Information push model training and information push method, device, equipment and medium

Abstract

The embodiment of the present invention discloses an information push model training and information push method, device, equipment and medium, wherein the method includes: inputting user shopping behavior sample data into a multi-task learning network, and inputting cross-category item shopping behavior sample data into a cross-category expert network and a gating network of the cross-category expert network; generating input information of each task tower network in the multi-task learning network based on the output result of the cross-category expert network and the output result of its gating network, and inputting the input information into each task tower network; fusing the output result of each task tower network with the output result of the adaptive expert network in the multi-task learning network, and updating the parameters of the multi-task learning network, the cross-category expert network and its gating network according to the fusion result, so as to train and obtain the target information push model. The solution realizes pushing more information of items of different categories to users, and improves the richness of the item information browsed by users.

Description

Technical Field

[0001] Embodiments of the present invention relate to the field of artificial intelligence technology, and in particular to information push model training and information push methods, devices, equipment and media. Background Art

[0002] In online shopping platforms, item information is usually pushed in the platform display interface for users to refer to during the shopping process. Among the pushed item information, most are the item information push results generated based on the user's click and purchase records. Different users have different shopping preferences, and the category categories of item information pushed to each user are also limited. In order to achieve the growth of item sales and user volume, shopping platforms will consider pushing more item information of different categories to users, helping users obtain more reference information and find new needs. At present, pushing more cross-category item information to users mainly depends on the business experience of operators, and the pushed item information is determined according to the cross-category item marketing strategy formulated by operators.

[0003] However, in the process of implementing the present invention, it was found that there are at least the following technical problems in the prior art: on the one hand, the business experience of operators in guiding the push of cross-category item information has certain limitations, and the categories of items involved are relatively small; on the other hand, as the business direction changes, operators need to constantly adjust their strategies, which has high labor costs and the degree of automation of information push strategy adjustment needs to be improved. Summary of the invention

[0004] The embodiments of the present invention provide an information push model training and information push method, so as to push more information of items of different categories to users and improve the richness of the item information browsed by users.

[0005] In a first aspect, an embodiment of the present invention provides an information push model training method, the method comprising:

[0006] Acquire user shopping behavior sample data, input the user shopping behavior sample data into a multi-task learning network, and input the cross-category item shopping behavior sample data in the user shopping behavior sample data into a cross-category expert network and a gating network of the cross-category expert network, wherein the cross-category expert network is used to extract feature information of cross-category items, and the gating network of the cross-category expert network is used to determine weight parameters of the output results of the cross-category expert network;

[0007] Generating input information of each task tower network in the multi-task learning network based on the output result of the cross-category expert network and the output result of the gating network of the cross-category expert network, and inputting the input information into each task tower network;

[0008] The output results of each task tower network are fused with the output results of the adaptive expert network in the multi-task learning network, and a first loss function value is determined according to the fusion result. Based on the first loss function value, the parameters of the multi-task learning network, the cross-category expert network and the gating network of the cross-category expert network are updated to train a target information push model.

[0009] In a second aspect, an embodiment of the present invention further provides an information push method, the method comprising:

[0010] Obtaining a behavior data sample of the target user interacting with item information on the target information display platform, wherein the behavior data sample includes a behavior data sample of the target user interacting with cross-category item information, and the cross-category item is an item that the target user has not clicked or purchased within a preset historical time period;

[0011] Inputting the behavior data sample into a target information push model including a multi-task learning network, a cross-category expert network, and a gating network of a cross-category expert network obtained by training the information push model training method described in any embodiment;

[0012] The item information to be pushed is determined based on the output result of the target information push model, and the item information to be pushed is pushed and displayed to the target user.

[0013] In a third aspect, an embodiment of the present invention further provides an information push model training device, the device comprising:

[0014] A sample input module, used to obtain user shopping behavior sample data, input the user shopping behavior sample data into a multi-task learning network, and input the cross-category item shopping behavior sample data in the user shopping behavior sample data into a cross-category expert network and a gating network of the cross-category expert network, wherein the cross-category expert network is used to extract feature information of cross-category items, and the gating network of the cross-category expert network is used to determine weight parameters of the output results of the cross-category expert network;

[0015] A sample operation module, used for generating input information of each task tower network in the multi-task learning network based on the output result of the cross-category expert network and the output result of the gating network of the cross-category expert network, and inputting the input information into each task tower network;

[0016] A model training module is used to fuse the output results of the task tower networks with the output results of the adaptive expert network in the multi-task learning network, and determine the first loss function value based on the fusion result, and update the parameters of the multi-task learning network, the cross-category expert network and the gating network of the cross-category expert network based on the first loss function value to train and obtain a target information push model.

[0017] In a fourth aspect, an embodiment of the present invention further provides an information push device, the device comprising:

[0018] A data acquisition module is used to acquire behavioral data samples of target users interacting with item information on a target information display platform, wherein the behavioral data samples include behavioral data samples of the target users interacting with cross-category item information, and the cross-category items are items that the target users have not clicked or purchased within a preset historical time period;

[0019] A data analysis module, used for inputting the behavior data sample into a target information push model including a multi-task learning network, a cross-category expert network and a gating network of a cross-category expert network obtained by training the information push model training method described in any embodiment;

[0020] The information push module is used to determine the item information to be pushed based on the output result of the target information push model, and push and display the item information to be pushed to the target user.

[0021] In a fifth aspect, an embodiment of the present invention further provides a computer device, the computer device comprising:

[0022] one or more processors;

[0023] A memory for storing one or more programs;

[0024] When the one or more programs are executed by the one or more processors, the one or more processors implement the information push model training method or information push method provided by any embodiment of the present invention.

[0025] In a sixth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an information push model training method or an information push method as provided in any embodiment of the present invention.

[0026] The embodiments of the above invention have the following advantages or beneficial effects:

[0027] In an embodiment of the present invention, by acquiring user shopping behavior sample data, the user shopping behavior sample data is input into a multi-task learning network, and the cross-category item shopping behavior sample data in the user shopping behavior sample data is input into a cross-category expert network and a gating network of the cross-category expert network, that is, a cross-category expert network and a gating network of the cross-category expert network are added to the multi-task learning network for information push; then, based on the output results of the cross-category expert network and the output results of the gating network of the cross-category expert network, input information of each task tower network in the multi-task learning network is generated, and the input information is input into each task tower network. , even if the cross-category expert network can be learned in the task learning in the multi-task learning network to extract the feature information of cross-category items; finally, the output results of each task tower network are fused with the output results of the adaptive expert network in the multi-task learning network, and the first loss function value is determined based on the fusion result, and the parameters of the multi-task learning network, the cross-category expert network and the gating network of the cross-category expert network are updated based on the first loss function value to train the target information push model. The push information corresponding to the output result of the target information push model can contain more cross-category item information to meet the user's cross-category information needs. The technical solution of this embodiment solves the problem that the push of cross-category item information based on the business experience and business strategy guidance of the operator in the prior art has certain limitations, involves fewer categories of items and has a low degree of automation, and realizes data analysis based on the user's shopping behavior sample data, establishes a target information push model, can push more different categories of item information to users, and improve the richness of the item information browsed by users. At the same time, it improves the degree of automation of dynamically adjusting the target push information. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a flow chart of an information push model training method provided in Embodiment 1 of the present invention;

[0029] Figure 2 This is a schematic diagram of a multi-task learning model structure for information push provided in the first embodiment of the present invention;

[0030] Figure 3 is a flow chart of an information push model training method provided by Embodiment 2 of the present invention;

[0031] Figure 4 This is a schematic diagram of a multi-task learning model structure for information push provided by Embodiment 2 of the present invention;

[0032] Figure 5 is a flow chart of an information push method provided by Embodiment 3 of the present invention;

[0033] Figure 6It is a structural diagram of an information push model training device provided by Embodiment 4 of the present invention;

[0034] Figure 7 It is a structural diagram of an information push device provided by Embodiment 5 of the present invention;

[0035] Figure 8 It is a structural diagram of a computer device provided in Embodiment 6 of the present invention. DETAILED DESCRIPTION

[0036] The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for ease of description, only parts related to the present invention, rather than all structures, are shown in the accompanying drawings.

[0037] On shopping platforms or other information display platforms, some item information is usually pushed and displayed to users based on their information browsing records. Users can selectively interact with the pushed and displayed item information, such as adding the items corresponding to the item information to shopping, collection or purchase. At present, according to the business needs of the platform, some item information that is not directly related to the user's information browsing record may be pushed to users, that is, category information that is relatively new to the user. If the user interacts with the item information of the new category, it satisfies the user's more abundant information browsing needs and helps the business development of the platform. At present, in the process of information push, the pushed item information and the rights and interests information related to the item information of the novel category, such as item discount information, are mostly determined based on the operating strategy of the platform operator. On this basis, the scheme in the embodiment proposes a neural network model training scheme for information push, which is used to automatically learn the user's browsing possibility of the item information of the new category, so as to provide users with richer item information.

[0038] Embodiment 1

[0039] Figure 1 This is a flowchart of an information push model training method provided in Embodiment 1 of the present invention. This embodiment can be applied to the case of training a multi-task learning network for pushing item information on a shopping platform or other information browsing platform. The method can be executed by an information push model training device, which can be implemented by software and / or hardware and integrated into a computer device with application development function.

[0040] like Figure 1 As shown, the information push model training method includes the following steps:

[0041] S110, obtaining user shopping behavior sample data, inputting the user shopping behavior sample data into a multi-task learning network, and inputting the cross-category item shopping behavior sample data in the user shopping behavior sample data into a cross-category expert network and a gating network of the cross-category expert network.

[0042] Among them, the user shopping behavior sample data is obtained, that is, the operation behavior data of the user after seeing the pushed item information on the shopping platform, including only swiping over the item push information without clicking to view the detailed information content, clicking to view the item push information content, collecting the viewed item push information content, adding the item corresponding to the item push information content to the shopping cart, and purchasing the item corresponding to the item push information content. Usually, the behavior of not clicking on the item push information is used as a negative sample, and the behavior of clicking to browse or purchase the item push information is used as a positive sample. Further, the sample label can also be set according to the specific model learning task. For example, in the learning task model for pushing the information of the items that may be purchased, only browsing and generating a purchase behavior can be used as a positive sample, and even if the pushed item information is browsed and added to the shopping cart, it can only be used as a negative sample in the learning task model for pushing the information of the items that may be purchased. Further, in this embodiment, the user shopping behavior sample data can also be set with item discount information (such as whether there is a coupon), item discount sensitivity and user discount sensitivity features, so that the information push model can push relevant discount information to the user after the sample feature learning. Among them, item discount sensitivity refers to the discount level of items on daily basis or holidays, and user discount sensitivity refers to the discount level of items when users make item purchases. In addition, the characteristics of user shopping behavior sample data also include user shopping behavior, item category and other characteristics. Based on the learning of the above characteristics, it is helpful to improve the click-through rate of pushed item information and the conversion rate of item information.

[0043] In particular, in this embodiment, the user shopping behavior sample data also includes cross-category item shopping behavior sample data. In the cross-category item shopping behavior sample data, the behavior of pushing cross-category item information but not clicking to browse is a negative sample, and the behavior of pushing cross-category item information and generating information clicking to browse or cross-category item purchase is a positive sample.

[0044] Cross-category items are the categories of items for which users have not interacted with them in a certain historical period (such as 3 months, 6 months or 1 year). Cross-category items may be new item needs for users. The cross-category expert network is an expert network used to extract feature information of cross-category items. The gating network of the cross-category expert network is a network used to determine the weight parameters of the output results of the cross-category expert network.

[0045] In this embodiment, the multi-task learning network refers to a multi-gate mixture of experts (MMoE) multi-task learning model. The multi-task model can improve the learning efficiency and quality of each task by learning the connections and differences between different tasks. MMoE contains multiple expert networks (Experts), and then a gate network (Gate) is introduced for each learning task (task). Gate is to learn different combination modes of Experts for each task, that is, to adaptively weight the output of Experts.

[0046] In this embodiment, the structure of the multi-task learning network with the cross-category expert learning network and the gating network of the cross-category expert network can be referred to. Figure 2 The schematic diagram shown. Figure 2 In the figure, Expert4 is a newly added cross-category expert learning network, and GateC is the gating network of the newly added cross-category expert learning network. Expert0-Expert3 are all expert networks for Task A and Task B. That is, the gating network of Task A is GateA, the tower network of Task A is TowerA, and the corresponding expert networks of Task A include Expert0-Expert3; the gating network of Task B is GateB, the tower network of Task B is TowerB, and the corresponding expert networks of Task B include Expert0-Expert3. The structure of each tower network and each gating network can be a 2-3 layer multi-layer perceptron (MLP). Task A and Task B are only in Figure 2 The two tasks given as an example, such as the learning task of pushing information about items that may be purchased and the learning task of pushing information about items that may be clicked and browsed, can also include more learning tasks, such as the learning task of pushing information about items that may be collected.

[0047] Specifically, in the process of inputting the user shopping behavior sample data into the corresponding network module, the Input layer first outputs the acquired sample data to the Embedding layer (embedding layer) to represent the sample data as a vector. Then, each sample vector is input into the corresponding network structure. That is, the vectors of all samples are input into the adaptive expert network (ExpertX), Expert0-Expert3, GateA and GateB, and the vectors of the cross-category shopping behavior samples in all samples are input into Expert4 and GateC. Among them, the adaptive expert network (ExpertX) is a network used to automatically update the hyperparameters (export_weight) in the expert networks Expert0-Expert4.

[0048] S120. Generate input information of each task tower network in the multi-task learning network based on the output results of the cross-category expert network and the output results of the gating network of the cross-category expert network, and input the input information into each task tower network.

[0049] In this step, the output results of the expert network corresponding to each task are input into the tower network of each task under the constraint of the task gating network, and the output results of the cross-category expert network are input into the tower network of each task under the constraint of the output results of its gating network.

[0050] Specifically, the first input information of each task tower network can be determined based on the output results of the cross-category expert network and the output results of the gating network of the cross-category expert network. Specifically, the product of the output result of the cross-category expert network Expert4 and the output result of the gating network GateC can be calculated as the first input information.

[0051] Then, the output results of the gated networks of each task in the multi-task learning network and the output results of the expert networks corresponding to each task are calculated according to the preset operation rules to obtain the second input information of each task tower network. Specifically, the output results of the gated networks GateA and GateB of each task in the multi-task learning network and the output results of the corresponding expert networks Expert0-Expert3 are calculated according to the preset operation rules to obtain the second input information of each task tower network. For example, the output result of GateA is a vector (0.1, 0.2, 0.3, 0.4), and the output result of GateB is a vector (0.3, 0.1, 0.2, 0.4) The output results of Expert0-Expert3 are (E1, E2, E3, E4) respectively, that is, the result of 0.1*E1+0.2*E2+0.3*E3+0.4*E4 is used as the second input information of TowerA, and the result of 0.3*E1+0.1*E2+0.2*E3+0.4*E4 is used as the second input information of TowerB.

[0052] Finally, the first input information is superimposed with the second input information of each task tower network to obtain the input information of each task tower network in the multi-task learning network. That is, TowerA and TowerB can both learn the features of cross-category items.

[0053] S130. Fusing the output results of the task tower networks with the output results of the adaptive expert network in the multi-task learning network, and determining a first loss function value based on the fusion result, and updating the parameters of the multi-task learning network, the cross-category expert network, and the gating network of the cross-category expert network based on the first loss function value to train and obtain a target information push model.

[0054] The first loss function is determined based on the loss function of each task tower network, that is, based on the loss functions LossA and LossB of TowerA and TowerB, for example, the first loss function is obtained by interpolating LossA and LossB. It can be expressed as, first loss function = αLossA + βLossB, α and β are preset difference coefficients.

[0055] Furthermore, in order to reflect the goal of cross-category item information learning, this embodiment proposes a category gating constraint in the loss function and designs a cross-category coefficient γ0, so as to provide more adequate learning for cross-category samples. The improved loss function formula is as follows: Among them, p a and p b The probability of the results calculated by the tower network after feature learning in Task A and Task B respectively.

[0056] After determining the value of the first loss function, if the value cannot meet the preset convergence conditions, the parameters can be reversely updated according to the value of the first loss function to adjust the parameters of each expert network and the gating network in the network until the value of the first loss function meets the preset convergence conditions. A multi-task learning network model for information push is obtained, which includes a cross-category expert network and a gating network corresponding to the cross-category expert network.

[0057] The technical solution of this embodiment obtains user shopping behavior sample data, inputs the user shopping behavior sample data into the multi-task learning network, and inputs the cross-category item shopping behavior sample data in the user shopping behavior sample data into the cross-category expert network and the gating network of the cross-category expert network, that is, adds the cross-category expert network and the gating network of the cross-category expert network to the multi-task learning network for information push; then, based on the output results of the cross-category expert network and the output results of the gating network, the input information of each task tower network in the multi-task learning network is generated, and the input information is input into each task tower network. In the network, even if the cross-category expert network can be learned in the task learning in the multi-task learning network to extract the feature information of cross-category items; finally, the output results of each task tower network are fused with the output results of the adaptive expert network in the multi-task learning network, and the first loss function value is determined based on the fusion result, and the parameters of the multi-task learning network, the cross-category expert network and the gated network are updated based on the first loss function value to train the target information push model. The push information corresponding to the output result of the target information push model can contain more cross-category item information to meet the user's cross-category information needs. The technical solution of this embodiment solves the problem that the push of cross-category item information based on the business experience and business strategy guidance of the operator in the prior art has certain limitations, involves fewer categories of items and has a low degree of automation, and realizes data analysis based on the user's shopping behavior sample data, establishes a target information push model, can push more different categories of item information to users, and improves the richness of the item information browsed by users. At the same time, it improves the degree of automation of dynamically adjusting the target push information.

[0058] Embodiment 2

[0059] Figure 3 This is a flowchart of an information push model training method provided in Embodiment 2 of the present invention. This embodiment and the information push model training method in the above embodiment belong to the same inventive concept, and the model structure is further optimized. The method can be executed by an information push model training device, which can be implemented by software and / or hardware and integrated into a computer device with application development function.

[0060] like Figure 3 As shown, the information push model training method includes the following steps:

[0061] S210. Obtain user shopping behavior sample data, and multiply the representation vector of each user shopping behavior sample data by the category vector of the category to which the item corresponding to the sample data belongs, to obtain an updated representation vector of the user shopping behavior sample data.

[0062] In this step, based on the feature engineering of the original push algorithm's refined ranking model, in order to focus more on the samples of cross-category items, each sample is multiplied by an additional category representation vector. The category representation vector is a vector determined by aggregating the vector representations of first-level category items in the conventional item classification scheme. Categories under the same first-level category use the same set of category vectors. As part of the model parameters, the category vector can be continuously updated with the model.

[0063] S220. Input the updated user shopping behavior sample data representation vector into the multi-task learning network, and input the cross-category item shopping behavior sample representation vector in the updated user shopping behavior sample data representation vector into the cross-category expert network and the gating network of the cross-category expert network.

[0064] The cross-category expert network is used to extract feature information of cross-category items, and the gating network is used to determine the weight parameters of the output results of the cross-category expert network.

[0065] S230, generating input information of each task tower network in the multi-task learning network based on the output results of the cross-category expert network and the output results of the gating network, and inputting the input information into each task tower network.

[0066] In this embodiment, a routing filter network is added to each gated network, so that each task tower network can filter the output results of the expert network, so that each task tower network can learn the features more specifically. Figure 4 The network structure shown.

[0067] based on Figure 4 The process of generating the input information of each task tower network in the multi-task learning network includes:

[0068] First, the output results of the control network of each task gate are compared with the gated screening value set by the routing screening network, and the values ​​greater than or equal to the gated screening value in the output results of the control network of each task gate are multiplied by the values ​​greater than or equal to 1, which is equivalent to keeping the original value or amplifying the value; the values ​​less than the gated screening value in the output results of the control network of each task gate are set to 0, which is equivalent to discarding the values ​​that do not reach the gated threshold, so as to update the output results of the control network of each task gate. For example, the output result of GateA is a vector (0.3, 0.1, 0.6), and the gated screening value is 0.5, then the output result of the updated GateA is a vector (0, 0, 1). The output results of the updated task gated networks are respectively compared with the output results of the expert networks corresponding to each task according to the preset operation rules to obtain partial input information of each task tower network. According to the above update results, the partial input information of TowerA is 0*E1+0*E2+1*E3. This process is equivalent to improving the weight of a certain expert network.

[0069] In fact, the routing filtering network is equivalent to a sign function, and the condition of the sign function serves as the gated filtering value.

[0070] S240. Fusing the output results of each task tower network with the output results of the adaptive expert network in the multi-task learning network, and determining a first loss function value based on the fusion result, and updating the parameters of the multi-task learning network, the cross-category expert network, and the cross-category expert network gating network based on the first loss function value to train and obtain a target information push model.

[0071] exist Figure 4 It can be seen from the figure that the expert networks corresponding to tasks A and B are better than those corresponding to tasks A and B. Figure 2 The expert network corresponding to task A includes Expert0-Expert2, and the expert network corresponding to task B includes Expert1-Expert3. In addition, task A and task B also receive the results output by the cross-category expert network Expert4 respectively.

[0072] Among them, Expert0 and Expert3 use the output results as part of the input information of only one task tower network, which are called private expert networks, and Expert1, Expert2 and Expert4 use the output results as part of the input information of multiple task tower networks at the same time, which are called public expert networks. Then, in order to ensure that the frequency of parameter updates of private experts and public experts remains synchronized, the reverse update of parameters can be adopted according to the loss functions of TowerA and TowerB respectively, that is, the alternating update strategy. Specifically, the parameters of the multi-task learning network, the cross-category expert network and the gated network are updated based on the value of the first loss function, including: updating the parameters of the adaptive expert network and the gated network of each task in the gated network and the multi-task learning network according to the first loss function. For the private expert network in the multi-task learning network and the class expert network, the parameters are updated according to the loss function value of the task tower network that receives the private expert information input; that is, the parameters of Expert0 are updated according to the loss function of LossA, and the parameters of Expert3 are updated according to the loss function of LossB. For the public expert network in the multi-task learning network and the cross-category expert network, in the process of updating the parameters of the public expert network twice in a row, the parameters are updated according to the loss function values ​​of different task tower networks, that is, the parameters of Expert1, Expert 2 and Expert 4 are updated once according to the loss function of LossA and once according to the loss function of LossB, that is, they are updated alternately.

[0073] Furthermore, for the selection of shared expert networks and private expert networks, any expert networks in Expert0-3 can be combined and set up to train multiple target information push models with different structures. The final private expert and public expert setting plan can be determined based on the effect of the model in the testing phase.

[0074] The effectiveness of the target information recommendation model can be evaluated using parameters such as category width and category novelty. Category width can be measured by the number of categories involved in the pushed item information, such as the number of first-level categories exposed per capita, the number of third-level categories exposed per capita, the number of first-level categories clicked per capita, and the number of third-level categories clicked per capita. Category novelty can be measured by the number of cross-category item information that users click to browse or the number of cross-category items purchased. For example, the number of orders for a cross-category item, the proportion of cross-category item information browsed by a user within a certain period of time to the total number of item information browsed. In addition, other measurement indicators can also be used, such as business-related KPI indicators.

[0075] The technical solution of this embodiment improves the target information push model structure in the above-mentioned embodiment, adds the processing process of sample vectors, and adds a routing screening network. It can also set the expert network as a private expert network and a public expert network according to the task characteristics, so that the network training and learning process is further inclined to cross-category samples, and each task learns effective features, which can make the updated target information push model more effective. The push information corresponding to the output result of the target information push model can contain more cross-category item information to meet the user's cross-category information needs. The technical solution of this embodiment solves the problem that the push of cross-category item information based on the business experience and business strategy guidance of the operator in the prior art has certain limitations, involves fewer categories of items and has a low degree of automation, and realizes data analysis based on the user's shopping behavior sample data, establishes a target information push model, can push more different categories of item information to users, and improve the richness of the item information browsed by users. At the same time, it improves the automation of dynamically adjusting the target push information.

[0076] In a preferred embodiment, during the sample acquisition or sample recall phase, the user shopping behavior data can be recalled according to a preset sample recall algorithm to obtain an initial recall data sample. Among them, the preset sample recall algorithm may include short-term interest information recall, medium- and long-term interest awakening, and cross-category item exploration recall. Short-term interest information recall is to extract information about items that users clicked to browse in the short term (such as a week or a month) but did not generate a purchase behavior as a recall trigger signal. Medium- and long-term interest recall is to use information that was clicked but did not generate a purchase behavior in the user's medium- and long-term (such as the past year) historical information browsing behavior as a recall trigger signal. The cross-category item exploration and recall process includes clustering analysis of users and building similar user groups based on the cluster analysis results; pushing and displaying items in the similar user group that have not been clicked or purchased by the current user within a preset historical time period and that have been clicked or purchased by other users except the current user as cross-category items of the current user; among the pushed and displayed cross-category items, items that have been clicked or purchased by the current user are used as positive samples of the cross-category item samples, and items that have not been clicked or purchased by the current user are used as negative samples of the cross-category item samples.

[0077] Furthermore, after the sample is recalled, the sample weight value can be updated for the cross-category item shopping behavior sample in the initial recalled data sample to obtain an updated recalled data sample; that is, the training and learning of the model further tends to the cross-category item shopping behavior sample.

[0078] Finally, the updated recall data samples are input into the preset neural network for preliminary screening of samples, and the sample screening is performed to obtain the user shopping behavior sample data, wherein, in the updated recall data samples, the sample label weight value of the cross-category item shopping behavior sample is greater than the sample label weight value of the non-cross-category item shopping behavior sample. The preset neural network for preliminary screening of samples can adopt the coarse-layer network structure in the existing central push algorithm architecture to achieve preliminary screening of recall samples. This is to accelerate the convergence process of the target information push model in the above embodiment.

[0079] Embodiment 3

[0080] Figure 5 This is a flowchart of an information push method provided in Embodiment 3 of the present invention. This embodiment can be applied to the case of pushing item information on a shopping platform or other information browsing platform. The method can be executed by an information push device, which can be implemented by software and / or hardware and integrated into a computer device with application development function.

[0081] like Figure 5 As shown, the information push method includes the following steps:

[0082] S310: Obtain a behavioral data sample of the target user interacting with item information on the target information display platform.

[0083] The data currently obtained is the behavior data of the target user interacting with the item information on the target information display platform within a certain historical period. The behavior of interacting with the item information includes just swiping over the item push information without clicking to view the detailed information content, clicking to view the item push information content, collecting the viewed item push information content, adding the item corresponding to the item push information content to the shopping cart, and purchasing the item corresponding to the item push information content, etc.

[0084] The behavior data samples include behavior data samples of the target user interacting with cross-category item information, and the cross-category items are items that the target user has clicked on or not purchased within a preset historical time period.

[0085] S320: Input the behavior data sample into a target information push model including a multi-task learning network, a cross-category expert network and a gating network of a cross-category expert network obtained by training the information push model training method described in any one of the embodiments.

[0086] Among them, the target information push model learns the cross-category expert network to extract the feature information of cross-category items. The push information corresponding to the output result of this model can contain more cross-category item information to meet users' needs for cross-category information.

[0087] S330: Determine the item information to be pushed based on the output result of the target information push model, and push and display the item information to be pushed to the target user.

[0088] Specifically, the output result of the target information push model includes the probability that multiple item information may be clicked or browsed or result in item purchase behavior. A certain amount of item information can be screened out and pushed to the target user based on the probability.

[0089] The form of pushing item information can be in the form of the item shopping cart pop-up layer, item settlement matching relationship layer, item purchase order details page matching layer or multiple content layer of the target information display platform, to push and display the item information to be pushed to the target user. Among them, the item shopping cart pop-up layer is the pop-up display interface of the shopping cart on the target information display platform; the item settlement matching relationship layer is the information display position that can be matched with the item in the item details display page; the item purchase order details page matching layer is the position where the item information is pushed in the order display page; the multiple content layer is the content display position with streaming media attributes such as articles or short video materials. By pushing item information to target users in multiple locations, the possibility of users clicking to browse item information is increased, allowing users to learn more about item information.

[0090] The technical solution of this embodiment, by obtaining user shopping behavior sample data and inputting the user shopping behavior sample data into the target information push model trained by the above-mentioned embodiment, can promote more cross-category item information to users based on the model output results, thereby meeting the user's cross-category information needs. The technical solution of this embodiment solves the problem that the push of cross-category item information based on the business experience and business strategy guidance of operators in the prior art has certain limitations, involves fewer categories of items and has a low degree of automation, and realizes data analysis based on the user's shopping behavior sample data, pushes more different categories of item information to users, improves the richness of the item information browsed by users, and at the same time, improves the degree of automation of dynamically adjusting the target push information.

[0091] Embodiment 4

[0092] Figure 6 This is a structural diagram of an information push model training device provided in Example 4 of the present invention. This embodiment can be applied to the case of training a multi-task learning network for pushing item information on a shopping platform or other information browsing platform. The device can be implemented by software and / or hardware and integrated into a computer terminal device with application development capabilities.

[0093] like Figure 6As shown, the information push model training device includes: a sample input module 410, a sample operation module 420 and a model training module 430.

[0094] The sample input module 410 is used to obtain user shopping behavior sample data, input the user shopping behavior sample data into the multi-task learning network, and input the cross-category item shopping behavior sample data in the user shopping behavior sample data into the cross-category expert network and the gating network of the cross-category expert network, wherein the cross-category expert network is used to extract feature information of cross-category items, and the gating network of the cross-category expert network is used to determine the weight parameters of the output results of the cross-category expert network; the sample operation module 420 is used to extract feature information of cross-category items based on the output results of the cross-category expert network and the cross-category item shopping behavior sample data of the cross-category expert network. The output result of the gating network of the category expert network generates input information of each task tower network in the multi-task learning network, and inputs the input information into each task tower network; the model training module 430 is used to fuse the output results of each task tower network with the output results of the adaptive expert network in the multi-task learning network, and determine the first loss function value based on the fusion result, and update the parameters of the multi-task learning network, the cross-category expert network and the gating network of the cross-category expert network based on the first loss function value to train and obtain the target information push model.

[0095] The technical solution of this embodiment obtains user shopping behavior sample data, inputs the user shopping behavior sample data into the multi-task learning network, and inputs the cross-category item shopping behavior sample data in the user shopping behavior sample data into the cross-category expert network and the gating network of the cross-category expert network, that is, adds the cross-category expert network and the gating network of the cross-category expert network to the multi-task learning network for information push; then, based on the output results of the cross-category expert network and the output results of the gating network of the cross-category expert network, the input information of each task tower network in the multi-task learning network is generated, and the input information is input into each task tower network. In the task tower network, even if the cross-category expert network can be learned in the task learning in the multi-task learning network to extract the feature information of cross-category items; finally, the output results of each task tower network are fused with the output results of the adaptive expert network in the multi-task learning network, and the first loss function value is determined based on the fusion result, and the parameters of the multi-task learning network, the cross-category expert network and the gated network are updated based on the first loss function value to train the target information push model. The push information corresponding to the output result of the target information push model can contain more cross-category item information to meet the user's cross-category information needs. The technical solution of this embodiment solves the problem that the push of cross-category item information based on the business experience and business strategy guidance of the operator in the prior art has certain limitations, involves fewer categories of items and has a low degree of automation, and realizes data analysis based on the user's shopping behavior sample data, establishes a target information push model, can push more different categories of item information to users, and improves the richness of the item information browsed by users. At the same time, it improves the degree of automation of dynamically adjusting the target push information.

[0096] Optionally, the sample operation module 420 is specifically used for:

[0097] Determine the first input information of each task tower network based on the output result of the cross-category expert network and the output result of the gating network of the cross-category expert network;

[0098] The output results of the gating network of each task in the multi-task learning network and the output results of the expert network corresponding to each task are calculated according to a preset operation rule to obtain the second input information of the tower network of each task;

[0099] The first input information and the second input information are superimposed to obtain input information of each task tower network in the multi-task learning network.

[0100] Optionally, an expert network that uses the output result as only part of the input information of one task tower network is called a private expert network, and an expert network that uses the output result as part of the input information of multiple task tower networks is called a public expert network. The model training module 430 is specifically used for:

[0101] updating parameters of the gating network and the adaptive expert network in the multi-task learning network and the gating network of each task according to the first loss function;

[0102] For the multi-task learning network and the private expert network in the cross-category expert network, updating parameters according to the loss function value of the task tower network that receives the private expert information input;

[0103] For the common expert network in the multi-task learning network and the cross-category expert network, in the process of updating the parameters of the common expert network twice in a row, the parameters are updated according to the loss function values ​​of different task tower networks.

[0104] Optionally, the gating network of each task in the multi-task learning network is provided with a routing screening network for screening the output results of the expert network corresponding to each task, and the sample operation module 420 can also be used for:

[0105] Compare the output results of the control networks of the task gates with the gating screening values ​​set by the routing screening network, and set the values ​​of the output results of the control networks of the task gates that are less than the gating screening values ​​to 0, so as to update the output results of the control networks of the task gates;

[0106] The updated output results of each of the task gated networks are respectively compared with the output results of the expert network corresponding to each task and the partial input information of each task tower network is calculated according to the preset operation rules.

[0107] Optionally, the sample input module 410 may also be used to multiply the representation vector of each user shopping behavior sample data by the category vector of the category to which the item corresponding to the sample data belongs before inputting the user shopping behavior sample data into the multi-task learning network.

[0108] Optionally, the information push model training device further includes a sample preprocessing module, including:

[0109] The sample recall submodule is used to recall the user's shopping behavior data according to a preset sample recall algorithm to obtain an initial recall data sample;

[0110] A sample weight updating submodule, used to update the sample weight value of the cross-category item shopping behavior sample in the initial recall data sample to obtain an updated recall data sample;

[0111] The sample pre-screening submodule is used to input the updated recall data samples into a preset neural network for preliminary screening of the samples, and perform sample screening to obtain the user shopping behavior sample data, wherein, in the updated recall data samples, the sample label weight value of the cross-category item shopping behavior sample is greater than the sample label weight value of the non-cross-category item shopping behavior sample.

[0112] Optionally, the sample recall submodule may also be used for:

[0113] Conduct cluster analysis on users and build similar user groups based on the cluster analysis results;

[0114] Push and display items in the similar user group that the current user has not browsed or purchased within a preset historical time period and that other users except the current user have clicked or purchased as cross-category items of the current user;

[0115] Among the cross-category items pushed for display, the items clicked or purchased by the current user are used as positive samples of the cross-category item samples, and the items not clicked or purchased by the current user are used as negative samples of the cross-category item samples.

[0116] Optionally, the user shopping behavior sample data includes item discount information, item discount sensitivity and user discount sensitivity characteristics.

[0117] The information push model training device provided in the embodiment of the present invention can execute the information push model training method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0118] Embodiment 5

[0119] Figure 7 This is a structural diagram of an information push device provided in Example 5 of the present invention. This embodiment can be applied to situations where item information is pushed on a shopping platform or other information browsing platform. The device can be implemented by software and / or hardware and integrated into a server device with application development capabilities.

[0120] like Figure 7 As shown, the information push device includes: a data acquisition module 510, a data analysis module 520 and an information push module 530.

[0121] Among them, the data acquisition module 510 is used to obtain behavioral data samples of the target user interacting with item information on the target information display platform, wherein the behavioral data samples include behavioral data samples of the target user interacting with cross-category item information, and the cross-category items are items that the target user has not clicked or purchased within a preset historical time period; the data analysis module 520 is used to input the behavioral data samples into a target information push model trained by the information push model training method described in any embodiment, which includes a multi-task learning network, a cross-category expert network and a gated network of a cross-category expert network; the information push module 530 is used to determine the item information to be pushed based on the output result of the target information push model, and push the item information to be pushed to the target user for display.

[0122] The technical solution of this embodiment, by obtaining user shopping behavior sample data and inputting the user shopping behavior sample data into the target information push model trained by the above-mentioned embodiment, can promote more cross-category item information to users based on the model output results to meet the user's cross-category information needs. The technical solution of this embodiment solves the problem that the push of cross-category item information based on the business experience and business strategy guidance of operators in the prior art has certain limitations, involves fewer categories of items and has a low degree of automation. It realizes data analysis based on the user's shopping behavior sample data, pushes more different categories of item information to users, improves the richness of the item information browsed by the user, and at the same time, improves the degree of automation of dynamically adjusting the target push information. Optionally, the information push module 530 is specifically used to:

[0123] The item information to be pushed is pushed and displayed to the target user in the item shopping cart pop-up layer, the item settlement matching relationship layer, the item purchase order detail page matching layer or the multiple content layer of the target information display platform.

[0124] The information push device provided in the embodiment of the present invention can execute the information push method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0125] Embodiment 6

[0126] Figure 8 A schematic diagram of the structure of a computer device provided in Embodiment 6 of the present invention. Figure 8 A block diagram of an exemplary computer device 12 suitable for use in implementing embodiments of the present invention is shown. Figure 8 The computer device 12 shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention. The computer device 12 can be any terminal device with computing capabilities, such as an intelligent controller and server, a mobile phone and other terminal devices.

[0127] like Figure 8 As shown, the computer device 12 is in the form of a general-purpose computing device. The components of the computer device 12 may include, but are not limited to: one or more processors or processing units 16, a system memory 28, and a bus 18 that connects various system components (including the system memory 28 and the processing unit 16).

[0128] Bus 18 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor or a local bus using any of a variety of bus architectures. By way of example, these architectures include, but are not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MAC) bus, an Enhanced ISA bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus.

[0129] The computer device 12 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by the computer device 12, including volatile and non-volatile media, removable and non-removable media.

[0130] The system memory 28 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. The computer device 12 may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the storage system 34 may be used to read and write non-removable, non-volatile magnetic media ( Figure 8 not shown, usually called a "hard drive"). Although Figure 8 Not shown in the figure, a disk drive for reading and writing to a removable non-volatile disk (e.g., a "floppy disk"), and an optical disk drive for reading and writing to a removable non-volatile optical disk (e.g., a CD-ROM, a DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to the bus 18 via one or more data medium interfaces. The system memory 28 may include at least one program product having a set (e.g., at least one) of program modules that are configured to perform the functions of the various embodiments of the present invention.

[0131] A program / utility 40 having a set (at least one) of program modules 42 may be stored, for example, in system memory 28, such program modules 42 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. Program modules 42 generally perform the functions and / or methods of the embodiments described herein.

[0132] The computer device 12 may also communicate with one or more external devices 14 (e.g., keyboards, pointing devices, displays 24, etc.), one or more devices that enable a user to interact with the computer device 12, and / or any device that enables the computer device 12 to communicate with one or more other computing devices (e.g., network cards, modems, etc.). Such communication may be performed via an input / output (I / O) interface 22. Furthermore, the computer device 12 may 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 20. As shown, the network adapter 20 communicates with the other modules of the computer device 12 via the bus 18. It should be understood that although Figure 8 Not shown, other hardware and / or software modules may be used in conjunction with computer device 12, including but not limited to microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0133] The processing unit 16 executes various functional applications and data processing by running the programs stored in the system memory 28, such as implementing the information push model training method or the information push method provided in the embodiment of the present invention.

[0134] Among them, the information push model training method includes:

[0135] Acquire user shopping behavior sample data, input the user shopping behavior sample data into a multi-task learning network, and input the cross-category item shopping behavior sample data in the user shopping behavior sample data into a cross-category expert network and a gating network of the cross-category expert network, wherein the cross-category expert network is used to extract feature information of cross-category items, and the gating network of the cross-category expert network is used to determine weight parameters of the output results of the cross-category expert network;

[0136] Generating input information of each task tower network in the multi-task learning network based on the output result of the cross-category expert network and the output result of the gating network of the cross-category expert network, and inputting the input information into each task tower network;

[0137] The output results of each task tower network are fused with the output results of the adaptive expert network in the multi-task learning network, and a first loss function value is determined according to the fusion result. Based on the first loss function value, the parameters of the multi-task learning network, the cross-category expert network and the gating network of the cross-category expert network are updated to train a target information push model.

[0138] The information push method is characterized in that the method comprises:

[0139] Obtaining a behavior data sample of the target user interacting with item information on the target information display platform, wherein the behavior data sample includes a behavior data sample of the target user interacting with cross-category item information, and the cross-category item is an item that the target user has not clicked or purchased within a preset historical time period;

[0140] Inputting the behavior data sample into a target information push model including a multi-task learning network, a cross-category expert network, and a gating network of a cross-category expert network obtained by training the information push model training method described in any embodiment;

[0141] The item information to be pushed is determined based on the output result of the target information push model, and the item information to be pushed is pushed and displayed to the target user.

[0142] Embodiment 7

[0143] This embodiment seven provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, it implements the information push model training method or information push method provided in any embodiment of the present invention.

[0144] Among them, the information push model training method includes:

[0145] Acquire user shopping behavior sample data, input the user shopping behavior sample data into a multi-task learning network, and input the cross-category item shopping behavior sample data in the user shopping behavior sample data into a cross-category expert network and a gating network of the cross-category expert network, wherein the cross-category expert network is used to extract feature information of cross-category items, and the gating network of the cross-category expert network is used to determine weight parameters of the output results of the cross-category expert network;

[0146] Generating input information of each task tower network in the multi-task learning network based on the output result of the cross-category expert network and the output result of the gating network of the cross-category expert network, and inputting the input information into each task tower network;

[0147] The output results of each task tower network are fused with the output results of the adaptive expert network in the multi-task learning network, and a first loss function value is determined according to the fusion result. Based on the first loss function value, the parameters of the multi-task learning network, the cross-category expert network and the gating network of the cross-category expert network are updated to train a target information push model.

[0148] The information push method is characterized in that the method comprises:

[0149] Obtaining a behavior data sample of the target user interacting with item information on the target information display platform, wherein the behavior data sample includes a behavior data sample of the target user interacting with cross-category item information, and the cross-category item is an item that the target user has not clicked or purchased within a preset historical time period;

[0150] Inputting the behavior data sample into a target information push model including a multi-task learning network, a cross-category expert network, and a gating network of a cross-category expert network obtained by training the information push model training method described in any embodiment;

[0151] The item information to be pushed is determined based on the output result of the target information push model, and the item information to be pushed is pushed and displayed to the target user.

[0152] The computer storage medium of the embodiment of the present invention can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to: an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer 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 of the above. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, device or device or used in combination with it.

[0153] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, which carry computer-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 of the above. Computer-readable signal media may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0154] The program code embodied on the computer 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.

[0155] Computer program code for performing the operation of the present invention may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0156] It should be understood by those skilled in the art that the modules or steps of the present invention described above can be implemented by a general-purpose computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, optionally, they can be implemented by a program code executable by a computer device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0157] Note that the above are only preferred embodiments of the present invention and the technical principles used. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and that various obvious changes, readjustments and substitutions can be made by those skilled in the art without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in more detail through the above embodiments, the present invention is not limited to the above embodiments, and may include more other equivalent embodiments without departing from the concept of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims

1. A method for training an information push model, characterized in that: The method comprises: Acquire user shopping behavior sample data, input the user shopping behavior sample data into a multi-task learning network, and input the cross-category item shopping behavior sample data in the user shopping behavior sample data into a cross-category expert network and a gating network of the cross-category expert network, wherein the cross-category expert network is used to extract feature information of cross-category items, and the gating network of the cross-category expert network is used to determine weight parameters of the output results of the cross-category expert network; Generating input information of each task tower network in the multi-task learning network based on the output result of the cross-category expert network and the output result of the gating network of the cross-category expert network, and inputting the input information into each task tower network; The output results of each task tower network are fused with the output results of the adaptive expert network in the multi-task learning network, and a first loss function value is determined according to the fusion result. Based on the first loss function value, the parameters of the multi-task learning network, the cross-category expert network and the gating network of the cross-category expert network are updated to train a target information push model.

2. The method according to claim 1, characterized in that The step of generating input information of each task tower network in the multi-task learning network based on the output result of the cross-category expert network and the output result of the gating network of the cross-category expert network includes: Determine the first input information of each task tower network based on the output result of the cross-category expert network and the output result of the gating network of the cross-category expert network; The output results of the gating network of each task in the multi-task learning network and the output results of the expert network corresponding to each task are calculated according to a preset operation rule to obtain the second input information of the tower network of each task; The first input information and the second input information are superimposed to obtain input information of each task tower network in the multi-task learning network.

3. The method according to claim 2, characterized in that An expert network that uses the output result as only part of the input information of one task tower network is called a private expert network, and an expert network that uses the output result as part of the input information of multiple task tower networks is called a public expert network. The updating of the parameters of the multi-task learning network, the cross-category expert network, and the gating network based on the first loss function value includes: updating parameters of the gating network and the adaptive expert network in the multi-task learning network and the gating network of each task according to the first loss function; For the multi-task learning network and the private expert network in the cross-category expert network, updating parameters according to the loss function value of the task tower network that receives the private expert information input; For the public expert network in the multi-task learning network and the cross-category expert network, in two consecutive processes of updating the parameters of the public expert network, the parameters are updated according to the loss function values ​​of different task tower networks; Wherein, the first loss function is determined according to the loss function of each task tower network.

4. The method according to claim 2, characterized in that: The gating network of each task in the multi-task learning network is provided with a layer of routing screening network for screening the output results of the expert network corresponding to each task, and the output results of the gating network of each task in the multi-task learning network and the output results of the expert network corresponding to each task are calculated according to preset operation rules to obtain the second input information of the gating network of each task, including: Compare the output results of the gating network of each task with the gating screening value set by the routing screening network, and set the values ​​of the output results of the gating network of each task that are less than the gating screening value to 0, so as to update the output results of the gating network of each task; The output results of the updated gated networks of the tasks are respectively compared with the output results of the expert networks corresponding to the tasks according to the preset operation rules to obtain partial input information of the tower networks of the tasks.

5. The method according to claim 1, characterized in that Before inputting the user shopping behavior sample data into the multi-task learning network, the method further includes: The representation vector of each of the user shopping behavior sample data is multiplied by the category vector of the category to which the item corresponding to the sample data belongs.

6. The method according to claim 1, characterized in that The obtaining of user shopping behavior sample data includes: Recall the user's shopping behavior data according to a preset sample recall algorithm to obtain an initial recall data sample; Updating sample weight values ​​for the cross-category item shopping behavior samples in the initial recalled data samples to obtain updated recalled data samples; The updated recall data samples are input into a preset neural network for preliminary screening of samples, and sample screening is performed to obtain the user shopping behavior sample data, wherein, in the updated recall data samples, the sample label weight value of the cross-category item shopping behavior sample is greater than the sample label weight value of the non-cross-category item shopping behavior sample.

7. The method according to claim 5, characterized in that The recall process of the cross-category item samples includes: Conduct cluster analysis on users and build similar user groups based on the cluster analysis results; Push and display items in the similar user group that the current user has not browsed or purchased within a preset historical time period and that other users except the current user have clicked or purchased as cross-category items of the current user; Among the cross-category items pushed for display, the items clicked or purchased by the current user are used as positive samples of the cross-category item samples, and the items not clicked or purchased by the current user are used as negative samples of the cross-category item samples.

8. The method according to any one of claims 1 to 7, characterized in that: The user shopping behavior sample data includes item discount information, item discount sensitivity and user discount sensitivity characteristics.

9. An information push method, characterized in that: The method comprises: Acquire user shopping behavior sample data of target users interacting with item information on a target information display platform, wherein the user shopping behavior sample data includes user shopping behavior sample data of the target users interacting with cross-category item information, and the cross-category items are items that the target users have not clicked or purchased within a preset historical time period; Inputting the behavior data sample into a target information push model including a multi-task learning network, a cross-category expert network and a gating network of a cross-category expert network obtained by training the information push model training method according to any one of claims 1 to 8; The item information to be pushed is determined based on the output result of the target information push model, and the item information to be pushed is pushed and displayed to the target user.

10. The method according to claim 9, characterized in that The pushing and displaying of the information of the item to be pushed to the target user includes: The item information to be pushed is pushed and displayed to the target user in the item shopping cart pop-up layer, the item settlement matching relationship layer, the item purchase order detail page matching layer or the multiple content layer of the target information display platform.

11. An information push model training device, characterized in that: The device comprises: A sample input module, used to obtain user shopping behavior sample data, input the user shopping behavior sample data into a multi-task learning network, and input the cross-category item shopping behavior sample data in the user shopping behavior sample data into a cross-category expert network and a gating network of the cross-category expert network, wherein the cross-category expert network is used to extract feature information of cross-category items, and the gating network of the cross-category expert network is used to determine weight parameters of the output results of the cross-category expert network; A sample operation module, used for generating input information of each task tower network in the multi-task learning network based on the output result of the cross-category expert network and the output result of the gating network of the cross-category expert network, and inputting the input information into each task tower network; A model training module is used to fuse the output results of the task tower networks with the output results of the adaptive expert network in the multi-task learning network, and determine the first loss function value based on the fusion result, and update the parameters of the multi-task learning network, the cross-category expert network and the gating network of the cross-category expert network based on the first loss function value to train and obtain a target information push model.

12. An information push device, characterized in that: The device comprises: A data acquisition module is used to acquire user shopping behavior sample data of target users interacting with item information on a target information display platform, wherein the user shopping behavior sample data includes user shopping behavior sample data of the target users interacting with cross-category item information, and the cross-category items are items that the target users have not clicked or purchased within a preset historical time period; A data analysis module, used for inputting the behavior data samples into a target information push model including a multi-task learning network, a cross-category expert network and a gating network of a cross-category expert network obtained by training the information push model training method according to any one of claims 1 to 8; The information push module is used to determine the item information to be pushed based on the output result of the target information push model, and push and display the item information to be pushed to the target user.

13. A computer device, characterized in that: The computer device comprises: one or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the information push model training method or the information push method as described in any one of claims 1-10.

14. 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 information push model training method or the information push method as described in any one of claims 1 to 10.

Images