Peer-inspired student performance prediction in interactive online question pools with graph neural network

1. A method for autonomously predicting student performance in interactive question pools, comprising:

Executing a data processing and feature extraction logical process execution module, comprising:

receiving a plurality of historical student question-answer score data records of a plurality of students on an interactive question pool having a plurality of questions;

receiving a plurality of pointing device movement data records containing data of one or more pointing device movements made by the students corresponding to the historical student question-answer score data records;

extracting from historical student question-answer score data records:

statistical student features reflecting the students' past performances on the questions; and

statistical question features indicating a question popularity of each of the questions and the average students' scores on each of the questions; and

extracting from the pointing device movement data records a plurality of interactive edge features representing one or more characteristics of the students' problem-solving behaviors;

executing a network of nodes construction logical process execution module, comprising:

performing an Edge2Node transformation to transform the interactive edge features into one or more interaction nodes; and

constructing a heterogeneous student-interaction-question (SIQ) network of one or more student nodes from the statistical student node features, one or more question nodes from the statistical question node features, and the interaction nodes; and

executing a prediction logical process execution module, comprising:

processing the SIQ network by a residual relational graph neural network (R 2 GCN) to predict a student's score on one or more unattempt questions in the interactive online question pool;

wherein the R 2 GCN comprises:

one or more parallel input layers for feature transformation of different types of nodes in the SIQ network into nodes of same shape;

one or more consequential relational graph neural network (R-GCN) layers for message passing;

one or more residual connections to one or more hidden states, the statistical student features, statistical question features, and interaction edge features for capturing different levels of information; and

an output layer for final prediction of a student's score on the unattempt questions in the interactive online question pool;

wherein the interactive edge features comprise:

time length between when a question is first shown to a student and when a first general click on the question is made by the student;

percentage of a time length of a first general click in an entire time length spent on answering a question by a student;

number of pointing device movement events when a question is first shown to a student and before a first general click on the question is made by the student;

percentage of pointing device movement events made when a question is first shown to a student and before a first general click on the questions is made by a student;

number of pointing device movement events made by a student on a question before the first general click ends;

total number of general clicks made on a question answered by a student;

average number of general clicks on made on a question answered by a student per second;

average time length between general clicks made on a question answered by a student;

median time length between general clicks made on a question answered by a student;

standard deviation of time length between general clicks made on a question answered by a student;

total pointing device trajectory length made by a student; and

point of time when a student answers a question;

wherein a general click is of a pair of necessary pointing device movement events in a pointing device movement trajectory when a student is interacting with a question.

2. The method of claim 1 , wherein the statistical student features comprise:

number of a student's total trials made on the questions;

number of a student's second trials made on the questions;

percentages of a student's total trials on questions of certain mathematical dimension, grade, and difficulty; and

mean scores of a student's first trials on questions of certain mathematical dimension, grade, and difficulty.

3. The method of claim 1 , wherein the statistical question features comprise:

a question's mathematical dimension;

a question's grade;

a question's difficulty;

total number of trials on a question made by the students;

total number of second trials on a question made by the students; and

percentage of trials made by the students on a question in each score level achieved by the students.

4. The method of claim 1 , wherein the R 2 GCN further comprises:

a message function for transmitting and aggregating messages from all neighboring nodes to a center node of the SIQ network in the message passing;

an averaging function for reducing messages transmitted on same type of edges;

a summing for reducing messages transmitted on different type of edges;

an update function for updating a center node's hidden state after each of the R-GCN layers; and

a readout function for transforming a final hidden state to the final prediction.

5. An apparatus for autonomously predicting student performance in interactive question pools, comprising:

one or more computer-readable media; and

one or more processors that are coupled to the one or more computer-readable media and that are configured to execute:

a data processing and feature extraction logical process execution module for:

receiving a plurality of historical student question-answer score data records of a plurality of students on an interactive question pool having a plurality of questions;

receiving a plurality of pointing device movement data records containing data of one or more pointing device movements made by the students corresponding to the historical student question-answer score data records;

extracting from historical student question-answer score data records:

statistical student features reflecting the students' past performances on the questions; and

statistical question features indicating a question popularity of each of the questions and the average students' scores on each of the questions; and

extracting from the pointing device movement data records a plurality of interactive edge features representing one or more characteristics of the students' problem-solving behaviors;

a network of nodes construction logical process execution module for:

performing an Edge2Node transformation to transform the interactive edge features into one or more interaction nodes; and

constructing a heterogeneous student-interaction-question (SIQ) network of one or more student nodes from the statistical student node features, one or more question nodes from the statistical question node features, and the interaction nodes; and

a prediction logical process execution module for:

processing the SIQ network by a residual relational graph neural network (R 2 GCN) to predict a student's score on one or more unattempt questions in the interactive online question pool;

wherein the R 2 GCN comprises:

one or more parallel input layers for feature transformation of different types of nodes in the SIQ network into nodes of same shape;

one or more consequential relational graph neural network (R-GCN) layers for message passing;

one or more residual connections to one or more hidden states, the statistical student features, statistical question features, and interaction edge features for capturing different levels of information; and

an output layer for final prediction of a student's score on the unattempt questions in the interactive online question pool;

wherein the interactive edge features comprise:

time length between when a question is first shown to a student and when a first general click on the question is made by the student;

percentage of a time length of a first general click in an entire time length spent on answering a question by a student;

number of pointing device movement events when a question is first shown to a student and before a first general click on the question is made by the student;

percentage of pointing device movement events made when a question is first shown to a student and before a first general click on the questions is made by a student;

number of pointing device movement events made by a student on a question before the first general click ends;

total number of general clicks made on a question answered by a student;

average number of general clicks on made on a question answered by a student per second;

average time length between general clicks made on a question answered by a student;

median time length between general clicks made on a question answered by a student;

standard deviation of time length between general clicks made on a question answered by a student;

total pointing device trajectory length made by a student; and

point of time when a student answers a question;

wherein a general click is of a pair of necessary pointing device movement events in a pointing device movement trajectory when a student is interacting with a question.

6. The apparatus of claim 5 , wherein the statistical student features comprise:

number of a student's total trials made on the questions;

number of a student's second trials made on the questions;

percentages of a student's total trials on questions of certain mathematical dimension, grade, and difficulty; and

mean scores of a student's first trials on questions of certain mathematical dimension, grade, and difficulty.

7. The apparatus of claim 5 , wherein the statistical question features comprise:

a question's mathematical dimension;

a question's grade;

a question's difficulty;

total number of trials on a question made by the students;

total number of second trials on a question made by the students; and

percentage of trials made by the students on a question in each score level achieved by the students.

8. The apparatus of claim 5 , wherein the R 2 GCN further comprises:

a message function for transmitting and aggregating messages from all neighboring nodes to a center node of the SIQ network in the message passing;

an averaging function for reducing messages transmitted on same type of edges;

a summing for reducing messages transmitted on different type of edges;

an update function for updating a center node's hidden state after each of the R-GCN layers; and

a readout function for transforming a final hidden state to the final prediction.