Doctor of Philosophy / Computing and Information Sciences / Doina Caragea / Increasing amounts of content on the Web means that users can select from a wide variety of items (i.e., items that concur with their tastes and requirements). The generation of personalized item suggestions to users has become a crucial functionality for many web applications as users benefit from being shown only items of potential interest to them. One popular solution to creating personalized item suggestions to users is recommender systems. Recommender systems can address the item recommendation task by utilizing past user preferences for items captured as either explicit or implicit user feedback. Numerous collaborative filtering (CF) approaches have been proposed in the literature to address the recommendation problem in the single-domain setting (user preferences from only one domain are used to recommend items). However, increasingly large datasets often prevent experimentation of every approach in order to choose the one that best fits an application domain. The work in this dissertation on the single-domain setting studies two CF algorithms, Adsorption and Matrix Factorization (MF), considered to be state-of-the-art approaches for implicit feedback and suggests that characteristics of a domain (e.g., close connections versus loose connections among users) or characteristics of data available (e.g., density of the feedback matrix) can be useful in selecting the most suitable CF approach to use for a particular recommendation problem. Furthermore, for Adsorption, a neighborhood-based approach, this work studies several ways to construct user neighborhoods based on similarity functions and on community detection approaches, and suggests that domain and data characteristics can also be useful in selecting the neighborhood approach to use for Adsorption. Finally, motivated by the need to decrease computational costs of recommendation algorithms, this work studies the effectiveness of using short-user histories and suggests that short-user histories can successfully replace long-user histories for recommendation tasks. Although most approaches for recommender systems use user preferences from only one domain, in many applications, user interests span items of various types (e.g., artists and tags). Each recommendation problem (e.g., recommending artists to users or recommending tags to users) can be considered unique domains, and user preferences from several domains can be used to improve accuracy in one domain, an area of research known as cross-domain recommender systems. The work in this dissertation on cross-domain recommender systems investigates several limitations of existing approaches and proposes three novel approaches (two Adsorption-based and one MF-based) to improve recommendation accuracy in one domain by leveraging knowledge from multiple domains with implicit feedback. The first approach performs aggregation of neighborhoods (WAN) from the source and target domains, and the neighborhoods are used with Adsorption to recommend target items. The second approach performs aggregation of target recommendations (WAR) from Adsorption computed using neighborhoods from the source and target domains. The third approach integrates latent user factors from source domains into the target through a regularized latent factor model (CIMF). Experimental results on six target recommendation tasks from two real-world applications suggest that the proposed approaches effectively improve target recommendation accuracy as compared to single-domain CF approaches and successfully utilize varying amounts of user overlap between source and target domains. Furthermore, under the assumption that tuning may not be possible for large recommendation problems, this work proposes an approach to calculate knowledge aggregation weights based on network alignment for WAN and WAR approaches, and results show the usefulness of the proposed solution. The results also suggest that the WAN and WAR approaches effectively address the cold-start user problem in the target domain.

As primeiras mídias sociais da internet surgiram há pouco mais de duas décadas, segunda metade dos anos 90. Em comparação com a evolução humana, isso seria algo como um milésimo de segundo de sua existência. Neste período, vários estudos procuram entender o comportamento e o agrupamento dos seres humanos nesta nova forma de comunicação. Teorias sobre formas de analisar as pessoas neste meio e como elas se agrupam e criam novos modos de comunicação e propagação de suas ideias florescem e iluminam este desconhecido caminho a ser criado e percorrido. Os métodos de identificação do comportamento humano criados antes das mídias sociais ganham uma nova forma de serem utilizados. Estudos sobre o \"eu\" (Belk, 1988), tribalismo (Cova, B., 1997), etnografia (Danzig, 1985), netnografia (Kozinets, 1998) e filtragem colaborativa (Golberg, Nichols, Oki e Terry, 1992) entram em cena para colocar uma luz no estudo das relações humanas no mundo digital. A internet revolucionou o modo de as pessoas interagirem e a evolução constante da tecnologia vem incessantemente gerando profundas implicações para o marketing. A rede mundial passou a ser um canal global pelo qual as empresas podem divulgar e vender seus produtos. No entanto, mesmo oferecendo um enorme potencial para as empresas, a internet aumentou a complexidade de identificar os clientes. Os usuários presentes nas mídias sociais estão menos interessados nos produtos e valorizam mais as identidades e os laços sociais gerados em torno de seus assuntos de interesse. Estas tribos eletrônicas ultrapassam as fronteiras geográficas e independem de raça, sexo e aspectos culturais de seus integrantes. Este trabalho apresenta um método para identificar tribos nas mídias sociais. O método foi aplicado na identificação da tribo de MMA (MixedMartialArts, em tradução livre, Artes Marciais Mistas) no Twitter. A validação foi realizada usando a plataforma de anúncios do Twitter, enviando durante 72 horas uma publicidade para mais de 600 mil usuários, divididos em grupo de controle e segmentações do Twitter e do método proposto DNA. O estudo comparou os resultados obtidos pelo método proposto DNA com os resultados do grupo de controle e da segmentação realizada pelo Twitter. Os resultados obtidos apontaram o aumento de interações dos usuários identificados como pertencentes a tribo de MMA, validando o método. / The first Internet social media emerged just over two decades ago, at the second half of 90\'s. Compared to human evolution, this would be something like a millisecond of its existence. In this period, several studies try to understand the behavior and grouping of human beings in this new form of communication. Theories about ways of analyzing people in this environment and how they group themselves and create new ways of communication and propagation their ideas flourish and illuminate this unknown pathway to be created and traveled. Methods of identifying human behavior created before social media receive a new way of being used. Studies on the \"self\" (Belk, 1988), tribalism (Cova, B., 1997), ethnography (Danzig, 1985), netnography (Kozinets, 1998) and collaborative filtering (Golberg, Nichols, Oki and Terry, 1992) come on the scene to shed light on the study of human relations in the digital world. The Internet has revolutionized people\'s way of interacting and the constant evolution of technology generates profound implications for the marketing. The worldwide network has become a global channel through which companies can disclose and sell their products. However, while offering tremendous potential to businesses, the Internet has increased the complexity of identifying customers. Users present in social media are less interested in products and value more the identities and social ties generated around their subjects of interest. These electronic tribes transcend the geographical borders and are independent of race, sex and cultural aspects of its members. This paper presents a method to identify tribes in social media. The method was applied in the identification of the MMA (Mixed Martial Arts) tribe on Twitter. The validation was done using the Twitter ads platform, sending 72 hours of advertisement for more than 600 thousand users, divided in control group and segmentations of Twitter and the proposed method. The study compared the results obtained by the proposed method with that of the control group and the segmentation created by Twitter. The obtained results pointed out the increase of interactions of the users identified as belonging to the MMA tribe validating the method.

Detaljhandeln i fysiska butiker är utsatt av konkurrens från en betydligt mer innovationsrik e-handel och har därför ett behov av att vidareutvecklas. Ett sätt för detaljhandeln att utvecklas är att utnyttja tekniker som visats fungera bra inom e-handeln. Rekommendationssystem som ger rekommendationer till sina användare har nått stora framgångar och används av i stort sett alla företag inom e-handeln. Den mest använda tekniken för att ta fram rekommendationer kallas för collaborative filtering. Inom detaljhandel används dock inte detta i någon större utsträckning. Det finns därför förhållandevis lite kunskap om vad kunder anser om rekommendationer i butik. Syftet med studien är därför att utvärdera hur ett rekommendationssystem baserat på collaborative filtering presterar i en fysisk butik. Utvärderingen sker genom att mäta träffsäkerheten på rekommendationerna kunder får i en butik samt vad kunderna anser om dessa. Studien ämnar även att ta reda på hur kunder förhåller sig till automatiska rekommendationer i butik. I studien används två forskningsmetodiker för att uppnå dess forskningsmål. Design science har tillämpats för att utvärdera hur ett rekommendationssystem baserat på collaborative filtering presterar i en fysisk butik. En prototyp baserat på collaborative filtering utvecklades för att generera rekommendationer. Prototypen användes sedan i ett användartest som genomfördes i en butiksmiljö. För att belysa hur kunder förhåller sig till automatiska rekommendationer i butik användes en enkätundersökning som utfördes i samband med studiens användartest. Studiens resultat visar att prototypen gav rekommendationer med en hög träffsäkerhet där deltagarna upplevde rekommendationerna som bra och relevanta. Resultaten visar även att deltagarna i studien var positivt inställda till att få rekommendationer i butik. Detta leder till slutsatsen att rekommendationssystem baserat på collaborative filtering kan prestera väl i butiker vilket ger en indikation om att detta kan vara ett sätt för butiker att vidareutveckla handeln. / Retail stores are challenged by competition from the more innovative retailers in e-commerce and thus needs to adapt and evolve in order to stay competitive. This could be accomplished by using technology which has been proven successful in e-commerce. Recommender systems that produces recommendations to its users has been used successfully and is used by essentially all businesses involved in e-commerce. The most common method employed in these recommender systems is called collaborative filtering. Recommender systems have however not yet found its way into retail stores to a greater extent. This has led to a gap in knowledge regarding customer’s opinions of recommendations in retail stores. The purpose of this study is therefore to evaluate how recommender system based on collaborative filtering performs when used in retail stores. The evaluation is performed by measuring the accuracy of the recommendations a customer receives in a retail store as well as what the customer thinks of the recommendation. This study also intends to explore and shed light on people’s opinions concerning automatic recommendations in retail stores. Two different research methods have been used in this study. Design science is being used in order to evaluate how a recommender system based on collaborative filtering performs when used in retail stores. A prototype based on collaborative filtering was developed in order to generate recommendations. The prototype was then used in a user-test taking place in a retail-like environment. In order to shed light on people’s opinions regarding automatic recommendations in retail stores a questionnaire was handed out to the participants in conjunction with the user-test. The results of the study show that the prototype could produce high accuracy recommendations where the participants perceived the recommendations as good and relevant. The results also show that the participants of the study have positive attitude and were in favor of receiving automatic recommendations in retail stores. This leads to the conclusion that recommendations based on collaborative filtering could indeed perform well in retail stores. This indicates that recommender systems using collaborative filtering is one possible way for retail stores to evolve their business.

Due to the growth of online shopping and services, various types of products can be recommended to an individual. After reviewing the current methods for cross-domain recommendations, we believe that there is a need to make different types of recommendations by relying on a common base, and that it is better to depend on a target customer’s information when building the base, because the customer is the one common element in all the purchases. Therefore, we suggest a recommender system (RS) that develops a personality profile for each product, and represents items by an aggregated vector of personality features of the people who have liked the items. We investigate two ways to build personality profiles for items (IPPs). The first way is called average-based IPPs, which represents each item with five attributes that reflect the average Big Five Personality values of the users who like it. The second way is named proportion-based IPPs, which consists of 15 attributes that aggregate the number of fans who have high, average and low Big Five values. The system functions like an item-based collaborative filtering recommender; that is, it recommends items similar to those the user liked. Our system demonstrates the highest recommendation quality in providing cross-domain recommendations, compared to traditional item-based collaborative filtering systems and content-based recommenders.

L'objectif de cette thèse est de développer des algorithmes d'apprentissage adaptés aux grandes masses de données. Dans un premier temps, nous considérons le problème de la classification avec un grand nombre de classes. Afin d'obtenir un algorithme adapté à la grande dimension, nous proposons un algorithme qui transforme le problème multi-classes en un problème de classification binaire que nous sous-échantillonnons de manière drastique. Afin de valider cette méthode, nous fournissons une analyse théorique et expérimentale détaillée.Dans la seconde partie, nous approchons le problème de l'apprentissage sur données distribuées en introduisant un cadre asynchrone pour le traitement des données. Nous appliquons ce cadre à deux applications phares : la factorisation de matrice pour les systèmes de recommandation en grande dimension et la classification binaire. / This thesis focuses on developing scalable algorithms for large scale machine learning. In this work, we present two perspectives to handle large data. First, we consider the problem of large-scale multiclass classification. We introduce the task of multiclass classification and the challenge of classifying with a large number of classes. To alleviate these challenges, we propose an algorithm which reduces the original multiclass problem to an equivalent binary one. Based on this reduction technique, we introduce a scalable method to tackle the multiclass classification problem for very large number of classes and perform detailed theoretical and empirical analyses.In the second part, we discuss the problem of distributed machine learning. In this domain, we introduce an asynchronous framework for performing distributed optimization. We present application of the proposed asynchronous framework on two popular domains: matrix factorization for large-scale recommender systems and large-scale binary classification. In the case of matrix factorization, we perform Stochastic Gradient Descent (SGD) in an asynchronous distributed manner. Whereas, in the case of large-scale binary classification we use a variant of SGD which uses variance reduction technique, SVRG as our optimization algorithm.

We demonstrate spectral learning can be combined with a random forest classifier to produce a hybrid recommender system capable of incorporating meta information. Spectral learning is supervised learning in which data is in the form of one or more networks. Responses are predicted from features obtained from the eigenvector decomposition of matrix representations of the networks. Spectral learning is based on the highest weight eigenvectors of natural Markov chain representations. A random forest is an ensemble technique for supervised learning whose internal predictive model can be interpreted as a nearest neighbor network. A hybrid recommender can be constructed by first deriving a network model from a recommender's similarity matrix then applying spectral learning techniques to produce a new network model. The response learned by the new version of the recommender can be meta information. This leads to a system capable of incorporating meta data into recommendations.

Recommender systems are used on many different websites today and are mechanisms that are supposed to accurately give personalized recommendations of items to a set of different users. An item can for example be movies on Netflix. The purpose of this paper is to implement an algorithm that fulfills five stated goals of the implementation. The goals are as followed: the algorithm should be easy to implement, be effective on query time, accurate on recommendations, put little expectations on users and alternations of algorithm should not have to be changed comprehensively. Slope One is a simplified version of linear regression and can be used to recommend items. By using the Netflix Prize data set from 2009 and the Root-Mean-Square-Error (RMSE) as an evaluator, Slope One generates an accuracy of 1.007 units. The Weighted Slope One, which takes the relevancy of items into the calculation, generates an accuracy of 0.990 units.  Adding Weighted Slope One to the Slope One implementation can be done without changing the fundamentals of the Slope One algorithm. It is nearly instantaneous to generate a recommendation of a movie with regular Slope One and Weighted Slope One. However, a precomputing stage is needed for the mechanism. In order to receive a recommendation of the implementation in this paper, the user must at least have rated two items. / Rekommendationssystem används idag på många olika hemsidor, och är en mekanism som har syftet att, med noggrannhet, ge en personlig rekommendation av objekt till en mängd olika användare. Ett objekt kan exempelvis vara en film från Netflix. Syftet med denna rapport är att implementera en algoritm som uppfyller fem olika implementationsmål. Målen är enligt följande: algoritmen ska vara enkel att implementera, ha en effektiv tid på dataförfrågan, ge noggranna rekommendationer, sätta låga förväntningar hos användaren samt ska algoritmen inte behöva omfattande förändring vid alternering.  Slope One är en förenklad version av linjär regression, och kan även användas till att rekommendera objekt. Genom att använda datamängden från Netflix Prize från 2009 och måttet Root-Mean-Square-Error (RMSE) som en utvärderare, kan Slope One generera en precision på 1.007 enheter. Den viktade Slope One, som tar hänsyn till varje föremåls relevans, genererar en precision på 0.990 enheter. När dessa två algoritmer kombineras, behövs inte större fundamentala ändringar i implementationen av Slope One. En rekommendation av något objekt kan genereras omedelbart med någon av de två algoritmerna, dock krävs det en förberäkningsfas i mekanismen. För att få en rekommendation av implementationen i denna rapport, måste användaren åtminstone ha värderat två objekt.

E-commerce is an ever growing industry as the internet infrastructure continues to evolve. The benefits from a recommendation system to any online retail store are several. It can help customers to find what they need as well as increase sales by enabling accurate targeted promotions. Among many techniques that can form recommendation systems, this thesis compares Collaborative Filtering against Association Rule Mining, both implemented in combination with clustering. The suggested implementations are designed with the cold start problem in mind and are evaluated with a data set from an online retail store which sells clothing. The results indicate that Collaborative Filtering is the preferable technique while associated rules may still offer business value to stakeholders. However, the strength of the results is undermined by the fact that only a single data set was used. / E-handel är en växande marknad i takt med att Internet utvecklas samtidigt som antalet användare ständigt ökar. Antalet fördelar från rekommendationssytem som e-butiker kan dra nytta av är flera. Samtidigt som det kan hjälpa kunder att hitta vad de letar efter kan det utgöra underlag för riktade kampanjer, något som kan öka försäljning. Det finns många olika tekniker som rekommendationssystem kan vara byggda utifrån. Detta examensarbete ställer fokus på de två teknikerna Collborative Filtering samt Association Rule Mining och jämför dessa sinsemellan. Båda metoderna kombinerades med klustring och utformades för att råda bot på kallstartsproblemet. De två föreslagna implementationerna testades sedan mot en riktig datamängd från en e-butik med kläder i sitt sortiment. Resultaten tyder på att Collborative Filtering är den överlägsna tekniken samtidigt som det fortfarande finns ett värde i associeringsregler. Att dra generella slutsatser försvåras dock av att enbart en datamängd användes.

Teknik och data är nyckeln till att Bonnier Business Media (BBM) ska kunna nå sina mål och leverera ytterligare tillväxt. Därför vill man ligga i framkant när det gäller att undersöka nya tekniker som kan förbättra plattformarna och göra dem mer tidsenliga. BBM har bland annat velat ta fram ett rekommendationssystem som ska användas till att göra innehållet individanpassat på webbplatserna och på ett effektivt sätt presentera detta så att de olika målgrupperna får den information de förväntar sig. Till exempel ska besökaren kunna få förslag på artiklar och evenemang som kan vara av intresse. Målet med detta examensarbete har varit att ta fram en prototyp för ett rekommendationssy- stem med tillhörande algoritmer. Prototypen skulle kunna användas som ett “koncepttest” för att undersöka möjligheten att skapa personliga rekommendationer till läsare på Veckans Affärers webbplats, va.se. Implementationen av rekommendationssystem som togs fram till BBM bestod av en objektbaserad kollaborativ filtrerings algoritm som använde besökarnas beteende, publiceringsdatum och popularitet på artiklarna och evenemangen för att skapa individuella rekommendationer. Efter genomförda tester och analyser visar resultatet att det är fullt möjligt att skapa personliga rekommendationer som har en högre precision än vad ett grundläggande rekommendationssystem, till exempel en popularitetslista, kan erbjuda. / Technology is the key for Bonnier Business Media (BBM) to reach their goals and deliver future growth. Therefore they want to be in the very forefront when it comes to exploring new technologies that can improve their platforms and make them more up to date. BBM has among other things aimed to develop a recommendation system that is supposed to make the content of their web sites personalized and in an efficient way present this so that the different target groups will get the information that they expect. For example the visitor should be able to get suggestions on articles and events that might be of interest. The aim of this thesis has been to develop a prototype of a recommendation system with associated algorithms. The prototype could be used as to examine the possibility to create personalized recommendations for the readers on BBM:s website va.se (Veckans Affärer). The implementation of the recommendation system that was developed for BBM consisted of an object-based collaborative filtering algorithm using visitor behavior, publication date and popularity of articles and events to create personalized recommendations. After com- pleting tests and analyzes the results show that it is possible to create recommendations with a higher precision than a basic recommendation system, like a popularity list, can of- fer.

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.