Research On Personalized Exercise Recommendation Method For Students Based On Federated Learning

In recent years,online education has developed rapidly,and students can expand their horizons and increase their knowledge level through a large number of learning resources.In the field of educational data mining,personalized exercise recommendation methods help students learn and explore knowledge better by analyzing students’ past practice records,digging out their learning patterns,and recommending exercises to students based on corresponding recommendation rules.The current exercise recommendation methods are mainly divided into two categories:collaborative filtering-based methods implement the recommendation for target students by mining the interaction information between students and exercises,but these methods do not take into account that the difficulty of the exercises is in line with students’ learning status,and the recommendation results are poorly interpretable;knowledge modeling-based methods implement the recommendation for target students by modeling students’ knowledge level,but these methods do not dig deeper into the relationship between students and relevant knowledge concepts,and the recommendation results are single.In practical scenarios,education data is usually distributed among different schools(or platforms)without being shared,and the models have limited access to the data,which affects the effectiveness of recommendations.To solve the above problems,this thesis proposes a personalized exercise recommendation method for students based on federated learning,the specific work content is as follows:(1)The student’s exercise recommendation should be appropriate to the student’s current knowledge level.This thesis designs a novel recommendation method.First,it uses deep knowledge tracking to model students’ knowledge status,then uses a collaborative filtering algorithm to find students with similar knowledge level vectors,further adjusts the knowledge state prediction for target students,and then establishes the relationship between knowledge concepts in the learning process of students.Finally,an exercise filter is used to filter out the exercises that meet the student’s criteria,so that it can meet the educational philosophy of teaching students by their aptitude.(2)The limited data set will affect the training of the model,and the recommendation result will not reach the ideal state.To address the problem of an " isolated data island ",a more reasonable way to train the model is needed.This thesis designs a federated learning framework,which is divided into two parts: client-side and server-side.For each participating client,after the training is completed by local data,the model parameters are encrypted and uploaded to the server,which redistributes the new global model parameters after aggregation and computation,and the client obtains and trains locally again until the end.After the collaborative training of multiple clients with federated learning,the accuracy of model recommendation is improved,and students will be recommended more suitable exercises for their situations.(3)The heterogeneity of data can have an impact on federated learning,so the problem of non-independent homogeneous distribution needs to be fully considered.This thesis designs a federated hierarchical attention mechanism,each layer of the neural network considers the similarity between the client model and the server model in the parameter space,which not only maximizes the mutual dissimilarity of different clients but also enhances the recommendation accuracy.In this thesis,accuracy,novelty,and diversity are used as evaluation metrics,and corresponding experiments are designed for each step of the above work.By analyzing the experimental results on the real data set,the exercise recommendation method proposed in this thesis has achieved better performance.