Research On Group Activity Recommendation For Mobile Scenarios

It has become a popular trend for mobile social applications to organize and participate in activities in the form of groups,and the focus of recommendation systems has expanded from individual-oriented "personalization" to group-oriented "consensus".Group member preference aggregations mostly use predefined fixed strategies,e.g.,simplest aggregation strategy,least misery strategy,maximum pleasure strategy,etc.These strategies are difficult to capture the complex and dynamic group decision process,and for the recently proposed neural network models,the interactions among members and the dynamic decision process are difficult to form inputs and characterize by training.When it comes to making recommendations for groups,how to capture users’ behavioral sequence patterns and preferences,and simulate the interaction and dynamic decision-making process of group members has become an urgent problem,and there are often time and space constraints for group activity recommendations in mobile scenarios,so this thesis focuses on the "group activity recommendation problem for mobile scenarios".For ad hoc groups,a consensus decision recommendation model for ad hoc groups is proposed,which integrates the individual preference perception of members in mobile scenarios with the group consensus decision process,enabling the model to more accurately map the consensus formation process of group members in realistic scenarios.The model contains two phases:individual preference prediction phase and group consensus decision phase.First of all,a spatio-temporal-aware CNN model is designed to get the individual preference prediction of group members;then the individual member prediction is used as the initial value to construct the fuzzy preference relationship matrix and calculate the group consensus degree,and the group consensus is formed by imposing a finite number of consensus progression iterations,and finally the consensus result is used as the group activity recommendation list.Real datasets such as Mafengwo and Meetup are used for validation,and the experimental results of various evaluation metrics show that the model is able to obtain a higher degree of member consensus for group activity recommendations compared with the current preference aggregation models.A mobile recommendation attention network model for persistent clusters is proposed for persistent groups.The model has two modules,since it takes into account the existence of fixed preferences for persistent groups and the fact that the decisions of other group members are influenced by the vocal members of the group.The first module focuses on the learning of group members’ feature expressions;the second module simulates the process of decision selection by other members under the influence of the member with more discourse power through the attention network,and then fuses the output decision of module one with the decision result of module two to obtain the comprehensive decision result of group recommendation.The experiments show that the proposed model outperforms the baseline method and improves the accuracy of the experimental results.