Research Of Knowledge Tracing Model Based On Stacked GRU Residual Network

With the rapid development of the Internet,large-scale online open courses have gained a broad audience,and some online education platforms have launched personalized instructional products to provide students with personalized services.How to correctly evaluate and track students' abilities is a difficult problem in personalized instruction.Knowledge tracking technology is to track the level of knowledge mastery of students based on their learning history and predict their future performance.This thesis proposes new knowledge tracking models in response to insufficient prediction accuracy,over-fitting,and low utilization rate of multiple features in traditional knowledge tracking technologies.My work is listed as follows:(1)A knowledge tracking model S-GRU-R based on stacked GRU residual network is proposed to solve the problem of knowledge tracking with one feature.Aiming at the over-fitting problem caused by excessive LSTM network parameters used in the traditional deep knowledge tracking model,this thesis uses the gated recurrent unit GRU instead of LSTM for sequence learning,and adds the number of GRU network layers to increase the capacity of sequence learning.This thesis uses residual connection to solve the problems of network degradation and gradient disappearance,which are caused by the increasing of the number of network layers.Since the S-GRU-R model only uses the characteristics of students' answer scores,this article also calls it a knowledge tracking model based on a single feature.S-GRU-R is tested on five data sets,and similar models comparison experiment and S-GRU-R model ablation experiment are set up,using AUC,RMSE and F1 values as evaluation indicators.Experimental results show that the S-GRU-R model proposed in this thesis can achieve the best performance on most data sets.(2)A knowledge tracing model S-GRU-R@MECF is proposed to solve the problem of knowledge tracing with multi-feature,for efficient exploitation of the learning process data of students.The S-GRU-R@MECF model first uses the LightGBM algorithm to select the most important features,and uses the cross feature method and the one-hot method to encode the selected features.Since the cross-feature encoding will cause the input data dimension to increase rapidly,the auto-encoder method is utilized to compress the input data,and the compressed data is used as input of the S-GRU-R model.The S-GRU-R@MECF model is tested on the data set Riiid.We compare the results with the LightBGM prediction model and DKT model.Three ablation experiments are set up to verify the effects of cross-encoding and auto-encoder compression.The experimental results show that the proposed multi-feature fusion knowledge tracking model can achieve optimal performance.