Recommender systems are an integral and successful application of machine learning in e-commerce industry and in everyday lives of online users. Recommendation algorithms are used extensively for news, musics, books, point of interests, or travel recommendation as well as in many other domains. Although much focus has been paid on improving recommendation quality, however, some real-world aspects are not considered: How to ensure that top-n recommendations are fair and not biased due to any popularity boosting events, such as awards for movies or songs? How to recommend items to entities by explicitly considering information from heterogeneous sources. What is the best way to model sequential recommendation systems as heterogeneous context-aware design, and learning on-the-fly from spatial, temporal and social contexts. Can we model attributes and heterogeneous relations in a heterogeneous information network? The goal of this thesis is to pave the way towards the next generation of realworld recommendation systems tackling fairness and information heterogeneity challenges to improve the user experience, while giving good recommendations. This thesis bridges techniques from recommendation and deep-learning techniques for representation learning by proposing novel techniques to address the above real-world problems. We focus on four directions: (1) model the effect of popularity bias over time on the consumption of items, (2) model the heterogeneous information associated with sequential history of users and social links for sequential recommendation, (3) model the heterogeneous links and rich content of nodes in an academic heterogeneous information network, and (4) learn semantics using topic modeling for nodes based on their content and heterogeneous links in a heterogeneous information network.

Recommendation systems typically use historical interactions between users and items topredict what other items can be of interest to a user. The recommendations are based onpatterns in how users interact similarly with items. This thesis investigates if it is possible toimprove the quality of the recommendations by including more information about the items inthe model that predicts the recommendations. More specifically, the use of deep learning toextract information from item images is investigated. To do this, two types of collaborativefiltering models, based on historic interactions, are implemented. These models are thencompared to different collaborative filtering models that either make use of user and itemattributes, or images of the items. Three pre-trained image classification models are used toextract useful item features from the item images. The models are trained and evaluated using adataset of historic transactions and item images from the online sports shop Stadium, given bythe thesis supervisor. The results show no noticeable improvement in performance for themodels using the images compared to the models without images. The model using the userand item attributes performs the best, indicating that the collaborative filtering models can beimproved by giving it more information than just the historic interactions. Possible ways tofurther investigate using the image feature vectors in collaborative filtering models, as well asusing them to create better item attributes, are discussed and suggested for future work.

Recommendation systems become more and more necessary due to the growing supply chain. Therefore, scientists are developing models that can serve different recommendation needs faster than before, and it is getting more complicated to choose the model for a specific case. In this thesis, there are three neural collaborative filtering methods compared regarding dataset fit. This research shows that there is no one-fits-all method. There is much space for improvement in all the areas: dataset selection and aggregation, method development and operation, and selective approaches for the analysis of the results. In the thesis, three contrasting datasets are chosen (Chess, Library, and LastFM), and three novel approaches are tested: recently released Dynamic Graph Collaborative Filtering (DGCF) and Dynamic Embeddings for Interaction Prediction (DeePRed) are compared to the Joint Dynamic User- Item Embeddings (JODIE) as the baseline. Results show DeePRed being a state-of-the-art model that outperforms other methods. It runs an epoch for a small dataset in less than a minute, shows great prediction accuracy in an average of 98% for small datasets. However, DGCF does not show accuracy improvement over JODIE but is significantly faster for an extensive dataset. / Rekommendationssystem blir mer och mer nödvändiga på grund av den växande försörjningskedjan. Därför utvecklar forskare modeller som kan tjäna olika rekommendationsbehov snabbare än tidigare och det blir mer och mer komplicerat att välja modell för ett specifikt fall. I denna avhandling finns det tre neurologiska samarbetsfiltreringsmetoder som jämförs avseende deras gran för olika datamängder. Denna forskning visar att det inte finns någon metod som passar alla och det finns mycket utrymme för förbättring inom alla områden: datasatsval och aggregering, metodutveckling och drift och selektiva metoder för analys av resultaten. I avhandlingen väljs tre kontrasterande datamängder (Chess, Library och LastFM) och tre nya metoder testas: nyligen släppt Dynamic Graph Collaborativefiltering (DGCF) och Dynamic Embedding for Interaction Prediction (DeePRed) jämförs med Joint Dynamic User-Item. Inbäddning (JODIE) som baslinje. Resultaten visar att (DeePRed) är en avancerad modell som överträffar andra metoder som snabba genom att köra en epok för liten dataset på mindre än en minut, vilket visar stor förutsägelsesnoggrannhet i genomsnitt 98% för små datamängder. Men (DGCF) visar inte förbättring av noggrannhet jämfört med (JODIE), men är betydligt snabbare för en stor dataset.

This study analyzes the recommendation systems from a user’s perspective and identifies five areas of concern in developing and using a recommendation system. The study’s methods are focus group discussions with Data scientists and Full-stack developers working in the industry. An online survey was distributed to several Facebook groups of various universities. The study results indicate that users have a strong desire to have their moral sensitivities under their control. The study also enables the system developers to understand the recommendations of the system affect the conflicting interests of various entities. / Den här studien analyserar rekommendationssystemen ur ett användarperspektiv, och identifierar fem relevanta områden att ha i åtanke i utvecklingen och användandet av ett rekommendationssystem. Studiens metoder består av fokusgruppsdiskussioner med datavetare och s.k. “full-stack-utvecklare” som arbetar inom IT-branschen. En online-enkät delades ut till flera Facebook-grupper tillhörande olika universitet. Studiens resultat indikerar att användare har en tydlig preferens att ha kontroll över sina moraliska perspektiv. Vidare tillåter även studien systemutvecklare att förstå att systemets rekommendationer påverkar intressekonflikter mellan olika enheter och intressenter.

Recommendation systems have been widely utilized in e-commerce settings to aid users through their shopping experiences. The principal advantage of these systems is their ability to narrow down the purchase options in addition to marketing items to customers. However, a number of challenges remain, notably those related to obtaining a clearer understanding of users, their profiles, and their preferences in terms of purchased items. Specifically, recommender systems based on collaborative filtering recommend items that have been rated by other users with preferences similar to those of the targeted users. Intuitively, the more information and ratings collected about the user, the more accurate are the recommendations such systems suggest. In a typical recommender systems database, the data are sparse. Sparsity occurs when the number of ratings obtained by the users is much lower than the number required to build a prediction model. This usually occurs because of the users’ reluctance to share their reviews, either due to privacy issues or an unwillingness to make the extra effort. Grey-sheep users pose another challenge. These are users who shared their reviews and ratings yet disagree with the majority in the systems. The current state-of-the-art typically treats these users as outliers and removes them from the system. Our goal is to determine whether keeping these users in the system may benefit learning. Thirdly, cold-start problems refer to the scenario whereby a new item or user enters the system and is another area of active research. In this case, the system will have no information about the new user or item, making it problematic to find a correlation with others in the system. This thesis addresses the three above-mentioned research challenges through the development of machine learning methods for use within the recommendation system setting. First, we focus on the label and data sparsity though the development of the Hybrid Cluster analysis and Classification learning (HCC-Learn) framework, combining supervised and unsupervised learning methods. We show that combining classification algorithms such as k-nearest neighbors and ensembles based on feature subspaces with cluster analysis algorithms such as expectation maximization, hierarchical clustering, canopy, k-means, and cascade k-means methods, generally produces high-quality results when applied to benchmark datasets. That is, cluster analysis clearly benefits the learning process, leading to high predictive accuracies for existing users. Second, to address the cold-start problem, we present the Popular Users Personalized Predictions (PUPP-DA) framework. This framework combines cluster analysis and active learning, or so-called user-in-the-loop, to assign new customers to the most appropriate groups in our framework. Based on our findings from the HCC-Learn framework, we employ the expectation maximization soft clustering technique to create our user segmentations in the PUPP-DA framework, and we further incorporate Convolutional Neural Networks into our design. Our results show the benefits of user segmentation based on soft clustering and the use of active learning to improve predictions for new users. Furthermore, our findings show that focusing on frequent or popular users clearly improves classification accuracy. In addition, we demonstrate that deep learning outperforms machine learning techniques, notably resulting in more accurate predictions for individual users. Thirdly, we address the grey-sheep problem in our Grey-sheep One-class Recommendations (GSOR) framework. The existence of grey-sheep users in the system results in a class imbalance whereby the majority of users will belong to one class and a small portion (grey-sheep users) will fall into the minority class. In this framework, we use one-class classification to provide a class structure for the training examples. As a pre-assessment stage, we assess the characteristics of grey-sheep users and study their impact on model accuracy. Next, as mentioned above, we utilize one-class learning, whereby we focus on the majority class to first learn the decision boundary in order to generate prediction lists for the grey-sheep (minority class). Our results indicate that including grey-sheep users in the training step, as opposed to treating them as outliers and removing them prior to learning, has a positive impact on the general predictive accuracy.

There exist many real-world applications such as recommendation systems, document retrieval, and computational biology where the correct ordering of instances is of equal or greater importance than predicting the exact value of some discrete or continuous outcome. Learning-to-Rank (LTR) refers to a group of algorithms that apply machine learning techniques to tackle these ranking problems. Despite their empirical success, most existing LTR models are not built to be robust to errors in labeling or annotation, distributional data shift, or adversarial data perturbations. To fill this gap, we develop four LTR frameworks that are robust to various types of perturbations. First, Pairwise Elastic Net Regression Ranking (PENRR) is an elastic-net-based regression method for drug sensitivity prediction. PENRR infers robust predictors of drug responses from patient genomic information. The special design of this model (comparing each drug with other drugs in the same cell line and comparing that drug with itself in other cell lines) significantly enhances the accuracy of the drug prediction model under limited data. This approach is also able to solve the problem of fitting on the insensitive drugs that is commonly encountered in regression-based models. Second, Regression-based Ranking by Pairwise Cluster Comparisons (RRPCC) is a ridge-regression-based method for ranking clusters of similar protein complex conformations generated by an underlying docking program (i.e., ClusPro). Rather than using regression to predict scores, which would equally penalize deviations for either low-quality and high-quality clusters, we seek to predict the difference of scores for any pair of clusters corresponding to the same complex. RRPCC combines these pairwise assessments to form a ranked list of clusters, from higher to lower quality. We apply RRPCC to clusters produced by the automated docking server ClusPro and, depending on the training/validation strategy, we show. improvement by 24%–100% in ranking acceptable or better quality clusters first, and by 15%–100% in ranking medium or better quality clusters first. Third, Distributionally Robust Multi-Output Regression Ranking (DRMRR) is a listwise LTR model that induces robustness into LTR problems using the Distributionally Robust Optimization framework. Contrasting to existing methods, the scoring function of DRMRR was designed as a multivariate mapping from a feature vector to a vector of deviation scores, which captures local context information and cross-document interactions. DRMRR employs ranking metrics (i.e., NDCG) in its output. Particularly, we used the notion of position deviation to define a vector of relevance score instead of a scalar one. We then adopted the DRO framework to minimize a worst-case expected multi-output loss function over a probabilistic ambiguity set that is defined by the Wasserstein metric. We also presented an equivalent convex reformulation of the DRO problem, which is shown to be tighter than the ones proposed by the previous studies. Fourth, Inversion Transformer-based Neural Ranking (ITNR) is a Transformer-based model to predict drug responses using RNAseq gene expression profiles, drug descriptors, and drug fingerprints. It utilizes a Context-Aware-Transformer architecture as its scoring function that ensures the modeling of inter-item dependencies. We also introduced a new loss function using the concept of Inversion and approximate permutation matrices. The accuracy and robustness of these LTR models are verified through three medical applications, namely cluster ranking in protein-protein docking, medical document retrieval, and drug response prediction.

With a rapid growth in smartwatch and smartwatch audio technologies, there is a lack of knowledge regarding user needs for smartwatch audio experiences and how those needs can be satisfied through user-centered design. Previous smartwatch user behavior studies suggest that audio app usage is not a primary use case for the smartwatch. However, audio applications are increasingly incorporated into smartwatches, which leads to the question of the apps’ purpose, validity, overlooked contexts and use cases. This thesis aims to understand what kind of audio experience(s) a user-centered design process might generate for the smartwatch. The design process generated insights from smartwatch users of audio applications, that were used as design guidelines for Context Awareness, Micro-interactions, and Device Ecosystem. The resulting prototype HeartBeats considers Context Awareness with heart rate music recommendations, Micro-interactions with one-handed song skipping and Quickplay music, and Device Ecosystem with speaker access and phone battery support. / Med en snabb teknisk utveckling av smartklockor och tillhörande ljudteknik finns det en kunskapsbrist om användarbehov och hur dessa kan tillfredsställas genom användarcentrerad design. Tidigare forskning om smartklocksanvändares beteenden tyder på att ljudapplikationer inte är ett huvudsakligt användningsområde för smartklockor. Ljudapplikationer implementeras dock allt mer i smartklockor, vilket leder till frågan om vilket värde de ger och om användningsområden möjligen har förbisetts. Den här uppsatsen syftar till att förstå vilka sorts ljudupplevelser en användarcentrerad designprocess skulle resultera i för smartklockor. Designprocessen resulterade i insikter om smartklocksanvändares beteenden med ljudapplikationer, vilket användes som designriktlinjer för kontextmedvetenhet, mikrointeraktioner och ekosystem av enheter. Den resulterande prototypen HeartBeats nyttjar kontextmedvetenhetgenom att rekommendera musik med användarens hjärtrytm i åtanke, mikrointeraktioner med en gest för att byta låt och snabbstart av musik, samt ekosystem av enheter genom snabb åtkomst till klockhögtalare och stöd för att spara telefonbatteri.

This report investigates if it is possible to use the Indian Buffet Process (IBP), a stochastic process that defines a probability distribution, as part of a recommendation system. The report focuses on recommendation systems where one type of object, for instance movies, is recommended to another type of object, for instance users.         A concept of performing link prediction with IBP is presented, along with a method for performing inference. Three papers that are related to the subject are presented and their results are analyzed together with additional experiments on an implementation of the IBP.        The report arrives at the conclusion that it is possible to use IBP in a recommendation system when recommending one object to another. In order to use IBP priors in a recommendation system which include real-life datasets, the paper suggests the use of a coupled version of the IBP model and if possible perform inference with a parallel Gibbs sampling. / Denna rapport undersöker om det är möjligt att använda Indian Buffet Process (IBP), en stokatisk process som definierar en sannolikhetsfördelning, som en del av ett rekommendationssystem. Rapporten fokuserar på rekommendationssystem där en sorts objekt, exempelvis filmer, rekommenderas till en annan sorts objekt, exempelvis användare.         Ett sätt att förutse länkar, link prediction, mellan olika objekt med hjälp av IBP presenteras tillsammans med en metod för att dra statistiska slutsatser, inference. Tre rapporter som är relaterade till ämnet presenteras och deras resultat analyseras tillsammans med ytterligare experiment på en implementation av IBP.        Rapporten drar slutsatsen att det är möjligt att använda IBP i ett rekommendationssystem då systemet rekommenderar ett objekt till ett annat objekt. Rapporten föreslår en kopplad version av IBP för att kunna använda IBP i ett rekommendationssystem som arbetar på riktigt data samt att inference ska utföras med en parallell Gibbs sampling.

This thesis proposes a two-stage recommendation system for providing music recommendations based on seed playlists as inputs. The goal is to help businesses find relevant and brand-fit music to play in their venues. The problem of recommending music using machine learning has been investigated quite a bit in both academia and the industry, with collaborative filtering and content-based filtering being the major approaches used. One of the difficulties of creating a recommendation system is how to evaluate it. In this thesis, both a quantitative and a qualitative evaluation are made to determine how well the results correspond to the actual quality of recommendations. The application of recommending music to businesses also poses different problems than a service directed at end consumers, mostly related to how many track recommendations are needed. A two-stage approach was used with Stage 1 producing candidates and a Stage 2 model using a neural network comparing five tracks from the playlist with a candidate was used to rank said candidates. The results show that the Stage 2 model has substantially better results in both the qualitative and quantitative evaluation compared to Stage 1. The quality of the recommendations from the whole system is not completely satisfactory, and some possible reasons for this are discussed, including improving the Stage 1 candidate generator (which was not modified in the scope of this thesis). / Automatisk musikrekommendation med hjälp av maskininlärning har utforskats av både industrin och akademin genom åren, där två huvudsakliga metoder utkristalliserats: collaborative filtering samt content-based filtering. I det här arbetet har en content-based modell tagits fram, uppdelad i två stadier: Steg 1 som genererar kandidater som Steg 2 sedan ordnade om med hjälp av ett neuralt nätverk som jämförde 5 låtar i taget från en spellista med motsvarande kandidater genererade av Steg 1 En av svårigheterna med att skapa automatiska rekommendationer är utvärderingen av den. I det här arbetet har både en kvantitativ och kvalitativ studie utförts för att försäkra att resultaten motsvarar den faktiska kvaliten hos rekommendationerna. Slutmålet med att hjälpa företag med musikrekommendation ställer också unika problem att lösa i jämförelse med en tjänst för privatpersoner, framförallt relaterat till storleken på de returnerade rekommendationerna. Resultaten visade att Steg 2 lyckades rangordna rekommendationerna från Steg 1 på ett sätt som gav högre poäng i både den kvantitativa och kvalitativa utvärderingen av systemen. De slutgiltiga resultaten var inte helt tillfredsställande, och potentialla orsaker till detta diskuteras. Dessa inkluderar Steg 1 (som inte modifierades inom ramen för detta arbete). Utvärderingen visade dock att de kvantitativa utvärderingsramarna verkar motsvara den upplevda kvaliten hos rekommendationerna baserat på den kvalitativa utvärderingen.

Les systèmes de recommandation sont devenus une partie indispensable des services et des applications d’internet, en particulier dû à la surcharge de données provenant de nombreuses sources. Quel que soit le type, chaque système de recommandation a des défis fondamentaux à traiter. Dans ce travail, nous identifions trois défis communs, rencontrés par tous les types de systèmes de recommandation: les données, les modèles d'apprentissage et la protection de la vie privée. Nous élaborons différents problèmes qui peuvent être créés par des données inappropriées en mettant l'accent sur sa qualité et sa quantité. De plus, nous mettons en évidence l'importance des réseaux sociaux dans la mise à disposition publique de systèmes de recommandation contenant des données sur ses utilisateurs, afin d'améliorer la qualité des recommandations. Nous fournissons également les capacités d'inférence de données publiques liées à des données relatives aux utilisateurs. Dans notre travail, nous exploitons cette capacité à améliorer la qualité des recommandations, mais nous soutenons également qu'il en résulte des menaces d'atteinte à la vie privée des utilisateurs sur la base de leurs informations. Pour notre second défi, nous proposons une nouvelle version de la méthode des k plus proches voisins (knn, de l'anglais k-nearest neighbors), qui est une des méthodes d'apprentissage parmi les plus populaires pour les systèmes de recommandation. Notre solution, conçue pour exploiter la nature bipartie des ensembles de données utilisateur-élément, est évolutive, rapide et efficace pour la construction d'un graphe knn et tire sa motivation de la grande quantité de ressources utilisées par des calculs de similarité dans les calculs de knn. Notre algorithme KIFF utilise des expériences sur des jeux de données réelles provenant de divers domaines, pour démontrer sa rapidité et son efficacité lorsqu'il est comparé à des approches issues de l'état de l'art. Pour notre dernière contribution, nous fournissons un mécanisme permettant aux utilisateurs de dissimuler leur opinion sur des réseaux sociaux sans pour autant dissimuler leur identité. / Recommendation systems have gained tremendous popularity, both in academia and industry. They have evolved into many different varieties depending mostly on the techniques and ideas used in their implementation. This categorization also marks the boundary of their application domain. Regardless of the types of recommendation systems, they are complex and multi-disciplinary in nature, involving subjects like information retrieval, data cleansing and preprocessing, data mining etc. In our work, we identify three different challenges (among many possible) involved in the process of making recommendations and provide their solutions. We elaborate the challenges involved in obtaining user-demographic data, and processing it, to render it useful for making recommendations. The focus here is to make use of Online Social Networks to access publicly available user data, to help the recommendation systems. Using user-demographic data for the purpose of improving the personalized recommendations, has many other advantages, like dealing with the famous cold-start problem. It is also one of the founding pillars of hybrid recommendation systems. With the help of this work, we underline the importance of user’s publicly available information like tweets, posts, votes etc. to infer more private details about her. As the second challenge, we aim at improving the learning process of recommendation systems. Our goal is to provide a k-nearest neighbor method that deals with very large amount of datasets, surpassing billions of users. We propose a generic, fast and scalable k-NN graph construction algorithm that improves significantly the performance as compared to the state-of-the art approaches. Our idea is based on leveraging the bipartite nature of the underlying dataset, and use a preprocessing phase to reduce the number of similarity computations in later iterations. As a result, we gain a speed-up of 14 compared to other significant approaches from literature. Finally, we also consider the issue of privacy. Instead of directly viewing it under trivial recommendation systems, we analyze it on Online Social Networks. First, we reason how OSNs can be seen as a form of recommendation systems and how information dissemination is similar to broadcasting opinion/reviews in trivial recommendation systems. Following this parallelism, we identify privacy threat in information diffusion in OSNs and provide a privacy preserving algorithm for the same. Our algorithm Riposte quantifies the privacy in terms of differential privacy and with the help of experimental datasets, we demonstrate how Riposte maintains the desirable information diffusion properties of a network.