A method for predicting product conversion rate

Abstract

The present invention belongs to the field of Internet application technology, and specifically relates to a product conversion rate prediction method, comprising: obtaining original data through an e-commerce platform, and calculating user-product interaction features based on the original data; inputting the user-product interaction features into a multi-task product prediction model to train the multi-task product prediction model; obtaining attribute information of a target user and a target product, and calculating the target user-product interaction features, inputting the target user-product interaction features into the trained multi-task product prediction model to predict the target user's predicted conversion rate for the target product, the target user's predicted click-through conversion rate for the target product, and the target user's predicted delayed conversion rate for the target product, and using an EM algorithm to obtain the target user's final conversion rate for the target product. The present invention can accurately predict the conversion rate of products, accurately recommend products to users based on the final conversion rate, increase the transaction volume of the platform, and save users' browsing time.

Description

Technical Field

[0001] The present invention belongs to the field of Internet application technology, and particularly relates to a commodity conversion rate prediction method. Background Art

[0002] As the internet's demographic dividend fades and user growth slows, promoting user activity and increasing user stickiness have become key future development directions for e-commerce platforms. Product conversion rate, which refers to the probability that a user clicks on a product and ultimately purchases it, is one of the most important performance evaluation metrics for e-commerce platforms. Effectively predicting product conversion rates can accurately establish the relationship between user preferences and products. By effectively enhancing the user's actual purchasing experience through user profiling, it can significantly increase platform revenue. Therefore, an efficient conversion rate prediction model is crucial for e-commerce platforms to build core competitiveness.

[0003] In recent years, many scholars have conducted extensive research on product conversion rate (CVR) prediction. For example, the Alimama team proposed a CVR prediction model, ESMM, in their paper "Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate," published at SIGIR'2018. This model uses click-through rate and click-through conversion rate as auxiliary task models for product conversion rate, enabling product conversion rate training based on the entire sample space. This effectively addresses the two key issues of data sparsity and sample selection bias that face product conversion rate prediction in real-world scenarios. While ESMM addresses these issues, it fails to account for the delay between a user clicking a product and its order being placed, resulting in low accuracy and poor prediction results for product conversion rate. Summary of the Invention

[0004] In order to solve the problem in the prior art that the delay time between a user clicking a product and the product being ordered by the user is not taken into account, resulting in low accuracy and poor prediction effect in the prediction results of the product conversion rate, the present invention proposes a product conversion rate prediction method, which includes the following steps:

[0005] S1: Obtaining raw data through the e-commerce platform; the raw data includes: user attribute information, product attribute information, user historical click behavior information, and user historical conversion behavior information;

[0006] S2: Perform feature processing on the user attribute information and the product attribute information to obtain a user feature matrix and a product feature matrix, and then perform feature concatenation on the user feature matrix and the product feature matrix to obtain a user-product interaction feature matrix;

[0007] S3: Estimate the user's click-through rate, click-to-conversion rate, and delayed conversion rate based on the user's historical click behavior and conversion behavior information. Use the user's click-to-product rate as a click label, the user's click-to-product conversion rate as a click-to-conversion label, and the user's delayed conversion rate as a delayed conversion label.

[0008] S4: Constructing a multi-task product prediction model, and using the user-product interaction features in the user-product interaction feature matrix as input to train the multi-task product prediction model, wherein the multi-task product prediction model includes: a click-through rate learning task model, a click-through conversion rate learning task model, and a delayed conversion rate learning task model;

[0009] S5: Acquire attribute information of the target user and target product, perform feature extraction on the attribute information of the target user and target product to obtain feature vectors of the target user and target product, and concatenate the feature vectors of the target user and target product to obtain target user-product interaction features;

[0010] S6: Input the target user-item interaction features into the trained multi-task item prediction model to predict the target user's predicted click-through rate for the target item, the target user's predicted click-through conversion rate for the target item, and the target user's predicted delayed conversion rate for the target item.

[0011] S7: Calculate the final conversion rate of the target user for the target product using the EM algorithm based on the predicted conversion rate of the target user for the target product, the predicted click-through conversion rate of the target user for the target product, and the predicted delayed conversion rate of the target user for the target product.

[0012] The present invention has at least the following beneficial effects

[0013] The present invention collects user attribute information, product attribute information, user historical click behavior information records, user historical conversion behavior information records, and creates a user-product interaction feature matrix, which can accurately express the complex feature correlation between users and products. Through the constructed multi-task product prediction model, it can be based on the click-through rate learning task model, the click-through conversion rate learning task model and the delayed conversion rate learning task model as auxiliary task models for product conversion rate prediction, and can predict the conversion rate in the entire sample space, solving the data sparsity problem. By using the delayed conversion rate learning task model as the auxiliary task model of the conversion rate prediction model, it fully considers the delay time between the user clicking on the product and the user placing an order for the product, thereby improving the accuracy of the product conversion rate prediction results. Through the prediction results of the product conversion rate, the product is recommended to the corresponding users, thereby increasing the platform's transaction volume and saving users' browsing time. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0015] In order to better illustrate the technical solution of the present invention and make its advantages more concise and clear, the problem to be solved by the present invention will be specifically explained below, and then the specific implementation methods of the present invention will be further described in detail with reference to the accompanying drawings.

[0016] See also Figure 1 The present invention proposes a method for predicting commodity conversion rate, the method comprising:

[0017] S1: Obtaining original data through the e-commerce platform; the original data includes: user attribute information, product attribute information, user historical click behavior information, and user historical conversion behavior information.

[0018] The user attribute information includes: user's gender, user's age, user's consumption level, and whether the user is employed.

[0019] The commodity attribute information includes: the category to which the commodity belongs, the store to which the commodity belongs, and the brand to which the commodity belongs.

[0020] The user's historical click behavior information includes: the user's click behavior on the product, the user's click behavior on the category to which the product belongs, the user's click behavior on the store to which the product belongs, the user's click behavior on the brand to which the product belongs, and the time that has passed since the user clicked on the product.

[0021] User historical conversion behavior information includes: user's conversion behavior on products; and the time from user clicking on a product to conversion.

[0022] For the convenience of description and understanding, the present invention makes the following definitions:

[0023] Among them, y∈{0,1} indicates whether the user clicks on the product; z∈{0,1} indicates whether the user converts the product; e∈[0,∞] indicates the time after the user clicks on the product; d∈[0,∞] indicates the time from the user clicking on the product to the conversion; the user's conversion behavior for the product means that the user places an order after clicking on the product.

[0024] S2: Perform feature processing on the user attribute information and the product attribute information to obtain a user feature matrix and a product feature matrix, and perform feature concatenation on the user feature matrix and the product feature matrix to obtain a user-product interaction feature matrix, wherein the user-product interaction feature matrix includes:

[0025] UIFeatur={User asic ,Item basic}

[0026] UIFeatur represents the user-product interaction feature matrix, User asic Represents user features, Item basic Indicates product characteristics.

[0027] Preferably, performing feature processing on the user attribute information and the product attribute information to obtain the user feature matrix and the product feature matrix includes:

[0028] S21: Calculate the similarity between users based on user attribute information, create a user attribute network space with users as nodes and users whose similarity exceeds a set threshold as adjacent nodes; calculate the similarity between products based on product attribute information, create a product attribute network space with products as nodes and products whose similarity exceeds a set threshold as adjacent nodes;

[0029] Preferably, calculating the similarity between users based on user attribute information includes:

[0030]

[0031] Among them, Inter(u,v) represents the similarity between user u and user v; Interact k (u, v) represents the number of times user u and user v have the same k-th behavior, and maxInteract represents the total number of behaviors of user u and user v. For example, user A browsed item a 2 times, user A browsed item b 5 times, and user A browsed item c 8 times; user B browsed item a 1 time, user B browsed item b 9 times, and user B browsed item c 2 times.

[0032] Interact a (A,B)=1,Interact b(A,B)=5,Interact b (A,B)=2

[0033] maxInteract(A,B)=27

[0034]

[0035] Among them, user u and user v represent two different users.

[0036] Preferably, calculating the similarity between users based on user attribute information includes:

[0037] The similarity between users is calculated based on the user's gender, age, consumption level, and whether the user is employed.

[0038] Preferably, the calculating the similarity between products based on product attribute information includes:

[0039]

[0040] Among them, Sim(Item1,Item2) represents the similarity between Item1 and Item2; Belong k (Item1, Item2) indicates whether the commodity Item1 and the commodity Item2 have the Mth category belonging relationship. If so, it is 1, otherwise it is 0. For example, if Item1 belongs to store d and Item2 belongs to store d, then Sim(Item1, Item2) = 1; if Itrm1 belongs to store d and Item2 belongs to store e, then (Item1, Item2) = 0, where Item1 and Item2 are two different commodities.

[0041] S22: Input the user attribute network space into the UIGraph model for feature processing and calculation to obtain a user feature matrix, and input the product attribute network space into the UIGraph model for feature processing and calculation to obtain a product feature matrix, wherein the UIGraph model is a graph neural network.

[0042] Preferably, the UIGraph model includes: an input layer, an aggregation layer, and an output layer.

[0043] Preferably, the steps of calculating the product feature matrix / user feature matrix include:

[0044] Step 1: Input the user attribute network space / product attribute network space into the aggregation layer through the input layer;

[0045] Step 2: The aggregation layer samples neighboring nodes of nodes in the user attribute network space / product attribute network space;

[0046] Preferably, sampling neighbor nodes of a node in the user attribute network space / product attribute network space to obtain a neighbor node sampling result of the node includes:

[0047]

[0048] in, Indicates the neighbor node sampling results of node x, Aggregate k represents the maximum aggregation function of the kth layer of the aggregation layer, Represents the neighbor nodes of node x, where x represents a node in the user attribute network space / product attribute network space.

[0049] Step 3: Aggregate the node's neighbor node sampling results and use them as the node feature representation of the node. Build the user feature matrix / product feature matrix based on the node feature representation of the node.

[0050] Preferably, aggregating the neighbor node sampling results of the node includes:

[0051]

[0052]

[0053] Where σ represents the non-threaded activation function, W k is the parameter of the kth layer of the aggregation layer, CONCAT represents the connection function, Represents the node feature representation of node x output by the k-1th layer of the aggregation layer, Indicates the neighbor node sampling result of node x, X user / X Item Represents the user feature matrix / item feature matrix, σ includes: ReLU activation function.

[0054] S3: Estimate the user's click-through rate, click-to-conversion rate, and delayed conversion rate based on the user's historical click behavior and conversion behavior information. Use the user's click-to-product rate as a click label, the user's click-to-product conversion rate as a click-to-conversion label, and the user's delayed conversion rate as a delayed conversion label.

[0055] Preferably, the user's click rate for a product includes:

[0056] p ctr ' = P(y = 1 | X = x)

[0057] p ctr' represents the click rate of users on the product, y = 1 represents the user clicking on the product, x represents the product and user samples, and X represents the user-product interaction characteristics of the product and the user.

[0058] Preferably, the user's click-through conversion rate for a product includes:

[0059] p ctcvr '=P(z=1,Y=1|X=x)

[0060] p ctcvr ' represents the user's click rate on the product, y = 1 represents the user clicking on the product, z = 1 represents the user converting the product, x represents the product and user samples, and X represents the user-product interaction characteristics of the product and user.

[0061] Preferably, the delayed conversion rate of users to products includes:

[0062] p dec '=p(d|z=1,y=1,e,X=x)

[0063] p dec ' represents the delayed conversion rate of users to products, y = 1 represents the user clicking on the product, z = 1 represents the user converting to the product, e∈[0,∞] represents the time after the user clicks on the product; d∈[0,∞] represents the time from the user clicking on the product to the conversion.

[0064] S4: Construct a multi-task product prediction model, and use the user-product interaction features in the user-product interaction feature matrix as input to train the multi-task product prediction model, wherein the multi-task product prediction model includes: a click-through rate learning task model, a click-through conversion rate learning task model, and a delayed conversion rate learning task model.

[0065] Preferably, the training of the multi-task commodity prediction model includes:

[0066] The user-item interaction features in the user-item interaction feature matrix are used as training samples, and the click labels are used as constraints for the click-through rate learning task model to train the click-through rate learning task model.

[0067] The user-item interaction features in the user-item interaction feature matrix are used as training samples, and the click-through conversion labels are used as constraints for the click-through conversion rate learning task model to train the click-through conversion rate learning task model.

[0068] The user-item interaction features in the user-item interaction feature matrix are used as training samples, and the delayed conversion labels are used as constraints of the delayed conversion rate learning task model to train the delayed conversion rate learning task model.

[0069] Preferably, the training process of the click-through rate learning task model includes:

[0070] The click-through rate learning task model is based on the DCN neural network and includes: a first input layer, a first hidden layer, and a first output layer;

[0071] The first input layer is used to input the user-product interaction features in the user-product interaction feature matrix into the first hidden layer;

[0072] The first hidden layer includes: a first cross layer, a first deep layer, and a first combination layer;

[0073] Perform feature correlation processing on the user-product interaction features through the first cross layer to obtain the first correlation feature;

[0074] Preferably, the first association feature includes:

[0075] O t 0=X

[0076]

[0077] Among them, O t m represents the first sub-association feature, X represents the user-item interaction feature in the user-item interaction feature matrix; M t m-1 represents the weight parameter of the first cross layer m-1 layer, b t m-1 Represents the bias parameter of the m-1th layer of the first cross layer.

[0078] The first deep layer performs feature correlation processing on the user-product interaction features to obtain the second correlation feature;

[0079] Preferably, the second association feature includes:

[0080] h t 0=X

[0081] h t m =f(W t m-1 h t m-1 +b t m-1 )

[0082] Among them, h t m represents the second sub-association feature, X represents the user-item interaction feature in the user-item interaction feature matrix; W t m-1represents the weight of the first deep layer m-1 layer, b t m-1 Represents the bias parameter of the m-1th layer of the first deep layer; f represents the ReLU activation function.

[0083] The first associated feature and the second associated feature are connected through the first combination layer, and the connected features are calculated through the activation function to obtain the user's predicted click rate for the product;

[0084] Preferably, the predicted click rate of the user for the product includes:

[0085] P ctr =σ([O t m ,h t m]W logitst )

[0086] Among them, P ctr represents the predicted click rate of users on products, W logitsv represents the weight parameter of the first combination layer, σ represents the sigmoid function, h t m Represents the second sub-association feature, O t m Indicates the first sub-association feature.

[0087] The parameters of the click-through rate learning task model are adjusted according to the user's predicted click-through rate and click label of the product to complete the model training.

[0088] Preferably, adjusting the parameters of the click-through rate learning task model according to the user's predicted click-through rate and click labels for the product includes: adjusting the parameters of the learning task model according to the mean square error of the user's predicted click-through rate and click labels for the product.

[0089] Preferably, the method for adjusting the parameters of the click-through rate learning task model according to the user's predicted click-through rate of the product and the click label calculation includes: adjusting the parameters of the click-through rate learning task model according to the cross entropy function.

[0090] Preferably, the training process of the click-through conversion rate learning task model includes:

[0091] The click-through conversion rate learning task model is based on the DCN neural network and includes: a second input layer, a second hidden layer, and a second output layer;

[0092] The second input layer is used to input the user-product interaction features in the user-product interaction feature matrix into the second hidden layer;

[0093] The second hidden layer includes: a second cross layer, a second deep layer, and a second combination layer;

[0094] The second cross layer performs feature association processing on the user-product interaction features to obtain the first sub-association feature;

[0095] Preferably, the first sub-association feature includes:

[0096] O v 0=X

[0097]

[0098] Among them, O v m represents the first sub-association feature, X represents the user-item interaction feature in the user-item interaction feature matrix; W v m-1 Indicates the weight parameter of the second cross layer m-1 layer, b v m-1 Represents the bias parameter of the m-1th layer of the second cross layer.

[0099] The second deep layer performs feature association processing on the user-product interaction features to obtain the second sub-association features;

[0100] Preferably, the second sub-association feature includes:

[0101] h v 0=X

[0102] h v m =f(M v m-1 h v m-1 +b v m-1 )

[0103] Among them, h v m represents the second correlation feature, X represents the user-item interaction feature in the user-item interaction feature matrix; M v m-1 represents the weight parameter of the m-1th layer of the second deep layer, b v m-1 represents the bias parameter of the m-1th layer of the second deep layer; f represents the ReLU activation function.

[0104] The first sub-correlation feature and the second sub-correlation feature are connected through the second combination layer, and the connected features are calculated through the activation function to obtain the predicted click-through conversion rate of the user for the product;

[0105] The predicted click-through conversion rate of the user for the product includes:

[0106] P ctcvr =σ([O v m ,h v m ]W logitsv )

[0107] Among them, P ctcvr represents the predicted click-through rate of users for products, W logitst represents the weight parameter of the second combination layer, σ represents the sigmoid function, O v m represents the first sub-correlation feature, h v m Represents the second associated feature.

[0108] The parameters of the click-through conversion rate learning task model are adjusted based on the user's predicted click-through conversion rate and click-through conversion label for the product to complete the model training.

[0109] Preferably, adjusting the parameters of the click-through conversion rate learning task model according to the user's predicted click-through conversion rate for the product and the click-through conversion label includes: adjusting the parameters of the learning task model according to the mean square error of the user's predicted click-through conversion rate for the product and the click-through conversion label.

[0110] Preferably, the training process of the delayed conversion rate learning task model includes:

[0111] The delayed conversion rate learning task model is based on a DNN neural network: including an input layer, a hidden layer and an output layer;

[0112] The input layer is used to input the user-product interaction features in the user-product interaction feature matrix into the hidden layer;

[0113] The hidden layer is used to extract features of user-product interaction features in multiple dimensions to obtain a multi-dimensional feature vector, and calculate the user's predicted delayed conversion rate for the product based on the multi-dimensional feature vector;

[0114] Preferably, the predicted delayed conversion rate of the user for the product includes:

[0115] X 0 =X

[0116] X m+1 =σ(W m+1 X m +b m+1 )

[0117] z=X m+1

[0118]

[0119] Among them, p dec represents the predicted delayed conversion rate of users to products, X represents the user-product interaction feature vector in the original user-product interaction feature matrix, z is equal to the output of the last hidden layer, X m represents the output of the hidden layer m, σ represents the Relu activation function, W m+1 represents the weight parameter of the hidden layer m+1, b m+1 Represents the bias parameter of the hidden layer (m+1).

[0120] The parameters of the delayed conversion rate learning task model are adjusted according to the user's predicted delayed conversion rate and delayed conversion label for the product to complete the model training. The delayed conversion rate learning model can accurately calculate the number of days the user's delayed conversion of the product, thereby improving the accuracy of the conversion rate. At the same time, the click-through rate learning task model, click-to-conversion rate learning task model, and delayed conversion rate learning task model designed by the present invention all use the entire sample space as training samples, without data sparsity problems, and the delayed conversion rate calculated by the present invention is very accurate.

[0121] Preferably, adjusting the parameters of the delayed conversion rate learning task model according to the user's predicted delayed conversion rate and delayed conversion label of the product includes: adjusting the parameters of the learning task model according to the mean square error of the user's predicted delayed conversion rate and delayed conversion label of the product.

[0122] S5: Acquire attribute information of the target user and target product, perform feature extraction on the attribute information of the target user and target product to obtain feature vectors of the target user and target product, and concatenate the feature vectors of the target user and target product to obtain target user-product interaction features.

[0123] S6: Input the target user-item interaction features into the trained multi-task item prediction model to predict the target user's predicted click-through rate for the target item, the target user's predicted click-through conversion rate for the target item, and the target user's predicted delayed conversion rate for the target item.

[0124] S7: Calculate the final conversion rate of the target user for the target product using the EM algorithm based on the predicted conversion rate of the target user for the target product, the predicted click-through conversion rate of the target user for the target product, and the predicted delayed conversion rate of the target user for the target product.

[0125] Preferably, the step of calculating the final conversion rate of the target user to the target product includes:

[0126] S71: The predicted conversion rate of the target user for the target product, the predicted click-through conversion rate of the target user for the target product, and the predicted delayed conversion rate of the target user for the target product are used as initial parameters of the EM algorithm, and the expected value is calculated using the EM algorithm.

[0127] Preferably, the expected value includes:

[0128]

[0129] Among them, w i represents the expected value of the target product by the i-th target user, I1 represents the target user clicking on the target product, I2 represents the target user not clicking on the target product, and p ctcvri represents the predicted click-through conversion rate of the target product by the i-th target user, p ctri represents the predicted conversion rate of the target product by the i-th target user, p deci (t j ) represents the predicted delayed conversion rate of the target product by the i-th target user on the j-th day, T+1 represents the maximum number of days for the target user to delay conversion to the target product, t represents time, t j Indicates the time is the jth day.

[0130] S72: Construct the log-likelihood loss function of the EM algorithm based on the expected value.

[0131] Preferably, the log-likelihood loss function of the EM algorithm includes:

[0132]

[0133] Among them, w i represents the expected value of the target product by the i-th target user, p ctcvri represents the predicted click-through conversion rate of the target product by the i-th target user, p ctri represents the predicted conversion rate of the target product by the i-th target user, p deci (t j ) represents the predicted delayed conversion rate of the target product by the i-th target user on the j-th day, y i ∈{0,1} indicates whether the i-th target user clicks on the target product, z i ∈{0,1} indicates whether the i-th target user converts to the target product, e i ∈[0,∞] represents the time after the i-th target user clicks on the target product, d i ∈[0,∞] represents the time from the i-th target user clicking on the target product to the conversion, t represents time, p deci (t j) represents the predicted delayed conversion rate of the target product by the i-th target user on the j-th day, and T+1 represents the maximum number of days for the target user to delay conversion to the target product.

[0134] S73: Maximize the convergence of the parameters of the EM algorithm. When the log-likelihood loss function is maximized, the expected value is used as the final conversion rate of the target user to the target product.

[0135] Preferably, maximizing the convergence of the parameters of the EM algorithm includes:

[0136] θ′=argmax(θ)

[0137] Among them, θ′ represents the maximum parameter of EM algorithm, θ represents the initial parameter of EM algorithm, θ=(p deci (t j ), p ctcvri , p ctri ), p ctcvri represents the predicted click-through conversion rate of the target product by the i-th target user, p ctri represents the predicted conversion rate of the target product by the i-th target user, p deci (t j ) represents the predicted delayed conversion rate of the target product by the i-th target user on the j-th day, argmax() represents the maximization function, and the EM algorithm can effectively remove the characteristic noise of the target user and target product and the noise in the multi-task prediction model, thereby improving the accuracy of the final conversion rate.

[0138] By predicting the final conversion rate of target products by target users, the products can be recommended to the corresponding users, thereby increasing the transaction volume of the platform and saving users' browsing time.

[0139] Other embodiments of the present disclosure will readily occur to those skilled in the art after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow from the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the claims.

[0140] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for predicting commodity conversion rate, characterized in that: The method comprises: S1: Obtaining raw data through the e-commerce platform; the raw data includes: user attribute information, product attribute information, user historical click behavior information, and user historical conversion behavior information; S2: Perform feature processing on the user attribute information and the product attribute information to obtain a user feature matrix and a product feature matrix, and then perform feature concatenation on the user feature matrix and the product feature matrix to obtain a user-product interaction feature matrix; S3: Estimate the user's click-through rate, click-to-conversion rate, and delayed conversion rate based on the user's historical click behavior and conversion behavior information. Use the user's click-to-product rate as a click label, the user's click-to-product conversion rate as a click-to-conversion label, and the user's delayed conversion rate as a delayed conversion label. S4: Constructing a multi-task product prediction model, and using the user-product interaction features in the user-product interaction feature matrix as input to train the multi-task product prediction model, wherein the multi-task product prediction model includes: a click-through rate learning task model, a click-through conversion rate learning task model, and a delayed conversion rate learning task model; S5: Acquire attribute information of the target user and target product, perform feature extraction on the attribute information of the target user and target product to obtain feature vectors of the target user and target product, and concatenate the feature vectors of the target user and target product to obtain target user-product interaction features; S6: Input the target user-item interaction features into the trained multi-task item prediction model to predict the target user's predicted click-through rate for the target item, the target user's predicted click-through conversion rate for the target item, and the target user's predicted delayed conversion rate for the target item. S7: Calculate the final conversion rate of the target user for the target product using the EM algorithm based on the predicted conversion rate of the target user for the target product, the predicted click-through conversion rate of the target user for the target product, and the predicted delayed conversion rate of the target user for the target product.

2. A commodity conversion rate prediction method according to claim 1, characterized in that: The feature processing of the user attribute information and the product attribute information to obtain the user feature matrix and the product feature matrix includes: S21: Calculate the similarity between users based on user attribute information, create a user attribute network space with users as nodes and users whose similarity exceeds a set threshold as adjacent nodes; calculate the similarity between products based on product attribute information, create a product attribute network space with products as nodes and products whose similarity exceeds a set threshold as adjacent nodes; S22: Input the user attribute network space into the UIGraph model for feature processing and calculation to obtain a user feature matrix, and input the product attribute network space into the UIGraph model for feature processing and calculation to obtain a product feature matrix, wherein the UIGraph model is a graph neural network.

3. A commodity conversion rate prediction method according to claim 1, characterized in that: The training of the multi-task commodity prediction model includes: The user-item interaction features in the user-item interaction feature matrix are used as training samples, and the click labels are used as constraints for the click-through rate learning task model to train the click-through rate learning task model. The user-item interaction features in the user-item interaction feature matrix are used as training samples, and the click-through conversion labels are used as constraints for the click-through conversion rate learning task model to train the click-through conversion rate learning task model. The user-item interaction features in the user-item interaction feature matrix are used as training samples, and the delayed conversion labels are used as constraints of the delayed conversion rate learning task model to train the delayed conversion rate learning task model.

4. A commodity conversion rate prediction method according to claim 3, characterized in that: The training process of the click-through rate learning task model includes: The click-through rate learning task model is based on the DCN neural network and includes: a first input layer, a first hidden layer, and a first output layer; The first input layer is used to input the user-product interaction features in the user-product interaction feature matrix into the first hidden layer; The first hidden layer includes: a first cross layer, a first deep layer, and a first combination layer; Perform feature association processing on the user-product interaction feature vector through the first cross layer to obtain the first association feature; The first deep layer performs feature correlation processing on the user-product interaction features to obtain the second correlation feature; The first associated feature and the second associated feature are connected through the first combination layer, and the connected feature vector is calculated through the activation function to obtain the user's predicted click rate for the product; The parameters of the click-through rate learning task model are adjusted according to the user's predicted click-through rate and click label of the product to complete the model training.

5. A commodity conversion rate prediction method according to claim 3, characterized in that: The training process of the click-through conversion rate learning task model includes: The click-through conversion rate learning task model is based on the DCN neural network and includes: a second input layer, a second hidden layer, and a second output layer; The second input layer is used to input the user-product interaction features in the user-product interaction feature matrix into the second hidden layer; The second hidden layer includes: a second cross layer, a second deep layer, and a second combination layer; Perform feature association processing on the user-item interaction feature vector through the second cross layer to obtain the first sub-association feature; The second deep layer performs feature association processing on the user-item interaction feature vector to obtain the second sub-association feature; The first sub-correlation feature and the second sub-correlation feature are connected through the second combination layer, and the connected feature vector is calculated through the activation function to obtain the user's predicted click-through conversion rate for the product; The parameters of the click-through conversion rate learning task model are adjusted based on the user's predicted click-through conversion rate and click-through conversion label for the product to complete the model training.

6. A commodity conversion rate prediction method according to claim 3, characterized in that: The training process of the delayed conversion rate learning task model includes: The delayed conversion rate learning task model is based on a DNN neural network: including an input layer, a hidden layer and an output layer; The input layer is used to input the user-product interaction feature vector in the user-product interaction feature matrix into the hidden layer; The hidden layer is used to extract features of user-product interaction features in multiple dimensions to obtain a multi-dimensional feature vector, and calculate the user's predicted delayed conversion rate for the product based on the multi-dimensional feature vector; The parameters of the delayed conversion rate learning task model are adjusted based on the user's predicted delayed conversion rate and delayed conversion label for the product to complete the model training.

7. A commodity conversion rate prediction method according to claim 1, characterized in that: The steps for calculating the final conversion rate of the target user to the target product include: S71: using the predicted conversion rate of the target user for the target product, the predicted click-through conversion rate of the target user for the target product, and the predicted delayed conversion rate of the target user for the target product as initial parameters of the EM algorithm, and calculating the expected value through the EM algorithm; S72: Construct the log-likelihood loss function of the EM algorithm based on the expected value; S73: Maximize the convergence of the parameters of the EM algorithm. When the log-likelihood loss function is maximized, The expected value is taken as the final conversion rate of target users to target products.

8. A commodity conversion rate prediction method according to claim 7, characterized in that: The expected values ​​include: Among them, w i represents the expected value of the target product by the i-th target user, I1 represents the target user clicking on the target product, and I2 represents the target user not clicking on the target product. represents the predicted click-through conversion rate of the target product by the i-th target user, represents the predicted conversion rate of the target product by the i-th target user, represents the predicted delayed conversion rate of the target product by the i-th target user on the j-th day, T+1 represents the maximum number of days for the target user to delay conversion to the target product, t represents the time, t j Indicates the time is the jth day.

9. A commodity conversion rate prediction method according to claim 7, characterized in that: The log-likelihood loss function of the EM algorithm includes: Among them, w i represents the expected value of the target product by the i-th target user, represents the predicted click-through conversion rate of the target product by the i-th target user, represents the predicted conversion rate of the target product by the i-th target user, represents the predicted delayed conversion rate of the target product by the i-th target user on the j-th day, y i ∈{0,1} indicates whether the i-th target user clicks on the target product, z i ∈{0,1} indicates whether the i-th target user converts to the target product, e i ∈[0,∞] represents the time after the i-th target user clicks on the target product, d i ∈[0,∞] represents the time from the i-th target user clicking on the target product to the conversion, t represents the time, It represents the predicted delayed conversion rate of the target product by the i-th target user on the j-th day, and T+1 represents the maximum number of days for the target user to delay conversion to the target product.

10. A commodity conversion rate prediction method according to claim 7, characterized in that: The maximizing convergence of the parameters of the EM algorithm includes: θ′=argmax(θ) Among them, θ′ represents the maximum parameter of the EM algorithm, θ represents the initial parameter of the EM algorithm, represents the predicted click-through conversion rate of the target product by the i-th target user, represents the predicted conversion rate of the target product by the i-th target user, It represents the predicted delayed conversion rate of the target product by the i-th target user on the j-th day, and argmax() represents the maximization function.

Images