Construction of GCNN-based intelligent recommendation model for answering teachers in online learning system

3.1 Time-sensitive teacher recommendations based on Q&A

In the current online learning systems, Q&A teacher recommendations are generally based on the correlation between student and instructor users’ concerns and historical responses, or the sociality between student and instructor users. However, this approach ignores the problem that the teacher’s knowledge base and expertise grow dynamically over time [26]. Therefore, for a certain type of question, a certain teacher user cannot give an effective answer in the present, but in the future, that teacher user may be competent enough to answer that type of question. To address the above problem, the study proposes a time-sensitive online learning Q&A teacher recommendation model (A Time Sensitive Online Learning Q&A Teacher Recommendation Model, TSRM) based on the basic structure shown in Figure 1.

Figure 1

The basic structure of TSRM.

In the TSRM model shown in Figure 1, the number of likes of answers is taken as an important index to evaluate the quality of answers, and more likes indicate higher quality of answers and prove the higher level of Q&A teacher users. The relationship between users’ co-response and co-concern is analyzed, and a multi-relationship co-response network is constructed based on the analysis results, and a sequence of Q&A text is constructed by time series. The semantic features of the constructed text sequences are extracted by Bidirectional Encoder Representations from Transformers (BERT) model, the topology of the constructed answer network is learned to obtain the empty domain social features, and the time series on the quiz text The topology of the constructed response network is learned to obtain the spatial social features, and the change in the number of likes on the time series of Q&A texts is learned to obtain the time domain features. Finally, the extracted features are classified so that each node of the network can be predicted to determine whether teacher users are better at answering certain types of questions. Based on the GCNN idea, the undirected graph of the multi-relational co-answer network constructed in the study is noted as G = ( V , E ) , where E denotes the set of edges in the topological graph, and V denotes the set of all network nodes, i.e., the set of users. Based on GCNN, it can predict whether a user can answer a question well, as shown in formulas (1)–(3) [27].

where X ∈ ℝ n ⋅ u is the feature matrix of users in the network, where n is the total number of users and u is the feature dimension; Y U denotes the untagged users; and p ( y i ∣ X V ) is the label distribution of users i , which are considered independent of each other in the study. The TSRM model predicts the user label distribution as q ( y i ∣ X V ) , then the objective function is set as equation (2).

The greater the similarity between the predicted and actual distributions of the model, the greater the objective function. The study introduces the relative entropy method to calculate the similarity between the predicted and actual distributions of the model. In the study, the question-answering teacher recommendation problem is viewed as a binary classification problem of graph nodes. Therefore, the goal of the TSRM model is to minimize the error between the predicted value y ˆ i and the true value y i . Therefore, the study introduces cross-entropy loss to construct the objective function, which is given by equation (3).

In the above, the time-domain features are generally obtained by the Long Short-Term Memory (LSTM) artificial neural network model, whose basic structure is shown in Figure 2.

Figure 2

The basic structure of LSTM.

In Figure 2, we can see that the LSTM unit contains 3 internal gate structures: forgetting, input, and output gate. The LSTM is able to record long-term information through the gating mechanism, which has a good application in different scenarios. However, the structure of LSTM is relatively complex and the training time is long, so it is simplified to obtain Gated Recurrent Unit (GRU). Since GRU is simplified from LSTM, it has the advantages of LSTM while its structure is simpler and involves fewer parameters, so the training efficiency is higher. Based on the above reasons, the study uses GRU to obtain time-domain features. In this process, four main parts are involved, which are reset gate, update gate, and output gate. Based on LSTM, its calculation is specifically shown in formulas (4)–(8) [28]. Among them, the structure of the update gate can be expressed as equation (4), which is able to control the extent of the state information generated during the last training when it enters this training.

where σ is the nonlinear classification function, the Sigmoid function used in the study; W is the weight matrix in the training process; X t denotes the Q&A feature at the moment of t ; b is the bias vector in the training process; h t denotes the hidden state at the moment of t , which can be calculated using equation (5).

where c t denotes the stored state inside the model. The forgetting gate can be expressed as equation (6), which is able to control the degree of elimination of the state information generated during the last training.

At the moment of t , the storage state inside the model can be expressed as equation (7).

In equation (6), W is the weight matrix from the input layer to the implied layer. h t can be expressed as equation (8).

Finally, GCNN is used to obtain the social features in the null domain. The specific process is as follows: first, the text features of user i are extracted using GRU and input to the GCNN layer as the initial vector H i 0 ; then, the feature vector matrix H 0 = [ H 1 0 , H 2 0 , … , H n 0 ] is put into the GCNN layer together with the adjacency matrix of the graph by repeating the above operation for all users, so as to extract the airspace social features from the multi-relational common network. The above process is shown in Figure 3.

Figure 3

GCNN-based process for obtaining spatial social features.

The TSRM model is designed to provide students with the most suitable Q&A teacher. To do this, the model will take into account the multi-dimensional characteristics of students and teachers, and learn how to match teachers and students most effectively through intelligent algorithms. First, data about students and teachers need to be collected, and after the data are collected, it is pre-processed to clean, normalize, and encode it to ensure that the data are suitable for model processing. Then, natural language processing and machine learning techniques are used to extract feature vectors from students’ and teachers’ Q&A records and learning behavior logs. Next a graph structure is constructed, with nodes representing students and teachers and edges representing possible matching relationships between them, and a graph neural network is used to extract complex relationships and patterns between nodes. Finally, node features and graph structure are used to predict the matching score.

3.2 Teacher recommendation model based on long- and short-term interest in answering questions

In different periods of time, as time goes by, the learning content of student users will change, and therefore the types of problems that student users focus on will also shift to some extent. However, in current intelligent recommendation methods, this is often ignored, and the recommended teachers for answering questions are more fixed, which affects the efficiency of solving students’ problem [29]. And with the change in time, the user’s concern and good problem solving will also change somewhat, i.e., an interest shift occurs. If the user has a long time of interest shift, it is called long-term user shift; accordingly, if the user has a short time of interest shift, it is called short-term user shift. To remedy the shortcomings of traditional intelligent recommendation methods, a TSRM based on Short- and Long-Term Interest for Answering Questions (LSTR) is proposed. The model is divided into three parts: question vector feature extraction, user vector representation, and intelligent recommendation based on the extracted question vector features and user vector. In the problem vector feature extraction part, the study uses the BERT model to obtain the word embeddings of problem titles and labels. Based on BERT, the calculation process is shown in formulas (9)–(12) [30]. Let a question be q , the title text of the question is represented as q t , and the label text of the question is represented as q l , then the title word embedding of the question X Q t can be calculated using equation (9).

where BERT ( q t ) denotes the word embedding vector obtained by pre-training the title text of the problem q in the BERT model. The labeled word embedding of this problem X Q l can be calculated using equation (10).

where BERT ( q l ) denotes the word embedding vector obtained by pre-training the labeled text of the problem q in the BERT model. The feature vector of this problem X Q can be calculated using equation (11).

where ⊕ denotes vector splicing. The word embedding of the question title is stitched with the word embedding of the question label to obtain the vector representation of the question. The above process is represented in Figure 4.

Figure 4

The process of obtaining vector representation of problems.

In addition, the social relationships of the users in the online learning system are analyzed. Users who have a relationship such as a common concern, a common answer, or a concern are considered to have a social relationship with each other. If a user is interested in a certain type of question, it is assumed that users with whom the user has a social relationship may also be interested in that type of question. Based on this theory, the list of questions that a user is interested in is considered as an attribute of the user, and each user is considered as a node, and the relationship between users in terms of concern or common answer is considered as an edge, thus building a multi-relationship common answer network with attributes. Based on GCNN, we extract the features of each node in the network, which can be regarded as the long-term interest of the user. Let the user’s concern question text be p , and the word embedding features of the topic text are extracted by BERT model as shown in equation (12).

Then, the GraphSage model is used to process and analyze the word embedding features to obtain the long-term interest features of users X U l . On the basis of the time series of user’s responses, we can analyze the change pattern of user’s interests and process the analysis results to obtain the short-term interests of users. The short-term interest features of users can be extracted by RNN, the LSTM and GRU are both based on the improvement of RNN. The GRU model has a simpler structure and is more efficient in training, as can be seen in Section 3.1. Therefore, the study uses the GRU model to extract the short-term interest features of users X U s . A complete user feature vector is obtained by stitching the long-term interest features obtained grounded on GCNN with the short-term interest features obtained in line with RNN. The above process is shown in Figure 5.

Figure 5

User vector representation process.

Based on the obtained question vector and user vector, the answer teacher recommendation is implemented. The study uses a Deep Structured Semantic Model (DSSM) for result prediction. The extracted user feature vector X U and question vector X Q are added into two fully connected layers, respectively, and mapped to the low-dimensional space to obtain the low-dimensional vectors. The cosine similarity method is utilized, the similarity between the user features and the problem features S ( X Q , X U ) is able to be calculated. The teachers with the highest number of likes on the question are considered as a positive sample, and the teacher users who did not answer the question are considered as a negative sample. Based on the random sampling method with Softmax, it is able to convert the similarity between the question and the positive sample to the posterior probability of finding a positive sample for that question P ( X U ∣ X Q ) . During the training process, the loss function in equation (13) needs to be minimized [31].

The above contents can be represented as Figure 6. The LSTR model is completed by combining the above contents. Combining the LSTR model with the TSRM model, efficient and high-precision teacher recommendation for answering questions can be achieved. Compared with the existing learning-based networks, the model uses graph structure to represent and process data, which is particularly suitable for processing graph structure data and can capture complex relationships and patterns between nodes. At the same time, it is able to process non-Euclidean structured data, such as social network data, namely, the relationship between students and teachers. This capability makes it suitable for recommendation systems, especially in situations where interactions between users need to be captured and exploited. And the model can solve the problem of teacher recommendation in online learning system, which is not covered by the traditional learning network.

Figure 6

The basic structure of deep structured semantic model.

4.1 Performance analysis of TSRM model

On the premise of obtaining consent, data were obtained from an online learning system of a university to construct a dataset, which included student dimension information, teacher dimension information, classroom content information, and interactive dimension information. Student dimension information includes students’ basic information, learning behavior, and historical question answering request. Teacher dimension information includes teachers’ basic information, historical evaluation, and historical Q&A information. The classroom content information includes the basic information of the course, the difficulty of the course, the students’ participation, and the degree of completion. Interactive dimension information includes historical question answering records between students and teachers. The dataset is divided into two subsets according to the ratio of 7:3, which are noted as the training set and the testing set, for the training and testing of the model. The detailed configuration parameters of hardware and software used in the study are shown in Table 1.

Among the current online learning systems, there are two main state-of-the-art methods for teacher recommendation, which are LDA-PageRank model (LP) and WHITS model. First, the performance of TSRM model, LP model, and WHITS model is compared. First, the TSRM, LP, and WHITS model are trained using the training set, and Figure 7 shows the error variation and loss value variation of the three models during the training process. In Figure 7, we can see that the error and loss values of the three models decrease with the increase in the number of model iterations, and after the number of model iterations reaches a certain level, the error and loss values of the models stabilize and do not change significantly. Among them, the TSRM model needs 99 iterations to reach the optimal state, which is 48 and 103 times less than the LP model and WHITS model, respectively. In Figure 7(a), the error values of LP and WHITS are 0.02 and 0.03 higher than those of TSRM (0.02). In Figure 7(b), the loss values of TSRM (0.3) are 0.4 and 0.8 lower than those of LP and WHITS, respectively.

Figure 7

Error changes and loss value changes of three models during training. (a) Error. (b) Loss.

The performance of the TSRM model, LP model, and WHITS model was tested using the test set, and they are evaluated using F1 values and Recall values, as shown in Figure 8. In Figure 8, it can be seen that the F1 value and Recall value of the models increase as the number of iterations increases. In Figure 8(a), the TSRM’s F1 value is 96.5%, which is 1.1 and 1.4% higher than that of the LP and the WHITS. In Figure 8(b), the TSRM’s Recall value is 96.3%, which is 0.5 and 1.0% higher than the other two models.

Figure 8

F1 and Recall values of TSRM model, LP model, and WHITS model. (a) F1. (b) Recall.

The recommendation accuracy of the TSRM, the LP, and the WHITS model are compared on the training set and the test set, respectively. Figure 9 is the recommendation accuracy of the three models. The models’ accuracy values increase with the iterations on both the training and test sets. In Figure 9(a), the accuracy variation of the TSRM model, LP model, and WHITS model on the training set is shown. The TSRM accuracy is 98.5%, which is 0.5 and 1.2% higher than the LP and the WHITS. In Figure 9(b), the accuracy variation of the TSRM model, LP model, and WHITS model on the test set is shown. The accuracy of the TSRM is 99.5%. It is 0.4 and 0.9% beyond than those of other two. In summary, the TSRM model has better performance in the work of Q&A teacher recommendation for online learning systems.

Figure 9

Change in accuracy of models on training and testing sets. (a) Training set. (b) Test set.

4.2 Performance analysis of LSTR model

The LSTR model was tested and compared with the LP model and the WHITS model using the training and test sets described above to demonstrate the LSTR model performance. The variation in MAE values and RMSE values of the LSTR model, LP model, and WHITS model on the training set are exhibited in Figure10. In Figure 10(a), the mean MAE of the LSTR is 3.7%. It is 0.9 and 1.4% lower than LP model and the WHITS, respectively. In Figure 10(b), the mean value of RMSE for the LSTR is 3.3%. It is 0.6 and 1.2% lower than that of the LP and the WHITS, respectively.

Figure 10

Changes in MAE and RMSE values of the model. (a) MAE. (b) RMSE.

In the test set, the recommended success rate and computational complexity test of LSTR model, LP model, and WHITS model are shown in Figure 11. It can be seen that with the increase in the number of model iterations, the recommendation success rate of LSTR model, LP model, and WHITS model is also increasing. As can be seen in Figure 11(a), the LSTR model has a recommendation success rate of 98.4%, which is 4.2 and 4.9% higher than the LP model and WHITS model, respectively. As can be seen from Figure 11(b), the computational complexity of the LSTR model has been optimized very effectively, and it is 63.7 and 86.5% ahead of the LP model and WHITS model, respectively

Figure 11

The recommendation success rate of LSTR model, LP model, and WHITS model. (a) Recommendation success rate. (b) Computational complexity.

The comprehensive performance of LSTR model with LP model and WHITS model was evaluated by ROC curve. Figure 12 displays the evaluation results. The AUC value of the LSTR model reaches 0.925, which is 0.014 and 0.020 higher than that of the LP model and the WHITS model.

Figure 12

Comprehensive performance evaluation of LSTR model, LP model, and WHITS model.

The mean reciprocal ranking (MRR) and the normalized discounted cumulative gain (NDCG) of the three models were tested, and the test results are shown in Table 2. As can be seen from Table 2, the average reciprocal rank of LSTR model is 0.23, which is higher than LP model and WHITS model by 0.05 and 0.08, respectively. The cumulative normalized loss gain of LSTR model is 0.25, which is 0.08 and 0.09 higher than that of LP model and WHITS model, respectively.

In summary, the LSTR model and the TSRM model constructed by the study have high performance and can effectively carry out the recommendation of the answering teacher in the online learning system, thus improving the efficiency of solving students’ problem, improving their learning effect, and contributing to the development of education informatization in higher education institutions.