Research On Prediction Algorithm Of Problem Solving Test Results Based On Process Data Mining

How to evaluate students’ personal ability reliably and effectively through tests has always been a hot spot in the educational circles.At present,students’ personal ability is mainly measured according to item response theory or cognitive diagnosis theory.This method is difficult to accurately predict students’ ability to solve specific problems in the real situation.With the development of information technology,it is found that interactive scene dynamic tasks can better meet the test needs of problemsolving ability.At the same time,computer-based evaluation means that the process data in the process of scene task solving can be collected,and we can reasonably assume that the sequence of individual behavior events is related to the result.In this case,the steps for students to identify problems,plan and implement solutions,and readjust strategies according to real-time progress are as important as the result.Therefore,process data can be used as a predictor to more accurately predict whether students can successfully solve problems.This thesis chooses the climate control problem of CPS project in pisa2012 as the research background.Based on the log data such as action sequence and time series in the process of answering this topic,it predicts whether students can successfully solve this problem in the end.The specific work is as follows:(1)Firstly,starting from the answer and response results of the topic,this thesis focuses on the sorting of algorithms such as feature extraction and prediction of education process data,mainly including multi-dimensional scaling method and self coding method,then discusses the principle and experimental simulation of these algorithms,and compares and analyzes the prediction results of action sequence feature sets extracted based on these algorithms.(2)Secondly,aiming at the deficiency of the single model used in the prediction part of previous studies,the gradient lifting decision tree,convolutional neural network and other algorithms are proposed to predict based on the extracted action sequence features.In view of the shortcomings of previous studies that only considered based on action features,this thesis proposes to use key node timestamp to extract and transform time features again,and also use the algorithm above to predict and compare.Experimental results show that the accuracy and F score of the prediction algorithm based on motion features reach 85%,which is 2.2% higher than the generalized linear prediction model used in previous studies.On the basis of 81.3% accuracy,the running time of the prediction model based on time feature is nearly half shorter than that of the action feature prediction model,and the efficiency of the algorithm is greatly improved,which proves that the model based on time feature set can also achieve good prediction effect in a relatively short time.(3)Finally,on the basis of previous researches on self-coding algorithm,Transformer model based on attention mechanism is proposed to be applied to process data mining,and the prediction and analysis are carried out from the two dimensions of action and time.The experimental results show that the accuracy of classification prediction using Transformer model is 88.9%,which is 5.2% higher than the autocoding algorithm.Transformer model has the characteristics of parallel computing,only one-tenth of the former running time can achieve higher prediction accuracy,and the algorithm efficiency is further improved,which verifies the feasibility of Transformer model in the field of process data mining.