This paper investigates the trade-offs between effectiveness, fairness, and privacy in job recommender systems that suggest job seekers for specific job postings. We assume that the system cannot be trusted with sensitive attributes of job seekers (e.g., nationality, gender), yet it must still ensure fair representation of all sensitive groups in the recommendation list. This creates a fundamental tension between fairness and privacy. To address this, we propose a novel framework that enhances fairness even when sensitive attributes are perturbed using a local differential privacy mechanism. Extensive experiments on real-world job platform data demonstrate that our framework can maintain fairness, even under strong privacy constraints.

We present a novel comparative framework for sentiment analysis in product reviews that combines structured semantic similarity methods with Large Language Models (LLMs). We introduce two new algorithms, iSense and xSense, which leverage lexical knowledge graphs and an ensemble-based strategy to detect product safety issues through semantic proximity. Unlike purely neural models, our approach enables transparent and interpretable sentiment classification aligned with regulatory vocabularies. We evaluate the methods on both synthetic and real-world datasets, including over 1.8 million Amazon reviews in the ”Toys and Games” category. Experimental results demonstrate that xSense not only matches but in some cases surpasses the performance of state-of-the-art LLMs such as BERT, achieving an F1 score of 0.97. These findings establish structured semantic reasoning as a competitive alternative for domain-specific sentiment analysis, with implications for consumer safety analytics and hybrid AI systems.

In the intensive care units of the hospitals, the critically ill patients follow complex treatments consisting of drug combinations to avoid mortality and have fast recovery. However, the drug combinations of a patient's treatment may cause unwanted side effects. In this paper, we apply Principal Component Analysis (PCA) over patients' treatment medical data, so that we can identify similar clinical cases to the target patient and the effective and safe drug combinations that these patients received. Moreover, we employ from internet drug databases side information regarding the clinical trials of new drugs and their unwanted side effects they may have with other drugs. Our goal is the reduction of the unwanted side effects of the recommended drug combinations by replacing some drugs with their safer substitutes. Our experimental results have shown the effectiveness and safety of our PCA method in terms of drug recommendations compared with classic algorithms such as SVD, NMF and user-KNN.

This tutorial is part of chapters six and eight from my recent book "Deep Recommender Systems with Python", which can be found in https://www.panagiotissymeonidis.com/enbook/index.html. In particular, this tutorial is a comprehensive exploration of graph-based methods to enhance recommender systems. The presentation is structured around key topics, starting with the basics of graphs and recommender systems, and progresses to advanced techniques involving Graph Neural Networks (GNNs) and their applications. It offers a rich blend of theory and practice. It is important for the Recommender Systems community because it analyzes graph-based algorithms to address the similarity search and recommendation problem. Both problems affect our everyday experience, while searching for knowledge on a topic with challenging issues such as data scalability, noise, and sparsity. We can deal with all these challenges by applying advanced graph-based methods. In this tutorial, we provide a detailed step-by-step analysis with an integrated toy example in three jupyter notebook sessions, which helps the audience clearly understand the main differences of basic graph-based methods.

Nowadays the recommender systems represent an essential component in a huge variety of applications, and their continuous and rapid spread has allowed them to reach relevant fields like that of the healthcare. In such a delicate environment, it is extremely important to take into account all the variables that could occur, as the main objective is to support the doctors in prescribing accurate and safe treatments to patients affected by complex diseases. To address the aforementioned challenge, we decided to extend the Matrix Co-Factorization (MCF) method so that to allow it to include additional auxiliary data. A Post Hoc Re-Ranking technique has also been implemented to penalise the drugs that occur frequently in the adversarial drug-drug interactions knowdledge graph, therefore, to reduce the number of drugs in the final recommendations that could be dangerous for the patients' health. Different experiments have been used to assess the performances of our extended MCF. The results accomplished by our proposed method have also been compared against state-of-the-art models (such as C2PF, and PCRL) with a real-life dataset (MIMIC III). Our experiments depict that our method outperforms sufficiently the other methods in terms of accuracy and safety.

Polypharmacy is the prescription of many drugs for a critically-ill patient, which may have a higher risk of adverse drug reactions among them. Such patients usually follow a treatment that consists of multiple different drugs, which increases the risk of unwanted side-effects and can be occasionally harmful. In this paper, we address the safe drug prescription issues, by integrating the patient's Electronic Health Records (EHRs) with an adversarial Drug-Drug Interaction (DDI) knowledge graph. We study the task of predicting drugs to patients by making use of the real life data set MIMIC III. Our goal is to predict the next drug combination for a patient's therapy and at the same time minimize the drug unwanted side effects. We feed the input data to 6 different algorithms that we later compare in terms of effectiveness and safety in recommending drug combinations. Furthermore, we use the Post Hoc Re-rank technique together with Singular Value Decomposition algorithm to incorporate information in our prediction model from unwanted drug-drug interactions knowledge graph. Our experiments have shown that with a slight loss in the efficacy of the recommendation algorithm, we are able to reduce the toxicity score of the suggested drug combinations.

Medicine recommendation is denoted as the task of predicting drug combinations for patients' therapies with complex diseases (i.e., cancer, diabetes, etc.). These patients often follow a treatment that consists of multiple drugs simultaneously, focusing at different human targets such as genes, proteins, etc. Previous research has already integrated the patients' Electronic Health Records (EHRs) with an adversarial Drug-Drug Interaction (DDI) knowledge graph to predict the next drug combination for a patient's therapy and minimize the drug side effects. However, they miss to consider additional valuable information that comes from synergistic Drug-Drug interaction knowledge graphs. In this paper, we integrate an EHR graph, which incorporates the patient, the disease, the therapy, and the drug information, with a Synergistic and/or an Adversarial DDI knowledge graph to recommend both accurate and safe medication. By identifying those drugs which can act synergistically and/or adversely, we are able to improve either the efficacy of the patient's therapy or minimize the toxicity and drug side effects. We have run experiments with two real-life medical data sets. Our results show that we can assist doctors to prescribe effective and safe medication for the patients' treatment.

Drug recommendation is denoted as the task of predicting drug combinations for patients' therapies with complex diseases (i.e., thrombosis, diabetes, etc.). These patients usually suffer from polypharmacy, and consequently various drug drug interactions. In this paper, we integrate the patients' Electronic Health Records (EHRs) with an adversarial Drug-Drug Interaction (DDI) knowledge graph to predict the next drug combination for a patient's therapy and minimize the drug side effects. In particular, we integrate an EHR graph, which incorporates the patient, the disease, the therapy, and the drug information, with an Adversarial DDI knowledge graph to recommend both accurate and safe medication. We also predict mortality and the time to death of critically-ill patients, to identify clinically meaningful predictors (e.g., harmful drug combinations). By identifying those drugs which can act adversarially, we are able to improve either the efficacy of the patient's therapy or minimize the toxicity and drug side effects. We have run experiments with a real-life medical data set. Our results show that we can assist doctors to prescribe effective and safe medication for the patients' treatment.

Abstract There exist many research works that strive to answer the question “what news article is a user going to click next given his profile”. These works take into account the time dimension to reveal users’ preferences over time. However, few works exploit adequately the information that is hidden inside user sessions. User sessions include a list of user interactions with items within a short period of time such as 30 min, and can reveal her very last intentions. In this paper, we combine intra- with inter-session item transition probabilities to reveal the short- and long-term intentions of individuals. Thus, we are able to better capture the similarities among items that are co-selected inside a user session but also within any two consecutive sessions. We have evaluated experimentally our method and compare it against state-of-the-art algorithms on three real-life datasets. We demonstrate the superiority of our method over its competitors.

The global response to the pandemic introduced by COVID-19 is unprecedented. Scientists develop methods, which analyze data to identify an effective treatment that uncovers possible responses to the SARS-COV-2 virus. However, our global response should be based on knowledge exchange and collaboration among countries. In this paper, we present a recommender system for treatment recommendations, which exploits similar patterns among patients of different clinical studies, and recommends them health interventions (such as to provide oxygen therapy) and drugs (e.g., Remdesivir) based on their symptoms' or diseases' similarity with patients of other similar clinical studies. Our approach can also provide explanations along with recommended treatments to assist doctors in understanding the reasons behind a suggested drug or health intervention. We also perform experiments to identify the effectiveness of our system in terms of recommendation accuracy. Our results demonstrate that our system is able to minimize the false positive and false negative prediction rates. Finally, we provide web links to download both (i) our program's setup file and (ii) our Neo4j database file.

Patients with complex diseases (i.e., cancer, diabetes, etc.) often follow a therapeutic that consists of multiple drugs, focusing at different human targets such as genes, proteins, etc. There is already related work in medical research for drug-target prediction and drug re-purposing. In this paper, we try to provide both accurate and explainable drug recommendations. In particular, we develop models to help doctors screen candidate drugs and their possible substitutes more comprehensively, by providing also robust explanations. To do this, we build a heterogeneous information network to capture the latent associations between patients, their therapeutics, the drugs used, and diseases nodes by using a meta path-based similarity measure. Based on previous similar patients' historical drug treatments, we can provide personalized drug recommendations along with explanations to support critical medical decisions. Demo code for our hybrid meta-based explanations can be found here. We have performed experiments on three real life data sets, which show that we can increase drastically the explainability of our drug recommendations by using more historical data, whereas the recommendations' accuracy still remains at a high level.

The task “recommend a video to watch next?” has been in the focus of recommender systems’ research for a long time. However, adequately exploiting the clues hidden in the sequences of actions of user sessions in order to reveal users’ short-term intentions moved only recently into the focus of research. Based on a real-world application scenario, in this paper, we propose a Markov Chain-based transition probability matrix to efficiently reveal the short-term preferences of individuals. We experimentally evaluated our proposed method by comparing it against state-of-the-art algorithms in an offline as well as a live evaluation setting. In both cases our method not only demonstrated its superiority over its competitors, but exposed a clearly stronger engagement of users on the platform. In the online setting, our method improved the click-through rate by up to 93.61%. This paper therefore contributes real-world evidence for improving the recommendation effectiveness, by considering sequence-awareness, since capturing the short-term preferences of users is crucial in the light of items with a short life span such as tv programs (news, tv shows, etc.).

Rating-based summary statistics are ubiquitous in e-commerce, and often are crucial components in personalized recommendation mechanisms. Especially visual rating summarizations have been identified as important means to explain, why an item is presented or proposed to an user. Largely left unexplored, however, is the issue to what extent the descriptives of these rating summary statistics influence decision making of the online consumer. Therefore, we conducted a series of two conjoint experiments to explore how different summarizations of rating distributions (i.e., in the form of number of ratings, mean, variance, skewness, bimodality, or origin of the ratings) impact users' decision making. In a first study with over 200 participants, we identified that users are primarily guided by the mean and the number of ratings, and - to lesser degree - by the variance and origin of a rating. When probing the maximizing behavioral tendencies of our participants, other sensitivities regarding the summary of rating distributions became apparent. We thus instrumented a follow-up eye-tracking study to explore in more detail, how the choices of participants vary in terms of their decision making strategies. This second round with over 40 additional participants supported our hypothesis that users, who usually experience higher decision difficulty, follow compensatory decision strategies, and focus more on the decisions they make. We conclude by outlining how the results of these studies can guide algorithm development, and counterbalance presumable biases in implicit user feedback.

Conventional approaches to matrix factorisation (MF) typically rely on a centralised collection of user data for building a MF model. This approach introduces an increased risk when it comes to user privacy. In this short paper we propose an alternative, user-centric, privacy enhanced, decentralised approach to MF. Our method pushes the computation of the recommendation model to the user's device, and eliminates the need to exchange sensitive personal information; instead only the loss gradients of local (device-based) MF models need to be shared. Moreover, users can select the amount and type of information to be shared, for enhanced privacy. We demonstrate the effectiveness of this approach by considering different levels of user privacy in comparison with state-of-the-art alternatives.

Collaborative filtering systems heavily depend on user feedback expressed in product ratings to select and rank items to recommend. These summary statistics of rating values carry two important descriptors about the assessed items, namely the total number of ratings and the mean rating value. In this study we explore how these two signals influence the decisions of online users based on choice-based conjoint experiments. Results show that users are more inclined to follow the mean indicator as opposed to the total number of ratings. Empirical results can serve as an input to developing algorithms that foster items with a, consequently, higher probability of choice based on their rating summarizations or their it explainability due to these ratings when ranking recommendations.

rrecsys is a novel library in R for developing and assessing recommendation algorithms. In this demo, we extend rrecsys with functions for visual analytics of recommendation performance, that is one of the strong capabilities of the R environment. In particular, we show how the library can be used to depict dataset characteristics, train and test recommendation algorithms and to visually assess, for instance, their capability to exploit long-tail items for making correct predictions.

CheckInShop is an app that employs sensors to capture the user preferences in physical stores and provide either micro-location marketing or product recommendations. By utilizing iBeacon technology and with the exploitation of a mobile app, we keep track of customer's preferences in physical stores. Then, based on the time that a product is viewed by a customer and his micro-location inside the store, we send to him either product offers or recommendations of similar products to the ones he is looking at. These recommendations are accurate because they are provided at the right time and in the right place. A video that demonstrates our system can be found in the following link: https://www.youtube.com/watch?v=Z99IMCHowAA

This turorial offers a rich blend of theory and practice regarding dimensionality reduction methods, to address the information overload problem in recommender systems. This problem affects our everyday experience while searching for knowledge on a topic. Naive Collaborative Filtering cannot deal with challenging issues such as scalability, noise, and sparsity. We can deal with all the aforementioned challenges by applying matrix and tensor decomposition methods. These methods have been proven to be the most accurate (i.e., Netflix prize) and efficient for handling big data. For each method (SVD, SVD++, timeSVD++, HOSVD, CUR, etc.) we will provide a detailed theoretical mathematical background and a step-by-step analysis, by using an integrated toy example, which runs throughout all parts of the tutorial, helping the audience to understand clearly the differences among factorisation methods.

Recently, location-based social networks (LBSNs) gave the opportunity to users to share geo-tagged information along with photos, videos, and SMSs. Recommender systems can exploit this geographic information to provide much more accurate and reliable recommendations to users. In this paper, we present and compare 16 real life LBSNs, bringing into surface their advantages/ disadvantages, their special functionalities, and their impact in the mobile social Web. Moreover, we describe and compare extensively 43 state-of-the-art recommendation algorithms for LBSNs. We categorize these algorithms according to: personalization type, recommendation type, data factors/features, problem modeling methodology, and data representation. In addition to the above categorizations which cannot cover all algorithms in an integrated way, we also propose a hybrid k-partite graph taxonomy to categorize them based on the number of the involved k-partite graphs. Finally, we compare the recommendation algorithms with respect to their evaluation methodology (i.e., datasets and metrics) and we highlight new perspectives for future work in LBSNs.

Social tagging systems (STSs) allow users to annotate information items (songs, pictures, etc.) to provide them item/tag or even user recommendations. STSs consist of three main types of entities: 1) users; 2) items; and 3) tags. These data usually are represented by a three-order tensor, on which Tucker decomposition (TD) models are performed, such as higher order singular value decomposition. However, TD models require cubic computations for the tensor decomposition. Furthermore, TD models suffer from sparsity that incurs in social tagging data. Thus, TD models have limited applicability to large-scale datasets, due to their computational complexity and data sparsity. In this paper, we use two different ways to compute similarity/distance between tags (i.e., the term frequency - inverse document frequency vector space model and the semantic similarity of tags using the ontology of WordNet). Moreover, to reduce the size of the tensor's dimensions and its data sparsity, we use clustering methods (i.e., ${k}$ -means, spectral clustering, etc.) for discovering tag clusters, which are the intermediaries between a user's profile and items. Thus, instead of inserting the tag dimension in the tensor, we insert the tag cluster dimension, which is smaller and has less noise, resulting to better item recommendation accuracy. We perform experimental comparison of the proposed method against a state-of-the-art item recommendation algorithm with two real datasets (Last.fm and BibSonomy). Our results show significant improvements in terms of effectiveness and efficiency.

Online social networks have attracted users' attention in the last decade. Recommendation services constitute a critical functionality of such social platforms: users receive recommendations about resources (documents, pieces of music) and potential friends (people with the same interests). Recently, technological progressions in smart phones enabled the exploitation of geographical data information in social networks. Users can now receive recommendations about new Points of Interest (POIs), and new activities in POIs. Eventually, Location-based Social Networks (LBSNs) may become the 'Next Big Thing' of the Internet industry. This paper surveys the related work and current state-of-the-art algorithms in LBSNs. We also provide three new perspectives that concern recommendations in LBSNs: time-awareness, user's privacy issues, and explainability of recommendations. We present the latest work in LBSNs by comparing real systems and by categorizing them in multiple ways (platforms, personalization, etc.).

In this paper, we propose a new model to integrate additional data, which is obtained from geospatial resources other than original data set in order to improve Location/Activity recommendations. The data set that is used in this work is a GPS trajectory of some users, which is gathered over 2 years. In order to have more accurate predictions and recommendations, we present a model that injects additional information to the main data set and we aim to apply a mathematical method on the merged data. On the merged data set, singular value decomposition technique is applied to extract latent relations. Several tests have been conducted, and the results of our proposed method are compared with a similar work for the same data set.

Online Social Rating Networks (SRNs) such as Epinions and Flixter, allow users to form several implicit social networks, through their daily interactions like co-commenting on the same products, or similarly co-rating products. The majority of earlier work in Rating Prediction and Recommendation of products (e.g. Collaborative Filtering) mainly takes into account ratings of users on products. However, in SRNs users can also built their explicit social network by adding each other as friends. In this paper, we propose Social-Union, a method which combines similarity matrices derived from heterogeneous (unipartite and bipartite) explicit or implicit SRNs. Moreover, we propose an effective weighting strategy of SRNs influence based on their structured density. We also generalize our model for combining multiple social networks. We perform an extensive experimental comparison of the proposed method against existing rating prediction and product recommendation algorithms, using synthetic and two real data sets (Epinions and Flixter). Our experimental results show that our Social-Union algorithm is more effective in predicting rating and recommending products in SRNs.

Online social networks (OSNs) like Facebook, and Myspace recommend new friends to registered users based on local features of the graph (i.e. based on the number of common friends that two users share). However, OSNs do not exploit the whole structure of the network. Instead, they consider only pathways of maximum length 2 between a user and his candidate friends. On the other hand, there are global approaches, which detect the overall path structure in a network, being computationally prohibitive for huge-size social networks. In this paper, we define a basic node similarity measure that captures effectively local graph features. We also exploit global graph features introducing transitive node similarity. Moreover, we derive variants of our method that apply in signed networks. We perform extensive experimental comparison of the proposed method against existing recommendation algorithms using synthetic and real data sets (Facebook, Hi5 and Epinions). Our experimental results show that our FriendTNS algorithm outperforms other approaches in terms of accuracy and it is also time efficient. We show that a significant accuracy improvement can be gained by using information about both positive and negative edges.

Social tagging is becoming increasingly popular in music information retrieval (MIR). It allows users to tag music items like songs, albums, or artists. Social tags are valuable to MIR, because they comprise a multifaced source of information about genre, style, mood, users' opinion, or instrumentation. In this paper, we examine the problem of personalized music recommendation based on social tags. We propose the modeling of social tagging data with three-order tensors, which capture cubic (three-way) correlations between users-tags-music items. The discovery of latent structure in this model is performed with the Higher Order Singular Value Decomposition (HOSVD), which helps to provide accurate and personalized recommendations, i.e., adapted to the particular users' preferences. To address the sparsity that incurs in social tagging data and further improve the quality of recommendation, we propose to enhance the model with a tag-propagation scheme that uses similarity values computed between the music items based on audio features. As a result, the proposed model effectively combines both information about social tags and audio features. The performance of the proposed method is examined experimentally with real data from Last.fm. Our results indicate the superiority of the proposed approach compared to existing methods that suppress the cubic relationships that are inherent in social tagging data. Additionally, our results suggest that the combination of social tagging data with audio features is preferable than the sole use of the former.

Social tagging is the process by which many users add metadata in the form of keywords, to annotate and categorize items (songs, pictures, Web links, products, etc.). Social tagging systems (STSs) can provide three different types of recommendations: They can recommend 1) tags to users, based on what tags other users have used for the same items, 2) items to users, based on tags they have in common with other similar users, and 3) users with common social interest, based on common tags on similar items. However, users may have different interests for an item, and items may have multiple facets. In contrast to the current recommendation algorithms, our approach develops a unified framework to model the three types of entities that exist in a social tagging system: users, items, and tags. These data are modeled by a 3-order tensor, on which multiway latent semantic analysis and dimensionality reduction is performed using both the higher order singular value decomposition (HOSVD) method and the kernel-SVD smoothing technique. We perform experimental comparison of the proposed method against state-of-the-art recommendation algorithms with two real data sets (Last.fm and BibSonomy). Our results show significant improvements in terms of effectiveness measured through recall/precision.

Providing justification to a recommendation gives credibility to a recommender system. Some recommender systems (Amazon.com etc.) try to explain their recommendations, in an effort to regain customer acceptance and trust. But their explanations are poor, because they are based solely on rating data, ignoring the content data. Our prototype system MoviExplain is a movie recommender system that provides both accurate and justifiable recommendations.

Social tagging is the process by which many users add metadata in the form of keywords, to annotate and categorize information items (songs, pictures, web links, products etc.). Collaborative tagging systems recommend tags to users based on what tags other users have used for the same items, aiming to develop a common consensus about which tags best describe an item. However, they fail to provide appropriate tag recommendations, because: (i) users may have different interests for an information item and (ii) information items may have multiple facets. In contrast to the current tag recommendation algorithms, our approach develops a unified framework to model the three types of entities that exist in a social tagging system: users, items and tags. These data is represented by a 3-order tensor, on which latent semantic analysis and dimensionality reduction is performed using the Higher Order Singular Value Decomposition (HOSVD) technique. We perform experimental comparison of the proposed method against two state-of-the-art tag recommendations algorithms with two real data sets (Last.fm and BibSonomy). Our results show significant improvements in terms of effectiveness measured through recall/precision.

Recommender systems are gaining widespread acceptance in e-commerce applications to confront the ldquoinformation overloadrdquo problem. Providing justification to a recommendation gives credibility to a recommender system. Some recommender systems (Amazon.com, etc.) try to explain their recommendations, in an effort to regain customer acceptance and trust. However, their explanations are not sufficient, because they are based solely on rating or navigational data, ignoring the content data. Several systems have proposed the combination of content data with rating data to provide more accurate recommendations, but they cannot provide qualitative justifications. In this paper, we propose a novel approach that attains both accurate and justifiable recommendations. We construct a feature profile for the users to reveal their favorite features. Moreover, we group users into biclusters (i.e., groups of users which exhibit highly correlated ratings on groups of items) to exploit partial matching between the preferences of the target user and each group of users. We have evaluated the quality of our justifications with an objective metric in two real data sets (Reuters and MovieLens), showing the superiority of the proposed method over existing approaches.

Collaborative filtering (CF) systems are gaining widespread acceptance in recommender systems and e-commerce applications. These systems combine information retrieval and data mining techniques to provide recommendations for products, based on suggestions of users with similar preferences. Nearest-neighbor CF process is influenced by several factors, which were not examined carefully in past work. In this paper, we bring to surface these factors in order to identify existing false beliefs. Moreover, by being able to view the "big picture" from the CF process, we propose new approaches that substantially improve the performance of CF algorithms. For instance, we obtain more than 40% percent increase in precision in comparison to widely-used CF algorithms. We perform an extensive experimental evaluation, with several real data sets, and produce results that invalidate some existing beliefs and illustrate the superiority of the proposed extensions