Research On Online Learning Resource Recommendation Algorithm

With the popularization of computer and network,the information age is coming.People's life has been changed with each passing day,and the traditional education mode has also been impacted.However,massive resources and user information data will also lead to information overload,and the traditional recommendation algorithm will have data sparsity and cold start problems.In addition,a large number of information in the comment text will be ignored when processing the data.It is difficult for most online learning platforms to find their own learning resources that are interesting and suitable for them by relying on simple keyword search and popular recommendation functions.On the other hand,due to the differences of learners' cognitive ability,learning style and knowledge background,it is also difficult for learners to quickly and effectively recommend learning resources that they are interested in.If these problems cannot be solved,for learners and resource providers of online platform learning,these will be great challenges in the future.The existing research models of online learning resource recommendation algorithms are generally based on traditional recommendation models such as collaborative filtering and content-based.The calculation of similarity is excessively simple.For a large and complex data platform,more accurate recommendation of learning resources to users will be a major challenge.The main disadvantages of the model are as follows: firstly,it cannot automatically combine features,and manual combination of features will bring more work.Secondly,the main body of online learning resource recommendation is learners and courses,and the degree of relevance usually represents the degree of learners' preference for courses,while the relevance between learners and courses is ignored.Third,the number of layers of depth neural network determines the depth of feature extraction,but with the increase of depth,there are many problems related to optimization,which may result in gradient dissipation or gradient explosion.Fourthly,for the data set of online learning resource recommendation algorithm research,it usually contains continuous features and category features.The learning platform will also have text features of students' comments,and text features will also have a certain impact on the accuracy of learning resource recommendation.In face of massive learning resources and a large number of online learning user information,how to effectively recommend learning resources of interest to learners and how to mine hidden features rarely found by learners from historical data are the main research content of this paper.Neural network technology plays an important and effective role in feature extraction.This paper analyzes the existing recommendation model and finds that Wide & Deep is a better recommendation model.Therefore,combining the advantages of the Wide & Deep recommendation model,this paper proposes an online learning resource recommendation algorithm based on the Wide & Deep model.Firstly,aiming at the problem that part of the traditional online learning recommendation model cannot automatically cross feature learning and high-order feature interaction,which results in high labor cost and weak generalization ability,an online learning resource deep learning recommendation algorithm is proposed to solve the problem of extracting higher-dimensional feature information.Secondly,aiming at the problem of traditional online learning recommendation and the lack of feature extraction for text features in existing models,an improved recommendation algorithm for online learning resources based on Wide&Deep model is proposed.The main research can be summarized as follows:(1)Propose an online learning resource recommendation algorithm based on Wide&Deep residual interaction.Based on the Wide&Deep model proposed by Google,this paper adds an attention mechanism to extract the interaction characteristics between users and projects,and introduces the idea of residual network to deepen the fully connected network layer,which can effectively extract the robust feature information to ensure the model.After that,two experiments are set up.Firstly,the data of the online learning platform is trained on this model and compared with the ordinary neural network recommendation algorithm.Secondly,the network of the depth model part is set to different depths to compare the experimental results.(2)Propose an online learning resource improvement recommendation algorithm based on Wide&Deep model.Firstly,based on Wide&Deep model,text features are treated as category features.By adding embedding layer,data is divided into training set and test set with a certain proportion,and different epochs are set,then training is carried out.Secondly,an improved hybrid model based on Wide&Deep model is proposed.Text features in data are pre-trained by ELMo language model to generate context-related word vectors.Then,downstream tasks are trained by Wide&Deep model.Then,the feature vectors are extracted and auto-crossing features are input as embedding to further complete.Finally,the user's rating of the recommended course is predicted.The experimental results show that the Wide&Deep based residual interactive online learning resource recommendation algorithm proposed in Chapter 3 has better performance than the ordinary neural network recommendation algorithm,and then sets the residual network module to a different depth from the ordinary neural network.When the depth of the ordinary deep neural network is very high,the network is further deepened and gradient diffusion occurs.The performance of residual network increases with depth.The online learning improvement recommendation algorithm based on Wide&Deep proposed in Chapter 4 performs better than the basic model and improves the accuracy by 3.7%~5.4%.This shows that the improved method proposed in this paper fully encodes the information of text characteristics and has a great effect on improving the accuracy of recommendation.Although the performance of the recommendation algorithm proposed in this paper has been improved to some extent,but in the experimental process,the time cost is high,so how to carry out more efficient automatic training of the model still needs to be improved.At present,there are many excellent neural networks,such as reinforcement learning,DeepFM,and so on,which perform well in the recommendation system.It is believed that the application in the recommendation of online learning resources will have a better effect,which will also help to promote more in-depth research.