京都大学 / 新制・課程博士 / 博士(情報学) / 甲第24036号 / 情博第792号 / 新制||情||134(附属図書館) / 京都大学大学院情報学研究科社会情報学専攻 / (主査)准教授 馬 強, 教授 田島 敬史, 教授 森 信介 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

This thesis deals with the field of recommendation systems using deep neural networks and their use in book recommendation. There are the main traditional recommender systems analysed and their representations are summarized, as well as systems with more advanced techniques based on machine learning. The core of the thesis is to use convolutional neural networks for natural language processing and create a hybrid book recommendation system. Suggested system includes matrix factorization and make recommendation based on user ratings and book metadata, including texts descriptions. I designed two models, one with bag-of-words technique and one with convolutional neural network. Both of them defeat baseline methods. On the created data set, that was created from the Goodreads, model with CNN beats model with BOW.

在現在的社會裡，手機應用程式已經被人們接受與廣泛地利用，然而目前市面上的手機 App 推薦系統，多以使用者實際使用與回報作為參考，若有惡意行為軟體，在使用者介面後竊取使用者資料，這些推薦系統是難以查知其行為的，因此我們提出了 AppReco，一套可以系統化的推薦 iOS App 的推薦系統，而且不需要使用者去實際操作、執行 App。 整個分析流程包括三個步驟：(1) 透過無監督式學習法的隱含狄利克雷分布(Latent Dirichlet Allocation, LDA)做出主題模型，再使用增長層級式自我組織映射圖(Growing Hierarchical Self-Organizing Map, GHSOM)進行分群。(2)使用靜態分析程式碼，去找出其應用程式所執行的行為。(3)透過我們的評分公式對於這些 App，進行評分。 在分群 App 方面，AppReco 使用這些應用程式的官方敘述來進行分群，讓擁有類似屬性的手機應用程式群聚在一起；在檢視 App 方面，AppReco 透過靜態分析這些 App 的程式碼，來計算其使用行為的多寡；在推薦 App 方面，AppReco 分析類似屬性的 App 與其執行的行為，最後推薦使用者使用較少敏感行為(如使用廣告、使用個人資料、使用社群軟體開發包等)的 App。 而本研究使用在 Apple App Store 上面數千個在各個類別中的前兩百名 App 做為我們的實驗資料集來進行實驗。 / Mobile applications have been widely used in life and become dominant software applications nowadays. However there are lack of systematic recommendation systems that can be leveraged in advance without users’ evaluations. We present AppReco, a systematic recommendation system of iOS mobile applications that can evaluate mobile applications without executions. AppReco evaluates apps that have similar interests with static binary analysis, revealing their behaviors according to the embedded functions in the executable. The analysis consists of three stages: (1) unsupervised learning on app descriptions with Latent Dirichlet Allocation for topic discovery and Growing Hierarchical Self-organizing Maps for hierarchical clustering, (2) static binary analysis on executables to discover embedded system calls and (3) ranking common-topic applications from their matched behavior patterns. To find apps that have similar interests, AppReco discovers (unsupervised) topics in official descriptions and clusters apps that have common topics as similar-interest apps. To evaluate apps, AppReco adopts static binary analysis on their executables to count invoked system calls and reveal embedded functions. To recommend apps, AppReco analyzes similar-interest apps with their behaviors of executables, and recommend apps that have less sensitive behaviors such as commercial advertisements, privacy information access, and internet connections, to users. We report our analysis against thousands of iOS apps in the Apple app store including most of the listed top 200 applications in each category.

Baigiamajame magistro darbe yra nagrinėjamos pagrindinės rekomendacinių sistemų socialiniuose tinkluose problemos: rekomendacijų tikslumas, pasyvūs vartotojai, neaiškios rekomendacijos. Iškeliami pagrindiniai tikslai, kuriais siekiama išanalizuoti rekomendacinių sistemų veikimo principus ir pasiūlyti metodą rekomendacinės sistemos efektyvumui ir rekomendacijų tikslumui gerinti. Analizuojama užsienio literatūra, atliekamas eksperimentas su realia duomenų baze. Analizuoti rezultatai vertinami skirtingų vartotojų grupių (pasyvūs, aktyvūs, tarpiniai vartotojai), analizuojamas kiekvienos grupės poveikis bendram rezultatui. Aprašytos aiškios rekomendacijos ir pateikti jų pavyzdžiai. Rezultatai vertinti trimis skirtingais kriterijais: PAP, PAVP ir ĮA. Pirmi du rodo skaičiavimo nuokrypius, trečiasis – įverčių apimtį. Atlikti penki bandymai skirtingomis sąlygomis ir pateikti kiekvieno bandymo bei bendri rezultatai. Išsikelti tikslai yra įgyvendinti, nes analizuojant gautus rezultatus, tiek bendras rekomendacijų tikslumas, tiek pasyvių vartotojų pagerėjo. Bendras pagerėjo 4% ir 11%, pasyvių vartotojų – 11% ir 17%. Darbe aprašomas modelis gali būti nesunkiai pritaikomas rekomendacinei sistemai. Šis baigiamasis darbas gali būti naudojamas kaip pagrindas ar literatūros šaltinis tolesniems tyrimams Lietuvoje. / The thesis analyses the main problems of recommender systems in social networks: the accuracy of recommendations, cold start users, uncertain recommendations. The main goals are raised, which are used to analyze the principles of processes in recommender systems, and to offer a novel method to improve the efficiency and the accuracy of recommendations. The foreign articles are discussed and the new method is implemented on an existing data set. The results are evaluated according to the different groups of users (cold start users, heavy raters and intermediate users) and the effect of each group to the main result is analyzed. The transparent recommendations are explained with the examples. The results are evaluated using three different metrics: MAE, MAUE and RC. The first two determine the deviation of the calculations the third determines the coverage of the ratings. Five experiments were made with different conditions and the results of each are presented along with the general results. The held purposes were accomplished because the accuracy of recommendations increased for all users and for cold start users as well. The benefit of the accuracy for all users is 4% and 11%, for cold start users 11% and 17%. The model described in the thesis can be easily incorporated to the recommender system. This thesis can be used as the basis of future work of recommender systems in Lithuania.

選擇到哪家餐廳用餐的問題，不論旅遊或家居都經常會遇到。大多數的人會先上網，尋找符合自己喜好且評價好的美食。然而網際網路發達，在人人都可上網分享的情況下，造成資訊氾濫超載。使得使用者上網瀏覽資料時，很容易找到不切合需求的資訊。解決此資訊超載的方法之一是餐廳推薦系統。儘管目前有很多的推薦應用程式或是分享平台，諸如TripAdvisor、iPeen愛評網、foursquare…等等，資料豐富但卻沒有針對個人偏好做推薦。 　　本研究有鑒於許多人在品嘗美食之前，會先拍照並在Facebook或Instagram打卡做紀錄、分享給朋友，打卡的次數可能意味著此餐廳的熱門度。且使用者選擇的美食類型偏好也可能受到聚餐目的的影響。因此開發出一款結合社群互動圈以及考量用餐情境的餐廳推薦系統。此系統先利用使用者所選擇的聚餐場合、價位、餐廳類型、熱門商圈等元素篩選出合適的餐廳，再利用Facebook打卡資料取得與使用者偏好相似的好友，依據好友的相似度推算出使用者對餐廳的喜好程度，推薦符合使用者興趣及需求的餐廳，協助使用者能夠更容易地找到自己所喜好的店家。 　　本研究的實作系統，經過評估測試，結果發現結合社群互動圈及考量用餐情境的個人化推薦能讓使用者更容易找到自己所喜好的餐廳，而在推薦內容中顯示好友對餐廳的評論，更有效的幫助使用者作決策。未來本推薦系統所使用之結合情境元素所設計的模式亦可應用至其他領域的推薦平台，如旅遊景點推薦或旅遊住宿推薦。 / Most people face the issue of deciding which restaurant to eat. Searching through the Internet is the first step that people usually do. However the rapid growth of information has overloaded the Internet users, it makes difficult to find the most appropriate information for decision-making. Certainly there are several restaurant recommender systems have been developed to solve the problem, such as TripAdvisor, iPeen, foursquare, etc; but few systems provide personalized and context-based recommendations. 　　The research intends to develop a restaurant recommender system that considers the factors of social network and context. Nowadays, when people eat, they like to take a picture and check in on Facebook or Instagram to share with friends, the numbers of check-in for a restaurant may mean the restaurant’s popularity. In addition, the gathering purpose and personal preferences may also affect the users’ decisions. Therefore the recommender system first used the variables of eating criteria such as place, price, types of food, eating environments to filter restaurants. The system then got the user’s similar friends from check-in data of Facebook. Through calculating friends’ similarity and their preference of restaurants, the recommender system finds the most fitted ones for the user to choose from. 　　The afterward system’s users testing data prove that this personalized and context-based recommendation system provides better information to help the user make their decisions. The same model can be replicated to other domain of recommender platforms.

Sistemas de recomendação estão se tornando ferramentas indispensáveis para diversos websites, que buscam oferecer ao seu usuário uma experiência personalizada e simplificada, e sua adoção se deve principalmente devido ao grande volume de dados disponíveis, advindos de diferentes fontes e contendo informações diversificadas, aumentando a necessidade e a complexidade de se extrair valor desses dados. Com o surgimento de redes sociais online os usuários passaram a expressar seus gostos e preferências além de estabelecer relações com outros usuários, podendo estes serem seus amigos, parentes, ídolos, etc. Estas possibilidades encontradas em redes sociais motivou o presente trabalho a interpretar a comunidade científica como uma rede social, utilizando relações de coautoria, colaboração em projetos, orientações, além de citações de trabalhos e, consequentemente, citações aos respectivos autores. O objetivo deste projeto foi propor um sistema de recomendação de trabalhos científicos combinando informações sociais e informações bibliométricas, no que diz respeito a artigos citados em publicações, caraterizando-se como um facilitador para auxiliar os pesquisadores a responderem perguntas como: Quais artigos interessantes da minha área eu ainda não tenho conhecimento? e Quais artigos podem auxiliar em trabalhos que tenho em desenvolvimento? Para atingir o objetivo proposto foram desenvolvidas duas abordagens de recomendação. A primeira abordagem teve como premissa que o tempo em que as relações entre os autores foi estabelecida é determinante para selecionar os autores mais próximos (ou similares), ou seja, as relações mais recentes tendem a ser mais relevantes que as relações mais antigas. Já a segunda técnica combinou o resultados das diferentes técnicas implementadas (tanto a proposta quanto técnicas da literatura correlata) para gerar novas recomendações de maneira híbrida. Os resultados mostraram que a solução baseada no tempo apresentou resultados superiores às estratégias correlatas quando se possui mais informações sobre o autor, ou seja, autores que possuem diversas relações de coautoria e um conjunto de artigos citados elevado tendem a obter resultados melhores quando comparados aos autores que possuem poucas relações e citaram poucos artigos. Já a solução híbrida, que combina os resultados dos diversos recomendadores, apresentou uma cobertura de recomendações superior às demais, pelo fato de combinar os pontos fortes de cada uma das técnicas, encontrando recomendações relevantes no conjunto de testes em mais de 57% dos casos / Recommender systems are becoming indispensable tools on websites, in order to offer a simplified and personalized experience to their users, and its adoption is due to the fact that the volume of data available has increased and also comes from different sources with different types of information. Thus, it is challenge and necessary tools for helping to extract more valuable information from these data. The arise of online social networks allowed users to express their tastes and preferences and establish relationships with other users, such as friends, relatives, idols, etc. Those possibilities found in social networks motivated this work to interpret the scientific community as a social network, providing the ability to use co-authorship relations, collaboration in projects, tutoring relations, as well as paper citations and thus citations from their authors. The goal of this project was to propose a papers recommender system combining social and bibliometric information, regarding cited articles on published papers, being characterized as a facilitator to help researchers to answer questions such as: \"What interesting articles in my area I still have no knowledge of?\" and \"Which articles can assist in the project I am developing?\". The first algorithm proposed used the time when the coauthorship relations among authors were established as a determining parameter to choose which authors are more similar, meaning that relations established in recent time are more relevant than those that are older. The second algorithm combines the results from different implemented algorithms to determine which would be the ideal weight of each algorithm on the recommendation result, using a linear regression on the recommendations scores. The results showed that the time based solution achieved a better performance for the authors with higher amount of information available, i.e., if the author has many coauthorship relations and cited many papers, the results are better when compared with authors that does not have many relations and cited articles. On the other hand, the hybrid solution which combines the results from different recommendations approaches presented a higher coverage compared with others, due to the fact that it combines the strengths of each one of the algorithms, finding recommendation for users on 57% of the cases.

"Recommender systems are being widely applied in many E-commerce sites to suggest products, services, and information items to potential users. Collabora-tive filtering systems, the most successful recommender system technology to date, help people make choices based on the opinions of other people. While collaborative filtering systems have been a substantial success, there are sev-eral problems that researchers and commercial applications have identified: the early rater problem, the sparsity problem, and the large scale problem. Moreover, existing collaborative filtering systems do not consider data re-cency. For this reason, if a user's preferences have changed over time, the sys-tems might not recognize it quickly. This thesis studies how to apply data re-cency to collaborative filtering systems to get more predictive accuracy. We define the data recency problem as the negative impact of old data on the pre-dictive accuracy of collaborative filtering systems. In order to mitigate this shortcoming, the combinations of time-based forgetting mechanisms, pruning and non-pruning strategies and linear and kernel functions, are utilized to ap-ply weights. A clustering technique is employed to detect the user's changing preferences. We apply our research approach to the DeliBook dataset. The goal of our experiments is to show that our algorithm that incorporates tempo-ral factors provides better recommendations than existing methods."

Sistemas de recomendação estão se tornando ferramentas indispensáveis para diversos websites, que buscam oferecer ao seu usuário uma experiência personalizada e simplificada, e sua adoção se deve principalmente devido ao grande volume de dados disponíveis, advindos de diferentes fontes e contendo informações diversificadas, aumentando a necessidade e a complexidade de se extrair valor desses dados. Com o surgimento de redes sociais online os usuários passaram a expressar seus gostos e preferências além de estabelecer relações com outros usuários, podendo estes serem seus amigos, parentes, ídolos, etc. Estas possibilidades encontradas em redes sociais motivou o presente trabalho a interpretar a comunidade científica como uma rede social, utilizando relações de coautoria, colaboração em projetos, orientações, além de citações de trabalhos e, consequentemente, citações aos respectivos autores. O objetivo deste projeto foi propor um sistema de recomendação de trabalhos científicos combinando informações sociais e informações bibliométricas, no que diz respeito a artigos citados em publicações, caraterizando-se como um facilitador para auxiliar os pesquisadores a responderem perguntas como: Quais artigos interessantes da minha área eu ainda não tenho conhecimento? e Quais artigos podem auxiliar em trabalhos que tenho em desenvolvimento? Para atingir o objetivo proposto foram desenvolvidas duas abordagens de recomendação. A primeira abordagem teve como premissa que o tempo em que as relações entre os autores foi estabelecida é determinante para selecionar os autores mais próximos (ou similares), ou seja, as relações mais recentes tendem a ser mais relevantes que as relações mais antigas. Já a segunda técnica combinou o resultados das diferentes técnicas implementadas (tanto a proposta quanto técnicas da literatura correlata) para gerar novas recomendações de maneira híbrida. Os resultados mostraram que a solução baseada no tempo apresentou resultados superiores às estratégias correlatas quando se possui mais informações sobre o autor, ou seja, autores que possuem diversas relações de coautoria e um conjunto de artigos citados elevado tendem a obter resultados melhores quando comparados aos autores que possuem poucas relações e citaram poucos artigos. Já a solução híbrida, que combina os resultados dos diversos recomendadores, apresentou uma cobertura de recomendações superior às demais, pelo fato de combinar os pontos fortes de cada uma das técnicas, encontrando recomendações relevantes no conjunto de testes em mais de 57% dos casos / Recommender systems are becoming indispensable tools on websites, in order to offer a simplified and personalized experience to their users, and its adoption is due to the fact that the volume of data available has increased and also comes from different sources with different types of information. Thus, it is challenge and necessary tools for helping to extract more valuable information from these data. The arise of online social networks allowed users to express their tastes and preferences and establish relationships with other users, such as friends, relatives, idols, etc. Those possibilities found in social networks motivated this work to interpret the scientific community as a social network, providing the ability to use co-authorship relations, collaboration in projects, tutoring relations, as well as paper citations and thus citations from their authors. The goal of this project was to propose a papers recommender system combining social and bibliometric information, regarding cited articles on published papers, being characterized as a facilitator to help researchers to answer questions such as: \"What interesting articles in my area I still have no knowledge of?\" and \"Which articles can assist in the project I am developing?\". The first algorithm proposed used the time when the coauthorship relations among authors were established as a determining parameter to choose which authors are more similar, meaning that relations established in recent time are more relevant than those that are older. The second algorithm combines the results from different implemented algorithms to determine which would be the ideal weight of each algorithm on the recommendation result, using a linear regression on the recommendations scores. The results showed that the time based solution achieved a better performance for the authors with higher amount of information available, i.e., if the author has many coauthorship relations and cited many papers, the results are better when compared with authors that does not have many relations and cited articles. On the other hand, the hybrid solution which combines the results from different recommendations approaches presented a higher coverage compared with others, due to the fact that it combines the strengths of each one of the algorithms, finding recommendation for users on 57% of the cases.

In recent years, web has experienced a tremendous growth concerning users and content. As a result information overload problem has always been always one of the main discussion topics. The aim has always been to find the most desired solution in order to help users when they find it increasingly difficult to locate the accurate information at the right time. Recommender systems developed to address this need by helping users to find relevant information among huge amounts of data and they have now become a ubiquitous attribute to many websites. A recommender system guides users in their decisions by predicting their preferences while they are searching, shopping or generally surfing, based on their preferences collected from past as well as the preferences of other users. Until now, recommender systems has been vastly used in almost all professional e-commerce websites, selling or offering different variety of items from movies and music to clothes and foods. This thesis will present and explore different recommender system algorithms such as User-User Collaborative and Item-Item Collaborative filtering using open source library Apache mahout. Algorithms will be developed in order to evaluate the performance of these collaborative filtering algorithms. They will be compared and their performance will be measured in detail by using evaluation metrics such as RMSE and MAE and similarity algorithms such as Pearson and Loglikelihood.

När en kund går in i en e-butik för första gången får kunden inte anpassade rekommendationer. Anpassade rekommendationer bygger på information om användare från tidigare interaktion med e-butiken genom ett rekommendationssystem (RS). För nya användare har dock inte systemet den information som behövs för att kunna skapa anpassade rekommendationer, detta problem kallas för kallstart. För att bemöta nya användare i kallstart är RS beroende av deras explicita feedback. Konversationsrekommendationsystem (KRS) möjliggör explicit insamling från användare genom konversation. Explicit information kan lämnas av användare direkt i e-butiken. Informationen kan bestå av egenskaper användare föredrar hos produkter men även mer personlig information. Har inte användare förtroende till systemet kan det bidra till att användare inte vill dela med sig av information. För att användare ska vilja lämna explicit information kan KRS bemöta kallstart med att försöka skapa en känsla av förtroende hos användare. Studien är utförd med en designorienterad forskningsansats. Designelement identifierades genom en litteraturstudie som sedan implementerades i en prototyp för att undersöka om dessa kunde skapa en känsla av förtroende för systemet. Prototypen utvärderades därefter utifrån kriterierna trovärdighet, enkel användning och risk som fångar detta förtroende. Resultatet i denna studie indikerar att användare vill få anpassade rekommendationer presenterade och är därför villiga att bidra med den information systemet behöver. Studien presenterar åtta designförslag för hur KRS kan designas för att skapa en känsla av förtroende hos användare med en dialog mellan systemet och användare, möjlighet att ge effektiv feedback och systemets transparens. / När en kund går in i en e-butik för första gången får kunden inte anpassade rekommendationer. Anpassade rekommendationer bygger på information om användare från tidigare interaktion med e-butiken genom ett rekommendationssystem (RS). För nya användare har dock inte systemet den information som behövs för att kunna skapa anpassade rekommendationer, detta problem kallas för kallstart. För att bemöta nya användare i kallstart är RS beroende av deras explicita feedback. Konversationsrekommendationsystem (KRS) möjliggör explicit insamling från användare genom konversation. Explicit information kan lämnas av användare direkt i e-butiken. Informationen kan bestå av egenskaper användare föredrar hos produkter men även mer personlig information. Har inte användare förtroende till systemet kan det bidra till att användare inte vill dela med sig av information. För att användare ska vilja lämna explicit information kan KRS bemöta kallstart med att försöka skapa en känsla av förtroende hos användare. Studien är utförd med en designorienterad forskningsansats. Designelement identifierades genom en litteraturstudie som sedan implementerades i en prototyp för att undersöka om dessa kunde skapa en känsla av förtroende för systemet. Prototypen utvärderades därefter utifrån kriterierna trovärdighet, enkel användning och risk som fångar detta förtroende. Resultatet i denna studie indikerar att användare vill få anpassade rekommendationer presenterade och är därför villiga att bidra med den information systemet behöver. Studien presenterar åtta designförslag för hur KRS kan designas för att skapa en känsla av förtroende hos användare med en dialog mellan systemet och användare, möjlighet att ge effektiv feedback och systemets transparens.