Research And Application Of Curriculum Recommendation System Based On Deep Reinforcement Learning

With the development of the Internet,the amount of information in cyberspace is constantly expanding,and people’s demand for various information is also increasing.In this information age,the recommendation system has become an important interactive system,which can actively provide users with the information they need,or provide auxiliary information for optimized retrieval when users search.In order to enable users to quickly receive the most valuable and suitable information in massive data,the performance of recommendation systems is also constantly being upgraded.In the era of machine learning,many researchers have introduced machine learning algorithms into traditional recommendation systems.Two important branches,reinforcement learning and deep learning,have had more research in their combination with recommendation systems in recent years.The course recommendation system introduced in this paper is a research direction of recommendation systems.The key point of this recommendation task is different from the common recommendation task,which is that it is a sequential recommendation task that changes the recommendation results with the course process,and combines reinforcement learning to provide good interpretability for the recommendation results.To solve the problem of the course recommendation system,this paper designs a personalized and sequential recommendation system based on the student learning history record through the student tablet database.The main research contents are as follows:(1)Firstly,due to the scarcity of structured data in the relevant field,this paper proposes a modeling method for unstructured historical record data to process massive unlabeled data in this direction as training samples.This scheme combines the basic Doc2 vec word embedding method and the multi-class network supervision mechanism,sets the loss function in the form of a multi-task learning model,and updates the parameters of the two networks at the same time.Meanwhile,the combination of Doc2 vec and BLSTM network improves the semantic expression ability of sentence vectors.Finally,it generates item sentence vectors with classification information to better reflect the different feature information of unstructured data,which is mainly used for the subject of course content in this paper.At the same time,the vectors based on Doc2 vec and item2 vec will enter the recommendation system as item feature representation vectors at the same time,so as to simultaneously represent the semantic information and temporal information of the item.The experimental results show that this modeling scheme improves the classification accuracy of item vectors relative to other baseline models,improves its semantic expression,and can better undertake the subsequent recommendation work.(2)On the core recommendation function,this paper designs a deep reinforcement learning-based recommendation system based on the reinforcement learning model,mainly using the DDPG network-based reinforcement learning network.In the reinforcement learning structure,the Critic network adopts the context slot machine mechanism to better reflect the long and short-term recommendation mechanism of the recommendation task.This method will input the information of actions and states at the same time when the Critic network provides recommendations,and present different performance under different state lengths.Secondly,because the output of the last layer of the simple DQN network is not enough to meet the goal of multi-course recommendation,a scoring weight combined with a multi-class network screening method is used to realize the list-level recommendation classified by subject with a smaller network size.This paper combines the DDPG network structure and the context slot machine mechanism to model and test 4million data under the state setting of length 4.The experimental results show that compared with two groups of baseline methods,the method proposed in this paper has certain performance improvement in accuracy and average recommendation score.This course recommendation system is mainly used in the online course recommendation scenario,which improves the adaptability of the push content and the current course progress of students,and adopts unstructured word embedding methods,which also provides a foundation for adding various features in other tasks.