A method for predicting interactive behavior of digital teaching materials based on time series data
By using recurrent neural network model update and projection operations in interactive behavior prediction of digital textbooks, combined with static and dynamic embedding, the problem of failure to consider behavior interval time in traditional methods is solved, the prediction accuracy is improved, and the teacher's teaching process is improved digital textbook usage effect.
Patent Information
- Application Number
- CN202410575326.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-10
- Publication Date
- 2024-11-26
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional digital textbook interactive behavior prediction methods fail to fully consider the interval time between teachers' behavior, resulting in low accuracy of prediction results and affecting teachers' teaching quality.
Two interdependent recurrent neural network models (RNNU and RNNI) are used for embedding update and projection operations, combining static and dynamic embedding, interactive behavior prediction is performed through a fully connected linear layer, taking into account the teacher's personal static characteristics and behavior interval time.
It improves the accuracy of teachers' interactive behavior prediction, enhances the ability to predict teachers' behavior trajectories, and improves the effectiveness of using digital textbooks in the teaching process.
Smart Images

Figure CN118469071B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of behavior prediction, and in particular to a method for predicting interactive behavior of digital teaching materials based on time series data. Background Art
[0002] With the development of technology, more and more teachers will use smart devices (with digital teaching materials inside) to assist teaching in the classroom. However, ordinary smart devices require teachers to perform corresponding operations during classroom teaching before they receive corresponding instructions. For example, when using projection equipment for teaching, teachers usually need to click the mouse or other wireless devices to switch the screen, even if the digital teaching materials change accordingly. However, this method will make students pay more attention to the teacher's actions, and these actions will also affect the quality of the teacher's teaching.
[0003] For the above reasons, developers of digital textbooks have introduced related technologies of behavior prediction in the control of digital textbooks. The traditional method of command input (control command) in behavior prediction only generates embedding (command corresponding to the behavior) when the user takes a certain action, and does not explicitly model the future trajectory of the user / behavior in the embedding space.
[0004] In other words, since the corresponding control instructions are generated only when the user takes an action (similar to simple gesture control, the device generates the corresponding control instructions when a certain characteristic gesture is recognized), if the teacher does not take any new actions for a period of time, the control instructions will remain unchanged (or stop inputting), and the behavior predictions given will also be the same. However, what the teacher will do next is related to the time interval between the previous and next actions, so the prediction results using this method are less accurate.
[0005] That is, traditional generative embedding methods often only consider one of the static and dynamic features of the prediction object. However, in the actual teaching process, teachers' digital teaching material usage behavior is not only affected by static features such as their personal teaching style, but also related to the interval time between their behaviors. Ignoring any part of the teacher's characteristics will affect the accuracy of the prediction.
[0006] Therefore, there is an urgent need for a digital teaching material interactive behavior prediction method based on time series data that can fully consider the interval time between teachers' behaviors and improve the accuracy of teacher interactive behavior prediction results when teachers use digital teaching materials for teaching. Summary of the invention
[0007] The present invention provides a method for predicting interactive behavior of digital teaching materials based on time series data, which can fully consider the interval time between teachers' behaviors when teachers use digital teaching materials for teaching, and improve the accuracy of the prediction results of teachers' interactive behavior.
[0008] In order to solve the above technical problems, this application provides the following technical solutions:
[0009] A method for predicting interactive behavior of digital teaching materials based on time series data, including the following contents:
[0010] S1. Building the basic model: Building two interdependent recurrent neural network models, namely RNN U and RNN I , where RNN U For updating the teacher embedding, RNN I It is used to update the embedding of behavior; the embedding in this technology includes dynamic embedding and static embedding; static embedding uses one-hot encoding, which is represented as and u(t) represents the dynamic embedding of teacher u at time t, i(t) represents the dynamic embedding of behavior i at time t, and the feature vector f represents the features of each interaction;
[0011] S2. User update embedding and behavior update embedding: RNN U For updating the user embedding, RNN I Used to update the embedding of behavior. When teacher u interacts with behavior i at time t, RNN U Using the feature vector f at time t and the dynamic embedding of the previous behavior i at time t as input, RNN updates u(t). I Update i(t) using the feature vector f at time t and the dynamic embedding of the teacher u before time t as input;
[0012] S3. Perform projection after each update: After time t, after a period of time Δ, the embedding of time t+Δ is performed. Perform the projection operation. As Δ increases (i.e., Δ1<Δ2), the projected embedding will be farther away from the input embedding until a new interaction is observed, and then the embedding is updated, so as to achieve the prediction of the teacher's behavior back and forth;
[0013] S4. Interaction behavior prediction: Use the projected embedding u(t+Δ) of the teacher u at time t+Δ and the dynamic embedding i(t+Δ-) and static embedding of the previous behavior i at time t+Δ to predict the interaction behavior of u at time t+Δ. This process is performed using a fully connected linear layer.
[0014] S5. Model training: During training, each interaction model will minimize the model's loss value. The loss value represents the gap between the predicted value and the actual value. The smaller the loss value, the better the robustness of the model. The calculation formula for the total training loss value is as follows:
[0015]
[0016] λ U and λ I Represents a scaling parameter, which is used to control the loss value within a certain range.
[0017] The basic scheme principle and beneficial effects are as follows: teachers are divided into one set, and the functions of digital teaching materials are divided into another set. Then, when teachers are operating digital teaching materials, there is interaction between teachers and digital teaching materials. The key problem to be solved by prediction technology is to predict the possibility of teachers performing a certain behavior at a certain moment in the future through the historical interaction behavior between teachers and digital teaching materials.
[0018] In the present invention, two interdependent recurrent neural networks (RNNs) are used to perform embedding update and embedding projection operations. A recurrent neural network (RNN) is a neural network model that models time dependencies of time series data, and can depict the mutual connection between the previous and next data in the time series. Each input and output of RNN is a time series, and the input at each moment includes not only the input information x at the current moment, but also the input information x at the current moment. t , and also includes the implicit state s of the previous moment t-1 The information stored in it. RNN also has memory capabilities. Vectors encoded using one-hot encoding do not contain the current state of the encoded object, and when the real data set has millions of objects, the dimension of the one-hot vector will be very large, making the model difficult to train and expand. Therefore, it is used in the present invention to express static attributes, such as the long-term interests of users. If there are no such unchanging features, it may not be needed.
[0019] At each moment there is a hidden state {s1,s2,…,s t ,…} can store the information that has been involved in the calculation before time t and use it as the input for the next cycle, so that the time series can be predicted based on the current implicit state using the fully connected layer neural network. RNN no longer needs to recalculate the previous data each time it processes, which can greatly improve the calculation speed of RNN. Embedding at time t+Δ Performing a projection operation, as Δ increases (i.e., Δ1<Δ2), the projected embedding will be farther away from the input embedding, where the distance refers to the projection distance of the two interacting features.
[0020] When a user takes an action, the embeddings of the user and the action are updated. The update operation uses the interaction to update the interaction state of the user and the action, and the projection operation can use the previous observed state and the elapsed time to predict the future embedding of the user. Embedding is a way to encode the user and the action.
[0021] When predicting the interaction trajectory between teachers and digital teaching materials, the present invention fully considers the personal static characteristics of teachers (such as the school where the teacher is located, gender, teaching style, and subjects taught). After each update, a projection operation is performed, and the projection embedding will be farther and farther away from the input embedding until a new interaction is observed, and then the embedding update is performed. In this way, the prediction of the interaction between teachers and digital teaching materials is realized back and forth, so that when teachers use digital teaching materials for teaching, the interval time between teachers' behaviors is fully considered, thereby improving the accuracy of the prediction results of teachers' interactive behaviors.
[0022] Further, the calculation method of u(t) and i(t) is as follows:
[0023]
[0024]
[0025] W1 in Equation 1 and Equation 2 (u) 、W2 (u) 、W3 (u) 、W4 (u) Refers to RNN U The parameter matrix of (i) 、W2 (i) 、W3 (i) 、W4 (i) Representing RNN I The parameter matrix of u represents the time that teacher u interacts with any behavior i before time t; Δ i represents the time that behavior i interacts with any teacher u before time t; σ represents the nonlinear sigmoid function, which maps the input to the output range of (0,1).
[0026] The above method can improve the accuracy of teacher behavior prediction.
[0027] Further, the The calculation method is as follows:
[0028]
[0029] Where w = W p Δ,w is a temporal background vector, W P represents the parameter matrix.
[0030] Furthermore, in S4, represents the behavior embedding at time t+Δ, which is calculated as follows:
[0031]
[0032] Among them, W1, W2, W2, W2 represent parameter matrices; B represents the bias vector, which is used to control the activation of neurons. The parameter matrix and bias vector constitute the linear layer of the model.
[0033] The above method can fully consider the influence of the time interval between teachers' behaviors on the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 A flowchart of an embodiment of a method for predicting interactive behavior of digital teaching materials based on time series data;
[0035] Figure 2 This is a prediction model diagram of a digital teaching material interactive behavior prediction method based on time series data. DETAILED DESCRIPTION
[0036] The following is further described in detail through specific implementation methods:
[0037] A method for predicting interactive behavior of digital teaching materials based on time series data (such as Figure 1 ), including the following:
[0038] S1. Building a basic model: Building two interdependent recurrent neural network models (such as Figure 2 As shown), respectively, RNN U and RNN I , where RNN U For updating the teacher embedding, RNN I It is used to update the embedding of behavior; the embedding in this technology includes dynamic embedding and static embedding; static embedding uses one-hot encoding, which is represented as and u(t) represents the dynamic embedding of teacher u at time t, i(t) represents the dynamic embedding of behavior i at time t, and the feature vector f represents the features of each interaction;
[0039] S2. User update embedding and behavior update embedding: RNN U Used to update the user embedding (updating the embedding of user u at time t, which is equivalent to adding a new embedding to the original user set), RNN I It is used to update the embedding of the behavior (the embedding of behavior i at time t is updated, which is equivalent to adding a new embedding to the original item set). When teacher u interacts with behavior i at time t, RNN U Using the feature vector f at time t and the dynamic embedding of the previous behavior i at time t as input, RNN updates u(t). I Use the feature vector f at time t and the dynamic embedding of the teacher u before time t as input to update i(t). The calculation methods of u(t) and i(t) are as follows:
[0040]
[0041]
[0042] W1 in Equation 1 and Equation 2 (u) 、W2 (u) 、W3 (u) 、W4 (u) Refers to RNN U The parameter matrix of (i) 、W2 (i) 、W3 (i) 、W4 (i) Representing RNN I The parameter matrix of u represents the time that teacher u interacts with any behavior i before time t; Δ i represents the time that behavior i interacts with any teacher u before time t; σ represents the nonlinear sigmoid function, which maps the input to the output range of (0,1);
[0043] S3. Perform a projection operation after each update (that is, after the update operation, continue to predict the next interaction behavior based on the change in the interval time): After time t, after a period of time Δ, the model will embed the time t+Δ Perform the projection operation. As Δ increases, that is, Δ1<Δ2, the projected embedding will be farther away from the input embedding until a new interaction is observed, and then the embedding is updated, so as to achieve the prediction of the teacher's behavior back and forth. The calculation method is as follows:
[0044]
[0045] Where w = W p Δ,w is a temporal background vector, W P represents the parameter matrix;
[0046] S4. Interaction behavior prediction: The interactive behavior prediction of u at time t+Δ is realized by using the projected embedding u(t+Δ) of the teacher u at time t+Δ and the dynamic embedding i(t+Δ-) and static embedding of the previous behavior i at time t+Δ. This process is performed using a fully connected linear layer. represents the behavior embedding at time t+Δ, which is calculated as follows:
[0047]
[0048] Where W1, W2, W2, and W2 represent parameter matrices; B represents a bias vector, which is used to control the activation of neurons. The parameter matrix and the bias vector constitute the linear layer of the model.
[0049] S5. Model training: During training, each interaction model will minimize the model's loss value. The loss value represents the gap between the predicted value and the actual value. The smaller the loss value, the better the robustness of the model. The calculation formula for the total training loss value is as follows:
[0050]
[0051] λ U and λ I Represents a scaling parameter, which is used to control the loss value within a certain range.
[0052] When used specifically: Let the set of teachers be U = {u1,u2,…,u n}, where n represents the number of teachers; the digital teaching material function set is recorded as I = {i1,i2,…,i m}, where m represents the number of digital textbook functions. Then, an interaction between a teacher and a digital textbook can be recorded as e = (u, i, t), that is, teacher u interacted with function i of the digital textbook at time t. All historical interactions of teacher u are represented in chronological order as a set k represents that teacher u has k interaction events in the historical time interaction network, and the prediction problem can be transformed into a historical interaction sequence based on the teacher and the digital textbook. Predicting teachers' future behavior sequences Right now The interaction between n teachers and m digital teaching materials functions that changes over time is defined as a time interaction network in this embodiment.
[0053] In this embodiment, f is the interactive feature vector; u(t) represents the dynamic embedding of teacher u at time t, and i(t) represents the dynamic embedding of behavior i at time t. The teacher and behavior each have a static embedding and Some fixed characteristics that characterize teachers and behaviors respectively (characteristics that do not change naturally over time, such as the teacher's school, gender, teaching style, subjects taught, etc.).
[0054] That is, this embodiment constructs two interdependent recurrent neural network models, namely RNN U and RNN I , to perform embedding update and embedding projection operations to predict the interaction trajectory between teachers and digital teaching materials. In specific use, e is used to generate the dynamic embedding of teachers and behaviors at the corresponding moment. When teacher u interacts with behavior i at time t, RNN U The RNN will use the feature vector f at time t and the embedding of the behavior i before time t as input to update u(t). IThe feature vector f at time t and the embedding of the teacher u before time t are used as input to update i(t).
[0055] After each update, a projection operation is performed, that is, within the Δ time after time t, the embedding of time t+Δ As Δ(Δ1<Δ2) increases, the projected embedding will be farther away from the input embedding until a new interaction is observed, and then the embedding is updated, thus realizing the prediction of the interaction between teachers and digital teaching materials in a reciprocating manner.
[0056] In specific use, for example, when teacher A uses digital teaching materials to teach in class, when interacting with the digital teaching materials, there will be behavior X in advance, and then the digital teaching materials need to embed behavior Y after behavior X appears, so that the digital teaching materials can perform function Y (that is, the digital teaching materials need to predict the subsequent behavior Y that may appear after behavior X appears, that is, predict what behavior teacher A may have next). In the above scenario, the static embedding of teacher A and behavior X (which can be the type of subject taught by the teacher and the type of behavior, because the subject taught by the teacher may affect the teacher's preference and habit of using the digital teaching material function. If there is no such embedding in the specific scenario, it can be omitted) is represented as and In this embodiment, if teacher A has a record, then the subsequent processing is directly performed. If there is no record, it is necessary to perform a user update operation for teacher A. Similarly, the same is true for behavior X.
[0057] Then, after each update, a projection operation is performed. After a period of time Δ after time t, the embedding of time t+Δ is performed. During the projection operation, as Δ increases, that is, Δ1<Δ2, the projected embedding will be farther away from the input embedding until a new interaction is observed, and then the embedding is updated, so as to achieve the prediction of the teacher's behavior back and forth.
[0058] The above is only an embodiment of the present invention. The invention is not limited to the field involved in this implementation case. The common sense such as the known specific structure and characteristics in the scheme is not described too much here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can know all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to explain the content of the claims.
Claims
1. A method for predicting interactive behavior of digital teaching materials based on time series data. It is characterized in that It includes the following: S1. Building the basic model: Building two interdependent recurrent neural network models, namely RNN U and RNN I , where RNN U For updating the teacher embedding, RNN I It is used to update the embedding of the behavior; the embedding in this technology includes dynamic embedding and static embedding; static embedding uses one-hot encoding, and the teacher and behavior each have a static embedding and , respectively, characterize some fixed characteristics of teachers and behaviors, where fixed characteristics are those that do not change naturally over time; u ( t )express t Teacher of the moment u Dynamic embedding, i ( t )express t Moment Behavior i Dynamic embedding, feature vector f Represent the characteristics of each interaction; S2. User update embedding and behavior update embedding: RNN U For updating the user embedding, RNN I Embedding for updating behavior, when t Teacher of the moment u and behavior i When interaction occurs, RNN U use t The feature vector at time f and t Behavior before the moment i Dynamic embedding of as input update u ( t ), RNN I use t The feature vector at time f and t Teacher before the time u Dynamic embedding of as input update i ( t ); S3. Perform projection after each update: t After a certain time interval, Δ t +Δ time embedding Perform a projection operation. As Δ increases, the projected embedding will be farther away from the input embedding until a new interaction is observed, and then the embedding is updated. This is a reciprocating process to predict the teacher's behavior. S4. Interaction Behavior Prediction: Use t +ΔTime Teacher u Projection embedding and Behavior before the moment i Dynamic embedding i ( t +Δ-) and static embedding to achieve u exist t +Δ interactive behavior prediction, this process is carried out using a fully connected linear layer; S5. Model training: During training, each interaction model will minimize the model's loss value. The loss value represents the gap between the predicted value and the actual value. The smaller the loss value, the better the robustness of the model. The calculation formula for the total training loss value is as follows: (5) λ U and λ I Represents the scaling parameter, which is used to control the loss value within a certain range; Said u ( t )and i ( t ) is calculated as follows: (1) (2) In formula 1 and formula 2 W 1 (u) , W 2 (u) , W 3 (u) , W 4 (u) Refers to RNN U The parameter matrix of W 1 (i) , W 2 (i) , W 3 (i) , W 4 (i) Representing RNN I The parameter matrix of u Indicates teacher u exist t Before the moment and any behavior i the time at which the interaction took place; Δ i Expressing behavior i exist t Before time with any teacher u The time when the interaction occurs; σ represents the nonlinear sigmoid function, which maps the input to the output range of (0,1); The calculation method is as follows: (3) in w= , w is a time background vector, W P represents the parameter matrix; In S4, use express t The behavior embedding at time +Δ is calculated as follows: (4) in W 1 、W 2 、W 2 、W 2 represents the parameter matrix; B Represents the bias vector, which is used to control the activation of neurons. The parameter matrix and the bias vector constitute the linear layer of the model.
