rong et al. [33] have recently pointed out additional goals fairly explored in recommender systems explanations, such as understanding and usability, which should be specifically overseen in the case of SHAP-based explanations. Beyond these issues, it is important to take into account that while SHAP provides detailed interpretability, its use in large-scale recommendation systems leads to 20 Raciel Yera, Bapi Dutta, and Luis Mart´ ınez important computational costs. The complexity of SHAP-based methods increases with the number of features and model evaluations, making it challenging to deploy in real-time, high-dimensional contexts. These drawbacks particularly appear when dealing with sparse user-item interaction matrices, where the lack of data increases computational needs without guaranteeing trustable explanations. Moreover, the high dimensionality usually associated to recommender models also increments these challenges. Research efforts on the scalability of SHAP and its relevance in sparse contexts is then important to balance interpretability with efficiency in practical scenarios. 5 Case study This section discusses the initial steps that would be necessary to deploy the general framework discussed in the previous section, for providing SHAP-based explanations of the recommendation outputs. The selected scenario will be the group recommendation approach and dataset previously introduced by Perez-Almaguer et al. [28], for group restaurant recommendation in Havana City. Herein, the proposed group recommendation approach includes a hybridization between preference aggregation-based and recommendation-based approaches, as well as data imputation, feature weighting, and automatic selection of the aggregation methods [43]. Below, we subsequently discuss how the different steps of our framework introduced in the previous section, would be tailored to the current context. 1. Identify item features and associated output for individuals. The referred work characterizes restaurants according to five distinctive attributes based on TripAdvisor as information source, as described in Perez-Almaguer et al. [28]. These attributes and their respective possible values are: 1. Prize: cheap, average, expensive. 2. Food and service quality: low, medium, high. 3. Type of cuisine: Cuban, Spanish, Italian Basic, Italian Exclusive, Other. 4. Ranking: Top 20, Top 100, Other. 5. Popularity: High (200+ opinions), Medium (100+ opinions), Low (Other number of opinions). Herein, for generating top n recommendations the proposed content-based group recommendation approach [28] intermediately calculates an individual score for each group member, and subsequently calculates an aggregated score for the group. Both scores will be, respectively, regarded as the outputs for the calculation of SHAP-based feature contribution and ranking at the individual and the group level. 2. Feature contribution and ranking for each individual user For accomplishing this goal, we identify feature importance as it provides a highlevel summary of which variables the model relies on most to make predictions. In this case, we use CatBoost’s feature importance method with the prediction values SHAP for Individual and Group Recommendations 21 change type, which measures the average change in the model’s output when a feature’s value is varied, and for specific case of categorical features. This metric reflects the relative contribution of each feature across all samples in the dataset. While it does not show how a feature influences predictions, it is useful for discriminating the list of relevant features. It is also used for identifying redundant or irrelevant features, and serves as a valid starting point for further analysis. Figure 9 illustrates the feature importance for the individual recommendation score generated for each group member. Here it is clearly shown that Type of cuisine was the feature with the highest average impact in model output magnitude; having also Ranking and Food service quality a high importance. On the other side, Prize had an almost null impact. Fig. 9: Feature importance for the recommendation scores associated to the individual users in the groups As a second stage of this step, we will go through performing SHAP dependence plots across features. Particularly, it plots the SHAP value (i.e., the marginal contribution of a feature to the prediction) against the actual value of the feature, revealing whether a feature’s influence is positive or negative, and whether it is linear or non-linear. This visualization can also highlight interactions with other variables if a secondary color-coded feature is added. Unlike feature importance, which gives a global overview, SHAP dependence plots provide local interpretability by showing how specific input values translate into prediction changes across the dataset. Figure 10 illustrates the SHAP dependence plots the more relevant features for the recommendations (i.e. Type of cuisine vs. Ranking and Type of cuisine vs. Food service quality). Herein, it can be appreciated that Cuban cuisine out of top 100 restaurants, interestingly has the highest values and benefiting best the recommendation scores. Concerning Food service quality, it is surprisingly to notice that the best expected scores are associated to low quality services, even though for the three possible values of the attributes positive SHAP values are obtained. On the other hand, we can identified that for the remaining types of cuisine, negative SHAP 22 Raciel Yera, Bapi Dutta, and Luis Mart´ ınez (a) Type of cuisine vs. Ranking (b) Type of cuisine vs. Food service quality Fig. 10: SHAP dependence plots for the recommendation scores associated to the individual users in the groups values are obtained, indicating that such attribute values globally contribute to a lower score. 3. Feature ranking at the group level As previously discussed, one of the distinctive approach of the framework we are presenting across this contribution is the dual screening of explanations at the individual and group level. While in the previous step we explore feature ranking and SHAP dependence plots based on the individual recommendation scores, now we will be focused on performing the same graphics but regarding the group, aggregated scores. Fig. 11: Feature importance for the recommendation scores associated to the groups Figures 11 and 12 illustrate the feature importance and the SHAP dependence plots, associated to the group scores. Here we have obtained a feature importance graph that is different in relation to the individual case. For this group scenario, type of cuisine was associated to an even larger value, while food service quality SHAP for Individual and Group Recommendations 23 (a) Type of cuisine vs. Ranking (b) Type of cuisine vs. Food service quality Fig. 12: SHAP dependence plots for the recommendation scores associated to the groups was the second feature with the higher importance values. Furthermore, Ranking receives less importance for the group case, and finally in a similar way to individuals, Prize had the lower importance value. We think that this issue highlights the complexity of the group recommendation approach used [28], that models the group as more than the simpler sums of its individuals. In contrast, for the case of the SHAP dependence plots, we have found more notable differences in relation to the individual case, mainly regarding the magnitude of the features’ effect over the final score. Herein, the magnitude of the effect of the Cuban cuisine feature is lower in relation to the individual case. Also for Cuban cuisine, while for individuals the Other ranking had the higher impact, now for groups the impact is proportionally distributed across the different values of Ranking. Furthermore, in the case of Type of cuisine vs. Food service quality, while for individuals it can be appreciated similar contributions of the different food service qualities for each type of cuisine (i.e. points associated to each X-axis categories are close), in the case of group score the food service quality could lead to very different contributions, even associated to a same cuisine type (Figure 12). It is important to point out that the framework we are currently proposing, gives flexibility for performing more detailed analysis. For example, both in Steps 2 and 3 we have generated the graphics by taking into account the whole set of generated scores. Possibly, taken into account only the scores of some specific users or groups, would lead to even more interesting findings. 4. Generate the final explanations at individual and group level. Finally, the previously identified findings would lead to some explanations concerning the recommendation process. Some of the text-based explanations for adding trustworthiness to the recommendation process could be: •Type of cuisine and food service quality are the attributes that mainly moderate the recommendation process. 24 Raciel Yera, Bapi Dutta, and Luis Mart´ ınez •Cuban cuisine usually leads to high recommendation scores. •Some attribute values as Type of cuisine=Other, contributes particularly negative to the recommendation scores at the group level. •At the individual score case, the contribution of different types of cuisine is not sensitive to the ranking and quality values. Overall, the obtained results illustrate that our proposed two-stage framework is able to capture valuable knowledge both at the individual and group level, that can be used for building rich explanations. Our next future works will be focused on exploring the score of most specific users or groups, aiming on identifying more distinctive explanations of the obtained results. 6 Conclusions This contribution starts with the underexplored potential of SHAP values in explaining the outputs of both individual and group recommender systems, highlighting the limitations of current analyses in this area. In this regard, the analysis of background and related works on post-hoc explanations for recommender systems demonstrated the presence of several established research directions, including the generation of post-hoc explanations using data mining techniques such as association rule mining or decision tree induction [46], and the use of generative text-based explanations [2]. However, despite the use of SHAP by some researchers for this purpose, there is an important gap in the development of a comprehensive framework and methodological guidelines for conducting systematic research with this objective in mind. A PRISMA-based systematic review has been performed for supporting this statement. In the current work, we have proposed a general framework for explaining individual and group recommendations under the umbrella of the SHAP framework. It is composed of four main stages: 1) Identify the item features and their associated output, 2) Calculate the feature contribution and ranking for each individual user, 3) Generate the feature ranking at the group level, and 4) Generate the final explanations at individual and group level. Furthermore, special attention has been given to the necessity of establishing appropriate evaluation frameworks for both offline analysis and user studies. A case study has been developed, being focused on the use of the presented framework in a specific group recommendation scenario, showing how it would be able to add transparency to the individual and group recommendation scenario. Next future works will be focused on: 1) exploring the use of the presented procedure over specific users and groups instead of the whole available data, 2) evaluating the effect of the generated explanations across user studies, and 3) the benchmarking with other explainability techniques. SHAP for Individual and Group Recommendations 25 We hope that the current contribution will help to consolidate the research efforts to be made in the near future regarding the use of SHAP to support explanations for individual and group recommendations. Acknowledgements Raciel Yera is supported by the European Union’s Horizon Europe research and innovation program under the Marie Sklodowska-Curie Grant Agreement No.101106164. Bapi Dutta acknowledges the support of the Spanish Ministry of Science, Innovation and Universities, and the Spanish State Research Agency through the Ram´ on y Cajal Research grant (RYC2023045020-I), Spain. References 1. Adomavicius, G., Tuzhilin, A.: Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions. IEEE transactions on knowledge and data engineering 17(6), 734–749 (2005) 2. Ariza-Casabona, A., Boratto, L., Salam´ o, M.: A comparative analysis of text-based explainable recommender systems. In: Proceedings of the 18th ACM Conference on Recommender Systems, pp. 105–115 (2024) 3. Ariza-Casabona, A., Salam´ o, M., Boratto, L., Fenu, G.: Towards self-explaining sequenceaware recommendation. In: Proceedings of the 17th ACM Conference on Recommender Systems, pp. 904–911 (2023) 4. Ataei, P., Litchfield, A.: The state of big data reference architectures: a systematic literature review. IEEE Access (2022) 5. Barile, F., Draws, T., Inel, O., Rieger, A., Najafian, S., Ebrahimi Fard, A., Hada, R., Tintarev, N.: Evaluating explainable social choice-based aggregation strategies for group recommendation. User Modeling and User-Adapted Interaction 34(1), 1–58 (2024) 6. Chanson, A., Labroche, N., Verdeaux, W.: Towards local post-hoc recommender systems explanations. In: Proceedings of the 23rd International Workshop on Design, Optimization, Languages and Analytical Processing of Big Data (DOLAP) (2021) 7. Coba, L., Confalonieri, R., Zanker, M.: Recoxplainer: a library for development and offline evaluation of explainable recommender systems. IEEE Computational Intelligence Magazine 17(1), 46–58 (2022) 8. Contreras, D., Salam´ o, M., Boratto, L.: Integrating collaboration and leadership in conversational group recommender systems. ACM Transactions on Information Systems (TOIS) 39(4), 1–32 (2021) 9. Critical Appraisal Skills Programme (CASP): Casp qualitative studies checklist. https: //casp-uk.net/casp-tools-checklists/ (2018). Accessed: 2025-04-28 10. Dara, S., Chowdary, C.R., Kumar, C.: A survey on group recommender systems. Journal of Intelligent Information Systems 54(2), 271–295 (2020) 11. De Pessemier, T., Dooms, S., Martens, L.: Comparison of group recommendation algorithms. Multimedia tools and applications 72, 2497–2541 (2014) 12. D´ ıaz, S.B., Coussement, K., De Caigny, A.: From collaborative filtering to deep learning: Advancing recommender systems with longitudinal data in the financial services industry. European Journal of Operational Research (2025) 13. Ekstrand, M.D., Riedl, J.T., Konstan, J.A., et al.: Collaborative filtering recommender systems. Foundations and Trends®in Human–Computer Interaction 4(2), 81–173 (2011) 14. Felfernig, A., Boratto, L., Stettinger, M., Tkali, M.: Group Recommender Systems: An Introduction. Springer Publishing Company, Incorporated (2018) 15. Gatt, A., Krahmer, E.: Survey of the state of the art in natural language generation: Core tasks, applications and evaluation. Journal of Artificial Intelligence Research 61, 65–170 (2018)