Do-Rank: DCG optimization for learning-to-rank in tag-based item recommendation systems

Discounted Cumulative Gain (DCG) is a well-known ranking evaluation measure for models built with multiple relevance graded data. By handling tagging data used in recommendation systems as an ordinal relevance set of {negative,null,positive}, we propose to build a DCG based recommendation model. We present an efficient and novel learning-to-rank method by optimizing DCG for a recommendation model using the tagging data interpretation scheme. Evaluating the proposed method on real-world datasets, we demonstrate that the method is scalable and outperforms the benchmarking methods by generating a quality top-N item recommendation list.

Collaborative Filtering with Data Classification : A Combined Approach to Hotel Recommendation Systems

Recommendation systems aim to help users make decisions more efficiently. The most widely used method in recommendation systems is collaborative filtering, of which, a critical step is to analyze a user's preferences and make recommendations of products or services based on similarity analysis with other users' ratings. However, collaborative filtering is less usable for recommendation facing the "cold start" problem, i.e. few comments being given to products or services. To tackle this problem, we propose an improved method that combines collaborative filtering and data classification. We use hotel recommendation data to test the proposed method. The accuracy of the recommendation is determined by the rankings. Evaluations regarding the accuracies of Top-3 and Top-10 recommendation lists using the 10-fold cross-validation method and ROC curves are conducted. The results show that the Top-3 hotel recommendation list proposed by the combined method has the superiority of the recommendation performance than the Top-10 list under the cold start condition in most of the times.

'Movie Recommendation Systems Using An Artificial Immune System'

We apply the Artificial Immune System (AIS)technology to the collaborative Filtering (CF)technology when we build the movie recommendation system. Two different affinity measure algorithms of AIS, Kendall tau and Weighted Kappa, are used to calculate the correlation coefficients for this movie recommendation system. From the testing we think that Weighted Kappa is more suitable than Kendall tau for movie problems.

'Movie Recommendation Systems Using An Artificial Immune System'

We apply the Artificial Immune System (AIS)technology to the collaborative Filtering (CF)technology when we build the movie recommendation system. Two different affinity measure algorithms of AIS, Kendall tau and Weighted Kappa, are used to calculate the correlation coefficients for this movie recommendation system. From the testing we think that Weighted Kappa is more suitable than Kendall tau for movie problems.

A COMPARATIVE STUDY AND EVALUATION OF COLLABORATIVE RECOMMENDATION SYSTEMS

Consumers currently enjoy a surplus of goods (books, videos, music, or other items) available to purchase. While this surplus often allows a consumer to find a product tailored to their preferences or needs, the volume of items available may require considerable time or effort on the part of the user to find the most relevant item. Recommendation systems have become a common part of many online business that supply users books, videos, music, or other items to consumers. These systems attempt to provide assistance to consumers in finding the items that fit their preferences. This report presents an overview of recommendation systems. We will also briefly explore the history of recommendation systems and the large boost that was given to research in this field due to the Netflix Challenge. The classical methods for collaborative recommendation systems are reviewed and implemented, and an examination is performed contrasting the complexity and performance among the various models. Finally, current challenges and approaches are discussed.

A people-to-people recommendation system using Tensor Space Models

Existing recommendation systems often recommend products to users by capturing the item-to-item and user-to-user similarity measures. These types of recommendation systems become inefficient in people-to-people networks for people to people recommendation that require two way relationship. Also, existing recommendation methods use traditional two dimensional models to find inter relationships between alike users and items. It is not efficient enough to model the people-to-people network with two-dimensional models as the latent correlations between the people and their attributes are not utilized. In this paper, we propose a novel tensor decomposition-based recommendation method for recommending people-to-people based on users profiles and their interactions. The people-to-people network data is multi-dimensional data which when modeled using vector based methods tend to result in information loss as they capture either the interactions or the attributes of the users but not both the information. This paper utilizes tensor models that have the ability to correlate and find latent relationships between similar users based on both information, user interactions and user attributes, in order to generate recommendations. Empirical analysis is conducted on a real-life online dating dataset. As demonstrated in results, the use of tensor modeling and decomposition has enabled the identification of latent correlations between people based on their attributes and interactions in the network and quality recommendations have been derived using the 'alike' users concept.

A tag-based personalized item recommendation system using tensor modeling and topic model approaches

This research falls in the area of enhancing the quality of tag-based item recommendation systems. It aims to achieve this by employing a multi-dimensional user profile approach and by analyzing the semantic aspects of tags. Tag-based recommender systems have two characteristics that need to be carefully studied in order to build a reliable system. Firstly, the multi-dimensional correlation, called as tag assignment , should be appropriately modelled in order to create the user profiles [1]. Secondly, the semantics behind the tags should be considered properly as the flexibility with their design can cause semantic problems such as synonymy and polysemy [2]. This research proposes to address these two challenges for building a tag-based item recommendation system by employing tensor modeling as the multi-dimensional user profile approach, and the topic model as the semantic analysis approach. The first objective is to optimize the tensor model reconstruction and to improve the model performance in generating quality rec-ommendation. A novel Tensor-based Recommendation using Probabilistic Ranking (TRPR) method [3] has been developed. Results show this method to be scalable for large datasets and outperforming the benchmarking methods in terms of accuracy. The memory efficient loop implements the n-mode block-striped (matrix) product for tensor reconstruction as an approximation of the initial tensor. The probabilistic ranking calculates the probabil-ity of users to select candidate items using their tag preference list based on the entries generated from the reconstructed tensor. The second objective is to analyse the tag semantics and utilize the outcome in building the tensor model. This research proposes to investigate the problem using topic model approach to keep the tags nature as the “social vocabulary” [4]. For the tag assignment data, topics can be generated from the occurrences of tags given for an item. However there is only limited amount of tags availa-ble to represent items as collection of topics, since an item might have only been tagged by using several tags. Consequently, the generated topics might not able to represent the items appropriately. Furthermore, given that each tag can belong to any topics with various probability scores, the occurrence of tags cannot simply be mapped by the topics to build the tensor model. A standard weighting technique will not appropriately calculate the value of tagging activity since it will define the context of an item using a tag instead of a topic.

Towards user profiling for web recommendation

Collaborative recommendation is one of widely used recommendation systems, which recommend items to visitor on a basis of referring other's preference that is similar to current user. User profiling technique upon Web transaction data is able to capture such informative knowledge of user task or interest. With the discovered usage pattern information, it is likely to recommend Web users more preferred content or customize the Web presentation to visitors via collaborative recommendation. In addition, it is helpful to identify the underlying relationships among Web users, items as well as latent tasks during Web mining period. In this paper, we propose a Web recommendation framework based on user profiling technique. In this approach, we employ Probabilistic Latent Semantic Analysis (PLSA) to model the co-occurrence activities and develop a modified k-means clustering algorithm to build user profiles as the representatives of usage patterns. Moreover, the hidden task model is derived by characterizing the meaningful latent factor space. With the discovered user profiles, we then choose the most matched profile, which possesses the closely similar preference to current user and make collaborative recommendation based on the corresponding page weights appeared in the selected user profile. The preliminary experimental results performed on real world data sets show that the proposed approach is capable of making recommendation accurately and efficiently.

On the Incorporation of Psychologically-Driven 'Music' Preference Models for Music Recommendation

There are hundreds of millions of songs available to the public, necessitating the use of music recommendation systems to discover new music. Currently, such systems account for only the quantitative musical elements of songs, failing to consider aspects of human perception of music and alienating the listener’s individual preferences from recommendations. Our research investigated the relationships between perceptual elements of music, represented by the MUSIC model, with computational musical features generated through The Echo Nest, to determine how a psychological representation of music preference can be incorporated into recommendation systems to embody an individual’s music preferences. Our resultant model facilitates computation of MUSIC factors using The Echo Nest features, and can potentially be integrated into recommendation systems for improved performance.

Discovery & effective use of quality association rules in multi-level datasets

In today’s electronic world vast amounts of knowledge is stored within many datasets and databases. Often the default format of this data means that the knowledge within is not immediately accessible, but rather has to be mined and extracted. This requires automated tools and they need to be effective and efficient. Association rule mining is one approach to obtaining knowledge stored with datasets / databases which includes frequent patterns and association rules between the items / attributes of a dataset with varying levels of strength. However, this is also association rule mining’s downside; the number of rules that can be found is usually very big. In order to effectively use the association rules (and the knowledge within) the number of rules needs to be kept manageable, thus it is necessary to have a method to reduce the number of association rules. However, we do not want to lose knowledge through this process. Thus the idea of non-redundant association rule mining was born. A second issue with association rule mining is determining which ones are interesting. The standard approach has been to use support and confidence. But they have their limitations. Approaches which use information about the dataset’s structure to measure association rules are limited, but could yield useful association rules if tapped. Finally, while it is important to be able to get interesting association rules from a dataset in a manageable size, it is equally as important to be able to apply them in a practical way, where the knowledge they contain can be taken advantage of. Association rules show items / attributes that appear together frequently. Recommendation systems also look at patterns and items / attributes that occur together frequently in order to make a recommendation to a person. It should therefore be possible to bring the two together. In this thesis we look at these three issues and propose approaches to help. For discovering non-redundant rules we propose enhanced approaches to rule mining in multi-level datasets that will allow hierarchically redundant association rules to be identified and removed, without information loss. When it comes to discovering interesting association rules based on the dataset’s structure we propose three measures for use in multi-level datasets. Lastly, we propose and demonstrate an approach that allows for association rules to be practically and effectively used in a recommender system, while at the same time improving the recommender system’s performance. This especially becomes evident when looking at the user cold-start problem for a recommender system. In fact our proposal helps to solve this serious problem facing recommender systems.

A people-to-people matching system using graph mining techniques

A people-to-people matching system (or a match-making system) refers to a system in which users join with the objective of meeting other users with the common need. Some real-world examples of these systems are employer-employee (in job search networks), mentor-student (in university social networks), consume-to-consumer (in marketplaces) and male-female (in an online dating network). The network underlying in these systems consists of two groups of users, and the relationships between users need to be captured for developing an efficient match-making system. Most of the existing studies utilize information either about each of the users in isolation or their interaction separately, and develop recommender systems using the one form of information only. It is imperative to understand the linkages among the users in the network and use them in developing a match-making system. This study utilizes several social network analysis methods such as graph theory, small world phenomenon, centrality analysis, density analysis to gain insight into the entities and their relationships present in this network. This paper also proposes a new type of graph called “attributed bipartite graph”. By using these analyses and the proposed type of graph, an efficient hybrid recommender system is developed which generates recommendation for new users as well as shows improvement in accuracy over the baseline methods.

Mining sentiments from songs using latent dirichlet allocation

Song-selection and mood are interdependent. If we capture a song’s sentiment, we can determine the mood of the listener, which can serve as a basis for recommendation systems. Songs are generally classified according to genres, which don’t entirely reflect sentiments. Thus, we require an unsupervised scheme to mine them. Sentiments are classified into either two (positive/negative) or multiple (happy/angry/sad/...) classes, depending on the application. We are interested in analyzing the feelings invoked by a song, involving multi-class sentiments. To mine the hidden sentimental structure behind a song, in terms of “topics”, we consider its lyrics and use Latent Dirichlet Allocation (LDA). Each song is a mixture of moods. Topics mined by LDA can represent moods. Thus we get a scheme of collecting similar-mood songs. For validation, we use a dataset of songs containing 6 moods annotated by users of a particular website.

An economic approach to consumer product ratings

In three essays we examine user-generated product ratings with aggregation. While recommendation systems have been studied extensively, this simple type of recommendation system has been neglected, despite its prevalence in the field. We develop a novel theoretical model of user-generated ratings. This model improves upon previous work in three ways: it considers rational agents and allows them to abstain from rating when rating is costly; it incorporates rating aggregation (such as averaging ratings); and it considers the effect on rating strategies of multiple simultaneous raters. In the first essay we provide a partial characterization of equilibrium behavior. In the second essay we test this theoretical model in laboratory, and in the third we apply established behavioral models to the data generated in the lab. This study provides clues to the prevalence of extreme-valued ratings in field implementations. We show theoretically that in equilibrium, ratings distributions do not represent the value distributions of sincere ratings. Indeed, we show that if rating strategies follow a set of regularity conditions, then in equilibrium the rate at which players participate is increasing in the extremity of agents' valuations of the product. This theoretical prediction is realized in the lab. We also find that human subjects show a disproportionate predilection for sincere rating, and that when they do send insincere ratings, they are almost always in the direction of exaggeration. Both sincere and exaggerated ratings occur with great frequency despite the fact that such rating strategies are not in subjects' best interest. We therefore apply the behavioral concepts of quantal response equilibrium (QRE) and cursed equilibrium (CE) to the experimental data. Together, these theories explain the data significantly better than does a theory of rational, Bayesian behavior -- accurately predicting key comparative statics. However, the theories fail to predict the high rates of sincerity, and it is clear that a better theory is needed.

Adaptive Probabilistic Topic Models for Social Networks

Thesis (Master's)--University of Washington, 2012