推薦系統目的在於收集大眾對於資源的評價，經過推薦機制的整理，呈現給符合偏好的使用者，但是Supernode架構與傳統主從架構的不同，增加推薦系統收集、統計推薦指標的困難。在本篇論文中，將利用RDF(S)建置與Supernode、Peer、MP3相關的本體論，同時，依據Supernode與Peer在推薦系統中所扮演的功能與角色的不同，將不同的結構藍圖分別放置在Supernode與Peer之中，並且藉由Supernode與Peer中所描述的資源、屬性以及資源和資源之間的關係來解決在Supernode架構中推薦指標的收集、統計與更新等問題，以及降低推薦系統中"Cold Start"與"Spamming"這兩個問題所產生的影響。 / The purpose of recommendation system is to collect public opinions and provide synthetic recommended result for users. Using our proposed recommended mechanism, the recommendation system presents personal music recommended list for each user in peer node to match his preference. Because the complexity difference of network structure, we found that the rating collection and computation process will be more complex in supernode P2P network than in client-server WWW network. In this thesis, firstly we use RDF(S) to construct associated ontology schema for supernode, peer node, and MP3 music. Then, we put respective ontology schema in the suprenode and peer node based on their roles and functions within the recommendation system. The resources and attributes relationships between supernode and peer nodes provide the capacity to solve the rating collection, analysis, and updating problems. We also propose possible solutions for reducing the side effects of cold start and spamming problems in the recommendation system.

在西洋棋中，開局決定了棋局未來發展的基礎，棋手在開局階段局勢的好壞，會直接影響到接下來中局的發展，乃至全局的勝負。隨著西洋棋的演進，棋手們在比賽中進行各式各樣的棋步嘗試，發展出眾多經歷實戰考驗的開局，目前西洋棋的開局多達上千種變化，使得棋手在學習西洋棋的過程中，要花上大量的時間從眾多的西洋棋開局變化中，尋找適合自己的開局鑽研與使用。為幫助棋手在此階段的學習，本論文提出西洋棋開局推薦系統，從大數據協助學習的觀點，利用大量棋手們的開局經驗，對棋手做個人化的開局推薦。此系統以風格、棋力相似的棋手們所選用的開局為推薦基礎，並考量棋手習慣使用的下棋模式，推薦棋手善於發揮自身優勢、易於理解，並且投其所好的開局。為此，此西洋棋開局推薦系統包含風格分析、棋力評估、棋形截取，以及混合式推薦等部分。依據棋手過去的對局記錄，風格分析評估棋手下棋偏好冒險或保守的程度；棋力評估將傳統西洋棋棋力轉成可直觀比較棋手棋力程度差異之量表；棋形截取找出棋手習慣使用的下棋模式。最後，混合式推薦綜合考量上述三項因素，推薦出符合棋手棋風、棋力與下棋習慣模式的開局。 本論文以兩個實驗來評估風格分析與開局推薦系統的效果，在風格分析的實驗中，將風格分析方法評估棋手風格的結果與專家判斷的結果做比較；在開局推薦系統的實驗中，以棋手是否將會在比賽使用系統所推薦的開局來評估推薦效果。實驗結果顯示，風格分析對於世界冠軍棋手的風格評估幾乎與專家的判斷相同；開局推薦系統針對開局所設計的混合式推薦方法，推薦效果優於常見的推薦方法。 / The Opening is the fundamental phase of a chess game, and significantly affects the result of a competition. With the evolution of chess, there has been developed thousands of chess openings at present. This makes it difficult and time-consuming for chess players to find and learn the openings suitable for them. For helping players to learn chess in the opening, we provide Opening Recommendation System (OPRS), which considers chess players’ experiences and recommends chess openings that could be understandable and favorite for the players. For personalized recommendation, OPRS analyzes the playing style, translates chess rating, extracts the playing patterns, and then performs hybrid recommendation based on the features obtained. In the evaluation, the performance of the playing style analysis are demonstrated by comparing with the styles judged by chess experts for world chess championships. For OPRS, the evaluations are according to the openings the players use in the chess tournaments in the next years. The experiments show that OPRS achieves good accuracies of the playing style analysis and outperforms the competitive methods for chess opening recommendation.

近年來，智慧型手機(Smartphone)的銷量超過其他型式手機。智慧型手機具有更先進、更開放的行動作業系統，可允許使用者自行安裝應用程式軟體(Application)來擴充手機功能。目前市面上的應用程式數量非常龐大，在眾多的應用程式和有限的時間下，使用者不太可能將所有的應用程式下載試用，所以對使用者而言，找出自己所想要和需要的應用程式，是個困難的問題。推薦系統可依照使用者的喜好，或是準備推薦項目的相似程度來做推薦，讓使用者能較快得到想要的資訊，目前主要的方式有協同過濾(Collaborative Filtering, CF)、內容過濾(Content-Based Filtering, CBF)，還有結合前述兩種方式的混和式推薦(Hybrid Approach)。 本研究所使用的資料集是由政治大學資訊科學系所開發的實驗平台蒐集而來。資料以側錄的方式，將使用者實際操作手機應用程式的狀況記錄下來，其中包含了25位使用者和1125個應用程式。我們將原始資料集以三種方式整理成三個資料集：一、是否使用應用程式；二、使用應用程式的次數；三、使用應用程式的頻率，其值表示使用者在該應用程式的使用狀況。我們並將資料分成前段與後段時間兩部分，以前段時間的資料當作基準，推薦最多同群使用者使用的應用程式、同群使用者使用次數最多的應用程式，以及同群使用者最常使用的應用程式，然後以後段時間的資料做驗證，計算推薦結果的準確率與召回率加以比較。 我們使用知名的Information Theoretic Co-Clustering Algorithm和兩種基於Minimum Squared Residue Co-Clustering Algorithm的演算法將使用者與應用程式分群，利用分群結果做計算，推薦應用程式給使用者。實驗發現三種演算法在第一個資料集的準確率與召回率表現最好，此資料集以0和1的值，來紀錄使用者在各應用程式的使用狀況。實驗比較三個演算法的結果，在大部分的情況之下，一個基於Minimum Squared Residue Co-Clustering Algorithm的演算法，給出的結果較好。 此外，我們也發現應用程式開發者將應用程式上架提供下載時，以個人主觀想法對該應用程式定義其分類，與我們利用雙分群方法，以使用者實際操作的情況將應用程式分類的結果有些差異，或許在Google Play的分類上可做調整。 本研究提出推薦系統的框架具有彈性，未來可以使用不同的雙分群演算法做分群，也能套用其他的推薦方式。 / With the rapid evolution of smartphone devices, tens of thousands applications have been supplied on online stores such as App Store (operated by Apple Inc.) and Google Play (operated by Google Inc.). Since there are many applications, recommending applications to users becomes an important topic. In this thesis, we present a framework for using a co-clustering algorithm to recommend applications to users. Recommendations are a part of everyday life. People usually rely on some external knowledge to make informed decisions about a particular artifact or action. Using recommender systems is one of general approaches that help people make decisions. There are three common types of recommender systems, namely collaborative filtering, content-based filtering, and hybrid recommender systems. In this thesis, we use the dataset that was collected by a tool developed by the Department of Computer Science at the National Chengchi University. It recorded the users’ behavior when they were using their smartphones. We transform the original dataset into three types of datasets: 1) indicating whether a user used an application; 2) indicating the number of uses made by a user for an application; 3) indicating the frequency of uses made by a user for an application. Furthermore, we divide each dataset into two parts: The first part containing data for the early time period is used as the recommending base, and the second part containing data for the late time period is used for verifying the results. We utilize three famous co-clustering algorithms, which are the Information Theoretic Co-Clustering Algorithm and two algorithms based on the Minimum Squared Residue Co-Clustering Algorithm, in the proposed framework. According to the clusters given by a co-clustering algorithm, we recommend top five applications to each user by referring to the maximum number of users, the maximum number of uses, and the most frequently used applications that are in the same cluster. We calculate the precision and recall values to compare the results. From the experimental results, we find that the best result corresponds to the first type of dataset and also that one of the algorithms based on the Minimum Squared Residue Co-Clustering Algorithm is better than the other two algorithms in terms of the precision and recall values. From the clusters of applications, we obtain some interesting insights into the categories of applications. The categories of applications are set by their developers, but the users may not totally agree with the settings. There might be space for improvement for the categories of applications on the online store. In the future, we can utilize different co-clustering algorithms and other recommended methods in the proposed framework.

推薦系統的出現是為了解決訊息過載的問題，其需求隨著科技的進步、網路的普及而增加，相關技術也越發多樣且成熟。廣泛應用於各領域的統計模型也在技術的行列中。 推薦系統的運作仰賴使用者偏好訊息，而使用者對項目所組成的偏好空間往往十分巨大且不平衡，統計上需要相對複雜的隨機效應模型或混合效應模型來描繪這樣的變數結構，且通常需要計算效率相對低的反覆疊代過程來估計模型參數。因此Perry（2014）、Gao & Owen（2016）先後提出以動差法處理階層線性模型與兩因子隨機效應模型，是一種犧牲統計效率換取計算效率的做法。 本研究便是採用統計模型中的隨機效應模型法，分別以最大概似法和動差法估計參數，與同為協同過濾技術觀點的矩陣分解法進行分析比較。透過預測準確度和運算效率兩個層面，來評估各演算法在MoiveLens這筆資料上的推薦表現。 根據試驗結果歸納出隨機效應模型法無論以什麼樣的參數估計方式，在預測準確度的表現上都不如矩陣分解法來得好；但以動差法估計參數在穩定度上與矩陣分解法的表現差不多，且在運算效率上好很多。 / The recommender system (RS) appeared to solve the problem of information overload. The demand of the RS has increased with the advancement of technology and the popularity of the Internet, and related techniques have become more diverse and mature. The statistical models widely used in various fields are also in the list of techniques. The operation of the RS relies on user preference information, and the space of users’ preference to items is often large and unbalanced. Statistically, relatively complex random effects models or mixed effects models are needed to describe such variable structures, and often require a large number of iterations to estimate model parameters. Perry (2014), Gao & Owen (2016) proposed using the moment-based method to deal with hierarchical linear models and two-factor random effects models, respectively, expressing an idea of sacrificing statistical efficiency in exchange for computational efficiency. In this study, we analyze and compare the random effects model, using the maximum likelihood method and the moment-based method to estimate the parameters with the matrix factorization. Through the prediction accuracy and computational efficiency to evaluate the performance of each algorithm on the MoiveLens data. According to the experiment results, the random effects model is not as good as the matrix factorization in terms of the prediction accuracy no matter what kind of parameter estimation method is used; however, the performance of the moment-based parameter estimation is consistent with the matrix factorization in terms of the prediction stability, and much better in terms of the efficiency.

根據2010年人力銀行調查54%的尋找人才是以「工作分析」為主要依據，可見人才的遴選仍以經歷為主要因素。而近年來社會網絡與推薦系統普為應用於人才之找尋。 本研究實際以某國立大學EMBA學員資料，以同學與同事關係建置一個人脈社會網絡之人才推薦系統。本系統能依據使用者所輸入之人才搜尋條件，藉由距離相似度之運算，找出最接近的所需人才，並依距離相似度排序。其次，本系統可由各成員學歷，工作經歷所在之產業別，以及在組織中任職之功能別，來呈現人才之專業輪廓（Professional Profile），以作為決策者在遴選人才之依據。並提供所有關係路徑，以利使用者可進一步的諮詢路徑上成員對於推薦人選之評價。 本研究針對該校EMBA學員共計2,121人，應用資料探勘中群集分析建立推薦系統，有別於一般以關鍵字比對的搜尋方式，能找出與使用者需求條件相似度高的人才；並藉由人脈社會網路路徑，幫助使用者藉由自身的人脈評估推薦的結果。最後，本研究並提出結論、建議以及未來研究方向。 / Abstract According to the Job Bank survey in 2010, about 54% recruiters who search for specialist is mainly based on job analysis. This research is based on Social Network and Rcommendation system to build a relationship between the students and the colleagues with the personnel social network contacts, thus, a specialist recommendation system is constructed. First the system can compute the dissimilarity between the conditions users input and the background of people, find out the closest result required by sorting of similarity. Secondly, the professional profiles is established by the education background and work experiences (contain the various industries and position type), to serve as the basis for decision-makers in the selection of specialist. Besides, they can also inquire people from social network path for further appraisals of the candidate. The research is based on EMBA students totaled 2121 people, applying cluster analysis of data mining to build up the recommendation system, opposite to using key-word matching as a way to search people. Thus, the study can find the highest similar conditions demand of input. Via the associated social networks paths, to help users identify and use their own network to assess the recommend candidates. Finally, this study proposes conclusions, recommendations and future research directions. Keywords: Social Network , Similarity , Professional Profile , Specialist Recommendation System , Social Network Path

隨著網路的發達,各式各樣的資訊和商品也在網路上充斥著,使用者尋找資訊或是上網購物時,有的網站有推薦系統(recommender system)能提供使用者相關資訊或商品。若推薦系統能夠讓消費者所搜尋的相關資訊或商品能夠符合他們的習性時,便能讓消費者增加對系統的信賴程度,因此系統是否能準確預測出使用者的偏好就成為一個重要的課題。本研究使用兩筆資料,並以相關研究的三篇文獻進行分析和比較。這三篇文獻分別為IRT模型法(IRT model-based method)、相關係數法(correlation-coefficient method)、以及矩陣分解法(matrix factorization)。 在經過一連串的實證分析後,歸納出以下結論: 1. 模型法在預測方面雖然精確度不如其他兩種方法來的好,但是模型有解釋變數之間的關係以及預測機率的圖表展示,因此這個方法仍有存在的價值。 2. 相關係數法容易因為評分稀疏性的問題而無法預測,建議可以搭配內容式推薦系統的運作方式協助推薦。 3. 矩陣分解法在預測上雖然比IRT模型法還好，但分量的數字只是一個最佳化的結果，實際上無法解釋這些分量和數字的意義。 / With the growth of the internet, websites are full of a variety of information and products. When users find the information or surf the internet to shopping, some websites provide users recommender system to find with which related. Hence, whether the recommender system can predict the users' preference is an important topic. This study used two data,which are "Mondo" and "MovieLens", and we used three related references to analyze and compare them. The three references are following: IRT model-based method, Correlation-coefficient method, and Matrix factorization. After the data analysis, we get the following conclusions: 1. IRT model-based method is worse then other methods in predicting, but it can explain the relationship of variables and display the graph of predicting probabilities. Hence this method still has it's value. 2. Correlation-coefficient method is hard to predict because of sparsity. We can connect it with content filtering approach. 3. Although matrix factorization is better then IRT model-based method in predicting, the vectors is a result of optimization. It may be hard to explain the meaning of the vectors.

本研究提出一創新的旅遊推薦服務系統，以意象模型作為旅客意象(包含自我意象和情感需求)、景點意象、以及中小企業所提供服務之意象在系統裡的一致性表達。以上所提及之利益關係人的意象會經由數個系統模組進行建立與管理，並演化以反映出意象擁有者在真實世界的狀態。除此之外，本系統為動態運行，強調旅遊產業裡各個利益關係人角色之間的互動關係。每當互動發生，相關意象模型會進行混合，演繹出額外的意象屬性，以進行意象模型之調整。另外，基於顏色與情緒可相互對應的相關研究，我們將色彩理論運用於意象媒合與意象混合模組之中，藉此為旅客推薦符合其情感需求的旅遊景點或服務。本研究所提出一系列基於意象衍伸的系統化方法，可被應用於各種不同的領域。我們相信本研究可以為其它領域之實務應用與學術探討帶來顯著的貢獻。 / This research presents a recommendation service system that considers the image as a uniform representation of tourist images (include self-image and emotional needs), destinations, and local SMEs. Images carried by each stakeholder roles are modeled and managed by several system modules, and they also evolve to reflect the real time situations of each entity. In addition, the system is dynamic in terms of its emphasis on the relationships among these roles. When interactions occur, image mixing will be conducted to derive extra image attributes for the adjustments of the images. Besides, since colors can be mapped onto emotions, we use colors to operate the image matching and mixing process to find good matches of destinations for the recommendation. This image related approach we proposed is domain-independent. We believe our method could contribute to other areas of practical applications and academic studies.

隨著資訊科技的發展，網際網路成為個人獲得資訊的主要來源之一。但是過多的資訊產生資訊爆炸(information overload)的現象，人們除了要在眾多資訊中找尋想要的資訊外，還需要擔心所尋找到的資訊的品質是否良好。因此，推薦系統提供了一個良好的解決方法。推薦系統透過分群與推薦的技術來達到減少資訊量與推估使用者潛在興趣的目的。目前推薦系統多應用在單一維度的推薦，本論文希望藉由某一情境來探討多維度推薦的應用，所以選擇助理軟體來實現多維度推薦的應用。選擇助理軟體是由於其已經成為個人日常生活中時常使用的工具，且由於助理軟體管理個人日常生活中的大小事務，成為最貼近個人的工具。若專注在個人行事曆的安排上，我們可以發現個人行事曆安排牽涉到有人、事、時、地、物五個維度。因此我們以五維度做分群，透過合作推薦(Collaborative Recommender)的方式將可以達到個人潛在興趣的多維度(Multi-Dimensions)推薦。本研究將以行事歷排定為情境，來說明如何將五個維度的各種可能組合依照其契合個人興趣的程度來進行推薦，這將使得助理軟體的內容更加豐富，且能貼近使用者的需求，提供意想不到的資訊組合。 / With the development of information science and technology, assistant software becomes a tool which often uses in personal daily life, and because all kinds of affairs in personal daily life that assistant software is managed, so assistant software becomes a tool which personally close to people. Intelligent assistant software hopes to make assistant software have intelligence which is similar to the mankind. Just like a personal general secretary, arrange the most proper individualized journey. Further, it can combine the idea of Recommender system to recommend the journey of the potential interest while arranging in the personal journey. This research proposes an intelligent assistant software with five- dimensions include of people, thing, when, location and things, uses cooperative Recommender approach to reach multi-dimension recommendation of personal potential interest. This research will give example of meeting as the situation to explain how to make five-dimensions recommendation according to personal interest. This will make the content of assistant software more abundant, and can press close to the user's demand.

現今許多網路商店或電子商務將產品銷售給消費者的過程中，皆使用推薦系統的幫助來提高銷售量。如亞馬遜公司(Amazon)、Netflix，深入了解顧客的使用習慣，建構專屬的推薦系統並進行個性化的推薦商品給每一位顧客。 推薦系統應用的技術分為協同過濾和內容過濾兩大類，本研究旨在探討協同過濾推薦系統中潛在因子模型方法，利用矩陣分解法找出評分矩陣。在Koren等人(2009)中，將矩陣分解法的演算法大致分為兩種，隨機梯度下降法(Stochastic gradient descent)與交替最小平方法(Alternating least squares)。本研究主要研究目的有三項，一為比較交替最小平方法與隨機梯度下降法的預測能力，二為兩種矩陣分解演算法在加入偏誤項後的表現，三為先完成交替最小平方法與隨機梯度下降法，以其預測值對原始資料之遺失值進行資料插補，再利用奇異值分解法對完整資料做矩陣分解，觀察其前後方法的差異。 研究結果顯示，隨機梯度下降法所需的運算時間比交替最小平方法所需的運算時間少。另外，完成兩種矩陣分解演算法後，將預測值插補遺失值，進行奇異值分解的結果也顯示預測能力有提升。 / Recommender system has been largely used by Internet companies such Amazon and Netflix to make recommendations for Internet users. Techniques for recommender systems can be divided into content filtering approach and collaborative filtering approach. Matrix factorization is a popular method for collaborative filtering approach. It minimizes the object function through stochastic gradient descent and alternating least squares. This thesis has three goals. First, we compare the alternating least squares method and stochastic gradient descent method. Secondly, we compare the performance of matrix factorization method with and without the bias term. Thirdly, we combine singular value decomposition and matrix factorization. As expected, we found the stochastic gradient descent takes less time than the alternating least squares method, and the the matrix factorization method with bias term gives more accurate prediction. We also found that combining singular value decomposition with matrix factorization can improve the predictive accuracy.

「人」是公司中最重要的資產，而在知識密集的行業中，這樣的資產更顯得重要。由於網路技術的出現，網路人力銀行也成為另外一種人才招募的新興管道，但透過網路人力銀行所召募的人才素質並沒有傳統上透過公司員工推薦進來的人才可更進一步瞭解的好處。因此本研究透過一網路人才推薦信任制度，來加強線上人力銀行之人才篩選能力，希望透過此制度能繼續保有網路人力銀行在人才招募速度上的優勢，並能加強其篩選的能力。 本研究針對人才招募管道進行了文獻的探討，提出一人才推薦制度，以某國立大學EMBA之人才庫，透過成員間的學經歷背景相似度，推薦出擁有相同顯性工作能力的人才。讓人才招募單位可以得到推薦的人才，並可對其作信任評價的推論。接著利用實驗來求出雛型系統的一些關鍵參數，讓雛型系統運作得更完善以及更符合使用者的需求。 本雛型系統結合了網路人力銀行人才招募方式可快速地招募到大量員工的特點，及員工推薦人才招募方式可招募到更適切員工的特點。並透過FOAF格式的使用，將線上社會網絡的資料格式統一，有助於縮短整個人才信任推薦系統的建立時間。 / "Human Resource" is one of the most important assets of company, especially in knowledge-intensive industries. As network technologies developed, commercial job site has also become another kind of recruitment channel. But through this kind of channel, companies don’t have better chance to know new employee than traditional way. Therefore this study filters new employees by a Recommendation & Trust Inference mechanism. Hope that commercial job site would continue to keep the advantages of high efficiency in recruitment, and enhance its filtering capability at the same time. First, this study surveys literatures in recruitment channels. And it proposes a Recommendation & Trust Inference mechanism using a national university EMBA program member data as an example. The Recommendation mechanism recommend candidates having the same specialty by comparing their similarity of education and work experience. Furthermore, recruitment unit could use Trust Inference mechanism to get suitable candidates. Third, we conduct experiments to find the key parameters for the prototype system. Make the system able to work better and meet users’ needs. The prototype system combines the benefit of commercial job site which can quickly recruit a large number of employees and the feature providing more appropriate candidates for the company recommended by staff. Simultaneously by taking use of the FOAF format, we can unify the data format in online social network. The way mentioned above will effectively reduce the system set-up time.