A method, apparatus, electronic device, and storage medium for predicting recommendation probability

Abstract

The present application provides a method, apparatus, electronic device and storage medium for predicting recommendation probability. The method includes: obtaining the information content feature, user general feature, environment general feature, user bias feature and environment bias feature of the information to be recommended; inputting the information content feature, user general feature and environment general feature of the information to be recommended into a recommendation probability prediction model to obtain the recommendation probability of the information to be recommended; wherein, the recommendation probability prediction model is obtained by training a preset model with the information content feature, user general feature, environment general feature, user bias feature and environment bias feature of sample information as inputs, and the recommendation probability prediction model includes a plurality of task networks, a plurality of gate switch networks, a plurality of fusion modules and a plurality of task layers. The task networks are used to extract feature vectors from different dimensions, and the gate switch networks are used to control the weights of each task network. This method can solve the conflicts between tasks and improve the model effect.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of data processing, and in particular, to a method, device, electronic device, and storage medium for estimating recommendation probability. Background Art

[0002] In machine learning, estimation is a common application scenario. By estimating the degree of a user's interest in a product through a model, and then recommending the products that the user is interested in in a targeted manner, not only can the user experience be improved, but also the merchant's revenue can be increased. For example, before an online food delivery platform pushes an advertisement of a merchant to a user, it will first estimate the click-through rate, order placement rate, etc. of each user for the advertisement, and then select the users with a higher click-through rate or order placement rate as target users for advertisement placement. Or, for a certain user, the online food delivery platform will estimate the products with a higher click-through rate or order placement rate among many products, and then recommend these products to the user.

[0003] In the estimation scenario, multiple parameters (such as click-through rate, conversion rate, click-and-conversion rate, total transaction amount, etc.) are usually estimated, that is, multi-task estimation. Common multi-task estimation methods in related technologies include: hard parameter sharing method and soft parameter sharing method. Among them, the hard parameter sharing method is a method of sharing the bottom hidden layer network among different tasks to share parameters, while the soft parameter sharing method is a method in which parameters are not shared among multiple tasks, but the parameters of different tasks are restricted by the L2 (Euclidean norm) norm.

[0004] However, the hard parameter sharing method restricts the parameters learned by each task to be the same, and cannot well express the differences between tasks. At the same time, for dissimilar tasks, due to the large differences in the underlying representations, there may be conflicts when sharing parameters, resulting in poor model prediction effects. Although the soft parameter sharing method does not share the parameters between tasks, the parameters learned by each task are restricted within the L2 norm, which increases the network scale and the estimation cost. Summary of the Invention

[0005] The embodiments of the present application provide a method, device, electronic device, and storage medium for estimating recommendation probability, aiming to solve the problem of conflicts between tasks when multiple tasks share parameters.

[0006] In a first aspect of the embodiments of the present application, a method for estimating recommendation probability is provided, and the method includes:

[0007] Obtain the information content feature, user general feature, environment general feature, user bias feature, and environment bias feature of the information to be recommended;

[0008] Input the information content features, user general features, and environment general features of the information to be recommended into a recommendation probability prediction model to obtain the recommendation probabilities in multiple dimensions of the information to be recommended;

[0009] Among them, the recommendation probability prediction model is trained by inputting the information content features, user general features, environment general features, user bias features, and environment bias features of sample information into a preset model. The recommendation probability prediction model includes multiple task networks, multiple gate switch networks, multiple fusion modules, and multiple task layers. Each task network is used to obtain a feature vector in one dimension of the sample information. A gate switch network is used to output the weights of the feature vectors output by the multiple task networks in one dimension. A fusion module is used to fuse the feature vectors output by the multiple task networks according to the weights of the feature vectors output by the multiple task networks in one dimension. A task layer is used to obtain the corresponding recommendation probability according to the fused feature vectors in one dimension.

[0010] Optionally, the multiple task networks share the respective features input into the recommendation probability prediction model. Each task network includes a bias sub-network, a task sub-network, and a splicing layer. The bias sub-network takes the user bias features and environment bias features of the sample information as input. The task sub-network takes the user bias features and environment bias features of the sample information as input. The splicing layer is used to splice the features output by the bias sub-network and the task sub-network to obtain the features output by a task network.

[0011] Optionally, the recommendation probabilities in multiple dimensions include any combination of the following: click-through rate, conversion rate, click-and-conversion rate, total transaction amount; and / or

[0012] The environment bias features include at least one of the following: the geographical location where the user is located, the time when the user operation occurs, the client type, the type of the user terminal; and / or

[0013] The user bias features include at least one of the following: click-through rate, order placement rate, order placement amount, order placement quantity.

[0014] Optionally, the number of the information to be recommended is multiple; the method further includes:

[0015] Determine the recommendation probabilities in multiple dimensions of the multiple pieces of information to be recommended for the target user;

[0016] According to the recommendation probabilities in multiple dimensions of the multiple pieces of information to be recommended for the target user, determine the comprehensive recommendation probability of the multiple pieces of information to be recommended for the target user;

[0017] Send the multiple pieces of information to be recommended to the user terminal used by the target user in descending order of the comprehensive recommendation probability.

[0018] Optionally, the number of the pieces of information to be recommended is multiple; the method further includes:

[0019] Determine the recommendation probabilities of the multiple pieces of information to be recommended to the target user in multiple dimensions;

[0020] Send the multiple pieces of information to be recommended to the user terminal used by the target user in descending order for each dimension.

[0021] A second aspect of the embodiments of the present application provides a recommendation probability prediction device, and the device includes:

[0022] An obtaining module, configured to obtain the information content feature, user general feature, environment general feature, user bias feature, and environment bias feature of the information to be recommended;

[0023] An input module, configured to input the information content feature, user general feature, and environment general feature of the information to be recommended into a recommendation probability prediction model, and obtain the recommendation probabilities of the information to be recommended in multiple dimensions;

[0024] Wherein, the recommendation probability prediction model is trained by using the information content feature, user general feature, environment general feature, user bias feature, and environment bias feature of sample information as inputs for a preset model. The recommendation probability prediction model includes multiple task networks, multiple gate switch networks, multiple fusion modules, and multiple task layers. Each task network is configured to obtain a feature vector of a dimension of the sample information. A gate switch network is configured to output the weights of the feature vectors output by the multiple task networks in a dimension. A fusion module is configured to fuse the feature vectors output by the multiple task networks according to the weights of the feature vectors output by the multiple task networks in a dimension. A task layer is configured to obtain the corresponding recommendation probability according to the fused feature vector in a dimension.

[0025] Optionally, the multiple task networks share each feature input to the recommendation probability prediction model. Each task network includes a bias sub-network, a task sub-network, and a splicing layer. The bias sub-network uses the user bias feature and environment bias feature of the sample information as inputs. The task sub-network uses the user bias feature and environment bias feature of the sample information as inputs. The splicing layer is configured to splice the features output by the bias sub-network and the task sub-network to obtain the feature output by a task network.

[0026] Optionally, the recommendation probabilities in multiple dimensions include any combination of the following: click-through rate, conversion rate, click-through and conversion rate, total transaction amount; and / or

[0027] The environmental bias feature includes at least one of the following: the geographical location where the user is located, the time when the user operation occurs, the client type, the type of the user terminal; and / or

[0028] The user bias feature includes at least one of the following: click-through rate, order placement rate, order placement amount, order placement quantity.

[0029] Optionally, the number of pieces of information to be recommended is multiple; the apparatus further includes:

[0030] A first determination module, configured to determine recommendation probabilities of the multiple pieces of information to be recommended for a target user in multiple dimensions;

[0031] A second determination module, configured to determine a comprehensive recommendation probability of the multiple pieces of information to be recommended for the target user according to the recommendation probabilities of the multiple pieces of information to be recommended for the target user in multiple dimensions;

[0032] A first sending module, configured to send the multiple pieces of information to be recommended to a user terminal used by the target user in descending order of the comprehensive recommendation probability.

[0033] Optionally, the number of pieces of information to be recommended is multiple; the apparatus further includes:

[0034] A third determination module, configured to determine recommendation probabilities of the multiple pieces of information to be recommended for a target user in multiple dimensions;

[0035] A second sending module, configured to send the multiple pieces of information to be recommended to a user terminal used by the target user in descending order for each dimension respectively.

[0036] A third aspect of an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps in the recommendation probability estimation method described in the first aspect of the present application are implemented.

[0037] A fourth aspect of an embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes, the steps in the recommendation probability estimation method described in the first aspect of the present application are implemented.

[0038] By using the recommended probability prediction method provided in the embodiments of the present application, first obtain the information content features, user general features, environmental general features, user bias features, and environmental bias features of the information to be recommended; then input the information content features, user general features, and environmental general features of the information to be recommended into the recommended probability prediction model to obtain the recommended probabilities in multiple dimensions of the information to be recommended, so as to recommend the information to be recommended to the user according to the recommended probabilities. Among them, the recommended probability prediction model is trained by inputting the information content features, user general features, environmental general features, user bias features, and environmental bias features of the sample information into a preset model. The recommended probability prediction model includes multiple task networks, multiple gate switch networks, multiple fusion modules, and multiple task layers. Each task network is used to obtain a feature vector in one dimension of the sample information. A gate switch network is used to output the weights of the feature vectors output by each of the multiple task networks in one dimension. A fusion module is used to fuse the feature vectors output by each of the multiple task networks according to the weights of the feature vectors output by each of the multiple task networks in one dimension. A task layer is used to obtain the corresponding recommended probability according to the fused feature vector in one dimension.

[0039] The recommended probability prediction method of the present application has the following technical effects:

[0040] First, the network structures of two multi-task prediction models in related technologies are improved to obtain a recommended probability prediction model. In this recommended probability model, multiple task networks are set, and the parameters of the multiple task networks are not completely shared. Each task network is used to extract partial information from the input information, and the information of each task network is selectively utilized through a gate switch network, effectively solving the conflict problem between multiple tasks.

[0041] Second, by controlling the weights of the feature vectors output by each task network through the gate switch network, the difference of the input features of each task layer can be ensured, thereby improving the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0043] Figure 1 is a schematic diagram of the hard parameter sharing method in related technologies;

[0044] Figure 2 is a schematic diagram of the soft parameter sharing method in related technologies;

[0045] Figure 3 It is a schematic diagram of an implementation environment shown in an embodiment of the present application;

[0046] Figure 4 It is a flowchart of a recommended probability estimation method shown in an embodiment of the present application;

[0047] Figure 5 It is a schematic diagram of a recommended probability prediction model shown in an embodiment of the present application;

[0048] Figure 6 It is a schematic structural diagram of a task network shown in an embodiment of the present application;

[0049] Figure 7 It is a structural block diagram of a recommended probability estimation device provided in an embodiment of the present application;

[0050] Figure 8 It is a schematic diagram of an electronic device shown in an embodiment of the present application. Detailed implementation manners

[0051] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments in the present application without creative efforts shall fall within the protection scope of the present application.

[0052] Before elaborating on the recommended probability estimation method of the present application in detail, the multi-task estimation method in the related art will be briefly introduced below.

[0053] Figure 1 It is a schematic diagram of the hard parameter sharing method in the related art. Figure 2 It is a schematic diagram of the soft parameter sharing method in the related art.

[0054] In Figure 1 Among them, multiple tasks include click-through rate, conversion rate, and total transaction amount. The hard parameter sharing method adopts the method of sharing the underlying hidden layer network by multiple tasks, that is, the three tasks of click-through rate, conversion rate, and total transaction amount share the parameters of the two fully connected layers at the bottom. The problem with this method is that the parameters learned by each task are the same, and it is impossible to express the differences between tasks well. For dissimilar tasks, the underlying representations are quite different, and when sharing parameters, they are likely to conflict with each other or have too much noise, resulting in poor performance of the multi-task learning model.

[0055] In Figure 2In the case of soft parameter sharing, multiple tasks do not share parameters, but the L2 norm is added to the parameters of different tasks. That is, there are certain restrictions on the parameters of each task, and the difference is not large. Each task is not completely independent. For an introduction to the L2 norm, please refer to the introduction in the related art, which will not be elaborated in this application. The problem with this method is that it increases the network size and the cost of multi-task estimation.

[0056] In order to solve the problems existing in the related technologies, this application proposes a new network structure for the multi-task estimation model, which can effectively solve the problem of inter-task conflicts when multiple tasks share parameters.

[0057] Figure 3 This is a schematic diagram of an implementation environment shown in an embodiment of the present application. Figure 3 In the example, the information recommendation platform is connected to multiple user terminals (including user terminal 1-user terminal N) in communication to recommend information to the user terminals. The information recommendation platform is provided with a new multi-task estimation model newly proposed in this application, through which the recommendation probability for each user terminal is estimated to achieve targeted information recommendation. For example, when the information recommendation platform is a food delivery platform, and the information to be recommended is an advertisement of a merchant stationed on the food delivery platform, for the advertisement B to be recommended of merchant A, the food delivery platform obtains the recommendation probability of each user terminal for the advertisement B to be recommended through the multi-task estimation model, and then sends the advertisement B to be recommended to the user terminal with a higher recommendation probability. Or for a user terminal that has logged into the food delivery platform, the food delivery platform estimates the advertisement with a higher recommendation probability among multiple advertisements, and then pushes the advertisement with a higher recommendation probability to the user terminal.

[0058] The recommendation probability estimation method proposed in this application is applied to Figure 3 Multi-task estimation model in . Figure 4 This is a flow chart of a method for estimating the recommendation probability according to an embodiment of the present application. Figure 4 The recommendation probability estimation method of this application may include the following steps:

[0059] Step S11: obtaining information content features, user general features, environment general features, user bias features, and environment bias features of the information to be recommended.

[0060] In this embodiment, the information to be recommended refers to information to be recommended to the user. Taking the information recommendation platform as a food delivery platform as an example, the information to be recommended can be product information, store advertisements, platform activities, etc., and this embodiment does not impose specific restrictions on this.

[0061] The information content features refer to the features obtained by extracting features of the information to be recommended from multiple dimensions. Taking the information of a product as an example of the information to be recommended, the information content features can be the price of the product, the taste of the product (sweet, sour, bitter, spicy), the quality of the product, the affiliated store, the purchase method, whether there is a discount, etc. How to obtain the information content features from multiple dimensions can be set according to actual needs, and this embodiment does not make specific restrictions on this.

[0062] Both the user general features and the user bias features refer to user features, and their essential contents are the same. User features include at least one of the following: click-through rate, order placement rate, order placement amount, order placement quantity. Specifically, user features can include click-through rate, order placement rate, order placement amount, order placement quantity, the gender, age, activity level, consumption range, consumption frequency, taste label of the user, etc. User features can be obtained by analyzing the historical behavior data of the user. The historical behavior data of the user refers to all historical behavior records generated by the user during the process of using an information recommendation platform (such as a food delivery platform, etc.) or using other platforms associated with the information recommendation platform. The categories of user features and the methods for obtaining user features can be set according to actual needs, and this embodiment does not make specific restrictions on this.

[0063] Both the environment general features and the environment bias features refer to environment features, and their essential contents are the same. Environment features include at least one of the following: the geographical location where the user is located, the time when the user operation occurs, the client type, the type of the user terminal. Specifically, environment features can include basic environment features such as the geographical location when the user operates, the time point of the user operation, the client type, the type of the user terminal, and deep environment features obtained by further analyzing the basic environment features. For example, the deep features can be the city where the user operates, whether it is a weekend when the user operates, the holiday closest to the user operation time, etc. The categories of the basic environment features and the deep environment features can be set according to actual needs, and this embodiment does not make specific restrictions on this.

[0064] Step S12: Input the information content features, user general features, and environment general features of the information to be recommended into the recommendation probability prediction model to obtain the recommendation probabilities of multiple dimensions of the information to be recommended.

[0065] In this embodiment, the recommendation probabilities of multiple dimensions include any combination of the following: click-through rate, conversion rate, click-and-conversion rate, total transaction amount. The multi-task prediction model can be the recommendation probability prediction model. The recommendation probability prediction model is used to predict the recommendation probability of the information to be recommended. Among them, the recommendation probability can have multiple dimensions. For example, the click-through rate, conversion rate, click-and-conversion rate, total transaction amount, etc. are respectively the recommendation probabilities of four dimensions. The recommendation probability prediction model can estimate a single dimension or a combination of multiple dimensions.

[0066] In this embodiment, the information to be recommended can be recommended to the user based on the recommendation probability in one dimension or multiple dimensions. Taking the recommendation of the information to be recommended to the user based on the recommendation probability in one dimension as an example, the information to be recommended can be recommended to the user when the predicted click-through rate is higher than the preset click-through rate threshold, or the information to be recommended can be recommended to the user when the predicted conversion rate is higher than the preset conversion rate threshold. This embodiment does not make specific restrictions on this. Taking the recommendation of the information to be recommended to the user based on the recommendation probability in multiple dimensions as an example, the information to be recommended can be recommended to the user when the predicted click-through rate is higher than the preset click-through rate threshold and the conversion rate is higher than the preset conversion rate threshold, or the information to be recommended can be recommended to the user when the predicted click-through rate is higher than the preset click-through rate threshold and the click-and-conversion rate is higher than the preset click-and-conversion rate threshold. This embodiment does not make specific restrictions on this.

[0067] In this embodiment, predicting the recommendation probability in one dimension corresponds to one task, and predicting the recommendation probability in multiple dimensions corresponds to multiple tasks. The number of tasks can be set manually in advance according to actual needs. Multiple tasks share information content features, user features, and environmental features. Therefore, by inputting the information content features, user general features, and environmental general features of the information to be recommended into the recommendation probability prediction model, the recommendation probabilities of multiple dimensions of the information to be recommended can be obtained. For example, the click-through rate, conversion rate, click-and-conversion rate, etc. can be obtained.

[0068] In this embodiment, the recommendation probability prediction model is obtained by training a preset model with the information content features, user general features, environmental general features, user bias features, and environmental bias features of sample information as inputs. The recommendation probability prediction model includes multiple task networks, multiple gate switch networks, multiple fusion modules, and multiple task layers. Each task network is used to obtain a feature vector of one dimension of the sample information. A gate switch network is used to output the weights of the feature vectors output by the multiple task networks in one dimension. A fusion module is used to fuse the feature vectors output by the multiple task networks according to the weights of the feature vectors output by the multiple task networks in one dimension. A task layer is used to obtain the corresponding recommendation probability based on the fused feature vector in one dimension.

[0069] Figure 5 It is a schematic diagram of a recommendation probability prediction model shown in an embodiment of the present application. In Figure 5 it, there are 3 task networks (Task Network 1 - Task Network 3), 3 gate switch networks (Gate Switch Network 1 - Gate Switch Network 3), 3 fusion modules (Fusion Module 1 - Fusion Module 3), and 3 task layers (Task Layer 1 - Task Layer 3) provided in the recommendation probability prediction model. Figure 5Only the case where the number of task networks is 3 and the number of task layers is 3 is shown. Of course, the number of task networks and the number of task layers can be set according to actual needs, and this embodiment does not make specific restrictions on this.

[0070] The task network is used to convert the features input by the information input layer into feature vectors. One task network is used to extract feature vectors from one dimension. For example, task network 1 is used to extract feature vectors related to click behavior, task network 2 is used to extract feature vectors related to order placement behavior, task network 3 is used to extract feature vectors related to order placement behavior, etc.

[0071] A gate switch network is used to output the weights of the feature vectors output by each task network in one dimension. Taking Figure 5 as an example, gate switch network 1 is used to control the weights of task networks 1 - 3 to input feature vectors to task layer 1 respectively, gate switch network 2 is used to control the weights of task networks 1 - 3 to input feature vectors to task layer 2 respectively, and gate switch network 3 is used to control the weights of task networks 1 - 3 to input feature vectors to task layer 3 respectively.

[0072] A fusion module is used to fuse the feature vectors output by each task network according to the weights of the feature vectors output by each task network in one dimension. Taking Figure 5 as an example, fusion module 1 is used to fuse the values output by gate switch network 1 and the values output by task networks 1 - 3 to obtain the input features of task layer 1; fusion module 2 is used to fuse the values output by gate switch network 2 and the values output by task networks 1 - 3 to obtain the input features of task layer 2; fusion module 3 is used to fuse the values output by gate switch network 1 and the values output by task networks 1 - task networks 3 to obtain the input features of task layer 3. The fusion by the fusion module means multiplying the weights corresponding to each feature vector to obtain weighted feature vectors. The fusion module can fuse the output of the task network and the output of the gate switch network through softmax. Of course, other methods can also be used for fusion, and this embodiment does not make specific restrictions on the fusion method.

[0073] Specifically, assume that the feature vector output by task network 1 is T1, the feature vector output by task network 2 is T2, and the feature vector output by task network 3 is T3, that is, the feature vectors output by task networks 1 - 3 are {T1, T2, T3}. The weight of the gating network 1 for the output of task network 1 is w1, the weight of the gating network 2 for the output of task network 2 is w2, and the weight of the gating network 3 for the output of task network 3 is w3, that is, the weights of the gating network 1 for the outputs of task networks 1 - 3 are {w1, w2, w3}. Similarly, the weights of the gating network 2 for the outputs of task networks 1 - 3 are {w1', w2', w3'}, and the weights of the gating network 3 for the outputs of task networks 1 - 3 are {w1", w2", w3"}. The weighted feature vectors obtained by fusing {T1, T2, T3} and {w1, w2, w3} by fusion module 1 are {w1T1, w2T2, w3T3}, which are the input features of task layer 1; the weighted feature vectors obtained by fusing {T1, T2, T3} and {w1', w2', w3'} by fusion module 2 are {w1'T1, w2'T2, w3'T3}, which are the input features of task layer 2; the weighted feature vectors obtained by fusing {T1, T2, T3} and {w1", w2", w3"} by fusion module 3 are {w1"T1, w2"T2, w3"T3}, which are the input features of task layer 3.

[0074] The number of task layers is the same as the number of tasks and can be set arbitrarily according to actual needs. One task layer corresponds to one fusion module and one gating network. For example, task layer 1 corresponds to fusion module 1 and gating network 1, task layer 2 corresponds to fusion module 2 and gating network 2, and task layer 3 corresponds to fusion module 3 and gating network 3. One task layer is used to obtain the corresponding recommendation probability based on the fused feature vectors in one dimension. For example, in Figure 5 the recommendation probability output by task layer 1 is the click-through rate, the recommendation probability output by task layer 2 is the conversion rate, and the recommendation probability output by task layer 3 is the total transaction amount.

[0075] In this embodiment, according to the recommendation records of historical information recommended to users and the user feedback situation, the information content features, user features, and environmental features of the sample information can be obtained. Using these features to train the preset model, a recommendation probability prediction model can be obtained. For example, multiple sample users can be determined, the information of historical products recommended to each sample user can be obtained, and the feedback behaviors of the sample users for the historical products can be analyzed to obtain the information content features, sample user features, and environmental features of the sample product information. Using these features to train the preset model, a recommendation probability prediction model for predicting the probability of recommending products to users can be obtained.

[0076] Through the recommended probability prediction method of this embodiment, the information content features, user general features, environment general features, user bias features, and environment bias features of the information to be recommended are first obtained; then the information content features, user general features, and environment general features of the information to be recommended are input into the recommended probability prediction model to obtain the recommended probabilities in multiple dimensions of the information to be recommended, so as to recommend the information to be recommended to the user according to the recommended probability. Among them, the recommended probability prediction model is trained on a preset model with the information content features, user general features, environment general features, user bias features, and environment bias features of the sample information as inputs. The recommended probability prediction model includes multiple task networks, multiple gate switch networks, multiple fusion modules, and multiple task layers. Each task network is used to obtain a feature vector of a certain dimension of the sample information. A gate switch network is used to output the weights of the feature vectors output by each of the multiple task networks in a certain dimension. A fusion module is used to fuse the feature vectors output by each of the multiple task networks according to the weights of the feature vectors output by each of the multiple task networks in a certain dimension. A task layer is used to obtain the corresponding recommended probability according to the fused feature vector in a certain dimension.

[0077] The recommended probability prediction method of this application has the following technical effects:

[0078] First, the network structures of two multi-task prediction models in the related art are improved to obtain a recommended probability prediction model. Multiple task networks are set in the recommended probability model, and the parameters of the multiple task networks are not completely shared. Each task network is used to extract partial information from the input information, and the information of each task network is selectively utilized through the gate switch network, effectively solving the conflict problem between multiple tasks.

[0079] Second, by controlling the weights of the feature vectors output by each task network through the gate switch network, the difference in the input features of each task layer can be guaranteed, thereby improving the prediction accuracy of the model.

[0080] Combined with the above embodiments, in one implementation manner, the multiple task networks share each feature input into the recommended probability prediction model. Each task network includes a bias sub-network, a task sub-network, and a splicing layer. The bias sub-network takes the user bias features and environment bias features of the sample information as inputs. The task sub-network takes the user bias features and environment bias features of the sample information as inputs. The splicing layer is used to splice the features output by the bias sub-network and the task sub-network to obtain the feature output by a task network.

[0081] Figure 6 It is a schematic structural diagram of a task network shown in an embodiment of this application. Combined with Figure 6, in this embodiment, each task network includes a bias sub-network, a task sub-network, and a splicing layer. The bias sub-network uses environmental features and user features as inputs and processes the input features through a full connect network (FCN). The task sub-network uses environmental features, user features, and information content features as inputs and processes the input features through a full connect network (FCN). Among them, the number of layers of the full connect network (FCN) can be set according to actual needs. The splicing layer is used to splice the output of the bias sub-network and the output of the task sub-network to obtain the output of the task network.

[0082] In this embodiment, in each task network, the role of the task sub-network is to extract features of one dimension from the input features, while the role of the bias sub-network is to remove the bias features from the features output by the task sub-network to achieve correction, so as to ensure the accuracy of the finally extracted features. The splicing layer is used to remove the bias features output by the bias sub-network from the features output by the task sub-network, convert the finally obtained features into feature vectors, and input them into the fusion module.

[0083] In this embodiment, there is more or less an influence of bias features when estimating the recommendation probability. For example, for Android and iOS systems, the order amount on the Android system is generally higher. However, when a user visits a certain food delivery platform, their operating system is already fixed, and the merchants recommended by the food delivery platform should not be affected by the operating system. Therefore, the bias feature of the operating system should be removed to recommend merchants that the user is interested in.

[0084] Exemplarily, in combination with Figure 5 and Figure 6, if task network 1 is responsible for extracting features of the "click behavior" dimension, then the task subnetworks of task network 1 are responsible for extracting click behavior features from environmental features, user features, and information content features. The bias subnetworks of task network 1 are used to extract bias features irrelevant to click behavior from environmental features and user features (such as bias information like the user's historical order amount and conversion rate that has no impact on click-through rate). The splicing layer is used to remove bias features irrelevant to click behavior from the click behavior features. If task network 2 is responsible for extracting features of the "order placement behavior" dimension, then the task subnetworks of task network 2 are responsible for extracting order placement behavior features from environmental features, user features, and information content features. The bias subnetworks of task network 2 are used to extract bias features irrelevant to order placement behavior from environmental features and user features (such as location features and store name information displayed on the front end, which have an impact on click-through rate but little impact on order placement rate). The splicing layer is used to remove bias features irrelevant to order placement behavior from the order placement behavior features. If task network 3 is responsible for extracting features of the "order amount behavior" dimension, then the task subnetworks of task network 3 are responsible for extracting order amount behavior features from environmental features, user features, and information content features. The bias subnetworks of task network 1 are used to extract bias features irrelevant to order amount behavior from environmental features and user features (such as location features and store name information displayed on the front end, which have an impact on click-through rate but little impact on order amount). The splicing layer is used to remove bias features irrelevant to order amount behavior from the click behavior features.

[0085] In this embodiment, each task has certain bias information, and the bias information for each task is different. Therefore, bias subnetworks are designed in each task network, and the gate switch network extracts different bias information for each task, realizing the adaptive acquisition of bias information for different tasks.

[0086] Of course, the task network can also extract features of other dimensions, which can be specifically set according to actual needs, and this embodiment does not make specific restrictions on this. For the features of one dimension, its bias features are pre-set manually according to business requirements, and this embodiment does not make specific restrictions on the setting of bias features. The same feature may be a bias feature in one task but not in another task. Then, by controlling the weights of the output values of the task network by the gate switch network, the input features of the corresponding task layer can be controlled, and thus the adaptive acquisition of bias information for different tasks can be achieved.

[0087] In specific implementation, for task network X, the dimension of the extracted feature is Y. If there are two determined bias features (bias feature 1 and bias feature 2), then four task networks (task network 1 - task network 4) can be set. Task network 1 is used to extract features with dimension Y and exclude bias feature 1. Task network 2 is used to extract features with dimension Y and exclude bias feature 2. Task network 3 is used to extract features with dimension Y and exclude both bias feature 1 and bias feature 2. Task network 4 is used to extract features with dimension Y without excluding bias features. In this way, when a certain business layer needs features with dimension Y and excludes bias feature 1, the corresponding gate-switch network of this business layer can control the weight of the features output by task network 1 to be 1, and the weights of the features output by task networks 2 - 4 to be 0. When a certain business layer needs features with dimension Y and does not exclude bias features, the corresponding gate-switch network of this business layer can control the weight of the features output by task network 4 to be 1, and the weights of the features output by task networks 1 - 3 to be 0. Of course, the number of bias features and the number of task networks can both be set according to actual needs, and this embodiment does not make specific limitations in this regard.

[0088] In this embodiment, to solve the bias problem, the structure of the task network is improved. A bias sub-network is added on the basis of the task sub-network, and the bias features in the task sub-network are removed through the bias sub-network, which improves the accuracy of the features output by the task network and the accuracy of the recommended probability estimated by the task layer, enabling the information recommendation platform to better obtain the to-be-recommended information that the user is interested in and perform targeted recommendations. On the one hand, it improves the benefits of the information recommendation platform and the information provider, and on the other hand, it improves the user experience.

[0089] Combined with the above embodiments, in one implementation manner, the present application also provides a method for recommending to-be-recommended information. When the number of to-be-recommended information is multiple, the recommendation method may include:

[0090] Determine the recommendation probabilities of the multiple to-be-recommended information for the target user in multiple dimensions;

[0091] According to the recommendation probabilities of the multiple to-be-recommended information for the target user in multiple dimensions, determine the comprehensive recommendation probability of the multiple to-be-recommended information for the target user;

[0092] Send the multiple to-be-recommended information to the user terminal used by the target user in descending order of the comprehensive recommendation probability.

[0093] In this embodiment, the information to be recommended can be recommended to the user according to the level of the comprehensive recommendation probability. Specifically, if the recommendation probability prediction model outputs recommendation probabilities in multiple dimensions, then for each piece of information to be recommended, the multiple dimensions of the to-be-recommended probabilities estimated for the target user can be used to obtain the comprehensive recommendation probability according to a preset rule. In this way, when the number of pieces of information to be recommended is multiple, the information to be recommended can be sorted according to the descending order of the comprehensive recommendation probability, and then the information to be recommended can be sent to the user terminal used by the target user according to the sorting result.

[0094] In this embodiment, the information recommendation platform sorts the information to be recommended according to the descending order of the comprehensive recommendation probability, which can better obtain the information to be recommended that the user is interested in and make targeted recommendations, not only improving the benefits of the information recommendation platform and the information provider, but also enhancing the user experience.

[0095] Combined with the above embodiments, in one implementation manner, the present application further provides another method for recommending information to be recommended. When the number of pieces of information to be recommended is multiple, the recommendation method may further include:

[0096] Determine the recommendation probabilities of the multiple pieces of information to be recommended for the target user in multiple dimensions;

[0097] Send the multiple pieces of information to be recommended to the user terminal used by the target user in descending order for each dimension.

[0098] In this embodiment, the information to be recommended can also be recommended to the user by dimension. For example, for the dimension of click-through rate, the click-through rates corresponding to the multiple pieces of information to be recommended estimated for the target user can be sorted in descending order, and then the information to be recommended can be sent to the user terminal used by the target user according to the sorting result. For another example, for the dimension of conversion rate, the conversion rates corresponding to the multiple pieces of information to be recommended estimated for the target user can be sorted in descending order, and then the information to be recommended can be sent to the user terminal used by the target user according to the sorting result.

[0099] In the actual business scenario, some businesses aim at click-through rate, some aim at conversion rate, and some aim at total transaction amount. Therefore, through this method of recommending information to be recommended by dimension, the business requirements can be better met, and while improving the benefits of the information recommendation platform and the information provider, the user experience is also enhanced.

[0100] The recommendation probability estimation method of the present application has the following multiple technical effects:

[0101] 1. The network structure of two multi-task estimation models in related technologies was improved to obtain a recommendation probability prediction model. The recommendation probability model is equipped with multiple task networks, and the multiple task networks share the features of the information input layer, which enhances the generalization ability of the model.

[0102] Second, by controlling the weights of the feature vectors output by each task network through the gate switch network, the input features of each task layer can be differentiated, thereby improving the accuracy of model estimation.

[0103] Third, by controlling the differences between tasks based on parameter sharing, we can prevent inter-task conflicts or excessive noise caused by large differences in the underlying representations of different tasks, thereby improving the model's estimation effect.

[0104] Fourth, the structure of the task network has been improved. A bias subnetwork has been added based on the task subnetwork. The bias subnetwork removes the bias features in the task subnetwork, improves the accuracy of the features output by the task network, and improves the accuracy of the recommendation probability estimated by the task layer. This enables the information recommendation platform to better obtain the recommended information of interest to users and make targeted recommendations. On the one hand, it increases the profits of the information recommendation platform and the information providers, and on the other hand, it improves the user experience.

[0105] Based on the same inventive concept, an embodiment of the present application provides a recommendation probability estimation device 700 . Figure 7 This is a structural block diagram of a recommendation probability estimation device provided by an embodiment of the present application. Figure 7 As shown, the recommendation probability estimation device 700 includes:

[0106] An acquisition module 701 is configured to obtain information content features, user general features, environment general features, user bias features, and environment bias features of the information to be recommended;

[0107] An input module 702 is configured to input the information content features, user general features, and environment general features of the information to be recommended into a recommendation probability prediction model to obtain recommendation probabilities of the information to be recommended in multiple dimensions;

[0108] Among them, the recommended probability prediction model is obtained by training a preset model with the information content features of sample information, user general features, environment general features, user bias features, and environment bias features as inputs. The recommended probability prediction model includes multiple task networks, multiple gate switch networks, multiple fusion modules, and multiple task layers. Each task network is used to obtain a feature vector of a certain dimension of the sample information. A gate switch network is used to output the weights of the feature vectors output by the multiple task networks in a certain dimension. A fusion module is used to fuse the feature vectors output by the multiple task networks according to the weights of the feature vectors output by the multiple task networks in a certain dimension. A task layer is used to obtain the corresponding recommended probability according to the fused feature vectors in a certain dimension.

[0109] Optionally, the multiple task networks share the respective features input to the recommended probability prediction model. Each task network includes a bias sub-network, a task sub-network, and a splicing layer. The bias sub-network takes the user bias features and environment bias features of the sample information as inputs. The task sub-network takes the user bias features and environment bias features of the sample information as inputs. The splicing layer is used to splice the features output by the bias sub-network and the task sub-network to obtain the features output by a task network.

[0110] Optionally, the recommended probabilities of multiple dimensions include any combination of the following: click-through rate, conversion rate, click-and-conversion rate, total transaction amount; and / or

[0111] The environment bias features include at least one of the following: the geographical location where the user is located, the time when the user operation occurs, the client type, the type of the user terminal; and / or

[0112] The user bias features include at least one of the following: click-through rate, order placement rate, order placement amount, order placement quantity.

[0113] Optionally, the number of pieces of information to be recommended is multiple; the apparatus 700 further includes:

[0114] A first determination module, configured to determine the recommended probabilities of multiple dimensions of the multiple pieces of information to be recommended for a target user;

[0115] A second determination module, configured to determine the comprehensive recommended probability of the multiple pieces of information to be recommended for the target user according to the recommended probabilities of multiple dimensions of the multiple pieces of information to be recommended for the target user;

[0116] A first sending module, configured to send the multiple pieces of information to be recommended to the user terminal used by the target user in descending order of the comprehensive recommended probability.

[0117] Optionally, the number of pieces of information to be recommended is plural; the apparatus 700 further includes:

[0118] A third determination module, configured to determine recommendation probabilities of the plural pieces of information to be recommended for a target user in multiple dimensions;

[0119] A second sending module, configured to send the plural pieces of information to a user terminal used by the target user in descending order for each dimension.

[0120] Based on the same inventive concept, another embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps in the method described in any of the foregoing embodiments of the present application are implemented.

[0121] Based on the same inventive concept, another embodiment of the present application provides an electronic device 800, as Figure 8 shown. Figure 8 is a schematic diagram of an electronic device shown in an embodiment of the present application. The electronic device includes a memory 802, a processor 801, and a computer program stored on the memory and executable on the processor. When the processor executes, the steps in the method described in any of the foregoing embodiments of the present application are implemented.

[0122] For the apparatus embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For related parts, refer to the partial description of the method embodiment.

[0123] Each embodiment in this specification is described in a progressive manner. The key point of each embodiment is the difference from other embodiments. For the same and similar parts among the embodiments, refer to each other.

[0124] Those skilled in the art should understand that the embodiments of the present application may be provided as a method, an apparatus, or a computer program product. Therefore, the embodiments of the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the embodiments of the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0125] Embodiments of the present application are described with reference to the flowcharts and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing terminal devices generate a device for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or a device for implementing the functions specified in multiple blocks.

[0126] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal device to work in a specific manner, such that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device that implements the functions specified in one process Figure 1 one process or multiple processes and / or blocks Figure 1 or the functions specified in multiple blocks.

[0127] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device, such that a series of operation steps are executed on the computer or other programmable terminal device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in one process Figure 1 one process or multiple processes and / or blocks<� Figure 1 or the functions specified in multiple blocks.

[0128] Although the preferred embodiments of the embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.

[0129] Finally, it should also be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or terminal device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or terminal device. Without further limitation, an element defined by the statement "comprising an..." does not exclude the existence of additional identical elements in the process, method, article or terminal device comprising the said element.

[0130] The above has introduced in detail a method, apparatus, storage medium and electronic device for estimating recommendation probability provided by the present application. Specific examples are used in this text to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those of ordinary skill in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.

Claims

1. A method for predicting recommendation probability, characterized in that, The method includes: Obtaining the information content features, user general features, environmental general features, user bias features, and environmental bias features of the information to be recommended. The information to be recommended includes product information, store advertisements, and platform activities. The information content features include product price, product taste, product quality, the affiliated store, the purchase method, and whether there are any discounts; Inputting the information content features, user general features, and environmental general features of the information to be recommended into a recommendation probability prediction model to obtain the recommendation probabilities of multiple dimensions of the information to be recommended; Among them, the recommendation probability prediction model is trained by using the information content features, user general features, environmental general features, user bias features, and environmental bias features of sample information as inputs for a preset model. The recommendation probability prediction model includes multiple task networks, multiple gate switch networks, multiple fusion modules, and multiple task layers. The task network is used to convert the features input by the information input layer into feature vectors. Each task network is used to obtain a feature vector of a certain dimension of the sample information. A gate switch network is used to output the weights of the feature vectors output by each of the multiple task networks in a certain dimension. A fusion module is used to fuse the feature vectors output by each of the multiple task networks according to the weights of the feature vectors output by each of the multiple task networks in a certain dimension. The fusion by the fusion module means multiplying the weights corresponding to each feature vector to obtain weighted feature vectors. The fusion module fuses the output of the task network and the output of the gate switch network through softmax. The number of task layers is the same as the number of tasks. One task layer corresponds to one fusion module and one gate switch network. One task layer is used to obtain the corresponding recommendation probability according to the fused feature vectors in a certain dimension; The multiple task networks share the respective features input into the recommendation probability prediction model. Each task network includes a bias sub-network, a task sub-network, and a splicing layer. The bias sub-network takes the user bias features and environmental bias features of the sample information as inputs. The task sub-network takes the user bias features and environmental bias features of the sample information as inputs. The splicing layer is used to splice the features output by the bias sub-network and the task sub-network to obtain the features output by one task network.

2. The method according to claim 1, wherein The recommendation probabilities of multiple dimensions include any combination of the following: click-through rate, conversion rate, click-and-conversion rate, total transaction amount; and / or The environmental bias features include at least one of the following: the geographical location where the user is located, the time when the user operation occurs, the client type, the type of the user terminal; and / or The user bias features include at least one of the following: click-through rate, order placement rate, order placement amount, order placement quantity.

3. The method according to claim 1 or 2, characterized in that The number of the information to be recommended is multiple; the method further includes: Determining the recommendation probabilities of multiple dimensions of the multiple pieces of information to be recommended for the target user; Determining the comprehensive recommendation probability of the multiple pieces of information to be recommended for the target user according to the recommendation probabilities of multiple dimensions of the multiple pieces of information to be recommended for the target user; Send the multiple pieces of information to be recommended to the user terminal used by the target user in descending order of the comprehensive recommendation probability.

4. The method according to claim 1 or 2, characterized in that, The number of the pieces of information to be recommended is multiple; the method further includes: Determine the recommendation probabilities of the multiple pieces of information to be recommended to the target user in multiple dimensions. Send the multiple pieces of information to be recommended to the user terminal used by the target user in descending order for each dimension.

5. A recommended probability prediction device, characterized in that, The device includes: An obtaining module, configured to obtain the information content features, user general features, environment general features, user bias features, and environment bias features of the information to be recommended, where the information to be recommended includes product information, store advertisements, and platform activities, and the information content features include product price, product taste, product quality, the affiliated store, the purchase method, and whether there are discounts. An input module, configured to input the information content features, user general features, and environment general features of the information to be recommended into a recommendation probability prediction model to obtain the recommendation probabilities of the information to be recommended in multiple dimensions. Among them, the recommendation probability prediction model is obtained by training a preset model with the information content features, user general features, environment general features, user bias features, and environment bias features of sample information as inputs. The recommendation probability prediction model includes multiple task networks, multiple gate switch networks, multiple fusion modules, and multiple task layers. The task network is configured to transform the features input by the information input layer into feature vectors. Each task network is configured to obtain a feature vector of a certain dimension of the sample information. A gate switch network is configured to output the weights of the feature vectors output by the multiple task networks in a certain dimension. A fusion module is configured to fuse the feature vectors output by the multiple task networks according to the weights of the feature vectors output by the multiple task networks in a certain dimension. The fusion by the fusion module means multiplying the weights corresponding to each feature vector to obtain weighted feature vectors. The fusion module fuses the output of the task network and the output of the gate switch network through softmax. The number of task layers is the same as the number of tasks. One task layer corresponds to one fusion module and one gate switch network. One task layer is configured to obtain the corresponding recommendation probability according to the fused feature vectors in a certain dimension. The multiple task networks share the input features of the recommendation probability prediction model. Each task network includes a bias sub-network, a task sub-network, and a splicing layer. The bias sub-network takes the user bias features and environment bias features of the sample information as inputs. The task sub-network takes the user bias features and environment bias features of the sample information as inputs. The splicing layer is configured to splice the features output by the bias sub-network and the task sub-network to obtain the features output by a task network.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps in the recommendation probability estimation method according to any one of claims 1-4.

7. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes, it implements the steps in the recommendation probability estimation method according to any one of claims 1-4.

Images