In this project we investigate the use of artificial neural networks(ANNs) as the core prediction function of a recommender system. In the past, research concerned with recommender systems that use ANNs have mainly concentrated on using collaborative-based information. We look at the effects of adding content-based information and how altering the topology of the network itself affects the accuracy of the recommendations generated. In particular, we investigate a mixture of experts topology. We create two expert clusters in the hidden layer of the ANN, one for content-based data and another for collaborative-based data. This greatly reduces the number of connections between the input and hidden layers. Our experimental evaluation shows that this new architecture produces the same accuracy of recommendation as the fully connected configuration with a large decrease in the amount of time it takes to train the network. This decrease in time is a great advantage because of the need for recommender systems to provide real time results to the user.

Recommender systems (RSs) provide personalised suggestions of information or products relevant to user's needs. RSs are considered as powerful tools that help users to find interesting items matching their own taste. Although RSs have made substantial progress in theory and algorithm development and have achieved many commercial successes, how to utilise the widely available information on Online Social Networks (OSNs) has largely been overlooked. Noticing this gap in existing research on RSs and taking into account a user's selection being greatly influenced by his/her trusted friends and their opinions, this thesis proposes a novel personalised Recommender System framework, so-called Implicit Social Trust and Sentiment (ISTS) based RSs. The main motivation was to overcome the overlooked use of OSNs in Recommender Systems and to utilise the widely available information from such networks. This work also designs solutions to a number of challenges inherent to the RSs domain, such as accuracy, cold-start, diversity and coverage. ISTS improves the existing recommendation approaches by exploring a new source of data from friends' short posts in microbloggings. In the case of new users who have no previous preferences, ISTS maps the suggested recommendations into numerical rating scales by applying the three main components. The first component is measuring the implicit trust between friends based on their intercommunication activities and behaviour. Owing to the need to adapt friends' opinions, the implicit social trust model is designed to include the trusted friends and give them the highest weight of contribution in recommendation encounter. The second component is inferring the sentiment rating to reflect the knowledge behind friends' short posts, so-called micro-reviews. The sentiment behind micro-reviews is extracted using Sentiment Analysis (SA) techniques. To achieve the best sentiment representation, our approach considers the special natural environment in OSNs brief posts. Two Sentiment Analysis methodologies are used: a bag of words method and a probabilistic method. The third ISTS component is identifying the impact degree of friends' sentiments and their level of trust by using machine learning algorithms. Two types of machine learning algorithms are used: classification models and regressions models. The classification models include Naive Bayes, Logistic Regression and Decision Trees. Among the three classification models, Decision Trees show the best Mean absolute error (MAE) at 0.836. Support Vector Regression performed the best among all models at 0.45 of MAE. This thesis also proposes an approach with further improvement over ISTS, namely Hybrid Implicit Social Trust and Sentiment (H-ISTS). The enhanced approach applies improvements by optimising trust parameters to identify the impact of the features (re-tweets and followings/followers list) on recommendation results. Unlike the ISTS which allocates equal weight to trust features, H-ISTS provides different weights to determine the different effects of the two trust features. As a result, we found that H-ISTS improved the MAE to be 0.42 which is based on Support Vector Regression. Further, it increases the number of trust features from two to five features in order to include the influence of these features in rating predictions. The integration of the new approach H-ISTS with a Collaborative Filtering recommender system, in particular memory-based, is investigated next. Therefore, existing users with a history of ratings can receive recommendations by fusing their own tastes and their friends' preferences using the two type of memory-based methods: user-based and item-based. H-ISTSitem is the integration of H-ISTS and item-based which provides the lowest error at 0.7091. The experiments show that diversity is better achieved using the H-ISTSuser which is the integration of H-ISTS and user-based technique. To evaluate the performance of these approaches, two real social datasets are collected from Twitter. To verify the proposed framework, the experiments are conducted and the results are compared against the most relevant baselines which confirmed that RSs have been successfully improved using OSNs. These enhancements demonstrate the effectiveness and promises of the proposed approach in RSs.

In this thesis, I propose a method for establishing a personalized recommendation system for re-ranking web search results and recommending web contents. The method is based on personal reading interest which can be reflected by the user’s dwell time on each document or webpage. I acquire document-level dwell times via a customized web browser, or a mobile device. To obtain better precision, I also explore the possibility of tracking gaze position and facial expression, from which I can determine the attractiveness of different parts of a document. Inspired by idea of Google Knowledge Graph, I also establish a graph-based ontology to maintain a user profile to describe the user’s personal reading interest. Each node in the graph is a concept, which represents the user’s potential interest on this concept. I also use the dwell time to measure concept-level interest, which can be inferred from document-level user dwell times. The graph is generated based on the Wikipedia. According to the estimated concept-level user interest, my algorithm can estimate a user’s potential dwell time over a new document, based on which personalized webpage re-ranking can be carried out. I compare the rankings produced by my algorithm with rankings generated by popular commercial search engines and a recently proposed personalized ranking algorithm. The results clearly show the superiority of my method. I also use my personalized recommendation framework in other applications. A good example is personalized document summarization. The same knowledge graph is employed to estimate the weight of every word in a document; combining with a traditional document summarization algorithm which focused on text mining, I could generate a personalized summary which emphasize the user’s interest in the document. To deal with images and videos, I present a new image search and ranking algorithm for retrieving unannotated images by collaboratively mining online search results, which consists of online images and text search results. The online image search results are leveraged as reference examples to perform content-based image search over unannotated images. The online text search results are used to estimate individual reference images’ relevance to the search query as not all the online image search results are closely related to the query. Overall, the key contribution of my method lies in its ability to deal with unreliable online image search results through jointly mining visual and textual aspects of online search results. Through such collaborative mining, my algorithm infers the relevance of an online search result image to a text query. Once I estimate a query relevance score for each online image search result, I can selectively use query specific online search result images as reference examples for retrieving and ranking unannotated images. To explore the performance of my algorithm, I tested it both on a standard public image datasets and several modestly sized personal photo collections. I also compared the performance of my method with that of two peer methods. The results are very positive, which indicate that my algorithm is superior to existing content-based image search algorithms for retrieving and ranking unannotated images. Overall, the main advantage of my algorithm comes from its collaborative mining over online search results both in the visual and the textual domains. / published_or_final_version / Computer Science / Doctoral / Doctor of Philosophy

Since the Netflix Prize competition, latent factor models (LFMs) have become the comparison ``staples'' for many of the recent recommender methods. Meanwhile, it is still unclear to understand the impact of data preprocessing and updating algorithms on LFMs. The performance improvement of LFMs over baseline approaches, however, hovers at only low percentage numbers. Therefore, it is time for a better understanding of their real power beyond the overall root mean square error (RMSE), which as it happens, lies at a very compressed range, without providing too much chance for deeper insight. We introduce an experiment based handbook of LFMs and reveal data preprocessing and updating algorithms' power. We perform a detailed experimental study regarding the performance of classical staple LFMs on a classical dataset, Movielens 1M, that sheds light on a much more pronounced excellence of LFMs for particular categories of users and items, for RMSE and other measures. In particular, LFMs exhibit surprising and excellent advantages when handling several difficult user and item categories. By comparing the distributions of test ratings and predicted ratings, we show that the performance of LFMs is influenced by rating distribution. We then propose a method to estimate the performance of LFMs for a given rating dataset. Also, we provide a very simple, open-source library that implements staple LFMs achieving a similar performance as some very recent (2013) developments in LFMs, and at the same time being more transparent than some other libraries in wide use. / Graduate

Thesis (MCIS)--AUT University, 2009. / Includes bibliographical references. Also held in print ( leaves : ill. ; 30 cm.) in the Archive at the City Campus (T 006.312 CRI)

Recommender systems (RS) have become increasingly popular in many web applications for eliminating online information overload and making personalized suggestions to users. In recent years, user personality has been recognized as valuable info to build more personalized recommender systems. However, the existing personality-based recommender systems has mainly focused on revealing the impact of personality on the user's preference over a single item or an attribute, which may ignore the impact of personality on users' perceptions of recommender systems when multiple recommendations are returned at the same time. In addition, they have mostly relied on personality quiz to explicitly acquire users' personality, which unavoidably demands user efforts. From users' perspective, they may be unwilling to answer the quiz for the sake of saving efforts or protecting their privacy. The application of existing personality-based recommender systems will thus be limited in real life.;In this thesis, we aim at 1) incorporating personality into top-N (N > 1) recommendations, with emphases on personalizing recommendation diversity and improving the recommendation interface design, 2) deriving users' personality from their implicit behavior for augmenting the existing recommender systems.;Specifically, we first develop a generalized, dynamic diversity adjusting approach based on user personality with the goal of achieving personalized diversity tailored to individual users' intrinsic needs. In particular, personality is integrated into a greedy re-ranking process, by which we select the item that can best balance accuracy and personalized diversity at each step, and then produce the final recommendation list. In this approach, personality is both used to estimate each user's diversity preference and to alleviate the cold-start problem of collaborative filtering recommendations. The experimental results demonstrate that our personalized diversity-oriented approach significantly outperforms related methods (including both non-diversity-oriented and diversity-oriented methods) in terms of both accuracy and diversity metrics, especially in the cold-start setting.;In addition to the algorithm development, designing diversity-oriented interface has been proven helpful to augment users' perception of recommendation diversity. However, little work has been done to identify the impact of users' personality on their preference for different types of recommendation interfaces (e.g., the diversity-oriented interface and the non-diversity-oriented interface). In order to fill the gap, we conduct a within-subject user study. We concretely compare a diversity-oriented organization-based recommendation interface with the standard ranked list interface covering three product domains with different investment levels and users' purchase experiences (i.e., mobile phone, hotel and movie). We find that users' perceptions of different recommendation interface are influenced by the product types. More notably, we identify the important role of users' personality in influencing their preference for recommendation interfaces. For instance, introverted users tend to reuse the organization-based interface in the future than the standard ranked list. The results can hence be constructive for improving existing recommendation interface design by considering users' personality.;Although personality has been proven effective at enhancing the multiple recommendations, the effort of explicitly acquiring users' personality traits via psychological questionnaire is unavoidably high, which may impede the application of personality-based recommenders in real life. We hence propose a generalized method to derive users' personality from their implicit behavior and further improve the existing recommender systems. A preliminary experiment has been conducted in movie domain. More specifically, we first identify a set of behavioral features through experimental validation, and develop inference model based on Gaussian Process to unify these features for determining users' big-five personality traits. We then test the model in a collaborative filtering based recommending framework on two real-life movie datasets, which demonstrates that our implicit personality based recommending algorithm significantly outperforms related methods in terms of both rating prediction and ranking accuracy. The experimental results point out an effective solution to boost the applicability of personality-based recommender systems in online environment.

M.Sc. (Information Technology) / Recommender systems assist a system user to identify relevant content within a specific context. This is typically performed through an analysis of a system user’s rating habits and personal preferences and leveraging these to return one or a number of relevant recommendations. There are numerable contexts in which recommender systems can be applied, such as movies, tourism, books, and music. The need for recommender systems has become increasingly relevant, particularly on the Internet. This is mainly due to the exponential amount of content that is published online on a daily basis. It has thus become more time consuming and difficult to find pertinent information online, leading to information overload. The relevance of a recommender system, therefore, is to assist a system user to overcome the information overload problem by identifying pertinent information on their behalf. There has been much research done within the recommender system field and how such systems can best recommend items to an individual user. However, a growing and more recent research area is how recommender systems can be extended to recommend items to groups, known as group recommendation. The relevance of group recommendation is that many contexts of recommendation apply to both individuals and groups. For example, people often watch movies or visit tourist attractions as part of a group. Group recommendation is an inherently more complex form of recommendation than individual recommendation for a number of reasons. The first reason is that the rating habits and personal preferences of each system user within the group need to be considered. Additionally, these rating habits and personal preferences can be quite heterogeneous in nature. Therefore, group recommendation becomes complex because a satisfactory recommendation needs to be one which meets the preferences of each group member and not just a single group member. The second reason why group recommendation is considered to be more complex than individual recommendation is because a group not only includes multiple personal preferences, but also multiple personality types. This means that a group is more complex from a social perspective. Therefore, a satisfactory group recommendation needs to be one which considers the varying personality types and behaviours of the group. The purpose of this research is to present PerTrust, a generic framework for group recommendation with the purpose of providing a possible solution to the aforementioned issues noted above. The primary focus of PerTrust is how to leverage both personality and trust in overcoming these issues.

By now, people are accustomed to getting some personalized recommendations when they are finding movies to watch, music to listen, and so on. All of these recommendations come from recommender systems, and can aid the process of the decision making to avoid the problem of "information overload". Over the years, there has been much work done both in industry and academia on developing new approaches for recommender systems. However, there are still some hurdles in adapting recommender systems to a broader range of real-life applications. In the e-commerce environment especially with the so called high-risk products (also called high-cost or high-involvement products, such as digital cameras, computers, and cars), because a user does not buy the high-risk product very often, it is normal that s/he is not able to rate many products. For the same reason, the current buyer is often a new user because s/he would not afford to buy the same kind of high-risk product before. The traditional recommender techniques (such as user-based collaborative filtering and content-based methods) can thus not be effectively applicable in this environment, because they largely assume that the users have prior experiences with products. Thus, the "data sparsity" and "new users" are two typical challenging issues that the classical recommender systems cannot well address in high-risk product domains. In some recommender systems, a new user will be asked to indicate his/her preferences on some aspects in order to address the so called cold-start problem via collecting some preferences. Such collected preferences are usually not complete due to unfamilaring with the product domain, which are called partial preferences.;In this thesis, we propose to leverage some auxiliary data of online reviewers' opinions, so as to enrich the partial preferences. With the objective of developing more effective recommender systems for high-risk products in e-commerce, in our work, we have exerted to derive reviewers' preferences from the textual reviews they posted. Then, these recovered preferences are leveraged to estimate and supplement a new buyer's preference with which the product recommendation is produced. Firstly, we propose a novel clustering method based on Latent Class Regression model (LCRM), which is able to consider both the overall ratings and feature-level opinion values (as extracted from textual reviews) to infer individual reviewers' weight feature preferences, that represent the weights the user places on different product features. Secondly, we propose a method to estimate reviewers' value preferences (i.e., the user's preferences on the product's attribute values) by matching their review opinions to the corresponding attributes' static specifications. Thirdly, we investigate how to combine weight preferences and value preferences to model user preferences based on Multi-Attribute Utility Theory (MAUT) with the purpose of providing higher quality product recommendations. Particularly, it was shown from our experimental studies that the incorporation of review information can significantly enhance the recommendation accuracy, relative to those without considering reviews. As the practical implication, our proposed solutions can be usefully plugged into an online system to be adopted in real-ecommerce sites.

Personalized recommender systems aim to assist users in retrieving and accessing interesting items by automatically acquiring user preferences from the historical data and matching items with the preferences. In the last decade, recommendation services have gained great attention due to the problem of information overload. However, despite recent advances of personalization techniques, several critical issues in modern recommender systems have not been well studied. These issues include: (1) understanding the accessing patterns of users (i.e., how to effectively model users' accessing behaviors); (2) understanding the relations between users and other objects (i.e., how to comprehensively assess the complex correlations between users and entities in recommender systems); and (3) understanding the interest change of users (i.e., how to adaptively capture users' preference drift over time). To meet the needs of users in modern recommender systems, it is imperative to provide solutions to address the aforementioned issues and apply the solutions to real-world applications. The major goal of this dissertation is to provide integrated recommendation approaches to tackle the challenges of the current generation of recommender systems. In particular, three user-oriented aspects of recommendation techniques were studied, including understanding accessing patterns, understanding complex relations and understanding temporal dynamics. To this end, we made three research contributions. First, we presented various personalized user profiling algorithms to capture click behaviors of users from both coarse- and fine-grained granularities; second, we proposed graph-based recommendation models to describe the complex correlations in a recommender system; third, we studied temporal recommendation approaches in order to capture the preference changes of users, by considering both long-term and short-term user profiles. In addition, a versatile recommendation framework was proposed, in which the proposed recommendation techniques were seamlessly integrated. Different evaluation criteria were implemented in this framework for evaluating recommendation techniques in real-world recommendation applications. In summary, the frequent changes of user interests and item repository lead to a series of user-centric challenges that are not well addressed in the current generation of recommender systems. My work proposed reasonable solutions to these challenges and provided insights on how to address these challenges using a simple yet effective recommendation framework.

Indiana University-Purdue University Indianapolis (IUPUI) / Research on social networks and understanding the interactions of the users can be modeled as a task of graph mining, such as predicting nodes and edges in networks. Dealing with such unstructured data in large social networks has been a challenge for researchers in several years. Neural Networks have recently proven very successful in performing predictions on number of speech, image, and text data and have become the de facto method when dealing with such data in a large volume. Graph NeuralNetworks, however, have only recently become mature enough to be used in real large-scale graph prediction tasks, and require proper structure and data modeling to be viable and successful. In this research, we provide a new modeling of the social network which captures the attributes of the nodes from various dimensions. We also introduce the Neural Network architecture that is required for optimally utilizing the new data structure. Finally, in order to provide a hot-start for our model, we initialize the weights of the neural network using a pre-trained graph embedding method. We have also developed a new graph embedding algorithm. We will first explain how previous graph embedding methods are not optimal for all types of graphs, and then provide a solution on how to combat those limitations and come up with a new graph embedding method.