Click rate estimation method and system, electronic equipment and computer program product

Abstract

The invention relates to a click rate estimation method and system, electronic equipment and a computer program product, and belongs to the field of information retrieval. Comprising the following steps: learning long-term interest representation of a user by using a multi-head attention network and an attention mechanism; the behavior pattern representation of a user is captured by using a bidirectional recurrent neural network; a deep neural network and an attention mechanism are adopted to model interaction among the behavior pattern representation of the user, the user features and the user behavior time sequence, and deep instant interest representation and attention instant interest representation are obtained; splicing the deep instant interest representation and the attention instant interest representation to form an instant interest representation; and splicing the user long-term interest representation, the instant interest representation and the user characteristics, and inputting into a factorization machine and a multi-layer perceptron to predict the click rate of the user. The method aims at capturing long-term interests and instant interests of the user from user behaviors and time information, and excellent interest mining ability and click rate estimation effect are shown.

Description

Technical Field

[0001] The invention relates to a click rate prediction method, system, electronic equipment and computer program product, and belongs to the field of information retrieval. Background Art

[0002] Click-through rate estimation plays an important role in improving the total transaction volume of the platform in online recommendation services. In the study of click-through rate estimation, the effectiveness of user interest information has been proven by a large number of works to improve the performance of the model. Therefore, most existing methods mainly learn user interest information based on the relationship between user historical behaviors to make more accurate click-through rate estimation. However, although user interests have been taken into account, existing methods mainly focus on modeling users' long-term interests from historical behavior sequences, but fail to consider users' immediate interests. At the same time, the scarcity of push time data limits the model's ability to capture the complex relationship between push time and personalized user preferences, which in turn increases the difficulty of the model in mining users' immediate interests.

[0003] Therefore, a method is needed that can utilize users' historical behaviors and time information to simultaneously mine and integrate users' long-term interests and immediate interests. Summary of the invention

[0004] The present invention mainly provides a click rate prediction method, system, electronic device, and computer program product. The present invention takes into account the click rate prediction value after the user's dual interests, so that the prediction result has higher accuracy.

[0005] The technical solution of the present invention is: the present invention provides a click rate prediction method, the method comprising:

[0006] Step 1: Use a multi-head attention network and attention mechanism to learn the user's long-term interest representation from the user behavior sequence;

[0007] Step 2: Capture the user's behavior pattern representation from the user behavior sequence and its corresponding pushed user behavior time series by adopting a bidirectional recurrent neural network;

[0008] Step 3: Use a deep neural network and an attention mechanism to model the interaction between the user's behavior pattern representation, user characteristics, and user behavior time series, and obtain a deep instant interest representation and an attention instant interest representation;

[0009] Step 4: Concatenate the depth instant interest representation and the attention instant interest representation to form an instant interest representation;

[0010] Step 5: Concatenate the user's long-term interest representation, immediate interest representation, and user features, and input them into the factorization machine and the multi-layer perceptron to predict the user's click rate.

[0011] Furthermore, in step 1, the correlation between behaviors in the user behavior sequence is learned using formula (1) and formula (2), the correlation between each behavior and the target product is calculated to weight the attention representation of the historical behavior, and the weight is calculated using formula (3) and formula (4) to obtain the user's long-term interest representation I; the calculation formula of the user's long-term interest representation I is expressed as (4):

[0012]

[0013] B′=Concat(head1, head2,..., head H ) (2)

[0014]

[0015] Where Q h , K h , V h is the embedding vector of the behavior obtained by different linear mappings, W′ is a matrix of trainable parameters, b′ l is the lth element of B′, d is the embedding vector dimension, head H represents the Hth head calculated by the multi-head attention network, t a represents the embedding vector of the target product to be predicted, B′ represents the vector obtained by concatenating a total of H headers, and a l represents the weight of the first behavior in the user behavior sequence in the interest fusion process, b' l , b' n represents the vectors of the lth and nth rows in B′, and also represents the representation vectors of the lth and nth behaviors in the user behavior sequence obtained after the multi-head attention network of formulas (1) and (2). concat() represents the concatenation operation between vectors, n is the index increasing from 1 to L when summing, and L is the length of the user behavior sequence.

[0016] Furthermore, in step 2, a bidirectional recurrent neural network is used to process the user behavior sequence and the corresponding user behavior time series respectively, and the user's behavior pattern representation P is obtained by formula (5) and formula (6); the user's behavior pattern representation P is calculated as shown in formula (6):

[0017]

[0018] Where W P is a trainable parameter, b″ t ,s″ t are the t-th output of the recurrent neural network used to process the user behavior sequence and time series, respectively. βn represents the weight of the n-th behavior in the user behavior sequence in the process of representing the user's behavior pattern. The same is true for βt. s″ nrepresents the representation vector of the nth time in the output sequence obtained by processing the user time series with a bidirectional recurrent neural network, b″ n It represents the representation vector of the nth behavior in the output sequence obtained after processing the user behavior sequence with a bidirectional recurrent neural network.

[0019] Furthermore, in step 3, according to formula (7) and formula (8), the user features are aggregated based on the attention mechanism to obtain the user feature representation U, and the fused user feature representation, the embedding vector t of the push time and the behavior pattern representation P are input into the multi-layer perceptron to learn the deep instant interest representation S′ according to formula (9), and the push time is the user time series; the calculation formula of the deep instant interest representation S′ is expressed as (9):

[0020]

[0021] S′=FC2(ReLU(FC1([U;t;P]))) (9)

[0022] Where W u ,u k ,u ID They are trainable parameters, the embedding vector of the kth feature of the user and the embedding vector of the ID, FC1 and FC2 are fully connected layers, and γ k It represents the weight of the kth feature in the user feature fusion process, and K is the total number of user features.

[0023] Furthermore, in step 3, the user features and the push time are concatenated as a query, and the user behavior obtained in step 1 is embedded as a key according to formula (10) and formula (11), and the attention weight is calculated to obtain the attention immediate interest representation S″; the calculation formula of the attention immediate interest representation S″ is expressed as (11):

[0024]

[0025] Where W s It is a trainable parameter, the purpose of which is to perform a linear transformation on the input to change b l The dimension of b l is the initial representation vector of the lth behavior in the user behavior sequence, [U; t] is the concatenation vector of the user feature representation U and the target product representation vector t, δ l Represents weight.

[0026] Furthermore, in the step 4, the deep instant interest representation and the attention instant interest representation are concatenated to obtain the instant interest representation S; in the step 5, the user's long-term interest representation, the instant interest representation and the corresponding embedding vector of the user's characteristics are combined as input data for click-through rate estimation, and the click-through rate is predicted by inputting into the factor decomposition machine and the multi-layer perceptron network structure.

[0027] The present invention also provides a click rate prediction system, which includes: a module for executing the above-mentioned click rate prediction method.

[0028] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned click rate prediction method when executing the program.

[0029] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned click rate prediction method is implemented.

[0030] The present invention also provides a computer program product, including a computer program, and when the computer program is executed by a processor, the above-mentioned click rate prediction method is implemented.

[0031] The beneficial effects of the present invention are:

[0032] 1. The present invention captures behavior patterns from historical behaviors and their corresponding push times, and then models the interaction between behavior patterns, user characteristics, and push times, demonstrating excellent interest mining capabilities and click-through rate prediction effects;

[0033] 2. The present invention cleverly combines users’ long-term interests to enhance their ability to express immediate interests;

[0034] 3. The present invention can preset the maximum sequence length of the user's historical behavior sequence according to the experience value, improve the calculation efficiency of the algorithm, and reduce the time cost;

[0035] 4. The present invention can estimate the probability of user clicks at a reasonable time cost through self-attention network, attention mechanism, and modeling of behaviors and behavior time series. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0038] Example 1: Figure 1As shown, the present invention provides a click rate prediction method, the method comprising:

[0039] Step 1: Use a multi-head attention network and attention mechanism to learn the user's long-term interest representation from the user behavior sequence;

[0040] Before step 1 begins, the input data is converted into a low-dimensional dense vector through table lookup embedding, which includes product user features, user behavior sequence, and user behavior time series;

[0041] The input data includes: the behavior in the user's historical behavior sequence, the time in the user's historical behavior time series, and various discrete features of users and products; the timestamp is converted to discrete time representation, and this example uses the hour of the day as the time information. In this example, all embedding vector dimensions are set to 8.

[0042] Furthermore, in step 1, the correlation between behaviors in the user behavior sequence is learned using formula (1) and formula (2), the correlation between each behavior and the target product is calculated to weight the attention representation of the historical behavior, and the weight is calculated using formula (3) and formula (4) to obtain the user's long-term interest representation I; the calculation formula of the user's long-term interest representation I is expressed as (4):

[0043]

[0044] B′=Concat(head1,head2,…,head H ) (2)

[0045]

[0046] Where Q h ,K h ,V h is the embedding vector of the behavior obtained by different linear mappings, W′ is a matrix of trainable parameters, b′ l is the lth element of B′, d is the embedding vector dimension, head H represents the Hth head calculated by the multi-head attention network, t a represents the embedding vector of the target product to be predicted, B′ represents the vector obtained by concatenating a total of H headers, and a l represents the weight of the first behavior in the user behavior sequence in the interest fusion process, b' l , b' nrepresents the vector of the lth and nth rows in B′, and also represents the representation vector of the lth and nth behaviors in the user behavior sequence obtained after the multi-head attention network of formulas (1) and (2). concat() represents the concatenation operation between vectors. n is an index that increases from 1 to L when summing. L is the length of the user behavior sequence. This example stipulates that the length of the user historical behavior sequence does not exceed 20. For sequences with a length greater than 20, the earliest behaviors exceeding 20 are removed to maintain a one-to-one correspondence with the user historical behavior sequence.

[0047] Step 2: Capture the user's behavior pattern representation from the user behavior sequence and its corresponding pushed user behavior time series by adopting a bidirectional recurrent neural network;

[0048] Furthermore, in step 2, a bidirectional recurrent neural network is used to process the user behavior sequence and the corresponding user behavior time series respectively, and the user's behavior pattern representation P is obtained by formula (5) and formula (6); the user's behavior pattern representation P is calculated as shown in formula (6):

[0049]

[0050] Where W P is a trainable parameter, b″ t ,s″ t are the t-th output of the recurrent neural network used to process the user behavior sequence and time series, respectively. βn represents the weight of the n-th behavior in the user behavior sequence in the process of representing the user's behavior pattern. The same is true for βt. s″ n represents the representation vector of the nth time in the output sequence obtained by processing the user time series with a bidirectional recurrent neural network, b″ n It represents the representation vector of the nth behavior in the output sequence obtained after processing the user behavior sequence with a bidirectional recurrent neural network. This example uses a bidirectional recurrent neural network with long short-term memory units.

[0051] Step 3: Use a deep neural network and an attention mechanism to model the interaction between the user's behavior pattern representation, user characteristics, and user behavior time series, and obtain a deep instant interest representation and an attention instant interest representation;

[0052] Furthermore, in step 3, according to formula (7) and formula (8), the user features are aggregated based on the attention mechanism to obtain the user feature representation U, and the fused user feature representation, the embedding vector t of the push time and the behavior pattern representation P are input into the multi-layer perceptron to learn the deep instant interest representation S′ according to formula (9), and the push time is the user time series; the calculation formula of the deep instant interest representation S′ is expressed as (9):

[0053]

[0054] S′=FC2(ReLU(FC1([U;t;P]))) (9)

[0055] Where W u ,u k ,u ID They are trainable parameters, the embedding vector of the kth feature of the user and the embedding vector of the ID, FC1 and FC2 are fully connected layers, and γ k represents the weight of the kth feature in the user feature fusion process, K is the total number of user features, and all multi-layer perceptrons used in this example are three-layer structures, in which the length of the hidden layer is 16.

[0056] Furthermore, in step 3, the user features and the push time are concatenated as a query, and the user behavior obtained in step 1 is embedded as a key according to formula (10) and formula (11), and the attention weight is calculated to obtain the attention immediate interest representation S″; the calculation formula of the attention immediate interest representation S″ is expressed as (11):

[0057]

[0058] Where W s It is a trainable parameter, the purpose of which is to perform a linear transformation on the input to change b l The dimension of b l is the initial representation vector of the lth behavior in the user behavior sequence, [U; t] is the concatenation vector of the user feature representation U and the target product representation vector t, δ l Represents weight.

[0059] Step 4: Concatenate the depth instant interest representation and the attention instant interest representation to form an instant interest representation;

[0060] Step 5: Concatenate the user's long-term interest representation, immediate interest representation, and user features, and input them into the factorization machine and multi-layer perceptron (MLP) to predict the user's click rate.

[0061] Furthermore, in step 4, the deep instant interest representation and the attention instant interest representation are concatenated to obtain the instant interest representation S; in step 5, the user's long-term interest representation, the instant interest representation and the corresponding embedding vector of the user's characteristics are combined as input data for click-through rate estimation, and the click-through rate is predicted by inputting into the factor decomposition machine and the multi-layer perceptron network structure. In this example, the multi-layer perceptron structure used in step 5 is the same as that in step 3, but it is a network with different parameters.

[0062] Through the self-attention network, attention mechanism, and modeling of behaviors and behavior time series, the probability of users clicking on the target product can be estimated at a reasonable time cost.

[0063] There are also the following points to note when implementing it:

[0064] The length of the user's historical behavior sequence will affect the computational efficiency of the algorithm. When the historical behavior sequence is too long, additional time costs will be incurred. Those skilled in the art can preset the maximum sequence length based on their own experience.

[0065] In specific implementation, the method provided by the present invention can realize automatic operation process based on software technology.

[0066] The present invention also provides a click rate prediction system, the system comprising:

[0067] Long-term interest representation learning module, which is used to learn the user's long-term interest representation from the user behavior sequence using a multi-head attention network and attention mechanism;

[0068] A behavior pattern representation capture module, used to capture the user's behavior pattern representation from the user behavior sequence and its corresponding pushed user behavior time series by adopting a bidirectional recurrent neural network;

[0069] A deep instant interest representation and an attentive instant interest representation acquisition module, which is used to use a deep neural network and an attention mechanism to model the interaction between the user's behavior pattern representation, user characteristics, and the user's behavior time series, and obtain a deep instant interest representation and an attentive instant interest representation;

[0070] A concatenation module, used for concatenating the deep instant interest representation and the attention instant interest representation to form an instant interest representation;

[0071] The prediction module is used to concatenate the user's long-term interest representation, immediate interest representation and user characteristics, and input them into the factor decomposition machine and the multi-layer perceptron to predict the user's click rate.

[0072] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned click rate prediction method when executing the program.

[0073] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned click rate prediction method is implemented.

[0074] The present invention also provides a computer program product, including a computer program, and when the computer program is executed by a processor, the above-mentioned click rate prediction method is implemented.

[0075] The present invention captures the user's long-term interests and immediate interests from the user's historical behavior and time information. The extraction of long-term interests uses a self-attention network to facilitate the capture of correlations within the behavior sequence, and extracts the user's long-term interest representation from historical behaviors through an attention mechanism.

[0076] In order to model the user's immediate interest, the instant interest estimation in the present invention integrates user behavior patterns, user characteristics and advertising push time. Specifically, the present invention adopts two recurrent neural networks with bidirectional long short-term memory network units to capture the dynamic changes of behavior and time series respectively. These representations are then concatenated and transformed using linear transformation and weighted summation to represent the user's behavior pattern. In addition, deep neural networks and target attention networks are used to capture the interactions between user behavior patterns, user characteristics and push time, as they have a significant impact on instant interest estimation.

[0077] Figure 1 This is a flowchart of the present invention, in which the long-term interest extraction module first receives the input of the user behavior sequence, and mines the user's long-term interest representation from the relationship between behaviors through the multi-head self-attention network and the attention mechanism. The immediate interest extraction module uses two different bidirectional recurrent neural networks to process the user's historical behavior sequence and the user's historical behavior time series, respectively. After splicing and linear transformation, softmax is used for weighted summation to obtain the user behavior pattern representation, and a multi-layer perceptron and an attention mechanism are used to combine the push time to obtain the attention immediate interest representation and the deep immediate interest representation. Finally, the long-term interest representation, the immediate interest representation and the user features are spliced ​​and input into the multi-layer perceptron to obtain the predicted value of the click rate.

[0078] In order to verify the effect of the present invention, the present invention conducted the following experiments for verification. Evaluation indicators: AUC (Area Under Curve) and logloss, which are widely used in the field of click-through rate prediction, are used. Among them, the higher the AUC value, the better the performance of the method in the click-through rate prediction task, and the lower the logloss value, the better the performance of the method in the click-through rate prediction task. The data used comes from the industrial-level data of Didi's message push scenario and the public advertising click data provided by Alibaba Tianchi platform.

[0079] Table 1 shows the verification results of this algorithm and other algorithms on Didi industrial data and Tianchi advertising data

[0080]

[0081] The compared algorithms are all advanced algorithms in this field, and these algorithms are additionally introduced to consider the variants of push time. It can be seen from the attached table that the method proposed in the present invention has better performance than other algorithms on two real data. The present invention can make more full use of the information in the user behavior sequence and time.

[0082] The specific embodiments of the present invention are described in detail above in conjunction with the accompanying drawings, but the present invention is not limited to the above embodiments. Various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A click rate prediction method, characterized in that: The method comprises: Step 1: Use a multi-head attention network and attention mechanism to learn the user's long-term interest representation from the user behavior sequence; Step 2: Capture the user's behavior pattern representation from the user behavior sequence and its corresponding pushed user behavior time series by adopting a bidirectional recurrent neural network; Step 3: Use a deep neural network and an attention mechanism to model the interaction between the user's behavior pattern representation, user characteristics, and user behavior time series, and obtain a deep instant interest representation and an attention instant interest representation; Step 4: Concatenate the depth instant interest representation and the attention instant interest representation to form an instant interest representation; Step 5: Concatenate the user's long-term interest representation, immediate interest representation, and user features, and input them into the factorization machine and the multi-layer perceptron to predict the user's click rate.

2. A click rate estimation method according to claim 1, characterized in that: In step 1, the correlation between behaviors in the user behavior sequence is learned using formula (1) and formula (2), the correlation between each behavior and the target product is calculated to weight the attention representation of the historical behavior, and the weight is calculated using formula (3) and formula (4) to obtain the user's long-term interest representation I; the calculation formula of the user's long-term interest representation I is expressed as (4): B′=Concat(head1,head2,...,head H ) (2) Where Q h , K h , V h is the embedding vector of the behavior obtained by different linear mappings, W′ is a matrix of trainable parameters, b′ l is the lth element of B′, d is the embedding vector dimension, head H represents the Hth head calculated by the multi-head attention network, t a represents the embedding vector of the target product to be predicted, B′ represents the vector obtained by concatenating a total of H headers, and a l represents the weight of the first behavior in the user behavior sequence in the interest fusion process, b' l , b' n represents the vectors of the lth and nth rows in B′, and also represents the representation vectors of the lth and nth behaviors in the user behavior sequence obtained after the multi-head attention network of formulas (1) and (2). concat() represents the concatenation operation between vectors, n is the index increasing from 1 to L when summing, and L is the length of the user behavior sequence.

3. A click rate estimation method according to claim 1, characterized in that: In step 2, a bidirectional recurrent neural network is used to process the user behavior sequence and the corresponding user behavior time series respectively, and the user's behavior pattern representation P is obtained by formula (5) and formula (6); the user's behavior pattern representation P is calculated as shown in formula (6): Where W P is a trainable parameter, b″ t ,s″ t are the t-th output of the recurrent neural network used to process the user behavior sequence and time series, respectively. βn represents the weight of the n-th behavior in the user behavior sequence in the process of representing the user's behavior pattern. The same is true for βt. s″ n represents the representation vector of the nth time in the output sequence obtained by processing the user time series with a bidirectional recurrent neural network, b″ n It represents the representation vector of the nth behavior in the output sequence obtained after processing the user behavior sequence with a bidirectional recurrent neural network.

4. A click rate estimation method according to claim 1, characterized in that: In step 3, user features are aggregated based on the attention mechanism according to formula (7) and formula (8) to obtain the user feature representation U. The fused user feature representation, the embedding vector t of the push time and the behavior pattern representation P are input into the multi-layer perceptron to learn the deep instant interest representation S′ according to formula (9). The push time is the user time series. The calculation formula of the deep instant interest representation S′ is as shown in (9): S′=FC2(ReLU(FC1([U;t;P]))) (9) Where W u ,u k ,u ID They are trainable parameters, the embedding vector of the kth feature of the user and the embedding vector of the ID, FC1 and FC2 are fully connected layers, and γ k It represents the weight of the kth feature in the user feature fusion process, and K is the total number of user features.

5. A click rate estimation method according to claim 1, characterized in that: In step 3, the user features and push time are concatenated as a query, and the user behavior obtained in step 1 is embedded as a key according to formula (10) and formula (11), and the attention weight is calculated to obtain the attention immediate interest representation S″; The calculation formula of the immediate interest representation S″ is as shown in (11): Where W s It is a trainable parameter, the purpose of which is to perform a linear transformation on the input to change b l The dimension of b l is the initial representation vector of the lth behavior in the user behavior sequence, [U; t] is the concatenation vector of the user feature representation U and the target product representation vector t, δ l Represents weight.

6. A click rate estimation method according to claim 1, characterized in that: In the step 4, the deep instant interest representation and the attention instant interest representation are concatenated to obtain the instant interest representation S; in the step 5, the user's long-term interest representation, the instant interest representation and the corresponding embedding vector of the user's characteristics are combined as input data for click-through rate estimation, and the click-through rate is predicted by inputting into the factor decomposition machine and the multi-layer perceptron network structure.

7. A click rate prediction system, characterized in that: The system comprises: a module for executing the click rate prediction method according to any one of claims 1 to 6.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, a click rate prediction method as described in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, a click rate prediction method as described in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, a click rate prediction method as described in any one of claims 1 to 6 is implemented.

Images