Learning From User Behavior For Military Information Recommendation Service

With the continuous development of information technology,the total amount of information is also increasing,and the content as well as information types are more complex and diverse.Various types of military information systems are facing the challenge of information overload when they tend to integrate into each other.Information services can help people obtain effective information resources from the massive amount of information effectively.However,the passive information service in which users enter query words for retrieval is not efficient enough in some complex scenarios.Especially for military applications,when the battlefield environment is complex and rapidly changing,information retrieval cannot support users’ information needs efficiently.Active military information recommendation service is proposed for military applications.It comprehensively considers user characteristics,combat mission characteristics and so on.Based on the shared environment of information infrastructure,it can actively mine useful information from large-scale distributed and heterogeneous information pool and recommend high-quality information to current users.Considering the characteristics of the dynamic changes of user information needs,modeling the user’s interest and preference based on the user’s historical behavior,capturing the user’s dynamic intentions and providing users with information that meets their real-time needs are the key issues of the active military information recommendation service.The thesis aims to improve the quality of information recommendation service.Considering personalization and diversification in sequential recommendation,it studies how to mine information from users’ historical behaviors,capture users’ current personalized as well as diversified information needs,and improve sequential recommendation performance.At the same time,we also explore the application of the proposed sequential recommendation methods in this thesis in the active military information recommendation service.The thesis provides some new ideas for the sequential recommendation problem and a reference for the active military information recommendation service.Specifically,the main researches of this thesis are as follows:(1)We propose a Hybrid-Preference Neural Model for basket-sensitive item recommendation(HPNM)Current Basket-sensitive item recommendation approaches often ignore users’ longterm preferences reflected in their historical sequential baskets and only consider users’ short-term preferences.Also they neglect different item importances in a basket for modeling users’ intents.To solve the above issues,we introduce a Hybrid-Preference Neural Model(HPNM)for basket-sensitive item recommendation.First,in the item level,an attention mechanism is applied to distinguish the importances of items in each basket to generate an accurate basket representation,which can partly solve the issue of information loss when generating the basket representations.Then,in the basket level,HPNM utilizes Gated Recurrent Units(GRU)to model a user’s historical sequential baskets to generate the long-term preference.In addition,the representation of current basket is regarded as user’s short-term preference,which is then combined with user’s long-term preference as the final user preference for item prediction.We conduct comprehensive experiments on the Ta Feng and Foursquare datasets,finding that our proposal achieves state-of-the-art performance,returning the target items at a higher position in the recommendation list.(2)We propose a Collaborative Co-attention Network for Session-based Recommendation(CCN-SR)Existing session-based recommendation methods based on recurrent neural network can only model the one-way temporal transition relationship between items,ignoring the complex structural information in the user behavior sequence.The graph neural network based recommendation methods capture the structural information while ignoring the temporal information.Besides,in the graph neural network,the information is only transmitted between two adjacent nodes,which weakens the influence of context information on the recommendation results.We propose a Collaborative Co-attention Network for Session-based Recommendation(CCN-SR)model.The model combines the advantages of recurrent neural networks and graph neural networks to model the session behavior of users.At the same time,we propose a co-attention network to capture the relationship between time series information and structural information,and combine them to obtain more accurate representation of the user’s current intention.We design two co-attention strategies,namely parallel co-attention strategy and alternate co-attention strategy.We conduct experiments on two public datasets to verify the effectiveness of our proposal,and explore the role of different co-attention mechanisms.Experimental results show that our proposed model beats baseline models in terms of both Recall and MRR,and our model can achieve better performance in recommendation scenarios with different session lengths.(3)We propose a Multi-interest Diversification for End-to-end Sequential Recommendation(MDSR)Existing sequential recommendation methods only consider a user’s single interest and ignore the situation that the user’s multiple intentions may be reflected in his sequential behavior,resulting in the recommendations lack of diversity.In this paper,the recommendation based on user sequential behavior is redefined as a sequence-list generation task to model the relationship between recommended items,and a Multi-interest Diversification for End-to-end Sequential Recommendation(MDSR)model is proposed.The model solves the task of information recommendation by considering recommendation accuracy and diversity at the same time.We first design an implicit intent mining module to automatically capture multiple user intents reflected in the user behavior sequence,and then use an intent-aware diversity decoder to directly generate a recommendation list for multiple user intents.In order to support the learning of the implicit intent mining module and help the model consider the diversity of recommendations during the training process,we design an intent-aware diversity loss function,which evaluates the accuracy and diversity of recommendations based on the generated recommendation list.We conduct a large number of experiments on four benchmark datasets.The experimental results show that the recommendation list generated by our proposed model not only has high accuracy,but also contains more types of information to meet the various information needs of users.(4)We construct a prototype system for active information recommendation service and conduct simulation experimentsIn order to integrate the methods proposed in this thesis into military information services and promote the development of active military information recommendation services,we take the military information services as the background and construct a prototype system for active information recommendation service combining the personalized and diversified recommendation approaches studied in this thesis.The prototype system has the ability to analyze user behavior and identify user needs.It can also construct user portraits based on user historical behavior.When user behavior records update,it can dynamically adjust user portraits.Combined with the military application background,the information recommendation prototype system can provide service for different users.According to the roles,tasks and historical behaviors of different users,it can provide personalized information recommendations.Besides,when the tasks update into new stages,the prototype system can provide dynamic recommendations accordingly.We conduct simulation experiments to test the information recommendation performance of the prototype system on a large-scale dataset.The experimental results verify that the accuracy of the prototype system on both user-oriented personalized recommendation and taskoriented dynamic recommendation are all above 60%,and the time needed for making recommendations is less than one second.