Recommendation probability prediction method and device, computer storage medium and electronic device

Abstract

The present disclosure specifically discloses a recommendation probability prediction method and device, storage medium and electronic device. The method includes: obtaining user behavior data, determining user behavior characteristics based on the behavior data; inputting the characteristics of the information to be recommended and the user behavior characteristics into a multi-task recommendation probability prediction model, and outputting multiple target recommendation probabilities of the user. The multi-task recommendation probability prediction model includes at least an expert network, a gate network corresponding to the prediction target, a hidden layer network and a fusion network; the expert network is used to extract features of the input first target feature from multiple dimensions, the gate network is used to output the weight of the extracted feature under the corresponding prediction target and the second target feature input to the gate network is determined according to the corresponding prediction target, the hidden layer network is used to fuse the extracted features according to the weight output by the corresponding gate network, and the fusion network is used to fuse the recommendation probability output by the hidden layer network. The present disclosure can improve the estimation accuracy of the recommendation probability.

Description

Technical Field

[0001] The present disclosure relates to the field of computer technology, and more specifically, to a recommendation probability prediction method, a recommendation probability prediction device, a computer storage medium, and an electronic device. Background Art

[0002] With the development of computer technology, the application areas of machine learning are becoming more and more extensive. In many application scenarios, data predictions need to be made based on the development patterns of historical data. For example, in the field of personalized recommendations, in order to prioritize the content that users are most likely to click on, like, and pay for, a probability estimate is made for the displayed content, and the decision on whether to display the content or the order of display is made based on the probability estimate results.

[0003] In the related art, a probability prediction model with a single target can only predict one probability. If multiple probabilities are to be predicted at the same time, it is necessary to train multiple models and obtain multiple recommendation probabilities by calling multiple models at the same time. However, during the model training process, the training sample size requirements for different prediction targets are different, that is, there is a data sparsity problem. For example, there is a difference of one or several orders of magnitude in the sample size for click-through rate and conversion rate. The samples used to train the conversion rate model alone are insufficient, resulting in reduced model learning accuracy and affecting the accuracy of probability prediction. On the other hand, there is a deviation in the sample distribution between the prediction stage and the training stage of the conversion rate estimation of a single model. For example, there are samples that are directly converted without being clicked and have not been learned by the conversion rate model, resulting in inaccurate actual predictions. In addition, in the related art, multi-target prediction models are also used to perform probability predictions at the same time. However, there is mutual interference between different prediction targets, which affects the prediction accuracy.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0005] The purpose of the present disclosure is to provide a recommendation probability prediction method and device, a computer storage medium and an electronic device, thereby overcoming, at least to a certain extent, the problems of low prediction accuracy when simultaneously predicting the recommendation probabilities of multiple targets due to the limitations and defects of related technologies.

[0006] Other features and advantages of the present disclosure will become apparent from the following detailed description, or may be learned in part by practice of the present disclosure.

[0007] According to one aspect of the present disclosure, a recommendation probability prediction method is provided, comprising: obtaining user behavior data and determining user behavior characteristics based on the behavior data; inputting characteristics of the information to be recommended and the user behavior characteristics into a multi-task recommendation probability prediction model to output multiple target recommendation probabilities for the user, wherein the multi-task recommendation probability prediction model comprises at least an expert network, a gate network corresponding to the prediction target, a hidden layer network, and a fusion network; wherein the expert network is used to extract features of an input first target feature from multiple dimensions, the gate network is used to output the weight of the extracted feature under the corresponding prediction target, and the second target feature input to the gate network is determined according to the corresponding prediction target, the first target feature is included in the second target feature, the hidden layer network is used to fuse the extracted features according to the weight output by the corresponding gate network, and the fusion network is used to fuse the recommendation probabilities output by the hidden layer network.

[0008] In an exemplary embodiment of the present disclosure, the multi-task recommendation probability prediction model is trained using basic sample features determined by user historical behavior sample data and recommendation information sample data as input; wherein, the first target sample features input to the expert network are the common sample feature set among the second target sample features input to each gate network; and the second target sample features input to the gate network are the sample features unique to the corresponding prediction target of the gate network in the basic sample features and the common sample features.

[0009] In an exemplary embodiment of the present disclosure, the target recommendation probability includes click-through rate and conversion rate, the prediction target includes click-through rate and conversion rate, and the hidden layer network fuses the extracted features according to the weights output by the corresponding gate network, including: performing weighted summation on the extracted features according to the weights output by the corresponding gate network to obtain vector features with the same dimension as the expert network; the vector features are processed by the fully connected network and the classification network in the hidden layer network to output the click-through rate or conversion rate.

[0010] In an exemplary embodiment of the present disclosure, the target recommendation probability also includes click-through and conversion rate, and the fusion network performs a fusion process on the output features of each hidden layer network, including: multiplying the click-through rate and conversion rate by probability to determine the click-through and conversion rate.

[0011] In an exemplary embodiment of the present disclosure, during the training process of the multi-task recommendation probability prediction model, the construction of the loss function includes: determining a first cross-entropy loss based on the obtained click-through rate and sample features with click attributes; determining a second cross-entropy loss based on the obtained click-through rate and sample features with conversion attributes; and fusing the first cross-entropy loss and the second cross-entropy loss to obtain the loss function.

[0012] In an exemplary embodiment of the present disclosure, the expert network includes multiple types of deep learning models and the features output by each expert network are feature vectors with the same dimension.

[0013] In an exemplary embodiment of the present disclosure, the number of information to be recommended is multiple; after inputting the characteristics of the information to be recommended and the user behavior characteristics into the multi-task recommendation probability prediction model to output multiple target recommendation probabilities for the user, the method further includes: combining the multiple target recommendation probabilities according to a preset combination strategy to obtain a combined recommendation probability of each information to be recommended to the user; and pushing the target recommendation information whose combined recommendation probability is greater than a probability threshold to the user.

[0014] According to one aspect of the present disclosure, a recommendation probability prediction device is provided, which includes: an acquisition module for acquiring user behavior data and determining user behavior characteristics based on the behavior data; a probability prediction module for inputting characteristics of the information to be recommended and the user behavior characteristics into a multi-task recommendation probability prediction model to output multiple target recommendation probabilities of the user, wherein the multi-task recommendation probability prediction model includes at least an expert network, a gate network corresponding to the prediction target, a hidden layer network and a fusion network; wherein the expert network is used to extract features of the input first target features from multiple dimensions, the gate network is used to output the weights of the extracted features under the corresponding prediction targets, and the second target features input to the gate network are determined according to the corresponding prediction targets, the first target features are included in the second target features, the hidden layer network is used to fuse the extracted features according to the weights output by the corresponding gate network, and the fusion network is used to fuse the recommendation probabilities output by the hidden layer network.

[0015] According to one aspect of the present disclosure, a computer storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the computer program implements any one of the above-mentioned recommendation probability prediction methods.

[0016] According to one aspect of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform any one of the above-mentioned recommendation probability prediction methods by executing the executable instructions.

[0017] The recommendation probability prediction method in the exemplary embodiment of the present disclosure outputs multiple target recommendation probabilities for a user by inputting user behavior features and features of the information to be recommended into a multi-task recommendation probability prediction model. On the one hand, the weights of the expert network output features under the corresponding prediction targets are output by the gate network corresponding to the prediction targets. The importance of different feature combinations to different prediction targets is dynamically calculated using an attention mechanism and applied to different expert networks. This allows both prediction and training to adaptively determine the degree of dependence of different prediction targets on the input features and each expert network, mitigating mutual interference between prediction targets. Furthermore, since the second target feature input to the gate network is determined based on the corresponding prediction target, the gate network input under different prediction targets can be specified, enabling the gate network to better learn the weighted combinations of the expert network output features under different prediction targets, thereby improving the prediction accuracy of each prediction target. On the other hand, based on the potential probabilistic relationship between different prediction targets, the recommendation probabilities output by different hidden layer networks are fused through a fusion network, allowing the model to capture the relationship between the two during the learning process. This not only improves the model's prediction performance, but also increases the number of output predicted recommendation probabilities and allows for application in different scenarios.

[0018] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings, in which several embodiments of the present disclosure are shown by way of example and not limitation, wherein:

[0020] Figure 1 A flowchart of a method for predicting recommendation probability according to an exemplary embodiment of the present disclosure is shown;

[0021] Figure 2 A schematic structural diagram of a multi-task recommendation probability prediction model according to an exemplary embodiment of the present disclosure is shown;

[0022] Figure 3 A schematic structural diagram of a hidden layer network according to an exemplary embodiment of the present disclosure is shown;

[0023] Figure 4 A flowchart showing a process of fusing features extracted by various expert networks using a hidden layer network according to an exemplary embodiment of the present disclosure is shown;

[0024] Figure 5 A flowchart of constructing a loss function in a multi-task recommendation probability prediction model training process according to an exemplary embodiment of the present disclosure is shown;

[0025] Figure 6 A schematic structural diagram of a recommendation probability prediction device according to an exemplary embodiment of the present disclosure is shown;

[0026] Figure 7 A schematic diagram illustrating a storage medium according to an exemplary embodiment of the present disclosure; and

[0027] Figure 8 A block diagram of an electronic device according to an exemplary embodiment of the present disclosure is shown.

[0028] In the drawings, the same or corresponding reference numerals denote the same or corresponding parts. DETAILED DESCRIPTION

[0029] The exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, the exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these examples are provided so that this disclosure will be thorough and complete and will fully convey the concepts of the exemplary embodiments to those skilled in the art. Identical reference numerals in the figures represent identical or similar structures, and thus detailed descriptions thereof will be omitted.

[0030] In addition, the described features, structures or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, well-known structures, methods, devices, implementations or operations are not shown or described in detail to avoid obscuring various aspects of the present disclosure.

[0031] The blocks shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically separate entities. Specifically, these functional entities may be implemented in software, or in one or more software-hardened modules, or in different networks and / or processor devices and / or microcontroller devices.

[0032] With the development of computer and internet technologies, the need for information object recommendations is increasing in more and more application scenarios. For example, e-commerce platforms can recommend products or services to users, movie ticket purchasing platforms can recommend movies to users, gaming platforms can recommend games, gameplay, and equipment to users, and portals can recommend information to users. The accuracy of recommendation algorithms directly affects the user experience.

[0033] In the related art in this field, a probability estimation is performed on the content to be recommended to the user, and the recommendation or recommendation order is determined based on the result of the probability estimation, which involves click-through rate (CTR) and conversion rate (CVR), where CTR is the ratio of the number of times the content is clicked by the user to the total amount of content displayed, and CVR is the probability that the user stays on the content for more than a preset time, or CVR is the probability that the user will make a purchase. In order to make a probability estimation before sending the data to the user, the probabilities corresponding to various contents can be predicted through a machine learning model. For example, a deep learning model can be used to calculate CTR and CVR by inputting user behavior characteristics and features of the recommended information, and the order of recommendation of the recommended information can be determined based on the obtained probability values. The related art uses a combination of single-objective models to predict probabilities, but there are problems with data sparsity and sample selection bias, and the multi-objective model prediction cannot eliminate the mutual influence between the various prediction targets, and it is difficult to grasp the prediction accuracy of each prediction target at the same time, which greatly affects the accuracy of the model prediction.

[0034] Based on this, in an exemplary embodiment of the present disclosure, a recommendation probability prediction method is first provided. Figure 1 As shown, the recommendation probability prediction method includes the following steps:

[0035] Step S110: Obtaining user behavior data and determining user behavior characteristics based on the behavior data;

[0036] Step S120: Inputting the features of the information to be recommended and the user behavior features into the multi-task recommendation probability prediction model to output multiple target recommendation probabilities of the user;

[0037] Among them, the multi-task recommendation probability prediction model includes at least an expert network, a gate network corresponding to the prediction target, a hidden layer network and a fusion network; the expert network is used to extract features of the input first target features from multiple dimensions, the gate network is used to output the weights of the extracted features under the corresponding prediction targets and the second target features input to the gate network are determined according to the corresponding prediction targets, the hidden layer network is used to fuse the extracted features according to the weights output by the corresponding gate network, and the fusion network is used to fuse the recommendation probabilities output by the hidden layer network.

[0038] According to the recommendation probability prediction method in this example embodiment, on the one hand, the weights of the expert network output features corresponding to the prediction target are obtained by outputting the gate network under the corresponding prediction target, and the importance of different feature combinations to different prediction targets is dynamically calculated using the attention mechanism and applied to different expert networks. This allows both prediction and training to adaptively determine the degree of dependence of different prediction targets on the input features and each expert network, thereby alleviating mutual interference between targets. At the same time, since the second target feature input to the gate network is determined according to the corresponding prediction target, the gate network input under different prediction targets can be specified, which can better enable the gate network to learn the weighted combination of the expert network output features under different prediction targets, thereby improving the prediction accuracy of each prediction target. On the other hand, based on the potential probability relationship between different prediction targets, the recommendation probabilities output by different hidden layer networks are fused through the fusion network, so that the model captures the relationship between the two during the learning process, which not only improves the model prediction performance, but also increases the number of output prediction recommendation probabilities and can be applied to different scenarios.

[0039] The following further describes the recommendation probability prediction method in the exemplary embodiment of the present disclosure.

[0040] Figure 2 A schematic diagram of the structure of a multi-task recommendation probability prediction model according to an exemplary embodiment of the present disclosure is shown. Figure 2 As shown, the multi-task recommendation probability prediction model 200 of the present disclosure includes at least an expert network 210, a gate network 220 corresponding to the prediction target, a hidden layer network 230, and a fusion network 240 for fusing the recommendation probabilities output by each hidden layer network.

[0041] In an exemplary embodiment of the present disclosure, each expert network is used to extract features of the first target feature of the input from multiple dimensions. For example, based on the first target feature of the input, the output of the expert network i is is a set of n-dimensional feature vectors; each gate network is used to output the weight of each expert network output feature under the corresponding prediction target and the second target feature input to each gate network is determined according to the corresponding prediction target. For example, based on the second target feature input to a gate network, the gate network outputs a vector feature of length m The numerical value on each dimension of the vector feature identifies the weight of the output features of each expert network under the corresponding prediction target, wherein the first target feature is contained in the second target feature, that is, the first target feature input to the expert network is a common feature set formed by the common features of the second target features input to each gate network.

[0042] In an exemplary embodiment of the present disclosure, the multi-task recommendation probability prediction model 200 is trained using basic sample features determined by user historical behavior sample data and recommendation information sample data as input.

[0043] Among them, the basic sample features determined by the user's historical behavior sample data include basic features such as age, gender, occupation, work and rest habits, and behavioral features such as clicks, favorites, user ratings, number of orders, item consumption preferences, friendship characteristics, and adding to shopping carts; the basic sample features determined by the recommendation information sample data include content attribute features (such as game attributes, audio and video attributes, or shopping attributes), audience gender and age, gameplay rules, etc. The present disclosure can determine the input of the basic sample features for training the multi-task recommendation probability prediction model based on actual recommendation needs. For example, if the recommendation information is a movie, the input basic sample features may include comments, favorites, and movie watching; if the recommendation information is a product, the input basic sample features may include comments, browsing, and purchase; if the recommendation information is a game, the input basic sample features may include skins, characters, equipment, and skills, etc. Of course, user behavior characteristics must also be input. The present disclosure does not specifically limit the basic training samples used for training the model.

[0044] For further information, see Figure 2 , the second target sample features input to different gate networks are not exactly the same, but are determined according to the prediction target corresponding to the gate network. Specifically, the second target sample features input to each gate network are the sample features unique to the gate network corresponding to the prediction target in the basic sample features and the features common to each gate network. For example, in the live broadcast scenario, in addition to the features common to each gate network, the second sample features input to the gate network corresponding to the prediction target of click-through rate also include: user viewing live broadcast sequence, user viewing live broadcast time sequence, user viewing preference sequence features, etc., and these second sample features will not be input to other gate networks; the second sample features input to the gate network corresponding to the prediction target conversion rate also include: user paid live broadcast sequence, user paid live broadcast amount sequence and other features related to user payment preferences, and these sample features will not be input to other gate networks. Based on this, the present disclosure can specify the second target sample features input to each gate network according to the actual situation of the recommendation probability, making the model training more flexible and better allowing the gate network to learn the weight combination of different expert networks under different prediction targets.

[0045] Continue to see Figure 2 , the first target sample feature input to each expert network of the present disclosure is the same, which is the common sample feature among the second sample features input to each gate network, that is, the first target sample feature input to each expert network is the intersection of the second target sample feature sets input to each gate network.

[0046] For example, if the basic sample features input into the multi-task recommendation probability prediction model include {A1, A2, A3, A4, A5, A6}, where the second target sample features specific to the first prediction target are {A1, A2}, and the second target sample features specific to the second prediction target are {A5, A6}, then during the model training process, the second sample features shared by each gate network are {A3, A4}, then the second target sample features {A1, A2, A3, A4} are input into the gate network corresponding to the first prediction target, the second target sample features {A5, A6, A3, A4} are input into the gate network corresponding to the second prediction target, and the first target sample features {A3, A4} are input into each expert network.

[0047] In the exemplary embodiments of the present disclosure, each gate network can be a single-layer DNN (Deep Neural Networks) model, a DCN (Deep & Cross Network), a DIN (Deep Interest Network) network based on a general attention model, and the like; each expert network can be a multi-layer DNN network or a DIN network, which can learn the interaction information between features through the network. Of course, other deep learning networks can also be selected according to actual prediction needs, such as deepFM network (deep recommendation model), Wide & Deep network, etc. The present disclosure does not specifically limit the model types of each gate network and multiple expert networks.

[0048] In an exemplary embodiment of the present disclosure, multiple expert networks may include multiple types of deep learning networks and the features output by each expert network have the same dimension. By adopting multiple different network structures for multiple expert networks, the advantages of different networks can be utilized, making the output features of the expert networks more flexible and improving the prediction accuracy of the recommendation probability during continuous debugging.

[0049] In an exemplary embodiment of the present disclosure, Figure 3 A schematic diagram of the structure of a hidden layer network according to an exemplary embodiment of the present disclosure is shown. Figure 3As shown, the solid lines represent the structures of two hidden network M and N corresponding to different prediction targets. The hidden network M corresponding to the first prediction target includes: a weighted fusion network m1, which is used to fuse the features extracted by each expert network according to the weights output by the gate network 1; a fully connected network and a classification network m2, which are used to process the features fused by the gate weighted fusion network m1 and output them as the first prediction target (first recommendation probability). The processing process of the hidden network N corresponding to the second prediction target is the same as that of the hidden network M, and will not be repeated here. The fully connected network and classification network in the hidden network can be a DNN network and a softmax network.

[0050] In an exemplary embodiment of the present disclosure, the recommendation probability includes click-through rate and conversion rate, the prediction targets include click-through rate and conversion rate, and the hidden layer network fuses the features extracted by each expert according to the weights output by the corresponding gate network, and outputs the click-through rate and conversion rate respectively. Figure 4 FIG. 1 shows a flowchart of a hidden layer network fusing features extracted by various expert networks according to an exemplary embodiment of the present disclosure. Figure 4 As shown, the process includes:

[0051] Step S410 , performing weighted summation on the extracted features according to the weights output by the corresponding gate networks, to obtain vector features having the same dimension as those of each expert network.

[0052] Continuing with the above example, for a certain prediction target, the corresponding gate network output is a feature vector of length m The output of expert network i is According to the eigenvector G i , perform weighted summation on the features extracted by each expert network, and obtain the vector features of the corresponding gate network under the prediction target:

[0053]

[0054] In step S420 , the vector features are processed by the fully connected network and the classification network in the hidden layer network, and the click rate or conversion rate is output.

[0055] Since the output of this model is a multidimensional feature vector, the weighted sum of the vector features is processed by the fully connected network and classification network in the hidden layer network to output the click-through rate or conversion rate. For example, in the DNN network in the hidden layer network, the weighted sum of the vector features is forward propagated and finally processed by the softmax network, outputting a recommendation probability value.

[0056] In exemplary embodiments of the present disclosure, the target recommendation probability may also include click-through conversion rate. A fusion network is used to fuse the output features of each hidden layer network. Specifically, the click-through conversion rate (PCTCVR) can be determined based on the click-through rate (PCTR) and the conversion rate (PCVR), as shown in the following formula: PCTCVR = PCTR × PCVR. Based on this, and by comprehensively considering the relationship between the various probabilities, the model can simultaneously learn and predict three probability values, namely, the click-through rate, conversion rate, and click-through conversion rate.

[0057] In addition, see Figure 5 As shown in the figure, during the training of the multi-task recommendation probability prediction model, the construction of the loss function includes the following steps:

[0058] Step S510, determine the first cross entropy loss between the click rate and the sample features with click attributes; step S520, determine the second cross entropy loss between the click and conversion rate and the sample features with conversion attributes; step S530, fuse the first cross entropy loss and the second cross entropy loss to obtain a loss function. Among them, cross entropy is used to characterize the distance between the actual output probability and the expected output probability. The smaller the cross entropy value, the smaller the gap between the actual and expected. The present disclosure obtains a loss function by weighting the first cross entropy loss and the second cross entropy loss. The specific weighting process can be performed using the following formula:

[0059] L=αL1+(1-α)L2

[0060] Wherein, L1 is the first cross entropy loss, L2 is the second cross entropy loss, 0<α<1, α is used to control the proportional relationship between the two cross entropy losses, and the value can be determined according to the actual recommendation situation. The present disclosure does not specifically limit the specific determination process of α.

[0061] The multi-task recommendation probability prediction model disclosed in the present invention comprehensively considers the relationship between click-through conversion rate and click-through rate. In the process of constructing the final output recommendation probability and the loss function, the potential probability relationship between the two probability targets is considered. During the model learning process, the relationship between the two can be captured, thereby improving the accuracy of the model in simultaneously predicting conversion rate, click-through rate, and click-through conversion rate.

[0062] The working principle of the multi-task recommendation probability prediction model disclosed in this disclosure is explained below:

[0063] First, specify the expert network and the target features input to each gate network (including the first target feature input to the expert network and the second target feature input to the gate network);

[0064] Secondly, user behavior features and features of the information to be recommended are input into the multi-task recommendation probability prediction model. These features are divided into multiple target features at the input layer. Among them, the second target feature input to each gate network is the sample feature unique to the prediction target of each gate network among these features, and the first target feature input to each expert network is the common sample feature of the second sample features input to each gate network.

[0065] Next, each expert network extracts features from the first target feature of the input and outputs multi-dimensional vector features. Each gate network outputs the weight of each expert network under the corresponding prediction target according to the second target feature of the input.

[0066] Then, each hidden layer network fuses the features extracted by each expert network according to the weight output by the corresponding gate network, and outputs the recommendation probability respectively;

[0067] Finally, the recommendation probabilities output by each hidden layer network can be fused through a fusion network to obtain multiple target recommendation probabilities.

[0068] In addition, in an exemplary embodiment of the present disclosure, when there are multiple pieces of information to be recommended, the user behavior characteristics and the characteristics of the information to be recommended can be input into the multi-task recommendation probability prediction model and multiple target recommendation probabilities of the user can be output. Then, the multiple target recommendation probabilities can be combined according to a preset combination strategy to obtain the combined recommendation probability of each piece of information to be recommended to the user, and the target recommendation information whose combined recommendation probability is greater than the probability threshold can be pushed to the user.

[0069] Alternatively, the click-through rate can be used directly to predict the click-through rate. Alternatively, offline experiments can be used to combine the click-through rate and the click-through conversion rate, and the resulting value can be used for prediction. The combination method can be a linear combination or other mathematical combination method, which is not specifically limited in this application. The probability threshold can be set based on the actual recommendation situation, for example, only recommended information with a combined recommendation probability greater than 75% will be pushed to the user. For example, these recommended information can be pushed to the user in descending order of combined recommendation probability.

[0070] It should be noted that the recommendation probability prediction method disclosed herein can obtain multiple recommendation probabilities in different application scenarios, which can be flexibly used to evaluate different recommendation information. For example, in a live broadcast scenario, the click-through rate and conversion rate can be used to predict the user's preference for paying for the anchor, which can be used as one of the features to measure the anchor's value.

[0071] The multi-task recommendation probability prediction model disclosed herein can use scene data on click-through conversions to model, utilize multiple expert networks to fit existing features to obtain multiple feature combinations, and then use the weights output by the gate network to assign the weights of the output features of each expert network under different prediction targets for click-through rate and conversion rate. The gate networks between each prediction target have unique feature inputs, and the input features are not shared between the underlying networks. This improves the combination of the weights of the output features of different experts under each prediction target in each gate network, improves the expression focus of the gate network, and thus improves the accuracy of the overall model prediction.

[0072] In an exemplary embodiment of the present disclosure, a recommendation probability prediction device is also provided, referring to Figure 6 As shown, the recommendation probability prediction device 600 may include an acquisition module 610 and a probability prediction module 620. The acquisition module 610 is used to acquire user behavior data and determine user behavior characteristics based on the behavior data; the probability prediction module 620 is used to input the characteristics of the information to be recommended and the user behavior characteristics into a multi-task recommendation probability prediction model to output multiple target recommendation probabilities of the user. The multi-task recommendation probability prediction model includes at least an expert network, a gate network corresponding to the prediction target, a hidden layer network, and a fusion network.

[0073] Among them, the expert network is used to extract features of the input first target feature from multiple dimensions, the gate network is used to output the weight of the extracted feature under the corresponding prediction target and the second target feature input to the gate network is determined according to the corresponding prediction target, the first target feature is included in the second target feature, the hidden layer network is used to fuse the extracted features according to the weight output by the corresponding gate network, and the fusion network is used to fuse the recommendation probability output by the hidden layer network.

[0074] Since the functional modules of the recommendation probability prediction device of the exemplary embodiment of the present disclosure are the same as those in the above-mentioned invention embodiment of the recommendation probability prediction method, they will not be described in detail here.

[0075] It should be noted that although several modules or units of the recommendation probability prediction device are mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be concretized in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into multiple modules or units to be concretized.

[0076] In addition, in the exemplary embodiments of the present disclosure, a computer storage medium capable of implementing the above-mentioned method is also provided. A program product capable of implementing the above-mentioned method of this specification is stored on the computer storage medium. In some possible embodiments, various aspects of the present disclosure may also be implemented in the form of a program product, which includes program code. When the program product is executed on a terminal device, the program code is used to cause the terminal device to perform the steps according to the various exemplary embodiments of the present disclosure described in the "Exemplary Methods" section of this specification.

[0077] refer to Figure 7 As shown, a program product 700 for implementing the above method according to an exemplary embodiment of the present disclosure is described. The program product 700 may be a portable compact disc read-only memory (CD-ROM) and include program code, and may be run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto. In this document, a readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0078] The program product may be implemented in any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0079] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0080] The program code embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0081] The program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, and the like, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a stand-alone software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0082] In addition, in an exemplary embodiment of the present disclosure, an electronic device capable of implementing the above method is also provided. Those skilled in the art will appreciate that various aspects of the present disclosure can be implemented as a system, method, or program product. Therefore, various aspects of the present disclosure can be specifically implemented in the following forms, namely: a complete hardware embodiment, a complete software embodiment (including firmware, microcode, etc.), or an embodiment combining hardware and software aspects, which can be collectively referred to herein as a "circuit," "module," or "system."

[0083] Refer to the following Figure 8 800 according to this embodiment of the present disclosure will be described. Figure 8 The electronic device 800 shown is merely an example and should not limit the functions and scope of use of the embodiments of the present disclosure.

[0084] like Figure 8 As shown, electronic device 800 is implemented as a general-purpose computing device. Components of electronic device 800 may include, but are not limited to, the aforementioned at least one processing unit 810, the aforementioned at least one storage unit 820, a bus 830 connecting various system components (including storage unit 820 and processing unit 810), and a display unit 840.

[0085] The storage unit stores program codes, which can be executed by the processing unit 810, so that the processing unit 810 performs the steps according to various exemplary embodiments of the present disclosure described in the above “Exemplary Method” section of this specification.

[0086] The storage unit 820 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 8201 and / or a cache memory unit 8202 , and may further include a read-only memory unit (ROM) 8203 .

[0087] The storage unit 820 may also include a program / utility 8204 having a set (at least one) of program modules 8205, such program modules 8205 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0088] Bus 830 may represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0089] The electronic device 800 can also communicate with one or more external devices 900 (e.g., a keyboard, a pointing device, a Bluetooth device, etc.), one or more devices that enable a user to interact with the electronic device 800, and / or any device that enables the electronic device 800 to communicate with one or more other computing devices (e.g., a router, a modem, etc.). Such communication can occur via an input / output (I / O) interface 850. Furthermore, the electronic device 800 can also communicate with one or more networks (e.g., a local area network (LAN), a wide area network (WAN), and / or a public network such as the Internet) via a network adapter 860. As shown, the network adapter 860 communicates with other modules of the electronic device 800 via a bus 830. It should be understood that, although not shown in the figure, other hardware and / or software modules can be used in conjunction with the electronic device 800, including but not limited to microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0090] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the embodiments of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes a number of instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the embodiments of the present disclosure.

[0091] Furthermore, the figures above are merely illustrative of the processes included in the methods according to exemplary embodiments of the present disclosure and are not intended to be limiting. It is readily understood that the processes illustrated in the figures above do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0092] 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 disclosure 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.

[0093] 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 recommendation probability prediction method, characterized in that: include: Obtaining user behavior data and determining user behavior characteristics based on the behavior data; Inputting the features of the information to be recommended and the user behavior features into a multi-task recommendation probability prediction model to output multiple target recommendation probabilities for the user, wherein the multi-task recommendation probability prediction model includes at least an expert network, a gate network corresponding to the prediction target, a hidden layer network, and a fusion network, and is trained using basic sample features determined by user historical behavior sample data and recommendation information sample data as input; wherein, the features of the information to be recommended and the user behavior features are divided into a plurality of target features, including a first target feature and a second target feature; wherein the second target feature input to each gate network is determined according to the corresponding prediction target, and the first target feature input to the expert network is a common feature set formed by the common features of the second target features input to each gate network; The expert network extracts features of the input first target from multiple dimensions, the gate network outputs weights of the features extracted by the expert network under corresponding prediction targets, the hidden layer network fuses the features extracted by the expert network according to the weights output by the corresponding gate networks to obtain corresponding recommendation probabilities, and the fusion network fuses the recommendation probabilities output by the hidden layer network to obtain multiple target recommendation probabilities; The target recommendation probability includes click-through rate and conversion rate, the predicted target includes click-through rate and conversion rate, and the hidden layer network fuses the extracted features according to the weights output by the corresponding gate network to obtain the corresponding recommendation probability, including: Performing weighted summation on the extracted features according to the weights output by the corresponding gate network to obtain vector features with the same dimension as that of the expert network; The vector features are processed by the fully connected network and the classification network in the hidden layer network to output the click-through rate or conversion rate.

2. The method according to claim 1, characterized in that The first target sample feature input to the expert network is a common sample feature among the second target sample features input to each gate network; The second target sample features input to the gate network are the sample features specific to the prediction target corresponding to the gate network in the basic sample features and the common sample features.

3. The method according to claim 1, characterized in that The target recommendation probability also includes click-through conversion rate. The fusion network performs fusion processing on the output features of each hidden layer network, including: The click-through rate and conversion rate are probabilistically multiplied to determine the click-through rate and conversion rate.

4. The method according to claim 3, characterized in that During the training of the multi-task recommendation probability prediction model, the construction of the loss function includes: Determining a first cross entropy loss based on the obtained click rate and sample features with click attributes; Determine a second cross entropy loss based on the obtained click-through conversion rate and the sample features with conversion attributes; The loss function is obtained by fusing the first cross entropy loss and the second cross entropy loss.

5. The method according to any one of claims 1 to 4, characterized in that The expert network includes multiple types of deep learning models and the features output by each expert network are feature vectors with the same dimension.

6. The method according to claim 1, characterized in that The number of the information to be recommended is multiple; After inputting the features of the information to be recommended and the user behavior features into the multi-task recommendation probability prediction model to output multiple target recommendation probabilities for the user, the method further includes: Combining the multiple target recommendation probabilities according to a preset combination strategy to obtain a combined recommendation probability of each to-be-recommended information to the user; The target recommendation information whose combined recommendation probability is greater than a probability threshold is pushed to the user.

7. A recommendation probability prediction device, characterized in that: The recommendation probability prediction device includes: An acquisition module, configured to acquire user behavior data and determine user behavior characteristics based on the behavior data; A probability prediction module is configured to input the characteristics of the information to be recommended and the user behavior characteristics into a multi-task recommendation probability prediction model to output multiple target recommendation probabilities for the user. The multi-task recommendation probability prediction model includes at least an expert network, a gate network corresponding to the prediction target, a hidden layer network, and a fusion network. The multi-task recommendation probability prediction model is trained using basic sample features determined by user historical behavior sample data and recommendation information sample data as input; wherein, according to the features of the information to be recommended and the user behavior features, multiple target features are divided, including a first target feature and a second target feature; wherein, the second target feature input to each gate network is determined according to the corresponding prediction target, and the first target feature input to the expert network is a common feature set formed by the common features of the second target features input to each gate network; the probability prediction module is configured to execute: feature extraction of the input first target feature from multiple dimensions through the expert network, the gate network is used to output the weight of the feature extracted by the expert network under the corresponding prediction target, the hidden layer network is used to fuse the features extracted by the expert network according to the weight output by the corresponding gate network to obtain the corresponding recommendation probability, and the fusion network is used to fuse the recommendation probabilities output by the hidden layer network to obtain multiple target recommendation probabilities; The target recommendation probability includes click-through rate and conversion rate, the predicted target includes click-through rate and conversion rate, and the hidden layer network fuses the extracted features according to the weights output by the corresponding gate network to obtain the corresponding recommendation probability, including: Performing weighted summation on the extracted features according to the weights output by the corresponding gate network to obtain vector features with the same dimension as that of the expert network; The vector features are processed by the fully connected network and the classification network in the hidden layer network to output the click-through rate or conversion rate.

8. A storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for predicting recommendation probability according to any one of claims 1 to 6 is implemented.

9. An electronic device, characterized in that: include: processor; as well as a memory for storing executable instructions of the processor; The processor is configured to execute the recommendation probability prediction method according to any one of claims 1 to 6 by executing the executable instructions.

Images