Intelligent behavior prediction method and system based on time interest network

Abstract

The invention discloses an intelligent behavior prediction method and system based on a time interest network, and relates to the technical field of software engineering, and the method comprises the steps: receiving a behavior sequence, dividing the behavior sequence through different time periods, and obtaining a behavior sequence of each time period, inputting the behavior sequence of each time period into a pre-established multi-head self-attention mechanism and a position coding model, and outputting to obtain an interest feature vector of each time period; inputting the interest feature vector of each time period into a pre-established Bi-LSTM model, outputting to obtain an interest sequence result, sequentially inputting the interest sequence result into a pre-established residual network and a Dropout layer to obtain an optimized interest sequence result, inputting the optimized interest sequence result into a pre-established MLP layer, and outputting to obtain a behavior prediction result. And the precision and the pre-estimation capability of intelligent behavior prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of software engineering, and in particular to an intelligent behavior prediction method and system based on a temporal interest network. Background Art

[0002] The establishment of a behavior prediction model relies on the mining and analysis of a large amount of user driving behavior data. By analyzing the driver's historical behavior data, the user's potential interests and preferences under different driving conditions can be discovered, thereby improving the accuracy and personalization of the prediction.

[0003] In the prediction scenario of user driving behavior, the model composed of local activation units is now commonly used. At different time nodes, the driver's behavior will be different. For example, in the daytime when the light is sufficient, the driver does not need to use too many lights, and in the night when the light is insufficient, the driver will frequently use light signals. We can call this type of behavior driving behavior with strong environmental attributes. In some driving behaviors with obvious environmental attributes, the contextual feature is regarded as one of the ordinary sparse features, spliced ​​with other features, and then passed to the MLP layer for learning. This method does not consider the relationship between environmental factors and driving behavior in the user's historical driving behavior data, ignores the environmental attributes of driving behavior, and thus limits the learning and expression ability of the model. However, traditional behavior prediction models usually have some limitations. They usually simply regard the user's driving behavior as an independent item, ignoring the mutual influence of driving behavior and the evolution of user's interests. In addition, traditional models have not fully considered the contextual information of user driving behavior, which leads to inaccuracy and limitations in prediction. At present, research on modeling the sequence of user driving history behavior has been in full swing, but some intelligent behavior prediction systems based on interest drift do not consider the relationship between user driving behaviors. Summary of the invention

[0004] In order to solve the deficiencies mentioned in the above background technology, the purpose of the present invention is to provide an intelligent behavior prediction method and system based on a temporal interest network, which improves the accuracy and prediction ability of intelligent behavior prediction.

[0005] In the first aspect, the purpose of the present invention can be achieved by the following technical solution: an intelligent behavior prediction method based on a time interest network, the method comprising the following steps: Receive a behavior sequence, divide the behavior sequence into different time periods, obtain a behavior sequence for each time period, input the behavior sequence for each time period into a pre-established multi-head self-attention mechanism and position encoding model, and output an interest feature vector for each time period, wherein the behavior sequence is the driver's behavior actions within a preset time period; The interest feature vector of each time period is input into the pre-established Bi-LSTM model, and the interest sequence result is output. The interest sequence result is sequentially input into the pre-built residual network and Dropout layer to obtain the optimized interest sequence result. The optimized interest sequence result is input into the pre-built MLP layer to obtain the behavior prediction result.

[0006] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the process of dividing the behavior sequence into different time periods: Use the last action as a reference point , and set the length of the time window in segments from near to far, and express a certain behavior time point as .

[0007] In combination with the first aspect, in some implementations of the first aspect, the method further includes: using formula (1) to To express: (1) In the formula, is the time window, is the initial time window length, is the growth rate of the time window.

[0008] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established multi-head self-attention mechanism and position encoding model are as follows: Using self-attention with 2 attention heads, the calculation is as follows: Among them, the query matrix , Keyword Matrix and information matrix The calculation is as follows: (2) In formula (2), is included The linear matrix of weight vectors, represents the number of queries, keywords, and information vectors in each attention head, represents the input sequence, Respectively represent The query, keyword and information matrix of each head, Respectively represent Head The calculation of the query, Respectively represent Head The calculation of keywords, Respectively represent Head The calculation of information, Respectively represent query weight vector, Respectively represent keyword weight vector, Respectively represent Information weight vectors are obtained by multiplying the input sequence by different weight matrices; In the model, ; in represents the attention head, represents the attention head matrix, which is a collection of multiple matrices. refers to the first attention head matrix, It refers to An attention head matrix, where Represents the index of the header, It refers to An attention head matrix, where refers to the number of attention heads, Indicates The dimension of the attention head matrix is is the length of the input sequence, is the dimension of the attention head, and , represents the dimension of the hidden units embedded in the model, is the dimension of the hidden units in the model, is the length of the embedding; head The expression is: (3) in The function represents the self-attention mechanism, The function is used to convert a real vector into a probability distribution. Representation Matrix The transpose of Representation Matrix The transpose of The vectors of different attention heads are connected and input into a feedforward neural network for calculation. The calculation expression is as follows: (4) (5) In formula (4) and (5), It refers to a feed-forward neural network. Indicates shared Attention head, Respectively represent attention head vector, The function concatenates vectors. Represents the weight matrix, by The heads are flattened and connected, multiplied by the weight matrix, to obtain the final calculation structure of multi-head self-attention. Represents the output after the above operation, and then normalizes each layer layer by layer to calculate the User interests in a certain time period , where the variable Represents the average pooling process.

[0009] In combination with the first aspect, in certain implementations of the first aspect, the method also includes: inputting the behavior sequence of each time period into a pre-established multi-head self-attention mechanism and position encoding model by vectorizing the initial features of the input, converting the sparse high-dimensional features using one-hot encoding and muti-hot encoding, and mapping the high-dimensional sparse features to a low-dimensional vector space through an embedding layer, wherein the initial features include user behavior sequences, time features, user features, and target features.

[0010] In combination with the first aspect, in certain implementations of the first aspect, the method further includes: in the embedding layer, the high-dimensional sparse features generated by the input layer are converted into low-dimensional vectors through an embedding operation, and features from different sources are embedded for representation, wherein the features from different sources include user behavior sequences, time features, user features, and target features, and the feature embedding results from different sources are fused, including a combination of user behavior and time features, and a combination of user features and target features, so as to generate a final feature representation as the feature vector of interest for each time period.

[0011] In combination with the first aspect, in some implementations of the first aspect, the method further includes: the pre-established Bi-LSTM model is The hidden state The expression is as follows: (6) In formula (6), is the hidden state of the forward LSTM, is the hidden state of the reverse LSTM, and ⊕ represents the XOR operation on the two hidden state vectors. The resulting expression after the target is applied to the sequence of interest is as follows: (7) (8) In formula (7) and formula (8), Represent two attention heads respectively, Indicates elements, Respectively refer to Elements in The attention scores in the attention heads, Respectively refer to The user interest vector and hidden state vector of elements, Respectively The weight matrix of the attention head, Respectively Attention head target behavior and corresponding user embedding, Indicates total elements, represents the exponential function, Respectively The corresponding output of the attention head.

[0012] In a second aspect, in order to achieve the above-mentioned purpose, the present invention discloses an intelligent behavior prediction system based on a temporal interest network, comprising: A feature extraction module is used to receive a behavior sequence, divide the behavior sequence into different time periods, obtain a behavior sequence for each time period, input the behavior sequence for each time period into a pre-established multi-head self-attention mechanism and position encoding model, and output an interest feature vector for each time period, wherein the behavior sequence is the driver's behavior actions within a preset time period; The behavior prediction module is used to input the interest feature vector of each time period into the pre-established Bi-LSTM model, output the interest sequence result, input the interest sequence result into the pre-built residual network and Dropout layer in sequence, obtain the optimized interest sequence result, input the optimized interest sequence result into the pre-built MLP layer, and output the behavior prediction result.

[0013] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a terminal device is disclosed, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, an intelligent behavior prediction method based on a time interest network as described above is adopted.

[0014] In another aspect of the present invention, in order to achieve the above-mentioned purpose, a computer-readable storage medium is disclosed, in which a computer program is stored. When the computer program is loaded and executed by a processor, the intelligent behavior prediction method based on a time interest network as described above is adopted.

[0015] Beneficial effects of the present invention: The present invention combines the self-attention mechanism and Bi-LSTM, and optimizes the algorithm by adding a combination of residual network and fully connected layer. The introduction of random dropout regularization (Dropout) makes the neural network not overly dependent on local features, thereby improving the generalization ability as a whole and reducing the risk of overfitting. Finally, the method is used for training through a multi-layer perceptron to further improve the accuracy and prediction ability of driver behavior prediction.

[0016] In the intelligent behavior prediction method based on the temporal interest network designed by the present invention, the introduction of multi-head self-attention mechanism, position encoding and Bi-LSTM has significant effectiveness in improving the model performance; the intelligent behavior prediction method based on the temporal interest network designed by the present invention adds an active unit layer, which makes the user's behavior-related interests more prominent, and can bring better prediction results in driving behavior prediction; the intelligent behavior prediction method based on the temporal interest network designed by the present invention has a clear algorithm structure and low time complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative work. Figure 1 It is a schematic flow chart of the method of the present invention; Figure 2 It is a schematic diagram of the workflow of the present invention; Figure 3 It is a structural schematic diagram of the model training of the present invention; Figure 4 It is a schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] Embodiment 1: like Figure 1 As shown, an intelligent behavior prediction method based on a temporal interest network comprises the following steps: S101: receiving a behavior sequence, dividing the behavior sequence into different time periods to obtain a behavior sequence for each time period, inputting the behavior sequence for each time period into a pre-established multi-head self-attention mechanism and position encoding model, and outputting an interest feature vector for each time period, wherein the behavior sequence is the driver's behavior actions within a preset time period; The process of dividing the behavior sequence into different time periods: Use the last action as a reference point , and set the length of the time window in segments from near to far, and express a certain behavior time point as .

[0020] Using formula (1) To express: (1) In the formula, is the time window, is the initial time window length, is the growth rate of the time window.

[0021] The pre-built multi-head self-attention mechanism and position encoding model are as follows: Using self-attention with 2 attention heads, the calculation is as follows: Among them, the query matrix , Keyword Matrix and information matrix The calculation is as follows: (2) In formula (2), is included The linear matrix of weight vectors, represents the number of queries, keywords, and information vectors in each attention head, represents the input sequence, Respectively represent The query, keyword and information matrix of each head, Respectively represent Head The calculation of the query, Respectively represent Head The calculation of keywords, Respectively represent Head The calculation of information, Respectively represent query weight vector, Respectively represent keyword weight vector, Respectively represent Information weight vectors are obtained by multiplying the input sequence by different weight matrices; In the model, ; in represents the attention head, represents the attention head matrix, which is a collection of multiple matrices. refers to the first attention head matrix, It refers to An attention head matrix, where Represents the index of the header, It refers to An attention head matrix, where refers to the number of attention heads, Indicates The dimension of the attention head matrix is is the length of the input sequence, is the dimension of the attention head, and , represents the dimension of the hidden units embedded in the model, is the dimension of the hidden units in the model, is the length of the embedding; head The expression is: (3) in The function represents the self-attention mechanism, The function is used to convert a real vector into a probability distribution. Representation Matrix The transpose of Representation Matrix The transpose of The vectors of different attention heads are connected and input into a feedforward neural network for calculation. The calculation expression is as follows: (4) (5) In formula (4) and (5), It refers to a feed-forward neural network. Indicates shared A head of attention, Respectively represent attention head vector, The function concatenates vectors. represents the weight matrix. The heads are flattened and connected, multiplied by the weight matrix, to obtain the final calculation structure of multi-head self-attention. Represents the output after the above operation, and then normalizes each layer layer by layer to calculate the User interests in a certain time period , where the variable Represents the average pooling process.

[0022] The behavior sequence of each time period is input into the pre-established multi-head self-attention mechanism and position encoding model. The initial input features are vectorized, and the sparse high-dimensional features are converted using one-hot encoding and muti-hot encoding. The high-dimensional sparse features are mapped to the low-dimensional vector space through the embedding layer, wherein the initial features include user behavior sequence, time features, user features, and target features.

[0023] In the embedding layer, the high-dimensional sparse features generated by the input layer are converted into low-dimensional vectors through embedding operations, and the features from different sources are embedded and represented, wherein the features from different sources include user behavior sequences, time features, user features and target features. The feature embedding results from different sources are fused, including the combination of user behavior and time features, and the combination of user features and target features, so as to generate a final feature representation as the interest feature vector for each time period.

[0024] S102: Input the interest feature vector of each time period into the pre-established Bi-LSTM model, output the interest sequence result, input the interest sequence result into the pre-built residual network and Dropout layer in sequence, obtain the optimized interest sequence result, input the optimized interest sequence result into the pre-built MLP layer, and output the behavior prediction result.

[0025] The pre-built Bi-LSTM model has a The hidden state The expression is as follows: (6) In formula (6), is the hidden state of the forward LSTM, is the hidden state of the reverse LSTM, and ⊕ represents the XOR operation of the two hidden state vectors. The correlation between the interest of the activated unit and the target determines the intelligent behavior of the user, so the interest weight of the time period needs to be redistributed. Studies have shown that the attention mechanism is an effective weight allocation mechanism for soft alignment between the source and the target, as described previously using multi-head self-attention and Bi-LSTM to capture the interest sequence, and the result expression after the target is applied to the interest sequence is as follows: (7) (8) In formula (7) and formula (8), Represent two attention heads respectively, Indicates elements, Respectively refer to Elements in The attention scores in the attention heads, Respectively refer to The user interest vector and hidden state vector of elements, Respectively The weight matrix of the attention head, Respectively Attention head target behavior and corresponding user embedding, Indicates total elements, represents the exponential function, Respectively The corresponding output of the attention head.

[0026] like Figure 2 As shown, the user behavior sequence is first divided into different time periods to obtain the behavior sequence of each time period, and then the behavior sequence of each time period is input into the pre-established multi-head self-attention mechanism and position encoding model, and the interest feature vector of each time period is output, and then the interest feature vector of each time period is input into the pre-established Bi-LSTM model, and the interest sequence result is output, and then the interest sequence result is input into the pre-built residual network and Dropout layer in turn to obtain the optimized interest sequence result, and then the optimized interest sequence result is input into the pre-built MLP layer, and the behavior prediction result is output.

[0027] Specifically, Figure 3 As shown, the present invention is further described below by way of embodiments: The input layer vectorizes the initial input features (user historical behavior, target features, user features), and converts the sparse high-dimensional features using one-hot encoding and multi-hot encoding. These high-dimensional sparse features are mapped to a low-dimensional vector space through subsequent embedding layers to ensure that the features can be efficiently input into the model. The embedding layer converts the high-dimensional sparse features generated by the input layer into low-dimensional vectors through embedding operations, embeds the user's historical behavior, target features, and user features, and then fuses the feature embedding results from different sources, including the combination of user's historical behavior and target features, as well as the combination of user features and target features, to generate the final feature representation, and then introduces bias encoding to enter the interaction layer; The interaction layer uses the multi-head self-attention mechanism and Bi-LSTM. Multi-head self-attention MHSA is an extension based on the self-attention mechanism, which is often used to process sequence data. It uses multiple queries to extract different aspects of information from the input sequence. The number of queries is the number of heads in the structure, and the model can learn multiple different attention weight matrices to capture dependencies at different levels and in different aspects of the sequence. In the self-attention mechanism, each attention head produces an independent output, and then the outputs of multiple heads are fused by concatenation, addition and normalization, and feedforward neural networks to obtain the final representation. MHSA improves the model's ability to learn information by increasing the number of attention structure parameters. Self-attention with 2 attention heads is used in the model. The dimension of extracted information is increased, and by increasing the number of queries, the model can learn and embed information in multiple sub-representation spaces. The Bi-LSTM neural network contains two independent LSTM parts. In the forward layer and the reverse layer, they are used to extract features of the input sequence in forward and reverse order, respectively. Then, the first two output feature vectors are aggregated to obtain the final feature expression. Using past and future temporal information, Bi-LSTM is more efficient and performs better than a single LSTM in extracting text features. The model uses Bi-LSTM to learn the relationship between interests extracted in different time periods. A deep residual network block is introduced, and MLP is used and jump links are adopted to solve the gradient disappearance problem. After activation by the activation unit, the Dropout Layer and ResNet Dropout methods are used to discard them, which enhances the overall generalization ability of the model and reduces the risk of overfitting. The randomly inactivated neurons also reduce the overall computational effort of the module, which helps to speed up the training of the model.

[0028] The MLP output layer concatenates and flattens the output information of the interaction layer with the target features, and uses attention The data are concatenated and input into the fully connected neural network MLP layer to obtain the output and realize the prediction.

[0029] Embodiment 2: In the second aspect, as Figure 4 As shown, in order to achieve the above-mentioned purpose, the present invention discloses an intelligent behavior prediction system based on a time interest network, comprising: The feature extraction module 11 is used to receive the behavior sequence, divide the behavior sequence into different time periods, obtain the behavior sequence of each time period, input the behavior sequence of each time period into a pre-established multi-head self-attention mechanism and position encoding model, and output the interest feature vector of each time period, wherein the behavior sequence is the behavior action of the driver within a preset time; The behavior prediction module 12 is used to input the interest feature vector of each time period into a pre-established Bi-LSTM model, output an interest sequence result, input the interest sequence result into a pre-built residual network and Dropout layer in sequence, obtain an optimized interest sequence result, input the optimized interest sequence result into a pre-built MLP layer, and output a behavior prediction result.

[0030] Based on the same inventive concept, the present invention also provides a computer device, which includes: one or more processors, and a memory for storing one or more computer programs; the program includes program instructions, and the processor is used to execute the program instructions stored in the memory. The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, which is used to implement one or more instructions, specifically for loading and executing one or more instructions in a computer storage medium to implement the above method.

[0031] It should be further explained that, based on the same inventive concept, the present invention also provides a computer storage medium, on which a computer program is stored, and the computer program is executed by the processor to execute the above method. The storage medium can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electrical, magnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0032] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present disclosure. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0033] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited by the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present disclosure. Without departing from the spirit and scope of the present disclosure, the present disclosure may have various changes and improvements, and these changes and improvements fall within the scope of the present disclosure to be protected.

Claims

1. An intelligent behavior prediction method based on temporal interest network, characterized in that: The method comprises the following steps: Receive a behavior sequence, divide the behavior sequence into different time periods, obtain a behavior sequence for each time period, input the behavior sequence for each time period into a pre-established multi-head self-attention mechanism and position encoding model, and output an interest feature vector for each time period, wherein the behavior sequence is the driver's behavior actions within a preset time period; The interest feature vector of each time period is input into the pre-established Bi-LSTM model, and the interest sequence result is output. The interest sequence result is sequentially input into the pre-built residual network and Dropout layer to obtain the optimized interest sequence result. The optimized interest sequence result is input into the pre-built MLP layer to obtain the behavior prediction result.

2. According to claim 1, the intelligent behavior prediction method based on time interest network is characterized in that: The process of dividing the behavior sequence into different time periods: Use the last action as a reference point , and set the length of the time window in segments from near to far, and express a certain behavior time point as .

3. According to claim 2, the intelligent behavior prediction method based on time interest network is characterized in that: Using formula (1) Behavior time point To express: (1) In the formula, is the time window, is the initial time window length, is the growth rate of the time window.

4. According to claim 1, the intelligent behavior prediction method based on time interest network is characterized in that: The pre-established multi-head self-attention mechanism and position encoding model are as follows: Using self-attention with 2 attention heads, the calculation is as follows: Among them, the query matrix , Keyword Matrix and information matrix The calculation is as follows: (2) In formula (2), is included The linear matrix of weight vectors, represents the number of queries, keywords, and information vectors in each attention head, represents the input sequence, Respectively represent The query, keyword and information matrix of each head, Respectively represent Head The calculation of the query, Respectively represent Head The calculation of keywords, Respectively represent Head The calculation of information, Respectively represent query weight vector, Respectively represent keyword weight vector, Respectively represent Information weight vectors are obtained by multiplying the input sequence by different weight matrices; In the model, ; in represents the attention head, represents the attention head matrix, which is a collection of multiple matrices. refers to the first attention head matrix, It refers to An attention head matrix, where Represents the index of the header, It refers to An attention head matrix, where refers to the number of attention heads, Indicates The dimension of the attention head matrix is is the length of the input sequence, is the dimension of the attention head, and , represents the dimension of the hidden units embedded in the model, is the dimension of the hidden units in the model, is the length of the embedding; head The expression is: (3) in The function represents the self-attention mechanism, The function is used to convert a real vector into a probability distribution. Representation Matrix The transpose of Representation Matrix The transpose of The vectors of different attention heads are connected and input into a feedforward neural network for calculation. The calculation expression is as follows: (4) (5) In formula (4) and (5), It refers to a feed-forward neural network. Indicates shared A head of attention, Respectively represent attention head vector, The function concatenates vectors. Represents the weight matrix, by The heads are flattened and connected, multiplied by the weight matrix, to obtain the final calculation structure of multi-head self-attention. Represents the output after operation, and then normalizes each layer layer by layer to calculate the User interests in a certain time period , where the variable Represents the average pooling process.

5. According to claim 1, the intelligent behavior prediction method based on time interest network is characterized in that: The behavior sequence of each time period is input into a pre-established multi-head self-attention mechanism and position encoding model. The initial features of the input are vectorized, and the sparse high-dimensional features are converted using one-hot encoding and muti-hot encoding. The high-dimensional sparse features are mapped to a low-dimensional vector space through an embedding layer, wherein the initial features include user behavior sequences, time features, user features, and target features.

6. The intelligent behavior prediction method based on time interest network according to claim 5 is characterized in that: In the embedding layer, the high-dimensional sparse features generated by the input layer are converted into low-dimensional vectors through embedding operations, and the features from different sources are embedded and represented, wherein the features from different sources include user behavior sequences, time features, user features and target features. The feature embedding results from different sources are fused, including the combination of user behavior and time features, and the combination of user features and target features, so as to generate a final feature representation as the interest feature vector for each time period.

7. The intelligent behavior prediction method based on time interest network according to claim 1 is characterized in that: The pre-built Bi-LSTM model has a The hidden state The expression is as follows: (6) In formula (6), is the hidden state of the forward LSTM, is the hidden state of the reverse LSTM, ⊕ represents the XOR operation on the two hidden states; the resulting expression after the target behavior is applied to the sequence of interest is as follows: (7) (8) In formula (7) and formula (8), Represent two attention heads respectively, Indicates elements, Respectively refer to Elements in The attention scores in the attention heads, Respectively refer to The user interest vector and hidden state vector of elements, Respectively The weight matrix of the attention head, Respectively Attention head target behavior and corresponding user embedding, Indicates total elements, represents the exponential function, Respectively The corresponding output of the attention head.

8. An intelligent behavior prediction system based on a time interest network, using an intelligent behavior prediction method based on a time interest network as claimed in any one of claims 1 to 7, characterized in that: include: A feature extraction module is used to receive a behavior sequence, divide the behavior sequence into different time periods, obtain a behavior sequence for each time period, input the behavior sequence for each time period into a pre-established multi-head self-attention mechanism and position encoding model, and output an interest feature vector for each time period, wherein the behavior sequence is the driver's behavior actions within a preset time period; The behavior prediction module is used to input the interest feature vector of each time period into the pre-established Bi-LSTM model, output the interest sequence result, input the interest sequence result into the pre-built residual network and Dropout layer in sequence, obtain the optimized interest sequence result, input the optimized interest sequence result into the pre-built MLP layer, and output the behavior prediction result.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, an intelligent behavior prediction method based on a time interest network as described in any one of claims 1 to 7 is adopted.

10. A computer-readable storage medium having a computer program stored therein, characterized in that: When the computer program is loaded and executed by the processor, an intelligent behavior prediction method based on a time interest network as described in any one of claims 1 to 7 is adopted.

Images