Reserch On Group Recommendation In Social Tagging Systems

With the emergence of Web2.0, a large number of Social Tagging Systems (STS) appear in the Web. STS allow users to annotate the shared resources with personal tags through an open platform, and users can upload resources and share them with appropriate groups. Groups can satisfy the needs of user interaction and interest sharing, so in recent years they have attracted a large number of users. However, the explosive growth in the number of the group makes it difficult for Flickr users to find relevant groups they are really interested in. Thus, it raises a great demand of developing tools to help users find the desirable groups more easily. To solve this problem, group recommendation systems are proposed. Group recommendation system can help users find the useful group information, enhance user satisfaction and save their time. Therefore, it can attract more users to join the website and achieve a win-win situation for both service providers and users. In this paper, we focus on automatically recommending groups to users and resources, and take Flickr photo sharing website as an example to describe our research works. Although the proposed methods are designed for Flickr, they can also be extended to any STS with group information.In Flickr, users are allowed to upload images, annotate tags and share images with appropriate groups. Users, tags, images and groups are four types of entities in Flickr, which are correlated to one another and form a quaternary relationship. In this paper, in order to capture the quaternary relationship among users, tags, image and groups, we propose a Flickr group recommendation model based on quaternary semantic analysis, and the follwing works have been done through this model.(1) Group recommendation for Flickr users based on quaternary semantic analysisMany existing works on group recommendation for Flickr users have utilized only user-group binary relationship or user-tag-group ternary relationship to generate their recommendations. These approaches neglected the rich visual information associated with the images and some tags may have the problem of polysemy and ambiguit, so these approaches may not provide accurate recommendation. In our work, we show that ternary relationship is insufficient to provide accurate recommendations through examples and experiments. Instead, our approach explores both visual features and tags to reveal the latent semantic associations between users and groups for recommendation. We represent the quaternary relationship among users, tags, image classes and groups as a4-order tensor and further employ the Higher-Order Singular Value Decomposition (HOSVD) to reduce the dimensionality of the4-order tensor; accordingly, the group recommendation problem is casted as a latent semantic analysis problem between users and groups. Experiments on the dataset crawled from Flickr and comparisons with the ternary relationship model demonstrate the effectiveness of the proposed approach in terms of top-k and MAP. To the best of our knowledge, this is the first work to explore the use of the quaternary relationship for Flickr group recommendation tasks. We propose an improved sparse coding based space pyramid matching algorithm (ScKSPM), which assign different weight to sparse coding at different level and design a new spatial pyramid matching kernel for image classification. Experiments on Caltech101/256and Pascal VOC2006dataset show that, the method based on new spatial pyramid matching kernel can achieve higher accuracy in the image classification.(2) Group recommendation for Flickr images based on quaternary semantic analysisExisting works on group recommendation for Flickr images all utilized content-based recommendation algorithms and provide recommendation by comparing the features of an image to the features of groups. These approaches need to build a model for each concept or group, which may lead to a strong limitation of scalability. Moreover, these methods often neglected an important fact that there are many latent user interests, which may influence users’ choices. In order to discover the underlying interests of users and avoid training multiple models, in this paper, we propose a model based on4-order tensor that integrates images, tags, groups and users information in one framework. Particularly, we propose a tetradic decomposition algorithm that employs the HOSVD and kernel-SVD to reduce the dimensionality of the4-order tensor and reveal the underlying interests of users and the latent semantic association between images and groups. Furthermore, as new users, tags, images or groups are being introduced to the system, we use folding-in and Incremental SVD techniques to update our model incrementally in order to avoid the costly batch recomputation. We compare our algorithm with existing group recommendation algorithms on the dataset crawled from Flickr. The experimental results show significant improvements in terms of MAP.(3) A unified framework for Flickr group recommendation Many previous works on Flickr group recommendation only suggested groups to users or to images, but rarely exploited these two kinds of recommendations in the same framework. Nevertheless, in the practical applications, users in Flickr desire to know not merely the groups with which his images should be shared, but also the groups in which they might be interested. Motivated by the above observations, in this paper, we propose a unified framework, which can simultaneously suggest groups to users as well as to images and implement a prototype system. Our system integrates the four types of entities that exist in Flickr system (i.e. users, tags, images and groups) into a tetradic model and utilizes a tetradic decomposition method to reveal the latent semantic associations among them to enhance the performance of both two kinds of recommendations. The experiments on Flickr dataset show that the performance of our group recommendation approach outperforms some well-known methods that only recommend groups to images or to users in terms of MAP. To the best of our knowledge, this is the first work to integrate the group recommendation for Flickr users and images in the same framework.