Online learning recommendation method and system

Abstract

The present application relates to an online learning recommendation method and system. The method comprises: collecting interaction data of a user with an online learning system, the interaction data comprising data about exercises answered by the user and corresponding scores; performing hierarchical aggregation using the interaction data to obtain an overall vector based on a graph structure, wherein the graph structure comprises a plurality of nodes and edges, each node indicating a topic and exercises belonging to the topic, and each edge connecting two nodes and indicating that a node indicating a preceding topic points to a node indicating a subsequent topic through the edge; inputting the overall vector into a sequence machine learning model to train the sequence machine learning model; and recommending exercises to the user using the trained sequence machine learning model.

Description

Technical Field

[0001] This invention belongs to the field of machine learning technology, specifically relating to an online learning recommendation method and system based on graph neural networks. Background Technology

[0002] Adaptive learning systems aim to provide personalized recommendations to students. Online learning platforms (MOOCs) such as Udemy and Coursera are becoming increasingly popular on the internet. They provide adaptive learning systems with large amounts of data, especially deep learning models. Adaptive learning systems learn a student's potential knowledge state based on the deep learning model and recommend topics or exercises for the student to study next. However, existing technologies often use one-hot encoding of topics and exercises as input, ignoring the dependencies between exercises and topics. Summary of the Invention

[0003] According to one aspect of the present invention, an online learning recommendation method is provided, comprising:

[0004] Collect user interaction data with the online learning system, including data on user answers to exercises and corresponding scores;

[0005] Based on the graph structure, the interactive data is used for hierarchical aggregation to obtain an overall vector. The graph structure includes multiple nodes and edges. Each node indicates a topic and exercises belonging to that topic. Each edge connects two nodes, and the node indicating the previous topic points to the node indicating the subsequent topic through the edge.

[0006] The overall vector is input into the sequence machine learning model to train the sequence machine learning model; and

[0007] Use the trained sequence machine learning model to recommend exercises to users.

[0008] According to an embodiment of the present invention, the hierarchical aggregation includes:

[0009] For each topic, aggregate all practice vectors belonging to that topic;

[0010] For each topic, the topic's vector is updated and aggregated based on its prerequisite topics. Each topic's vector includes the exercises belonging to that topic and information about its prerequisite topics.

[0011] Aggregate the vectors of all topics into the overall vector.

[0012] According to an embodiment of the present invention, the sequence machine learning model includes long short-term memory and / or gated recurrent units.

[0013] According to an embodiment of the present invention, the online learning recommendation method further includes dividing the interaction data into multiple subsequences and performing the hierarchical aggregation for each subsequence, wherein each subsequence is formed by merging the subsequence at the previous time point and the answer record at the current time point.

[0014] According to an embodiment of the present invention, a graph neural network model is used to update and aggregate the vector of each topic through its predecessor topics.

[0015] According to an embodiment of the present invention, a graph attention network is used to aggregate the vectors of all topics into the overall vector.

[0016] According to another aspect of the present invention, an online learning recommendation system utilizing the above-described method is provided. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the technical solutions of this disclosure and form part of the specification. They are used together with the embodiments of this disclosure to explain the technical solutions of this disclosure and do not constitute a limitation on the technical solutions of this disclosure. The shapes and sizes of the components in the drawings do not reflect actual proportions and are only intended to illustrate the content of this disclosure.

[0018] Figure 1 An example of an online course is shown.

[0019] Figure 2 This illustrates a representation using a graph structure. Figure 1 The example shown is an online course.

[0020] Figure 3 A flowchart of an online learning recommendation method according to an embodiment of the present invention is shown. Detailed Implementation

[0021] In many e-learning systems, recommender systems use collaborative filtering. However, the efficiency of these algorithms decreases significantly with the increase in the number of data records. Deep learning models are often used in adaptive learning recommender systems. Deep learning models take all of a student's answer records as input and predict the accuracy of their answers, then recommend exercises to the student based on the predicted probabilities. Each student's answer record includes the exercises they answered and their scores. Common approaches use various deep sequence models, such as recurrent neural network (RNN) models, including Long Short-Term Memory (LSTM) units, which capture sequence information to generate hidden knowledge state vectors for prediction.

[0022] Existing recommender systems often fail to consider the dependencies between exercises and topics. Some exercises belong to the same topic. Some topics require prior learning and are prerequisites for certain topics. For example, when using one-hot encoding to represent exercise responses, this representation always has limited information content. Using only the exercise representation will lose relevant information between exercises and topics, as well as between topics themselves.

[0023] like Figure 1 As shown, online course 101 typically includes a list of topics 102, 103…108, and each topic contains some learning materials and multiple exercises. For example, topic 108 includes learning material 110 and exercises 111, 112, and 113. Furthermore, there may be some dependencies between topics; some topics may reference previous topics. For example, topic 103 references topic 102, topic 104 references topic 101, topic 107 references topics 105 and 106, and topic 108 references topics 104 and 107.

[0024] In view of the above problems, this invention proposes an online learning recommendation method and system based on a deep learning framework using adaptive learning with graph neural networks (GNNs). The GNNs in this framework can capture rich relational information between exercises and topics in typical online learning platforms. Therefore, the accuracy of the recommendation system is improved more effectively. Online learning platforms can utilize this invention for personalized recommendations and for modeling students' learning states.

[0025] In this invention, a directed acyclic graph (DAG) structure called a knowledge graph is used. Figure 2 This illustrates a representation using a graph structure. Figure 1 The example shown is an online course. The knowledge graph specifies which exercises are performed under each topic, and which prerequisite topics need to be learned before each topic.

[0026] like Figure 3 As shown, the online learning recommendation method according to an embodiment of the present invention includes: step S1, collecting interaction data between the user and the online learning system, the interaction data including data on the user's answers to exercises and corresponding scores; step S2, based on a graph structure, performing hierarchical aggregation using the interaction data to obtain an overall vector, wherein the graph structure includes multiple nodes and edges, each node indicating a topic and exercises belonging to that topic, each edge connecting two nodes, and a node indicating a previous topic pointing to a node indicating a subsequent topic through an edge; step S3, inputting the overall vector into a sequence machine learning model to train the sequence machine learning model; and step S4, using the trained sequence machine learning model to recommend exercises to the user.

[0027] In embodiments of this disclosure, the hierarchical aggregation includes: for each topic, aggregating all practice vectors belonging to that topic; for each topic, updating and aggregating the vectors of that topic through its predecessor topics, wherein the vector of each topic includes the practice belonging to that topic and information about the predecessor topics of that topic; and aggregating the vectors of all topics into the overall vector.

[0028] Some methods do acknowledge the importance of the pre-existing topic structure, but they fail to consider the representation of different levels of knowledge as input to deep sequence models. This invention, however, considers representing students' knowledge states through various aggregation methods, including graph neural networks (GNNs). This invention presents a deep learning framework that uses GNNs for hierarchical aggregation in a knowledge graph. This framework can capture more information than a simple practice representation. The framework of this invention aggregates information hierarchically to generate a vector of student knowledge states: (1) aggregation from the practice level to the topic level; (2) neighborhood aggregation at the topic level using a GNN; and (3) aggregation from the topic level to the global level. The aggregation method and the GNN used in these three steps are optional, as model accuracy may vary depending on the dataset. This invention provides developers with flexibility in model accuracy and model selection. For example, if the knowledge graph is dense and large, GraphSAGE can be chosen instead of GNN.

[0029] The following describes the operational flow of the deep learning framework of this invention. First, it explains how to collect data for training. Second, it describes the data structure used to represent the knowledge graph. Finally, it provides an example framework that hierarchically summarizes information and the structure of the knowledge graph.

[0030] Data collection

[0031] The online platform must record student interactions to train deep learning models. This data collection process is similar to existing technologies. Table 1 shows an example of the collected data. Each student's training sequence was transformed using SQL commands, where the table is grouped by UserID and then sorted in ascending order by time. Training Sequences

[0032] I={(q1,s1),(q2,s2),...,(q T s T )}

[0033] Let q represent a student record of length T, where q t and s t This represents the sum of scores for the exercises answered at time t. Note that s t It is standardized to a value between 0 and 1.

[0034] Table 1

[0035]

[0036] Knowledge Graph

[0037] Tables 2 and 3 are used to represent the knowledge graph. Table 2 stores the Exercise IDs of all exercises under each Topic ID. Table 3 stores the correspondence between the preceding Topic ID1 and the subsequent Topic ID2, serving as the edge set of the knowledge graph.

[0038] Table 2

[0039]

[0040] Table 3

[0041]

[0042] The knowledge graph is represented by G = (V, E), where V is the set of topics in the graph, E is the set of edges for all (u, v), and u is the precondition of topic v. This represents all the presuppositions of topic v in the diagram.

[0043] Hierarchical aggregation

[0044] First, the training sequence I = {(q1, s1), (q2, s2), ..., (q...} is... T s T This is converted into the input of a sequence machine learning model. Typically, a vector x is used. t To represent the answer record (q) t s t If the total number of exercises m in the course is small, and the score s t If x has only a correct or incorrect value, then x t ∈{0,1} 2m Set it to a one-hot vector. If the score s t The value of x ranges from 0 to 1. t =s t ·e t , with e t Multiply by the ratio s t View as an answer q t The vector of "knowledge acquisition" afterward, where e t ∈R d It means practice q t The embedding vector is d, where d is the dimension.

[0045] Existing techniques directly convert the transformed sequence I′={x1,x2,...,x T-1 The input sequence is a machine learning model. The framework of this invention first divides I′ into subsequences I′ of length t. t ={x1, x2, ..., x t Let t = 1, ..., T-1. Each subsequence represents an independent state at a specific time point. Each subsequence is formed by merging the subsequence from the previous time point and the answer record from the current time point.

[0046] For each subsequence I′ t The hierarchical aggregation specified in (1), (2), and (3) is executed. `Aggregate` represents a differential aggregation method, such as mean, sum, and maximum. `Update` and `Message` represent distinguishable features, such as a fully connected neural network. Users can set different aggregation methods based on efficiency and accuracy. We have three layers of aggregation:

[0047] 1. For each topic v, according to formula (1), aggregate all practice vectors x∈I′ belonging to v. t If the corresponding exercise for x belongs to topic v, then pointer function 1 v (x) returns 1. Otherwise, it returns 0.

[0048] 2. For each topic v, according to formula (2), through its presupposition topic v. To update and aggregate the vector representing topic v Equation (2) is a generalized GNN model. The model can be stacked up to L layers, where k = 1, ..., L-1 and ...

[0049] 3. Similar to step 1, according to formula (3), aggregate all topic vectors into a single vector h. G，t This aggregation can be performed using GAT. GAT can adaptively find the weights of topics in the knowledge graph, which is more interpretable and can be visualized.

[0050]

[0051]

[0052]

[0053] At this stage, the overall vector h G，t It will capture the rich information in the knowledge graph. Finally, input {h c，1 h G，2 , ..., h G，T-1 Further optimization of the sequence model is then performed.

[0054] Personalized recommendations

[0055] The following example illustrates how to train the entire neural network, and how to train the overall vector h. G，t Personalized recommendations for practice. For example, using a gated recurrent unit (GRU) as a sequence neural network model, according to formula (4), the overall vector sequence {h} is... G，1 h G，2 , ..., h G，T-1} is encoded into a new sequence {z1, z2, ..., z} T-1}, where d′ is the hidden dimension of GRU.

[0056]

[0057] According to formula (5), a simple fully connected neural network Θ is used to predict whether the student will answer the next exercise q correctly. t+1 The probability is given by , where || is the concatenation operation, t = 1, ..., T-1. Then the predicted probability can be... With real label y t+1 The comparisons are performed, and the binary cross-entropy loss in Equation (6) is used to train the entire network.

[0058]

[0059]

[0060] After the training phase, the model can be used to recommend a series of exercises, such as the first K exercises with the highest probability according to formula (7).

[0061]

[0062] This invention provides a deep learning framework for adaptive learning using graph neural networks (GNNs). The GNNs in this framework can capture rich relational information between exercises and topics in typical MOOCs. Therefore, the accuracy of recommendation systems is improved more effectively. Online learning platforms can utilize this invention for personalized recommendations and for modeling students' learning states.

[0063] It is understood that the above embodiments are merely exemplary implementations used to illustrate the principles of the present invention, and the present invention is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. An online learning recommendation method, comprising: Collect user interaction data with the online learning system, including data on user answers to exercises and corresponding scores; Based on the graph structure, the interactive data is used for hierarchical aggregation to obtain an overall vector. The graph structure includes multiple nodes and edges. Each node indicates a topic and exercises belonging to that topic. Each edge connects two nodes, and the node indicating the previous topic points to the node indicating the subsequent topic through the edge. The overall vector is input into the sequence machine learning model to train the sequence machine learning model; and The trained sequence machine learning model is used to recommend exercises to users. The hierarchical aggregation includes: For each topic, aggregate all practice vectors belonging to that topic; For each topic, the topic's vector is updated and aggregated based on its prerequisite topics. Each topic's vector includes the exercises belonging to that topic and information about its prerequisite topics. Aggregate the vectors of all topics into the overall vector. The above hierarchical aggregation is specifically as follows: For each topic According to formula (1), aggregate all those belonging to Practice vectors ,if The corresponding exercises belong to the topic. Then pointer functions Returns 1 otherwise, returns 0. （1） For each topic According to formula (2), through its presupposition... To update and aggregate representative topics vector Formula (2) is a generalized GNN model, and the model can be stacked to... Layer, in which and , According to formula (3), all topic vectors are aggregated into a single vector. :

2. The online learning recommendation method according to claim 1, wherein, The sequence machine learning model includes long short-term memory and / or gated recurrent units.

3. The online learning recommendation method according to claim 1 further includes dividing the interaction data into multiple subsequences and performing the hierarchical aggregation for each subsequence, wherein each subsequence is formed by merging the subsequence at the previous time point and the answer record at the current time point.

4. An online learning recommendation system utilizing the method of claim 1.

Images