Development of Group Recommendation in Collaborative Framework

Abstract

Recommender systems (RS) provide personalized suggestions to users regarding products and services. These suggestions are generated for individual users only. However, group activities are gaining popularity in several applications, such as movie recommendations, travel and e-tourists, etc. Therefore, there is a need for a group recommendation system that helps people provide better recommendations for group members instead of a single user. This thesis addresses cold start, data sparsity issues and group member satisfaction. An effective group recommendation approach named GR Slope One is introduced to incorporate dynamic changes in the user-item rating information. The proposed method works in two ways: group aggregate prediction and group aggregate model. Firstly, groups are formed using k􀀀means clustering, and later, the group aggregate prediction is performed based on the group’s individual prediction information. Secondly, the group aggregate model is determined by aggregating individual user preferences. A novel modelling technique (Max-after-threshold) is introduced. It adopts an aggregate prediction feature to provide better recommendations to the group. The proposed approach is compared with popular existing methods such as Matrix factorization (MF), BaseGRA, and Improved GRA. The deep Collaborative Filtering Approach suffers from the cold-start issue in the presence of a group recommender system. To overcome the issue, metadata information is considered while predicting the rating information. The approach presents a rating prediction for groups that leverage multi-layer perceptrons, general matrix factorization using metadata, and neural collaborative filtering techniques. The proposed approach is discussed in two steps. The first step is to learn from group-item interaction, perform one hot encoding for the group and item, and then utilize this information to perform dot product by applying the GMF layer. In the second step, learn group-item interactions using group and item metadata. This information is concatenated using the MLP layer. Finally, the GMF and MLP layers are combined to get the final prediction ratings. Recently, the attention mechanism approach has drawn attention to the group recommendations system. Certain unsolved problems, such as the weights of group members, are crucial during the recommendation process. Existing works consider all the members in the group to be given equal priority. Moreover, preference aggregation is not considered. Therefore, group recommendation using the attention mechanism can be exploited to address the issue of preference aggregation, which uses a neural attention network and a neural collaborative filtering framework. The attention component is used to capture the effect of every member within the group. In addition, a neural collaborative filtering framework is utilized to learn the group-item interactions in the data. It strengthens the performance of the group recommendations and their user recommendations. Group recommendation using the Attention mechanism addressed the preference aggregation(GRAM). It has certain limitations, such as restrictions on group sizes, ignoring group modelling strategies, and group satisfaction metrics. Therefore, it is essential t consider all the mentioned limitations to achieve better performance. Major restrictions enforced on the literature studies are: (1) small number of users, (2) a large number of groups, (3) median number of group participants and (4) various centrality techniques. To address these limitations, we propose a preference network-based approach. It performs prediction based on weighting individual users in linear preference. The weight computing is based on the node centrality score. Multiple centrality techniques are analyzed for score calculation. This work also introduced two new modeling strategies AV GMP and MPAV G. This study uses the group satisfaction metric (GSM) to evaluate member satisfaction and satisfaction error for a group (SEG) to improve member satisfaction and recommendations for groups. The proposed method outperforms baseline aggregation techniques. The experiments were conducted on standard datasets and validate the effectiveness of the proposed approaches.