Research On Intelligent Recommendation Methods Under The Information Shortage Environment

As an effective method to solve the problem that user’s demands cannot be satisfied because of the information redundancy problem,recommender systems have been widely investigated in recent years.Different from the traditional information retrieval technologies that require users to actively retrieve relevant information,recommender systems can provide users with personalized recommendations according to analyzing user’s historical interaction behaviors,so as to satisfy user’s information demands in different conditions and improve the efficiency of users for acquiring information.However,in some practical civil scenarios,such as users are not logged in or not registered,and many scenarios under the military environment,such as the operational commander’s job just moved or the battlefield situation changed sharply,user’s long-term historical interactions cannot be required or are not suitable for detecting the user intent in the current recommendation condition.Thus,session-based recommender systems are proposed to capture user’s timely preference and detect his/her information demands according to modeling user’s recent limited interactions,so as to provide users with accurate recommendations.However,there still remains many deficiencies in the previous session-based recommendation methods,such as: In the methods introducing external information,the neighbor sessions cannot be accurately located,resulting in a large deviation in the introduced auxiliary information;In the methods distinguishing item importances,the attention mechanism cannot overcome the influence brought by the unrelated items,resulting in the item importances not being able to accurately estimated;And in the graph neural network based methods,the long-distance dependency relation between items cannot be modeled and the network is easily over-fitting,resulting in that user’s realistic behavior pattern cannot be adequately modeled.In view of the above problems,we propose the corresponding intelligent recommendations models as follows:(1)An intent-guided intelligent recommendation model,which relies on user’s intent in the ongoing session to locate the related neighbor session as auxiliary information,so as to help model the current user preference.Specifically,we design a hybrid longand short-term preference encoder to generate the current user intent,and propose an intent-guided neighbor detector to retrieve and select the neighbor sessions as auxiliary information to help model the current session.(2)An intelligent recommendation model based on item importance extraction,which aims to accurately distinguish the importance of different items for capturing user’s main intent.Specifically,we design an importance extraction module by improving the selfattention mechanism to determine the priority of each item,and then distinguish the items according to the priorities to generate user’s global preference,which is then combined with user’s current interest as the final user preference.(3)A star graph neural network based intelligent recommendation model,which aims to solve the information loss problem and the over-fitting issue in the traditional graph neural networks for accurately modeling user’s realistic behavior pattern.Specifically,we design a graph neural network to achieve the connection between long-distance items,and adopt the highway networks to alleviate the over-fitting problem,so as to increase the layer number of graph neural networks for modeling the high-order relations between items.We conduct extensive experiments on multiple public datasets under the information shortage environment.The experimental results demonstrate that our proposed three intelligent recommendation methods can significantly outperform the state-of-the-art baselines,proving the effectiveness of our proposals on providing personalized recommendation service for users.