Product click rate determination method

Abstract

The present disclosure provides a product click rate determination method. The method comprises: obtaining a product data training sample set, wherein the product data training sample set comprises a plurality of training samples, each training sample comprising a plurality of product feature vectors and an interaction behavior sequence; for each training sample, processing the plurality of product feature vectors using a mean clustering model to obtain a plurality of cluster center vectors; determining a sampling probability corresponding to each cluster center vector according to a preset feature vector of a preset target product and a plurality of cluster center vectors corresponding to the interaction behavior sequence; sampling the interaction behavior sequence according to the plurality of sampling probabilities to obtain a plurality of interaction behavior subsequences; training a deep interest network using the preset feature vector and the plurality of interaction behavior subsequences to obtain a trained click rate prediction model; and inputting a product feature vector of a product to be predicted and an interaction behavior sequence of a user to be predicted into the click rate prediction model to output a prediction result.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of neural networks, and more particularly, to a product click rate determination method and device, electronic equipment, computer readable storage medium and computer program product. BACKGROUND

[0002] Click rate prediction is a new user behavior prediction technology in the development background of the information age, which is widely used in the Internet industry and is the basis of online applications such as online advertising and e-commerce recommendation. The purpose is to predict the potential needs of different users and recommend appropriate and high-quality click content.

[0003] In real-world scenarios, due to the rapid development of information technology, users accumulate more and more behavior records on the Internet platform, the length of the behavior sequence increases explosively, and the demand involved is diverse, making the user behavior record data present the characteristics of numerous quantity and extensive interest.

[0004] In the process of implementing the present disclosure, the inventors have found that the existing click rate prediction model has at least the following problem: the prediction accuracy is low when predicting the click rate of a user clicking on a product. SUMMARY

[0005] Therefore, the present disclosure provides a product click rate determination method and device, electronic equipment, computer readable storage medium and computer program product.

[0006] One aspect of an embodiment of the present disclosure provides a product click rate determination method, comprising:

[0007] Obtaining a product data training sample set, wherein the product data training sample set includes a plurality of training samples, each of the training samples includes a plurality of product feature vectors with a preset dimension and an interaction behavior sequence, and the interaction behavior sequence represents the time and order of interaction behavior between a user and a plurality of products;

[0008] For each of the training samples, processing the plurality of product feature vectors using a mean clustering model to obtain a plurality of cluster center vectors, wherein the mean clustering model is obtained by pre-training a K-means clustering model, and each of the cluster center vectors is a feature vector that can represent a plurality of product feature vectors in the same class;

[0009] According to the preset feature vector of the preset target product and the plurality of cluster center vectors corresponding to the interaction behavior sequence, determining a sampling probability corresponding to each of the cluster center vectors;

[0010] According to the plurality of sampling probabilities, performing sampling processing on the interaction behavior sequence to obtain a plurality of interaction behavior subsequences;

[0011] training the deep interest network by using the preset feature vector and the plurality of interaction behavior subsequences to obtain a trained click rate prediction model;

[0012] inputting a product feature vector of a product to be predicted and an interaction behavior sequence of a user to be predicted into the click rate prediction model, and outputting a prediction result, wherein the prediction result represents a probability of the user to be predicted clicking the product to be predicted.

[0013] According to an embodiment of the present disclosure, the plurality of product feature vectors are obtained in the following manner:

[0014] performing screening processing on the plurality of obtained product data to obtain screened product data;

[0015] respectively performing one-hot encoding processing on a plurality of categories of the product data in a case where a product attribute of the product data meets a first preset condition, to obtain a first one-hot vector corresponding to each of the categories, wherein the product attribute includes a product price, a product brand, or a purchase time;

[0016] performing splicing processing on the plurality of first one-hot vectors to obtain a spliced feature vector of the product data;

[0017] in a case where the product attribute of the product data does not meet the first preset condition, processing the product by using a bucketing method to obtain a category vector of a plurality of categories;

[0018] respectively performing one-hot encoding processing on the category vector of each category to obtain a second one-hot vector corresponding to each category;

[0019] performing splicing processing on the plurality of second one-hot vectors to obtain the spliced feature vector of the product data;

[0020] performing mapping processing on the spliced feature vector in a preset dimension to obtain a product feature vector.

[0021] According to an embodiment of the present disclosure, the screening processing on the plurality of obtained product data to obtain screened product data comprises:

[0022] determining an interaction frequency of the user with each product according to the interaction behavior sequence;

[0023] determining a product with an interaction frequency greater than a preset threshold as the screened product data.

[0024] According to an embodiment of the present disclosure, the product feature vector is obtained by processing product information of a product by using an embedding layer; and the sampling probability is obtained by processing by using a sampling module.

[0025] According to an embodiment of the present disclosure, the method further includes:

[0026] training the initial neural network using the samples in the optimization training sample set in sequence to obtain a first loss value corresponding to the optimization training sample, wherein the samples in the optimization training sample set include positive samples and negative samples, the positive samples include a plurality of the interaction behavior sub-sequences of the user, and the negative samples include a plurality of the interaction behavior sub-sequences of at least one other user;

[0027] in a case where the first loss value does not satisfy a first convergence threshold, updating the model parameters of the embedding layer, the model parameters of the sampling module, and the model parameters of the initial neural network using a stochastic gradient descent algorithm;

[0028] in a case where the first loss value satisfies the first convergence threshold, determining the updated model parameters of the embedding layer as target model parameters of the embedding layer, determining the updated model parameters of the sampling module as target model parameters of the sampling module, and determining the initial neural network with the updated model parameters as a trained contrastive learning target model.

[0029] According to an embodiment of the present disclosure, the first loss value L C is calculated as shown in the following formula:

[0030]

[0031] D ω (p, q) = exp(p T · W · q)

[0032] wherein p and q represent the interaction behavior sub-sequences sampled twice from the user; P l represents a joint probability distribution of the user sub-sequences sampled, i.e., (p, q) ~ P l ; represents the interaction behavior sub-sequences randomly sampled from the interaction behavior sequence of at least one other user; D ω is a discriminator with parameters ω in the initial neural network, defined as a log bilinear model; p T represents transposition of p.

[0033] According to an embodiment of the present disclosure, the processing of the plurality of product feature vectors using the mean clustering model to obtain a plurality of clustering center vectors includes:

[0034] processing the plurality of product feature vectors using the mean clustering model to obtain a plurality of product feature sub-vectors corresponding to each of the product feature vectors;

[0035] selecting, according to the preset selection condition, a plurality of target product feature sub-vectors from the plurality of product feature sub-vectors of the plurality of products;

[0036] obtaining a plurality of cluster center vectors according to the plurality of target product feature sub-vectors, wherein a number of the cluster center vectors is less than a number of the product feature vectors.

[0037] According to an embodiment of the present disclosure, the determining, according to the preset target product feature vector and the plurality of cluster center vectors corresponding to the interaction behavior sequence, a sampling probability corresponding to each of the cluster center vectors, comprises:

[0038] determining a similarity score of the preset target product feature vector and each of the cluster center vectors according to the preset target product feature vector and the interaction behavior sequence;

[0039] determining a relative time difference of the preset target product and each of the cluster center vectors according to the interaction behavior sequence;

[0040] determining a sampling probability corresponding to each of the products according to the plurality of similarity scores and the plurality of relative time differences.

[0041] According to an embodiment of the present disclosure, the determining, according to the interaction behavior sequence, a relative time difference of the preset target product and each of the cluster center vectors, comprises:

[0042] measuring a time sequence relationship of the plurality of cluster center vectors corresponding to the interaction behavior sequence by using the relative time difference to obtain a timestamp set, wherein each timestamp in the timestamp set corresponds to a time when a user interacts with one of the cluster center vectors;

[0043] determining the relative time difference of the preset target product and each of the cluster center vectors based on the timestamp set.

[0044] According to an embodiment of the present disclosure, the determining, according to the plurality of similarity scores and the plurality of relative time differences, a sampling probability corresponding to each of the products, comprises:

[0045] for each of the relative time differences, converting the relative time difference into a time sequence score;

[0046] calculating a sampling probability of each of the products according to the time sequence score and a similarity score corresponding to the relative time difference.

[0047] According to an embodiment of the present disclosure, the training, by using the preset target product feature vector and the plurality of interaction behavior sub-sequences, a deep interest network to obtain a trained click rate prediction model, comprises:

[0048] input the preset feature vector and the plurality of product feature vectors corresponding to each of the above-mentioned interaction behavior sub-sequences into an attention mechanism layer to obtain a plurality of correlation weights, wherein one of the above-mentioned correlation weights represents the correlation between the above-mentioned preset feature vector and one of the above-mentioned product feature vectors;

[0049] input the plurality of correlation weights into a pooling layer to obtain a target feature vector, wherein the above-mentioned target feature vector represents a vector of the relationship between the above-mentioned user and the above-mentioned preset target product;

[0050] input the above-mentioned target feature vector into a multi-layer perception layer to output a training prediction result;

[0051] calculate a second loss value according to the above-mentioned training prediction result and a loss function;

[0052] in the case where the second loss value does not satisfy a second convergence threshold, update the model parameters of the above-mentioned deep interest network by using a stochastic gradient descent algorithm;

[0053] in the case where the second loss value satisfies the second convergence threshold, determine the updated deep interest network as a trained click rate prediction model.

[0054] Another aspect of the embodiments of the present disclosure provides a product click rate determination apparatus, comprising:

[0055] an acquisition module configured to acquire a product data training sample set, wherein the above-mentioned product data training sample set comprises a plurality of training samples, each of the above-mentioned training samples comprises a plurality of product feature vectors with a preset dimension and an interaction behavior sequence, and the above-mentioned interaction behavior sequence represents the time and order of the interaction behavior between a user and a plurality of products;

[0056] a clustering module configured to, for each of the above-mentioned training samples, process the plurality of above-mentioned product feature vectors by using a mean clustering model to obtain a plurality of cluster center vectors, wherein the above-mentioned mean clustering model is obtained by pre-training a K-means clustering model, and each of the above-mentioned cluster center vectors is a feature vector that can represent a plurality of the above-mentioned product feature vectors in the same class;

[0057] a first determination module configured to determine a sampling probability corresponding to each of the above-mentioned cluster center vectors according to a preset feature vector of a preset target product and a plurality of the above-mentioned cluster center vectors corresponding to the above-mentioned interaction behavior sequence;

[0058] a second determination module configured to sample and process the above-mentioned interaction behavior sequence according to a plurality of the above-mentioned sampling probabilities to obtain a plurality of interaction behavior sub-sequences;

[0059] The training module is configured to train the deep interest network by using the preset feature vector and the plurality of interaction behavior subsequences, and obtain a trained click rate prediction model.

[0060] The prediction module is configured to input a product feature vector of a to-be-predicted product and an interaction behavior sequence of a to-be-predicted user into the click rate prediction model, and output a prediction result, where the prediction result represents a probability that the to-be-predicted user clicks the to-be-predicted product.

[0061] Another aspect of the embodiments of the present disclosure provides an electronic device, comprising: one or more processors; a memory configured to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described above.

[0062] Another aspect of the embodiments of the present disclosure provides a computer-readable storage medium, which stores computer-executable instructions, the instructions being configured to implement the method as described above when executed.

[0063] Another aspect of the embodiments of the present disclosure provides a computer program product, which comprises computer-executable instructions, the instructions being configured to implement the method as described above when executed.

[0064] According to the embodiments of the present disclosure, the plurality of product feature vectors are processed by the mean clustering model, so as to replace the specific product feature vectors with the determined clustering centers, which is conducive to eliminating the noise, so as to facilitate the calculation of the sampling probability of the preset target commodity and each clustering center, and then the interaction behavior subsequence with high relevance, less noise and high information retention can be generated based on the sampling probability, and the click rate prediction model trained by using the interaction behavior subsequence and the preset feature vector has high prediction accuracy when used. BRIEF DESCRIPTION OF DRAWINGS

[0065] The above and other objects, features and advantages of the present disclosure will become more apparent from the following description when taken in conjunction with the accompanying drawings, in which:

[0066] Figure 1 An exemplary system architecture of the product click rate determination method according to the embodiments of the present disclosure is schematically shown;

[0067] Figure 2 A flowchart A of the product click rate determination method according to the embodiments of the present disclosure is schematically shown;

[0068] Figure 3 A flowchart B of the product click rate determination method according to the embodiments of the present disclosure is schematically shown;

[0069] Figure 4A flowchart of a product feature vector acquisition process according to an embodiment of the present disclosure is schematically shown.

[0070] Figure 5 A flowchart of a product recommendation process according to an embodiment of the present disclosure is schematically shown.

[0071] Figure 6 A block diagram of a product click rate determination apparatus according to an embodiment of the present disclosure is schematically shown.

[0072] Figure 7 A block diagram of an electronic device implementing a product click rate determination method according to an embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0073] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. It should be understood, however, that the description which follows is merely exemplary and is not intended to limit the scope of the present disclosure. In the following detailed description of the embodiments of the present disclosure, numerous specific details are set forth in order to provide a thorough understanding of the present disclosure. However, it would be apparent to one skilled in the art that the present disclosure can be practiced without these specific details. In other instances, well-known structures and functions have not been described in detail in order to avoid obscuring aspects of the present disclosure.

[0074] The terms used herein are merely used to describe specific embodiments and are not intended to limit the present disclosure. The terms "include", "comprise" and the like used herein indicate the presence of the described features, operations, operations and / or components, but do not exclude the presence or addition of one or more other features, operations, operations or components.

[0075] All terms used herein, including technical and scientific terms, have meanings commonly understood by one of ordinary skill in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having meanings consistent with the context of the present specification, and should not be interpreted in an idealized or overly formal manner.

[0076] In the case of using expressions similar to "at least one of A, B, and C, etc.", it should be generally interpreted as including at least one of the items, unless otherwise defined (for example, "a system having at least one of A, B, and C" should include but not be limited to a system having A alone, a system having B alone, a system having C alone, a system having both A and B, a system having both A and C, a system having both B and C, and / or a system having A, B, and C, etc.).

[0077] In real scenarios, due to the rapid development of information technology, the user accumulates more and more behavior records on the Internet platform, the length of the behavior sequence increases explosively, and the demand involved is diverse, so that the user behavior record data presents the characteristics of numerous quantity and extensive interest. Among them, numerous quantity refers to the large number of user behavior records and large time span, which brings obvious performance challenges; extensive interest refers to the diverse user demand contained in the numerous behavior records, which makes it difficult for traditional data analysis methods to obtain the accurate interest of the user at present.

[0078] The inventors have found that most of the related click rate prediction methods intercept the recent user behavior sequence, such as using attention mechanism directly on long sequence data, which will bring huge time cost in the reasoning stage and easily affect the click rate prediction accuracy, and is difficult to be applied in the click rate prediction task with high time performance requirement.

[0079] Therefore, it is an urgent problem for those skilled in the art to design a method that can efficiently extract information and improve the click rate prediction accuracy.

[0080] Therefore, the embodiments of the present disclosure provide a product click rate determination method. The method comprises: obtaining a product data training sample set, wherein the product data training sample set comprises a plurality of training samples, each training sample comprises a plurality of product feature vectors with a preset dimension and an interaction behavior sequence, and the interaction behavior sequence represents the time and order of the interaction behavior between a user and a plurality of products; for each training sample, processing the plurality of product feature vectors by using a mean clustering model to obtain a plurality of cluster center vectors, wherein the mean clustering model is obtained by pre-training a K-means clustering model, and each cluster center vector is a feature vector that can represent a plurality of product feature vectors in the same class; determining a sampling probability corresponding to each cluster center vector according to a preset feature vector of a preset target product and a plurality of cluster center vectors corresponding to the interaction behavior sequence; sampling processing the interaction behavior sequence according to the plurality of sampling probabilities to obtain a plurality of interaction behavior subsequences; training a deep interest network by using the preset feature vector and the plurality of interaction behavior subsequences to obtain a trained click rate prediction model; inputting a product feature vector of a product to be predicted and an interaction behavior sequence of a user to be predicted into the click rate prediction model, and outputting a prediction result, wherein the prediction result represents the probability of the user to be predicted clicking the product to be predicted.

[0081] Figure 1 An exemplary system architecture 100 to which the product click rate determination method according to embodiments of the present disclosure can be applied is schematically shown. It should be noted that, Figure 1The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, in order to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other equipment, systems, environments or scenarios.

[0082] like Figure 1 As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, and 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the terminal devices 101, 102, and 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links.

[0083] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as click-through rate prediction applications, web browser applications, search applications, instant messaging tools, email clients, and / or social platform software, etc. (for example only).

[0084] Terminal devices 101, 102, and 103 can be various electronic devices with displays and web browsing capabilities, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0085] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using terminal devices 101, 102, and 103 (for example only). The backend management server can analyze and process data such as received click-through rate prediction requests, and feed back the processing results (such as click-through rates, web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0086] It should be noted that the product click rate determination method provided by the embodiments of the present disclosure can generally be executed by the server 105. Accordingly, the product click rate determination apparatus provided by the embodiments of the present disclosure can generally be arranged in the server 105. The product click rate determination method provided by the embodiments of the present disclosure can also be executed by a server or a server cluster different from the server 105 and capable of communicating with the terminal devices 101, 102, 103 and / or the server 105. Accordingly, the product click rate determination apparatus provided by the embodiments of the present disclosure can also be arranged in a server or a server cluster different from the server 105 and capable of communicating with the terminal devices 101, 102, 103 and / or the server 105. Alternatively, the product click rate determination method provided by the embodiments of the present disclosure can also be executed by the terminal devices 101, 102, or 103, or by other terminal devices different from the terminal devices 101, 102, or 103. Accordingly, the product click rate determination apparatus provided by the embodiments of the present disclosure can also be arranged in the terminal devices 101, 102, or 103, or in other terminal devices different from the terminal devices 101, 102, or 103.

[0087] It should be understood that Figure 1 The number of terminal devices, networks and servers in the system is merely illustrative. Any number of terminal devices, networks and servers can be provided according to implementation needs.

[0088] Figure 2 A flowchart A of a product click rate determination method according to an embodiment of the present disclosure is schematically shown. Figure 3 A flowchart B of a product click rate determination method according to an embodiment of the present disclosure is schematically shown.

[0089] As shown in Figure 2 and Figure 3 The product click rate determination method includes operations S201-S206.

[0090] In operation S201, a product data training sample set is obtained, where the product data training sample set includes a plurality of training samples, each training sample including a plurality of product feature vectors with a preset dimension and an interaction behavior sequence, the interaction behavior sequence representing the time and order of interaction behaviors of a user with a plurality of products.

[0091] In operation S202, for each training sample, the plurality of product feature vectors are processed using a mean clustering model to obtain a plurality of cluster center vectors, where the mean clustering model is obtained by pre-training a K-means clustering model, and each cluster center vector is a feature vector capable of representing a plurality of product feature vectors in the same class.

[0092] At operation S203, a sampling probability corresponding to each cluster center vector is determined according to the preset feature vector of the preset target product and the plurality of cluster center vectors corresponding to the interaction behavior sequence.

[0093] At operation S204, the interaction behavior sequence is sampled according to the plurality of sampling probabilities to obtain a plurality of interaction behavior subsequences.

[0094] At operation S205, the depth interest network is trained using the preset feature vector and the plurality of interaction behavior subsequences to obtain a trained click rate prediction model.

[0095] At operation S206, the product feature vector of the product to be predicted and the interaction behavior sequence of the user to be predicted are input into the click rate prediction model, and a prediction result is output, wherein the prediction result represents the probability of the user to be predicted clicking the product to be predicted.

[0096] According to an embodiment of the present disclosure, the product feature vector is generated according to product data, and the product data refers to information data of each product in the user interaction sequence, including but not limited to the type and price of the product, etc., which is used to extract the representation information of the product. The product feature vectors in the plurality of training samples can be the same.

[0097] According to an embodiment of the present disclosure, the interaction behavior sequence can refer to N interaction behaviors of a user u composing the behavior sequence of the user, which can be represented by the symbol .

[0098] According to an embodiment of the present disclosure, the obtained product data and interaction behavior sequence can be within a preset time period, and the preset time period can be specifically set according to actual needs, for example, it can be three years. The preset dimension can be determined according to actual conditions, for example, it can be set to 50 dimensions.

[0099] According to an embodiment of the present disclosure, when the mean clustering model is pre-trained, a plurality of product feature vectors of the present disclosure can be used as pre-training samples, or a plurality of product feature vectors different from the present disclosure can be used as pre-training samples.

[0100] According to an embodiment of the present disclosure, when the mean clustering model processes the product feature vector, a corresponding mapping table can also be generated, and the mapping table is used to determine the cluster center vector closest to the preset feature vector.

[0101] According to an embodiment of the present disclosure, the plurality of product feature vectors in the training sample are processed by using the mean clustering model to obtain a clustering center capable of representing the plurality of product feature vectors in the same class. The sampling probability of each clustering center vector corresponding to the interaction behavior sequence of the preset feature vector of the preset target product is determined according to the preset feature vector. Based on the plurality of sampling probabilities, a small number of related products are sampled from the interaction behavior sequence to form an interaction behavior sub-sequence with high relevance, less noise and high information retention. For example, the products in a sampling probability interval are combined into an interaction behavior sub-sequence according to the size of the sampling probability. The preset feature vector e t and the plurality of interaction behavior sub-sequences are used to train a deep interest network to obtain a trained click rate prediction model.

[0102] According to an embodiment of the present disclosure, when the click rate prediction model is used, the product feature vector of the product to be predicted and the interaction behavior sequence of the user to be predicted are input into the click rate prediction model, and a prediction result of the probability of the user to be predicted clicking the product to be predicted is output, for example, the probability of the user to be predicted clicking the product to be predicted is 92%.

[0103] According to an embodiment of the present disclosure, the plurality of product feature vectors are processed by using the mean clustering model, so that the specific product feature vectors are replaced by the determined clustering centers, which is beneficial to eliminate noise, so as to calculate the sampling probability of the preset target product and each clustering center, and then generate the interaction behavior sub-sequence with high relevance, less noise and high information retention based on the sampling probability. The click rate prediction model trained by using the interaction behavior sub-sequence and the preset feature vector has high prediction accuracy when used.

[0104] According to an embodiment of the present disclosure, the plurality of product data obtained are processed to obtain screened product data, including the following operations:

[0105] According to the interaction behavior sequence, the interaction times of the user and each product are determined. The product with an interaction time greater than a preset threshold is determined as the screened product data.

[0106] According to an embodiment of the present disclosure, the preset threshold can be specifically set according to actual needs. For example, in order to process long sequence data, the preset threshold can be set to 50.

[0107] According to an embodiment of the present disclosure, the product feature vector is obtained by processing the product information of the product by using an embedding layer; and the sampling probability is obtained by processing by using a sampling module.

[0108] According to an embodiment of the present disclosure, the product click rate determination method further includes the following operations:

[0109] The initial neural network is trained by using samples in the optimization training sample set in sequence to obtain a first loss value corresponding to the optimization training sample, wherein the samples in the optimization training sample set include positive samples and negative samples, and the positive samples include a plurality of interaction behavior subsequences of the user, and the negative samples include a plurality of interaction behavior subsequences of at least one other user.

[0110] In a case where the first loss value does not satisfy the first convergence threshold, the model parameters of the embedding layer, the model parameters of the sampling module and the model parameters of the initial neural network are updated by using a stochastic gradient descent algorithm. In a case where the first loss value satisfies the first convergence threshold, the updated model parameters of the embedding layer are determined as target model parameters of the embedding layer, the updated model parameters of the sampling module are determined as target model parameters of the sampling module, and the initial neural network after the model parameters are updated is determined as a trained contrastive learning target model.

[0111] According to an embodiment of the present disclosure, the auxiliary task constituted by the above operations refers to a contrastive learning method proposed in the field of image processing, the idea of which is to obtain negative samples different from original data by adding noise to the original data, changing details and the like, and then to make the model learn to distinguish the original data from the negative samples through self-supervised learning, so as to effectively enhance the learning effect of the model without the need for more data.

[0112] According to an embodiment of the present disclosure, the interests of different users are often quite different, while the interests of the same user are often stable and not very different. Based on this common sense assumption, the embodiment samples multiple times in the interaction behavior sequence of each user, and samples multiple interaction behavior subsequences from the same user as positive samples and samples multiple interaction behavior subsequences from different users as negative samples, and trains a self-supervised contrastive learning target model according to a similar form of an InfoNCE loss function.

[0113] According to an embodiment of the present disclosure, the first loss value L C is calculated as shown in formulas (1) and (2):

[0114]

[0115] D ω (p, q) = exp(p T · W · q) (2)

[0116] wherein p and q represent the interaction behavior subsequences obtained by sampling twice from the user; P l represents a joint probability distribution of the user subsequences obtained by sampling, i.e., (p, q) ~ P l ; represents the interaction behavior subsequence obtained by randomly sampling from the interaction behavior sequence of at least one other user; D ωis the discriminator in the initial neural network with parameters ω, defined as a log bilinear model; p T Transposes p.

[0117] According to embodiments of the present disclosure, in each round of training, each user is sampled twice as positive samples p and q, and five negative samples are sampled from the interaction behavior sequences of other users Self-supervised contrastive learning is performed, a random gradient descent algorithm is executed by an Adam optimizer to optimize the parameters ω, and the parameters of the embedding layer, the sampling module and other models are optimized and updated by a back propagation mechanism, so as to enhance the model's capture and distinction of user interests, help the click rate prediction model to strengthen the memory of personal interests of each user, and thus effectively improve the prediction accuracy.

[0118] As shown in Figure 3 , a plurality of product feature vectors are processed by using a mean clustering model to obtain a plurality of cluster center vectors, including the following operations:

[0119] A plurality of product feature sub-vectors corresponding to each product feature vector are obtained by processing the plurality of product feature vectors by using the mean clustering model. A plurality of target product feature sub-vectors are selected from the plurality of product feature sub-vectors of the plurality of products according to a preset selection condition. A plurality of cluster center vectors are obtained according to the plurality of target product feature sub-vectors, wherein the number of cluster center vectors is less than the number of product feature vectors.

[0120] According to embodiments of the present disclosure, the cluster center vector can be one of the plurality of product feature vectors, or a cluster center vector spliced from the plurality of product feature vectors. The spliced cluster center vector can better represent the plurality of product feature vectors in the class.

[0121] According to embodiments of the present disclosure, the K-means clustering algorithm is used for clustering training based on all product feature vectors, wherein the number C of cluster center vectors can be set to one-tenth of the total number of products. For M products, the loss function of the K-means clustering method is shown in formula (3):

[0122]

[0123] Where e i is the product feature vector of product i, c j is the cluster center vector corresponding to the plurality of products i of the same class.

[0124] According to an embodiment of the present disclosure, in each round of training, clustering training is first performed to ensure that the clustering center vectors always have good representativeness. After this step, a product-clustering center mapping table and K clustering center vectors can be obtained, and the mapping table and the K clustering center vectors obtained in the last round of training are stored for online deployment.

[0125] According to an embodiment of the present disclosure, when the clustering center vectors are spliced clustering center vectors, the preset selection condition can refer to selecting according to different product categories. For example, a preset number of product target product feature sub-vectors are selected from the product feature sub-vectors of the plurality of products of each category, and the preset number of product target product feature sub-vectors of each category are spliced to obtain one or more clustering center vectors of the category.

[0126] According to an embodiment of the present disclosure, when the clustering center vectors are one of the plurality of product feature vectors, one or more clustering center vectors that can represent different categories can be directly selected from the plurality of product feature vectors.

[0127] As shown in Figure 3 According to the preset feature vector of the preset target product and the plurality of clustering center vectors corresponding to the interaction behavior sequence, a sampling probability corresponding to each clustering center vector is determined, including the following operations:

[0128] According to the preset feature vector and the interaction behavior sequence, a similarity score of the preset feature vector and each clustering center vector is determined. According to the interaction behavior sequence, a relative time difference between the preset target product and each clustering center vector is determined. According to the plurality of similarity scores and the plurality of relative time differences, a sampling probability corresponding to each product is determined.

[0129] According to an embodiment of the present disclosure, according to the plurality of similarity scores and the plurality of relative time differences, a sampling probability corresponding to each product is determined, including the following operations:

[0130] For each relative time difference, the relative time difference is converted into a time sequence score. According to the time sequence score and the similarity score corresponding to the relative time difference, a sampling probability of each product is calculated.

[0131] According to an embodiment of the present disclosure, before the above operations are performed, the plurality of products i in the interaction behavior sequence are replaced by the plurality of clustering center vectors i that have been determined.

[0132] According to an embodiment of the present disclosure, given the preset feature vector e t of the preset target product, first, a weighted vector inner product is used to measure the similarity between each clustering center vector i in the user's interaction behavior sequence and the preset target product t, to obtain a similarity score r i = (W a ei )·(W b e t ) T wherein W a and W b are corresponding weights, and T is the time when the interaction behavior occurs.

[0133] According to an embodiment of the present disclosure, determining the relative time difference between the preset target product and each cluster center vector according to the interaction behavior sequence comprises the following operations:

[0134] The relative time difference is used to measure the time sequence relationship of the plurality of cluster center vectors corresponding to the interaction behavior sequence, to obtain a timestamp set, wherein each timestamp in the timestamp set corresponds to the time when the user interacts with a cluster center vector. The relative time difference between the preset target product and each cluster center vector is determined based on the timestamp set.

[0135] According to an embodiment of the present disclosure, the relative time difference is used to measure the time sequence relationship of each cluster center vector i in the interaction behavior sequence. Formally, for the interaction behavior sequence of the user u, the timestamp of each behavior can be obtained, and the timestamp set is recorded as

[0136] According to an embodiment of the present disclosure, in the interaction behavior sequence, the smallest time difference is recorded as Thus, the relative time difference between the cluster center vector i and the preset target product t in the interaction behavior sequence can be obtained The relative time difference is embedded into the time sequence score by an encoder, and then weighted and summed with the similarity score to obtain the final sampling score R i =r i +W c ·E(l i ), wherein W c is the weight of the summation. In the above manner, the sampling score of each cluster center vector i in the interaction behavior sequence of the user u for a given preset target product t can be calculated, denoted as Thus, by using a Softmax-like method, the sampling weight of each product i, i.e., the sampling probability, can be obtained, as shown in formula (4).

[0137]

[0138] According to an embodiment of the present disclosure, in the last round of training, the sampling weight of each user when the plurality of cluster centers are used as the interaction behavior sequence is saved. Thus, when online reasoning, the product to be predicted can be directly mapped to a certain cluster center, and the saved sampling weight can be directly searched, so that the overhead of calculating the sampling weight can be saved, the sampling probability and the subsequent click rate prediction in the user sequence are very time-consuming, and the time efficiency is ensured.

[0139] Figure 4 A flowchart of obtaining a product feature vector is shown.

[0140] As shown in Figure 4 , a plurality of product feature vectors are obtained through operations S401-S407.

[0141] In operation S401, the obtained plurality of product data is filtered to obtain filtered product data.

[0142] In operation S402, for the plurality of filtered product data, if the product attributes of the product data meet the first preset condition, the plurality of categories of the product data are respectively processed by one-hot encoding to obtain a first one-hot vector corresponding to each category, wherein the product attributes include product price, product brand, or purchase time.

[0143] In operation S403, the plurality of first one-hot vectors are spliced to obtain a spliced feature vector of the product data.

[0144] In operation S404, if the product attributes of the product data do not meet the first preset condition, the product is processed by a bucketing method to obtain a category vector of a plurality of categories.

[0145] In operation S405, the category vector of each category is processed by one-hot encoding to obtain a second one-hot vector corresponding to each category.

[0146] In operation S406, the plurality of second one-hot vectors are spliced to obtain a spliced feature vector of the product data.

[0147] In operation S407, the spliced feature vector is mapped to a preset dimension to obtain a product feature vector.

[0148] According to an embodiment of the present disclosure, data with extremely low frequency is difficult to learn, so the present disclosure can first remove product data with a frequency of occurrence in the interaction behavior sequence of all users.

[0149] According to an embodiment of the present disclosure, the first preset condition can refer to whether the product price, product brand, purchase time, or product category is continuous.

[0150] In an embodiment, when it is determined that the categories of the plurality of filtered product data are relatively discrete, one-hot encoding can be performed on each product data to obtain a corresponding first one-hot vector. For example, the categories of the filtered product data total 1882, and each product has a one-hot vector with a length of 1882 to represent its category.

[0151] In another embodiment, when it is determined that the product price, product brand, and purchase time of the plurality of screened product data are relatively continuous, the plurality of screened product data are first subjected to a bucketing process, for example, the plurality of product data are divided into a low-price zone, a medium-price zone, and a high-price zone to obtain converted category features, and then the converted category features are subjected to one-hot encoding to obtain a corresponding second one-hot vector.

[0152] According to an embodiment of the present disclosure, the one-hot feature vector of each product is spliced and mapped to a dense feature vector with a dimension of 50, i.e., a product feature vector, through an embedding layer, as the final representation of the product. Formally, the final representation of the product feature vector of product i is e i = E(x i ), where E is a weighted embedding layer encoder, x i is the spliced feature vector of the product.

[0153] According to an embodiment of the present disclosure, the preset feature vector and the plurality of interaction behavior sub-sequences are used to train a deep interest network to obtain a trained click rate prediction model, including the following operations:

[0154] The preset feature vector and the plurality of product feature vectors corresponding to each interaction behavior sub-sequence are input into an attention mechanism layer to obtain a plurality of correlation weights, wherein each correlation weight represents the correlation between the preset feature vector and one product feature vector.

[0155] The plurality of correlation weights are input into a pooling layer to obtain a target representation vector, wherein the target representation vector represents a vector of the relationship between the user and the preset target product. The target representation vector is input into a multilayer perceptron layer to output a training prediction result.

[0156] According to the training prediction result and a loss function, a second loss value is calculated. In a case where the second loss value does not satisfy a second convergence threshold, the model parameters of the deep interest network are updated using a stochastic gradient descent algorithm. In a case where the second loss value satisfies the second convergence threshold, the updated deep interest network is determined as the trained click rate prediction model.

[0157] According to an embodiment of the present disclosure, predicting whether a user is willing to click on a preset target product is the target task of the present disclosure. In the foregoing description, a user behavior short sequence has been sampled and obtained, and therefore the click rate prediction module only needs to use the short sequence and the preset feature vector of the preset target product to complete the prediction.

[0158] According to an embodiment of the present disclosure, in order to obtain a faster prediction speed and a more stable and accurate prediction effect, the present disclosure selects a DIN (Deep Interest Network) model based on an attention mechanism as the click rate prediction module. This model adopts an Embedding & MLP architecture. First, the preset feature vector e t and the cluster center vectors e i in the interaction behavior sub-sequence, then further obtains the relevance weight of the preset target product and each cluster center vector e i in the interaction behavior sub-sequence through the attention mechanism, and then sums and pools to obtain the joint vector representation of the user and the preset target product as the target representation vector, which is then input into a multilayer perception (MLP). Finally, the training of the click rate prediction model is completed through multiple training samples.

[0159] It should be noted that when the target representation vector is input into the multilayer perception, the target representation vector and the feature vector after the splicing and smoothing processing of the additional features of the preset target product can also be used as the final target representation vector. The additional features include but are not limited to the merchant location of the preset target product, the popularity of the product, and the promotion content of the product.

[0160] The click rate prediction model adopts a cross-entropy loss function to train and optimize the click rate prediction model, which is specifically shown in formula (5):

[0161]

[0162] Wherein, B is a product data training sample set with a data amount D; x is the target representation vector; p(x) ∈ [0, 1] represents the click probability predicted by the multilayer perception for a sample; and y is the final prediction label, which takes the value of 1 or 0, representing that the sample actually belongs to a positive example or a negative example, respectively.

[0163] According to an embodiment of the present disclosure, according to the loss function, the random gradient descent method is also used to perform gradient update in each round of training by using the Adam optimizer, and the training is performed until the second loss value meets the second convergence threshold.

[0164] Figure 5 A recommended flowchart of a product according to an embodiment of the present disclosure is schematically shown.

[0165] As Figure 5As shown, since the click rate prediction model has learned the user preferences implied in the long interaction behavior sequence during training, and saves a small number of parameters such as the corresponding cluster center vector and sampling probability, it can quickly sample the relevant interaction behavior sub-sequence during prediction and perform efficient and accurate click rate prediction, meeting the higher timeliness requirements in the online environment. Therefore, specific use can be performed through operations S501 to operation 504.

[0166] Operation S501, deploy the click rate prediction model to the e-commerce platform, so as to determine the click rate of each product through the above product click rate determination method.

[0167] Operation S502, the e-commerce platform can recommend to the user according to the predicted click rate of each product of the user, so as to improve the purchase quantity of the product.

[0168] Operation S503, the e-commerce platform can update the product feature vector and the interaction behavior sequence locally and retrain the click rate prediction model every interval of a preset time.

[0169] Operation S504, update the online deployed click rate prediction model, so that the click rate prediction model can pay attention to the latest interest of the user at any time, thereby ensuring the accuracy of the prediction.

[0170] According to an embodiment of the present disclosure, the interaction behavior sequence is determined by the following operation:

[0171] According to the average sequence length, the initial interaction behavior sequence of each user is intercepted or padded to obtain the interaction behavior sequence.

[0172] According to an embodiment of the present disclosure, the average length N of the plurality of initial interaction behavior sequences is recorded in advance, and the initial interaction behavior sequence is then truncated or padded to make all interaction behavior sequences N. Then, the cluster center vectors corresponding to all products in the initial interaction behavior sequence are spliced in turn to obtain the feature vector of each user as the final representation u of the user. i As shown in formula (6):

[0173]

[0174] After the above processing, the product feature vector and the interaction behavior sequence are processed into low-dimensional dense embedding vectors, which are convenient for parameter training and calculation. The clustering method makes the final representation of the user only contain the representative product feature vector, that is, the obtained cluster center vector, which is beneficial to eliminate noise, and uses the cluster center to replace the specific product in subsequent calculation, optimizes the storage and operation performance of online deployment.

[0175] Figure 6A block diagram of a product click rate determination apparatus is shown schematically.

[0176] As shown in Figure 6 The product click rate determination apparatus 600 includes an acquisition module 601, a clustering module 602, a first determination module 603, a second determination module 604, a training module 605, and a prediction module 606.

[0177] The acquisition module 601 is configured to acquire a product data training sample set, where the product data training sample set includes a plurality of training samples, each training sample includes a plurality of product feature vectors with a preset dimension and an interaction behavior sequence, and the interaction behavior sequence represents the time and order of interaction behaviors between a user and a plurality of products.

[0178] The clustering module 602 is configured to, for each training sample, process the plurality of product feature vectors by using a mean clustering model to obtain a plurality of cluster center vectors, where the mean clustering model is obtained by pre-training a K-means clustering model, and each cluster center vector is a feature vector that can represent a plurality of product feature vectors in the same class.

[0179] The first determination module 603 is configured to determine a sampling probability corresponding to each cluster center vector according to a preset feature vector of a preset target product and a plurality of cluster center vectors corresponding to the interaction behavior sequence.

[0180] The second determination module 604 is configured to sample and process the interaction behavior sequence according to the plurality of sampling probabilities to obtain a plurality of interaction behavior subsequences.

[0181] The training module 605 is configured to train a deep interest network by using the preset feature vector and the plurality of interaction behavior subsequences to obtain a trained click rate prediction model.

[0182] The prediction module 606 is configured to input a product feature vector of a product to be predicted and an interaction behavior sequence of a user to be predicted into the click rate prediction model, and output a prediction result, where the prediction result represents a probability of the user to be predicted clicking the product to be predicted.

[0183] According to embodiments of the present disclosure, the plurality of product feature vectors are processed by the mean clustering model to replace the specific product feature vectors with the determined cluster centers, which is beneficial to eliminate noise, so as to calculate the sampling probability of the preset target product and each cluster center, and further generate an interaction behavior subsequence with high relevance, less noise, and high information retention based on the sampling probability. The click rate prediction model trained by using the interaction behavior subsequence and the preset feature vector has high prediction accuracy when used.

[0184] According to an embodiment of the present disclosure, the product click rate determination apparatus 600 further comprises a screening module, a first encoding module, a first splicing module, a bucketing module, a second encoding module, a second splicing module, and a mapping module.

[0185] The screening module is configured to perform screening processing on the obtained product data to obtain screened product data.

[0186] The first encoding module is configured to, for each piece of screened product data, perform one-hot encoding processing on the multiple categories of product data respectively in a case where a product attribute of the product data meets a first preset condition, to obtain a first one-hot vector corresponding to each category, wherein the product attribute comprises a product price, a product brand, or a purchase time.

[0187] The first splicing module is configured to perform splicing processing on the multiple first one-hot vectors to obtain a spliced feature vector of the product data.

[0188] The bucketing module is configured to, in a case where the product attribute of the product data does not meet the first preset condition, perform processing on the product by using a bucketing method to obtain a category vector of multiple categories.

[0189] The second encoding module is configured to perform one-hot encoding processing on the category vector of each category to obtain a second one-hot vector corresponding to each category.

[0190] The second splicing module is configured to perform splicing processing on the multiple second one-hot vectors to obtain the spliced feature vector of the product data.

[0191] The mapping module is configured to perform mapping processing on the spliced feature vector in a preset dimension to obtain a product feature vector.

[0192] According to an embodiment of the present disclosure, the screening module comprises a first determination unit and a screening unit.

[0193] The first determination unit is configured to determine, according to the interaction behavior sequence, the number of interactions between the user and each product.

[0194] The screening unit is configured to determine, as the screened product data, a product whose number of interactions is greater than a preset threshold.

[0195] According to an embodiment of the present disclosure, the product feature vector is obtained by processing product information of the product by using an embedding layer; and the sampling probability is obtained by processing by using a sampling module.

[0196] According to an embodiment of the present disclosure, the product click rate determination apparatus 600 further comprises an optimization training module, a first updating module, and a second updating module.

[0197] The optimization training module is configured to train the initial neural network by using samples in the optimization training sample set in sequence to obtain a first loss value corresponding to the optimization training sample, wherein the samples in the optimization training sample set include positive samples and negative samples, and the positive samples include a plurality of interaction behavior sub-sequences of the user, and the negative samples include a plurality of interaction behavior sub-sequences of at least one other user.

[0198] The first updating module is configured to update the model parameters of the embedding layer, the model parameters of the sampling module, and the model parameters of the initial neural network by using a stochastic gradient descent algorithm in a case where the first loss value does not satisfy the first convergence threshold.

[0199] The second updating module is configured to determine the updated model parameters of the embedding layer as target model parameters of the embedding layer, determine the updated model parameters of the sampling module as target model parameters of the sampling module, and determine the initial neural network after the model parameters are updated as a trained contrastive learning target model in a case where the first loss value satisfies the first convergence threshold.

[0200] According to an embodiment of the present disclosure, the clustering module 602 includes a first obtaining unit, a selecting unit, and a second obtaining unit.

[0201] The first obtaining unit is configured to process the plurality of product feature vectors by using a mean clustering model to obtain a plurality of product feature sub-vectors corresponding to each product feature vector.

[0202] The selecting unit is configured to select a plurality of target product feature sub-vectors from the plurality of product feature sub-vectors of the plurality of products according to a preset selection condition.

[0203] The second obtaining unit is configured to obtain a plurality of clustering center vectors according to the plurality of target product feature sub-vectors, wherein the number of the clustering center vectors is less than the number of the product feature vectors.

[0204] According to an embodiment of the present disclosure, the first determining module 603 includes a second determining unit, a third determining unit, and a fourth determining unit.

[0205] The second determining unit is configured to determine a similarity score of the preset feature vector and each clustering center vector according to the preset feature vector and the interaction behavior sequence.

[0206] The third determining unit is configured to determine a relative time difference of the preset target product and each clustering center vector according to the interaction behavior sequence.

[0207] The fourth determining unit is configured to determine a sampling probability corresponding to each product according to the plurality of similarity scores and the plurality of relative time differences.

[0208] According to an embodiment of the present disclosure, the third determining unit includes an obtaining sub-unit and a determining sub-unit.

[0209] The obtaining sub-unit is configured to obtain a timestamp set by measuring a time sequence relationship of a plurality of cluster center vectors corresponding to the interaction behavior sequence with the relative time difference, wherein each timestamp in the timestamp set corresponds to a time when the user interacts with one cluster center vector.

[0210] The determining sub-unit is configured to determine the relative time difference between the preset target product and each cluster center vector based on the timestamp set.

[0211] According to an embodiment of the present disclosure, the fifth determining unit includes a converting sub-unit and a calculating sub-unit.

[0212] The converting sub-unit is configured to convert the relative time difference into a time sequence score for each relative time difference.

[0213] The calculating sub-unit is configured to calculate a sampling probability of each product according to the time sequence score and a similarity score corresponding to the relative time difference.

[0214] According to an embodiment of the present disclosure, the training module 605 includes an input unit, a pooling unit, a perception unit, a calculating unit, an updating unit, and a fifth determining unit.

[0215] The input unit is configured to input a preset feature vector and a plurality of product feature vectors corresponding to each interaction behavior sub-sequence into an attention mechanism layer to obtain a plurality of correlation weights, wherein one correlation weight represents a correlation between the preset feature vector and one product feature vector.

[0216] The pooling unit is configured to input the plurality of correlation weights into a pooling layer to obtain a target representation vector, wherein the target representation vector represents a vector of the relationship between the user and the preset target product.

[0217] The perception unit is configured to input the target representation vector into a multi-layer perception layer to output a training prediction result.

[0218] The calculating unit is configured to calculate a second loss value according to the training prediction result and a loss function.

[0219] The updating unit is configured to update the model parameters of the deep interest network by using a stochastic gradient descent algorithm in a case where the second loss value does not satisfy a second convergence threshold.

[0220] The fifth determining unit is configured to determine the updated deep interest network as the trained click rate prediction model in a case where the second loss value satisfies the second convergence threshold.

[0221] Any of the modules, units, sub-units, or at least part of any of them according to embodiments of the present disclosure can be implemented in one module. Any of the modules, units, sub-units according to embodiments of the present disclosure can be split into a plurality of modules. Any of the modules, units, sub-units according to embodiments of the present disclosure can be implemented at least in part as a hardware circuit, for example, a Field Programmable Gate Array (FPGA), a Programmable Logic Array (PLA), a System on Chip, a System on Substrate, a System on Package, an Application Specific Integrated Circuit (ASIC), or any other reasonable manner of hardware or firmware by integrating or packaging circuits, or in any one of software, hardware, and firmware, or in a proper combination of any of them. Alternatively, one or more of the modules, units, sub-units according to embodiments of the present disclosure can be implemented at least in part as computer program modules, which can perform corresponding functions when the computer program modules are run.

[0222] For example, any of the modules 601, 602, 603, 604, 605, and 606 can be combined in one module / unit / sub-unit, or any of them can be split into a plurality of modules / units / sub-units. Alternatively, at least part of the functions of one or more of the modules / units / sub-units can be combined with at least part of the functions of other modules / units / sub-units, and implemented in one module / unit / sub-unit. According to embodiments of the present disclosure, at least one of the modules 601, 602, 603, 604, 605, and 606 can be implemented at least in part as a hardware circuit, for example, a Field Programmable Gate Array (FPGA), a Programmable Logic Array (PLA), a System on Chip, a System on Substrate, a System on Package, an Application Specific Integrated Circuit (ASIC), or any other reasonable manner of hardware or firmware by integrating or packaging circuits, or in any one of software, hardware, and firmware, or in a proper combination of any of them. Alternatively, at least one of the modules 601, 602, 603, 604, 605, and 606 can be implemented at least in part as computer program modules, which can perform corresponding functions when the computer program modules are run.

[0223] It should be noted that the product click rate determination apparatus part in the embodiments of the present disclosure corresponds to the product click rate determination method part in the embodiments of the present disclosure, and the description of the product click rate determination apparatus part is specifically referred to the product click rate determination method part, which will not be repeated here.

[0224] Figure 7 A block diagram of an electronic device suitable for implementing the method described above according to an embodiment of the present disclosure is schematically shown. Figure 7 The electronic device shown is merely an example and should not bring any limitation to the function and use range of the embodiments of the present disclosure.

[0225] As shown in Figure 7 The electronic device 700 according to an embodiment of the present disclosure includes a processor 701, which can perform various appropriate actions and processes according to programs stored in a Read-Only Memory (ROM) 702 or loaded into a Random Access Memory (RAM) 703 from a storage portion 708. The processor 701 may, for example, include a general-purpose microprocessor (such as a CPU), an instruction set processor, and / or a related chipset, and / or a special-purpose microprocessor (such as an Application-Specific Integrated Circuit (ASIC)), etc. The processor 701 can also include an on-board memory for cache use. The processor 701 can include a single processing unit or multiple processing units for performing different actions of the method processes according to the embodiments of the present disclosure.

[0226] In the RAM 703, various programs and data required for the operation of the electronic device 700 are stored. The processor 701, the ROM 702, and the RAM 703 are connected to each other through a bus 704. The processor 701 performs various operations of the method processes according to the embodiments of the present disclosure by executing the programs in the ROM 702 and / or the RAM 703. It should be noted that the programs can also be stored in one or more memories other than the ROM 702 and the RAM 703. The processor 701 can also perform various operations of the method processes according to the embodiments of the present disclosure by executing the programs stored in the one or more memories.

[0227] According to an embodiment of the present disclosure, the electronic device 700 can further include an input / output (I / O) interface 705 that is also connected to the bus 704. The system 700 can further include one or more of the following components connected to the I / O interface 705: an input part 706 including a keyboard, a mouse, etc.; an output part 707 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage part 708 including a hard disk, etc.; and a communication part 709 including a network interface card such as a LAN card, a modem, etc. The communication part 709 performs communication processing via a network such as the Internet. A drive 710 is also connected to the I / O interface 705 as necessary. A removable medium 711 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is mounted on the drive 710 as necessary, so that a computer program read therefrom is installed in the storage part 708 as necessary.

[0228] According to an embodiment of the present disclosure, the method flow according to the embodiment of the present disclosure can be implemented as a computer software program. For example, the embodiment of the present disclosure includes a computer program product including a computer program carried on a computer-readable storage medium, the computer program containing program codes for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network by the communication part 709, and / or installed from the removable medium 711. When the computer program is executed by the processor 701, the above-described functions defined in the system implementing the embodiment of the present disclosure are performed. According to an embodiment of the present disclosure, the system, device, apparatus, module, unit, etc. described above can be implemented by computer program modules.

[0229] The present disclosure also provides a computer-readable storage medium, which can be included in the device / apparatus / system described in the above embodiments; or can exist separately without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, when the one or more programs are executed, the method according to the embodiment of the present disclosure is implemented.

[0230] According to an embodiment of the present disclosure, the computer readable storage medium can be a non-volatile computer readable storage medium. For example, it can include, but is not limited to, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a Read Only Memory (ROM), an Erasable Programmable Read Only Memory (EPROM or Flash memory), a portable compact disc read only memory (CD-ROM), an optical memory device, a magnetic memory device, or any suitable combination of the foregoing. In the present disclosure, a computer readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in connection with an instruction execution system, apparatus, or device.

[0231] For example, according to an embodiment of the present disclosure, the computer readable storage medium can include one or more memories of the ROM 702 and / or the RAM 703 described above and / or other than the ROM 702 and the RAM 703.

[0232] Embodiments of the present disclosure also include a computer program product, which includes a computer program containing program codes for executing the method provided by the embodiments of the present disclosure, and when the computer program product is run on an electronic device, the program codes are used to make the electronic device implement the product click rate determination method provided by the embodiments of the present disclosure.

[0233] When the computer program is executed by the processor 701, the above-mentioned functions defined in the system / apparatus of the embodiments of the present disclosure are performed. According to an embodiment of the present disclosure, the system, apparatus, module, unit, etc. described above can be implemented by computer program modules.

[0234] In one embodiment, the computer program can rely on a tangible storage medium such as an optical storage medium, a magnetic storage medium, etc. In another embodiment, the computer program can also be transmitted, distributed, and downloaded in the form of a signal on a network medium, and be downloaded and installed through the communication part 709, and / or be installed from the detachable medium 711. The program codes contained in the computer program can be transmitted by any appropriate network medium, including but not limited to wireless, wired, etc., or any suitable combination of the foregoing.

[0235] According to embodiments of the present disclosure, program code of a computer program provided by embodiments of the present disclosure can be written in any combination of one or more programming languages, and specifically, can be implemented using a high-level procedural and / or object-oriented programming language, and / or an assembly / machine language. Programming languages include, but are not limited to, Java, C++, python, "C" language, or similar programming languages. Program code can execute entirely on a user's computing device, partly on a user device, partly on a remote computing device, or entirely on a remote computing device or server. In the latter scenario, the remote computing device can be connected to the user's computing device through any kind of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device, such as through the Internet using an Internet Service Provider.

[0236] The flow diagrams and the block diagrams in the drawings are illustrations of architectures, functionalities, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flow diagrams or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flow diagrams, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or operations, or combinations of special purpose hardware and computer instructions.

[0237] Embodiments of the present disclosure have been described. However, these embodiments are merely intended to illustrate the present disclosure, and are not intended to limit the scope of the present disclosure. Although each of the embodiments is described above separately, this does not mean that the measures in each of the embodiments cannot be used advantageously in combination. The scope of the present disclosure is defined by the appended claims and their equivalents. Without departing from the scope of the present disclosure, those skilled in the art can make various substitutions and modifications, and these substitutions and modifications should all fall within the scope of the present disclosure.

Claims

1. A product click rate determination method, comprising: obtaining a product data training sample set, wherein the product data training sample set comprises a plurality of training samples, each of the training samples comprises a plurality of product feature vectors with a preset dimension and an interaction behavior sequence, and the interaction behavior sequence represents a time and an order of interaction behaviors of a user with a plurality of products; for each of the training samples, processing the plurality of product feature vectors by using a mean clustering model to obtain a plurality of cluster center vectors, wherein the mean clustering model is obtained by pre-training a K-means clustering model, and each of the cluster center vectors is a feature vector capable of representing a plurality of product feature vectors in the same category; determining a similarity score of a preset feature vector of a preset target product and each of the cluster center vectors according to the preset feature vector and the interaction behavior sequence; measuring a time sequence relationship of a plurality of cluster center vectors corresponding to the interaction behavior sequence by using a relative time difference to obtain a timestamp set, wherein each timestamp in the timestamp set corresponds to a time of interaction behaviors of a user with one of the cluster center vectors; determining a relative time difference between the preset target product and each of the cluster center vectors based on the timestamp set; embedding the relative time difference into a time sequence score by using an encoder, and then performing weighted summation of the time sequence score and the similarity score to obtain a sampling score, so as to determine a sampling probability of each product according to the sampling score; sampling and processing the interaction behavior sequence according to a plurality of the sampling probabilities to obtain a plurality of interaction behavior subsequences; training a deep interest network by using the preset feature vector and the plurality of interaction behavior subsequences to obtain a trained click rate prediction model; inputting a product feature vector of a product to be predicted and an interaction behavior sequence of a user to be predicted into the click rate prediction model to output a prediction result, wherein the prediction result represents a probability of the user to be predicted clicking the product to be predicted.

2. The method of claim 1, wherein, The plurality of product feature vectors are obtained by: performing screening processing on the obtained plurality of product data to obtain screened product data; for a plurality of the screened product data, when a product attribute of the product data meets a first preset condition, performing one-hot encoding processing on a plurality of categories of the product data respectively to obtain a first one-hot vector corresponding to each of the categories, wherein the product attribute comprises a product price, a product brand or a purchase time; performing splicing processing on a plurality of the first one-hot vectors to obtain a spliced feature vector of the product data; when the product attribute of the product data does not meet the first preset condition, processing the product by using a bucketing method to obtain a category vector of each category; performing one-hot encoding processing on the category vector of each category to obtain a second one-hot vector corresponding to each category; performing splicing processing on a plurality of the second one-hot vectors to obtain a spliced feature vector of the product data; performing mapping processing on the spliced feature vector in a preset dimension to obtain one of the product feature vectors.

3. The method of claim 2, wherein, The obtained multiple product data are filtered to obtain filtered product data, and the filtering includes: determining the number of interactions of the user with each product according to the interaction behavior sequence; determining a product with an interaction number greater than a preset threshold as the filtered product data. 4.The method of claim 1 or 2, wherein the product feature vector is obtained by processing product information of the product using an embedding layer. The sampling probability is obtained by processing by using the sampling module. The method further includes: training the initial neural network by using samples in an optimized training sample set in sequence to obtain a first loss value corresponding to the optimized training sample, wherein the samples in the optimized training sample set include positive samples and negative samples, the positive samples include multiple interaction behavior sub-sequences of the user, and the negative samples include multiple interaction behavior sub-sequences of at least one other user; in a case where the first loss value does not satisfy a first convergence threshold, updating model parameters of the embedding layer, model parameters of the sampling module, and model parameters of the initial neural network by using a stochastic gradient descent algorithm; in a case where the first loss value satisfies the first convergence threshold, determining the updated model parameters of the embedding layer as target model parameters of the embedding layer, determining the updated model parameters of the sampling module as target model parameters of the sampling module, and determining the initial neural network with the updated model parameters as a trained contrastive learning target model.

5. The method of claim 4, wherein, the first loss value The calculation is shown in Equations (1) and (2): （1） （2） wherein, and denotes the sub-sequence of interaction behaviors sampled twice by the user; denotes the joint probability distribution of the sub-sequence of users sampled, i.e. ; denotes the sub-sequence of interaction behaviors randomly sampled from the interaction behavior sequence of at least one other user; is a discriminator in the initial neural network with parameters , defined as a log bilinear model; denotes the transpose of .

6. The method of claim 1, wherein, The multiple product feature vectors are processed by using a mean clustering model to obtain multiple cluster center vectors, and the processing includes: processing multiple product feature vectors by using a mean clustering model to obtain multiple product feature sub-vectors corresponding to each product feature vector; selecting multiple target product feature sub-vectors from multiple product feature sub-vectors of multiple products according to a preset selection condition; obtaining multiple cluster center vectors according to multiple target product feature sub-vectors, wherein the number of the cluster center vectors is less than the number of the product feature vectors.

7. The method of claim 1, wherein, The preset feature vector and the multiple interaction behavior sub-sequences are used to train a deep interest network to obtain a trained click rate prediction model, and the training includes: inputting a preset feature vector and multiple product feature vectors corresponding to each interaction behavior sub-sequence into an attention mechanism layer to obtain multiple correlation weights, wherein one correlation weight represents the correlation between the preset feature vector and one product feature vector; inputting the multiple correlation weights into a pooling layer to obtain a target representation vector, wherein the target representation vector represents a vector of the relationship between the user and the preset target product; inputting the target representation vector into a multilayer perceptron layer to output a training prediction result; calculating a second loss value according to the training prediction result and a loss function; in a case where the second loss value does not satisfy a second convergence threshold, updating model parameters of the deep interest network by using a stochastic gradient descent algorithm; in a case where the second loss value satisfies the second convergence threshold, determining the updated deep interest network as the trained click rate prediction model.

Images