Model training method and device, information generation method and device, equipment and medium

Abstract

The embodiment of the invention discloses a model training method and device, an information generation method and device, equipment and a medium. A specific embodiment of the method comprises the steps of obtaining training data; generating portrait feature information corresponding to the first user portrait and behavior feature information corresponding to the first user value behavior information sequence by using the initial model input layer; executing a first generation step: generating at least one piece of interest feature information according to the portrait feature information and the behavior feature information; generating category interest feature information and category value gear feature information according to the at least one piece of interest feature information; generating model loss information; and in response to determining that the model loss information is smaller than a preset value, determining the initial category value gear information generation model as a category value gear information generation model. The implementation mode is related to artificial intelligence, the model is generated by using the category value gear information, and the corresponding category value gear information can be accurately generated under the category without the corresponding value behavior data of the user.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of computer technology, and in particular to a model training method, an information generation method, an apparatus, a device, and a medium. Background Art

[0002] At present, in the Internet-related fields, the category value level information corresponding to the user is a way to measure the user's value transfer ability. The usual method for generating category value level information is to use a conventional neural network model based on the target user's value transfer data set to determine the target user's category value level information under the target category.

[0003] However, the inventors have found that when the above method is used to generate category value level information, the following technical problems often occur:

[0004] For users who do not have a value circulation data set or whose value circulation data set has a small amount of data, conventional neural network models cannot effectively extract features and often lack category generalization capabilities, resulting in the generated category value level information being inaccurate.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the inventive concept and therefore it may contain information that does not form the prior art that is already known in this country to a person of ordinary skill in the art. Summary of the invention

[0006] The content of this disclosure is used to introduce concepts in a brief form, which will be described in detail in the detailed implementation section below. The content of this disclosure is not intended to identify the key features or essential features of the technical solution claimed for protection, nor is it intended to limit the scope of the technical solution claimed for protection.

[0007] Some embodiments of the present disclosure propose model training methods, information generation methods, devices, equipment and media to solve the technical problems mentioned in the above background technology section.

[0008] In a first aspect, some embodiments of the present disclosure provide a model training method, comprising: acquiring training data, wherein the training data comprises: a first user portrait for a first target user, a first user value behavior information sequence and a category tag; utilizing an initial model input layer to generate portrait feature information corresponding to the first user portrait and behavior feature information corresponding to the first user value behavior information sequence, wherein an initial category value level information generation model comprises: the initial model input layer, an initial value level information output layer and a loss function layer; utilizing the initial value level information output layer to perform the following first generation step: generating at least one interest feature information for at least one user interest dimension based on the portrait feature information and the behavior feature information; generating category interest feature information and category value level feature information for the category tag based on the at least one interest feature information; utilizing the loss function layer to generate model loss information for the category interest feature information and the category value level feature information; in response to determining that the model loss information indicates that model training is complete, determining the initial category value level information generation model as the category value level information generation model.

[0009] Optionally, the initial model input layer includes: an initial compression excitation layer; and the use of the initial model input layer to generate the portrait feature information corresponding to the first user portrait and the behavior feature information corresponding to the first user value behavior information sequence includes: using the initial model input layer to perform the following second generation step: generating the first initial portrait feature information for the first user portrait, and generating the first initial behavior feature information for the first user value behavior information sequence; performing low-dimensional embedding processing on the first initial portrait feature information and the first initial behavior feature information, respectively, to generate first embedded feature information and second embedded feature information; inputting the first embedded feature information into the initial compression excitation layer to generate the main portrait feature information; generating the portrait feature information and the behavior feature information according to the main portrait feature information and the second embedded feature information.

[0010] Optionally, the above-mentioned portrait feature information and the above-mentioned behavior feature information are generated according to the above-mentioned portrait main feature information and the above-mentioned second embedded feature information, including: generating second initial behavior feature information according to the above-mentioned second embedded feature information; generating second initial portrait feature information according to the above-mentioned portrait main feature information and the above-mentioned first embedded feature information; generating portrait feature information based on behavior perception and behavior feature information based on portrait perception according to the above-mentioned second initial behavior feature information and the above-mentioned second initial portrait feature information, as the above-mentioned portrait feature information and the above-mentioned behavior feature information, respectively.

[0011] Optionally, the above-mentioned generating the portrait feature information based on behavior perception and the behavior feature information based on portrait perception according to the above-mentioned second initial behavior feature information and the above-mentioned second initial portrait feature information includes: obtaining the item attribute information sequence corresponding to the historical interaction item information sequence and the historical interaction item information sequence, the item attribute information including: at least one item category and item value level information; generating the third embedded feature information for the above-mentioned item attribute information sequence and the above-mentioned historical interaction item information sequence; generating the fused embedded feature information according to the above-mentioned second initial behavior feature information, the above-mentioned second initial portrait feature information and the above-mentioned third embedded feature information; generating the portrait feature information and the behavior feature information for the above-mentioned fused embedded feature information by using the self-attention mechanism model.

[0012] Optionally, the at least one user interest dimension includes: a population interest dimension; and generating at least one interest feature information for at least one user interest dimension based on the portrait feature information and the behavior feature information, including: performing a linear transformation on the second initial portrait feature information and the portrait feature information to generate population-level interest feature information corresponding to the population interest dimension as interest feature information.

[0013] Optionally, the at least one user interest dimension also includes: a target level category interest dimension; and the generating of at least one interest feature information for at least one user interest dimension based on the portrait feature information and the behavior feature information, including: obtaining a category information set corresponding to the target level category; generating a category vector corresponding to each category information in the category information set to obtain a category vector set; and generating target level category interest feature information corresponding to the target level category interest dimension based on the category vector set and the behavior feature information as interest feature information.

[0014] Optionally, the at least one user interest dimension also includes: a user-level latent interest dimension; and the generating of at least one interest feature information for at least one user interest dimension based on the portrait feature information and the behavior feature information includes: generating user-level latent interest feature information for the behavior feature information as interest feature information using a capsule network based on the portrait feature information, wherein the user-level latent interest feature information is feature information corresponding to the user-level interest dimension, and the user-level latent interest feature information includes: latent interest feature information under at least one interest representation.

[0015] Optionally, the above-mentioned generating category interest feature information and category value level feature information for the above-mentioned category tag based on the above-mentioned at least one interest feature information includes: obtaining the category representation vector and the category value level vector corresponding to the above-mentioned category tag; splicing feature information of the above-mentioned population-level interest feature information, the above-mentioned target-level category interest feature information and the above-mentioned user-level latent interest feature information to generate spliced ​​feature information; determining the category interest feature information for the above-mentioned category tag whose similarity with the above-mentioned category representation vector in the above-mentioned spliced ​​feature information satisfies a preset similarity condition; and generating category value level feature information representing the above-mentioned first target user under the above-mentioned category tag based on the above-mentioned category interest feature information and the above-mentioned category value level vector.

[0016] Optionally, the above-mentioned training data also includes: actual category value level information; and the above-mentioned use of the above-mentioned loss function layer to generate model loss information for the above-mentioned category interest feature information and the above-mentioned category value level feature information, including: using the above-mentioned loss function layer to perform the following third generation step: generating category loss information between the above-mentioned category representation vector and the above-mentioned category interest feature information; generating value level loss information between the above-mentioned category value level feature information and the above-mentioned actual category value level information; determining the predetermined level category corresponding to the above-mentioned category tag as the predetermined level category information; determining the predetermined level category feature information corresponding to the above-mentioned predetermined level category information from the above-mentioned splicing feature information; generating interest loss information for the above-mentioned category interest feature information and the above-mentioned predetermined level category feature information; generating the above-mentioned model loss information according to the above-mentioned category loss information, the above-mentioned target level category feature information and the above-mentioned interest loss information.

[0017] In a second aspect, some embodiments of the present disclosure provide a model training device, including: a first acquisition unit, configured to acquire training data, wherein the training data includes: a first user portrait for a first target user, a first user value behavior information sequence and a category label; a first generation unit, configured to use an initial model input layer to generate portrait feature information corresponding to the first user portrait and behavior feature information corresponding to the first user value behavior information sequence, wherein the initial category value level information generation model includes: the initial model input layer, the initial value level information output layer and the loss function layer; an execution unit, configured to use the initial value level information The output layer executes the following first generation step: based on the above-mentioned portrait feature information and the above-mentioned behavior feature information, at least one interest feature information for at least one user interest dimension is generated; based on the above-mentioned at least one interest feature information, category interest feature information and category value level feature information for the above-mentioned category tag are generated; the second generation unit is configured to use the above-mentioned loss function layer to generate model loss information for the above-mentioned category interest feature information and the above-mentioned category value level feature information; the determination unit is configured to determine the above-mentioned initial category value level information generation model as the category value level information generation model in response to determining that the above-mentioned model loss information represents that the model training is completed.

[0018] Optionally, the above-mentioned initial model input layer includes: an initial compression excitation layer; and the first generation unit can be configured to: use the above-mentioned initial model input layer to perform the following second generation step: generate the first initial portrait feature information for the above-mentioned first user portrait, and generate the first initial behavior feature information for the above-mentioned first user value behavior information sequence; perform low-dimensional embedding processing on the above-mentioned first initial portrait feature information and the above-mentioned first initial behavior feature information respectively to generate first embedded feature information and second embedded feature information; input the above-mentioned first embedded feature information into the above-mentioned initial compression excitation layer to generate the main feature information of the portrait; generate the above-mentioned portrait feature information and the above-mentioned behavior feature information according to the above-mentioned main feature information of the portrait and the above-mentioned second embedded feature information.

[0019] Optionally, the first generation unit can be configured to: generate second initial behavior feature information based on the above-mentioned second embedded feature information; generate second initial portrait feature information based on the above-mentioned portrait main feature information and the above-mentioned first embedded feature information; generate portrait feature information based on behavior perception and behavior feature information based on portrait perception based on the above-mentioned second initial behavior feature information and the above-mentioned second initial portrait feature information, as the above-mentioned portrait feature information and the above-mentioned behavior feature information, respectively.

[0020] Optionally, the first generation unit can be configured to: obtain a historical interaction item information sequence and an item attribute information sequence corresponding to the historical interaction item information sequence, the item attribute information including: at least one item category and item value level information; generate third embedded feature information for the above-mentioned item attribute information sequence and the above-mentioned historical interaction item information sequence; generate fused embedded feature information based on the above-mentioned second initial behavior feature information, the above-mentioned second initial portrait feature information and the above-mentioned third embedded feature information; and use a self-attention mechanism model to generate portrait feature information and behavior feature information for the above-mentioned fused embedded feature information.

[0021] Optionally, the at least one user interest dimension includes: a population interest dimension; and the execution unit can be configured to: perform a linear transformation on the second initial portrait feature information and the portrait feature information to generate population-level interest feature information corresponding to the population interest dimension as interest feature information.

[0022] Optionally, the at least one user interest dimension also includes: a target level category interest dimension; and the execution unit can be configured to: obtain a category information set corresponding to the target level category; generate a category vector corresponding to each category information in the above category information set to obtain a category vector set; and generate target level category interest feature information corresponding to the above target level category interest dimension according to the above category vector set and the above behavioral feature information as interest feature information.

[0023] Optionally, the at least one user interest dimension also includes: a user-level latent interest dimension; and the execution unit can be configured to: generate user-level latent interest feature information for the above-mentioned behavioral feature information based on the above-mentioned portrait feature information using a capsule network as interest feature information, wherein the user-level latent interest feature information is feature information corresponding to the above-mentioned user-level interest dimension, and the user-level latent interest feature information includes: latent interest feature information under at least one interest representation.

[0024] Optionally, the execution unit can be configured to: obtain the category representation vector and the category value level vector corresponding to the above-mentioned category tag; perform feature information splicing on the above-mentioned population-level interest feature information, the above-mentioned target-level category interest feature information and the above-mentioned user-level latent interest feature information to generate spliced ​​feature information; determine the category interest feature information for the above-mentioned category tag whose similarity with the above-mentioned category representation vector in the above-mentioned spliced ​​feature information meets a preset similarity condition; and generate category value level feature information representing the above-mentioned first target user under the above-mentioned category tag based on the above-mentioned category interest feature information and the above-mentioned category value level vector.

[0025] Optionally, the above-mentioned training data also includes: actual category value level information; and the second generation unit can be configured to: use the above-mentioned loss function layer to perform the following third generation step: generate category loss information between the above-mentioned category representation vector and the above-mentioned category interest feature information; generate value level loss information between the above-mentioned category value level feature information and the above-mentioned actual category value level information; determine the predetermined level category corresponding to the above-mentioned category tag as the predetermined level category information; determine the predetermined level category feature information corresponding to the above-mentioned predetermined level category information from the above-mentioned splicing feature information; generate interest loss information for the above-mentioned category interest feature information and the above-mentioned predetermined level category feature information; generate the above-mentioned model loss information based on the above-mentioned category loss information, the above-mentioned target level category feature information and the above-mentioned interest loss information.

[0026] On the third aspect, some embodiments of the present disclosure provide an information generation method, including: obtaining target category information, a second user portrait for a second target user, and a second user value behavior information sequence, wherein the target category information is the category information of the category under the user's valueless behavior; inputting the target category information, the second user portrait and the second user value behavior information sequence into a pre-trained category value level information generation model to generate category value level information, wherein the category value level information represents the value level status of the second target user under the target category information, and the category value level information generation model is generated based on a model training method.

[0027] In a fourth aspect, some embodiments of the present disclosure provide an information generating device, comprising: a second acquisition unit, configured to acquire target category information, a second user portrait for a second target user, and a second user value behavior information sequence, wherein the target category information is category information of a category under user's valueless behavior; a third generation unit, configured to input the target category information, the second user portrait and the second user value behavior information sequence into a pre-trained category value level information generation model to generate category value level information, wherein the category value level information characterizes the value level status of the second target user under the target category information, and the category value level information generation model is generated based on a model training method.

[0028] In a fifth aspect, some embodiments of the present disclosure provide an electronic device comprising: one or more processors; a storage device on which one or more programs are stored, and when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any one of the implementation methods in the first and third aspects.

[0029] In a sixth aspect, some embodiments of the present disclosure provide a computer-readable medium having a computer program stored thereon, wherein when the program is executed by a processor, the method described in any one of the implementation modes of the first aspect and the third aspect is implemented.

[0030] In a seventh aspect, some embodiments of the present disclosure provide a computer program product, including a computer program, which, when executed by a processor, implements the method described in any one of the implementation modes of the first and third aspects above.

[0031] The above-mentioned embodiments of the present disclosure have the following beneficial effects: through the model training method of some embodiments of the present disclosure, using the category value level information generation model, the corresponding category value level information can be accurately generated under the category where the user has no corresponding value behavior data. Specifically, the reason why the relevant category value level information is not accurate enough is that for users who do not have a value circulation data set or have a small amount of corresponding data in the value circulation data set, the conventional neural network model cannot effectively extract features and often lacks category generalization ability, resulting in the generated category value level information being not accurate enough. Based on this, the model training method of some embodiments of the present disclosure first obtains training data. Among them, the above-mentioned training data includes: a first user portrait for a first target user, a first user value behavior information sequence, and a category tag. Here, through the first user portrait, the initial category value level information generation model can subsequently learn more user features in the first user portrait that are associated with the first target user. Through the first user value behavior information sequence, the initial category value level information generation model can subsequently learn more value behavior features corresponding to the first user value behavior information sequence. The obtained category tag is used for the disadvantage of subsequent loss of information. Then, the initial model input layer is used to generate the portrait feature information corresponding to the first user portrait and the behavior feature information corresponding to the first user value behavior information sequence. Among them, the initial category value level information generation model includes: the initial model input layer, the initial value level information output layer and the loss function layer. The obtained portrait feature information can comprehensively characterize the main feature information of the portrait of the first target user in the first user portrait. The behavior feature information can comprehensively characterize the main feature information of the value behavior of the first target user in the first user value behavior information sequence. Here, the obtained portrait feature information and behavior feature information can allow the initial value level information output layer to learn more important and targeted feature information, not learn useless related feature information, and can generate more accurate category interest feature information and category value level feature information. Then, using the above-mentioned initial value level information output layer, the following first generation step is performed: the first step, according to the above-mentioned portrait feature information and the above-mentioned behavior feature information, the main feature information of the user and the main value behavior feature information of the user in the learned portrait can be used to accurately generate at least one interest feature information for at least one user interest dimension. Here, the corresponding interest feature information is determined from multiple user interest dimensions, which can fully reflect the user's corresponding category interest information from various fine-grained levels. In the second step, based on the above at least one interest feature information, the category interest feature information and category value level feature information for the above category tag can be accurately generated. The category interest feature information can represent the category interest feature information of the first target user under the category tag. The category value level feature information can represent the value level information of the first target user under the category tag.Furthermore, by using the above-mentioned loss function layer, the model loss information for the above-mentioned category interest feature information and the above-mentioned category value level feature information can be accurately generated, so as to be used for determining whether the subsequent initial category value level information generation model has been trained. Finally, in response to determining that the above-mentioned model loss information represents that the model training is completed, the above-mentioned initial category value level information generation model is determined as the category value level information generation model. In summary, through the initial model input layer, the portrait feature information and behavior feature information can be accurately generated, so that the subsequent initial value level information output layer can learn more refined user important feature information. Then, through the initial value level information output layer, under the premise that at least one interest feature information corresponding to at least one user interest dimension can be accurately generated, the category interest feature information and the category value level feature information are accurately generated by matching with the category tag. Based on this, the loss function layer is finally used to generate the corresponding model loss information to determine whether the model training is completed. Therefore, the trained category value level information generation model can be used to accurately generate the corresponding category value level information under the category where the user has no corresponding value behavior data. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The above and other features, advantages and aspects of the embodiments of the present disclosure will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. Throughout the accompanying drawings, the same or similar reference numerals represent the same or similar elements. It should be understood that the drawings are schematic and that components and elements are not necessarily drawn to scale.

[0033] Figure 1 is a schematic diagram of an application scenario of a model training method according to some embodiments of the present disclosure;

[0034] Figure 2 is a flowchart of some embodiments of the model training method according to the present disclosure;

[0035] Figure 3 is a flowchart of other embodiments of the model training method according to the present disclosure;

[0036] Figure 4 is a flow chart of some embodiments of the information generation method according to the present disclosure;

[0037] Figure 5 is a schematic diagram of the structure of some embodiments of the model training device according to the present disclosure;

[0038] Figure 6 is a schematic diagram of the structure of some embodiments of the information generating device according to the present disclosure;

[0039] Figure 7 It is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION

[0040] Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be construed as being limited to the embodiments set forth herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0041] It should also be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present disclosure can be combined with each other.

[0042] It should be noted that the concepts such as "first" and "second" mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0043] It should be noted that the modifications of "one" and "plurality" mentioned in the present disclosure are illustrative rather than restrictive, and those skilled in the art should understand that unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0044] The names of the messages or information exchanged between multiple devices in the embodiments of the present disclosure are only used for illustrative purposes and are not used to limit the scope of these messages or information.

[0045] With regard to the collection, storage, and use of user personal information (such as user portraits and user value behavior information) involved in this disclosure, before performing the corresponding operations, the relevant organizations or individuals shall fulfill their obligations, including conducting personal information security impact assessments, fulfilling the obligation to inform the personal information subject, and obtaining the authorization and consent of the personal information subject in advance.

[0046] The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0047] Figure 1 It is a schematic diagram of an application scenario of the model training method according to some embodiments of the present disclosure.

[0048] exist Figure 1In the application scenario, the electronic device 101 can obtain training data 102. Among them, the above training data 102 includes: a first user portrait 1021 for the first target user, a first user value behavior information sequence 1022 and a category label 103. In this application scenario, the first user value behavior information sequence 1022 includes: a user purchase behavior sequence 10221 and a user add-on purchase behavior sequence 10222. The category label 103 can be "upper clothing". Then, the electronic device 101 can use the initial model input layer 105 to generate the portrait feature information 108 corresponding to the above first user portrait 1021 and the behavior feature information 109 corresponding to the above first user value behavior information sequence 1022. Among them, the initial category value level information generation model 104 includes: the above initial model input layer 105, the initial value level information output layer 106 and the loss function layer 107. Next, the electronic device 101 can use the initial value level information output layer 106 to perform the following first generation step: First, based on the portrait feature information 108 and the behavior feature information 109, generate at least one interest feature information 110 for at least one user interest dimension 102. In this application scenario, the at least one user interest dimension 102 includes: user interest 1021 based on the primary category, user interest 1022 based on the secondary category, and user interest 1024 based on the tertiary category. The at least one interest feature information 110 includes: first interest feature information 1101 corresponding to the user interest 1021 based on the primary category, second interest feature information 1102 corresponding to the user interest 1022 based on the secondary category, and third interest feature information 1103 corresponding to the user interest 1023 based on the tertiary category. Second, based on the at least one interest feature information 110, generate category interest feature information 111 and category value level feature information 112 for the category tag 103. Furthermore, the electronic device 101 can use the loss function layer 107 to generate the model loss information 114 for the category interest feature information 111 and the category value level feature information 112. Finally, in response to determining that the model loss information 114 indicates that the model training is complete, the electronic device 101 can determine the initial category value level information generation model 104 as the category value level information generation model 113.

[0049] It should be noted that the electronic device 101 can be hardware or software. When the electronic device is hardware, it can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or it can be implemented as a single server or a single terminal device. When the electronic device is embodied as software, it can be installed in the hardware devices listed above. It can be implemented as multiple software or software modules for providing distributed services, for example, or it can be implemented as a single software or software module. No specific limitation is made here.

[0050] It should be understood that Figure 1 The number of electronic devices in the embodiment is only for illustration. Any number of electronic devices may be provided according to implementation requirements.

[0051] Continue to refer Figure 2 , shows a process 200 of some embodiments of the model training method according to the present disclosure. The model training method comprises the following steps:

[0052] Step 201, obtaining training data.

[0053] In some embodiments, the execution entity of the above model training method (for example Figure 1 The electronic device 101 shown can obtain training data through a wired connection or a wireless connection. The training data includes: a first user portrait for a first target user, a first user value behavior information sequence and a category tag. The training data is data used for subsequent training of an initial category value level information generation model. The first target user may be a user involved in the training data. The first user portrait may be a user portrait corresponding to the first target user. For example, the first user portrait may include but is not limited to at least one of the following: gender, age, region, and device. The first user value behavior information sequence may be a user behavior information sequence corresponding to the first target user within a target historical time period. The first user value behavior information sequence may be an information sequence under which the first target user performs a value behavior. In practice, for e-commerce scenarios, the value behavior may be but is not limited to at least one of the following: add-on behavior, purchase behavior. For example, the first user value behavior information sequence may be an add-on behavior information sequence of the first target user in the past 1 year and a purchase behavior information sequence in the past 3 years. The target historical time period may be the past 3 years. The category tag may be an item category of the value behavior items of the first target user in the next 7 days. The value behavior item may be a purchase item.

[0054] Step 202, using the initial model input layer, generates the portrait feature information corresponding to the first user portrait and the behavior feature information corresponding to the first user value behavior information sequence.

[0055] In some embodiments, the execution subject may use the initial model input layer to generate the portrait feature information corresponding to the first user portrait and the behavior feature information corresponding to the first user value behavior information sequence. The initial model input layer may be a model input layer that has not been trained yet. The model input layer may be a data input layer. For example, the model input layer may be a feature information extraction layer. For example, the feature extraction layer may be a multi-layer convolutional neural network. The portrait feature information may be feature information that characterizes the portrait features in the first user portrait. Specifically, the portrait feature information may also characterize feature information of the main portrait features in the first user portrait. The main portrait features may be features that are closely related to the subsequent category value level features. The behavior feature information may be feature information that characterizes the user value features in the first user value behavior information sequence. Specifically, the behavior feature information may also characterize feature information of the main user value features in the first user value behavior information sequence. The main user value features may be features that are closely related to the subsequent category value level features. In practice, both the portrait feature information and the behavior feature information may be information in the form of matrix vectors. The initial category value level information generation model includes: the initial model input layer, the initial value level information output layer, and the loss function layer. The initial category value level information generation model may be a category value level information generation model that has not yet been trained. The category value level information generation model may be a neural network model that generates category value level information. The category value level information may be the value level information corresponding to the user under the category. In practice, for the e-commerce field, the value level information may characterize the user's commodity purchasing power. For example, for the preset category, the category value level information corresponding to the user is {first level, second level, third level, fourth level, fifth level, sixth level, seventh level, eighth level, ninth level}. The commodity purchasing power corresponding to the first level is greater than the commodity purchasing power corresponding to the second level. The commodity purchasing power corresponding to the second level is greater than the commodity purchasing power corresponding to the third level. And so on, no further description is given. The initial value level information output layer may be a value level information output layer that has not yet been trained. In practice, the value level information output layer may include: an interest feature information generation layer, a category interest feature information generation layer, and a category value level feature information generation layer. The interest feature information generation layer may be a network layer for generating interest feature information. The category interest feature information generation layer may be a network layer for generating category interest feature information. The category value level feature information generation layer may be a network layer for generating category value level feature information. In practice, the interest feature information generation layer may be a multi-head attention mechanism model. The category interest feature information generation layer may be a convolutional layer. The category value level feature information generation layer may be a convolutional layer. The loss function layer may include at least one loss function.

[0056] In some optional implementations of some embodiments, the initial model input layer includes: an initial compression excitation layer. The initial compression excitation layer may be a compression excitation network layer whose parameters have not been trained. The compression excitation network layer may be a feature screening layer. The compression excitation network layer may be a neural network model that screens out key features from input features.

[0057] Optionally, the above-mentioned use of the initial model input layer to generate the portrait feature information corresponding to the above-mentioned first user portrait and the behavior feature information corresponding to the above-mentioned first user value behavior information sequence may include the following steps:

[0058] The execution subject may use the initial model input layer to perform the following second generation step:

[0059] Sub-step 1, generating the first initial portrait feature information for the above-mentioned first user portrait, and generating the first initial behavior feature information for the above-mentioned first user value behavior information sequence. Among them, the first initial portrait feature information can be feature information obtained by integrating the various portrait features in the first user portrait. The first initial portrait feature information can represent the feature semantics corresponding to the various portrait features in the first user portrait. Similarly, the first initial behavior feature information can be feature information obtained by integrating the various user value behavior features in the first user value behavior information sequence. The first initial portrait feature information can represent the feature semantics corresponding to the various user value behavior features in the first user value behavior information sequence. Both the first initial portrait feature information and the first initial behavior feature information can be feature information in the form of matrix vectors.

[0060] As an example, the above-mentioned execution entity can generate the first initial portrait feature information and the first initial behavior feature information through onehot encoding.

[0061] Sub-step 2, respectively performing low-dimensional embedding processing on the above-mentioned first initial portrait feature information and the above-mentioned first initial behavior feature information to generate first embedded feature information and second embedded feature information. The first embedded feature information and the second embedded feature information can be feature information in the form of matrix vectors. The dimension of the matrix vector corresponding to the first embedded feature information is smaller than the dimension of the matrix vector corresponding to the first initial portrait feature information. The dimension of the matrix vector corresponding to the second embedded feature information is smaller than the dimension of the matrix vector corresponding to the first initial behavior feature information.

[0062] As an example, the execution entity may utilize a low-dimensional embedding layer to perform low-dimensional embedding processing on the first initial portrait feature information and the first initial behavior feature information, respectively, to generate first embedded feature information and second embedded feature information.

[0063] Sub-step 3: input the first embedded feature information into the initial compression incentive layer to generate main feature information of the portrait, wherein the main feature information of the portrait may be feature information of the main portrait features in the first user portrait.

[0064] As an example, first, the above-mentioned execution subject can perform average pooling processing on the first embedded feature information to generate average pooling information. Then, the above-mentioned average pooling information is multiplied by the first parameter to obtain the first multiplication information. Then, the first multiplication information is added to the second parameter to obtain the first addition information. Furthermore, the first addition information is input into the RELU activation function to generate the first output result. Then, the first output result is multiplied by the third parameter to obtain the second multiplication result. Secondly, the second multiplication result is added to the fourth parameter to obtain the second addition information. Finally, the second addition information is input into the Sigmoid activation function to output the main feature information of the portrait. Among them, the first parameter, the second parameter, the third parameter and the fourth parameter can be parameters in the initial compression excitation layer, which are continuously updated as the model is trained.

[0065] Sub-step 4, generating the above-mentioned portrait feature information and the above-mentioned behavior feature information based on the above-mentioned portrait main feature information and the above-mentioned second embedded feature information.

[0066] As an example, the above-mentioned execution subject may determine the main feature information of the portrait as the portrait feature information, and determine the second embedded feature information as the behavior feature information.

[0067] Optionally, the generating of the portrait feature information and the behavior feature information according to the portrait main feature information and the second embedded feature information may include the following steps:

[0068] In the first step, second initial behavior feature information is generated according to the second embedded feature information, wherein the second initial behavior feature information may be feature information with smaller data dimension and less corresponding noise feature than the second embedded feature information.

[0069] As an example, the execution subject may perform average pooling on the second embedded feature information to generate pooled feature information as the second initial behavior feature information.

[0070] As another example, the execution entity may input the second embedded feature information into a convolutional layer to output the second initial behavior feature information.

[0071] The second step is to generate second initial image feature information based on the main feature information of the image and the first embedded feature information, wherein the second initial image feature information may be feature information with smaller data dimension and less corresponding noise features than the first embedded feature information.

[0072] As an example, first, the execution subject may perform a sameness OR process on the main feature information of the portrait and the first embedded feature information to generate a sameness OR process result, and then perform an average pooling process on the sameness OR process result to generate the second initial portrait feature information.

[0073] The third step is to generate, based on the second initial behavior feature information and the second initial portrait feature information, the portrait feature information based on behavior perception and the behavior feature information based on portrait perception as the portrait feature information and the behavior feature information, respectively. The portrait feature information based on behavior perception may be the portrait feature information generated after considering the influence of the behavior feature information on the portrait feature information. The behavior feature information based on portrait perception may be the behavior feature information generated after considering the influence of the portrait feature information on the behavior feature information.

[0074] As an example, the above-mentioned execution entity can input the second initial behavior feature information and the above-mentioned second initial portrait feature information into the Bert model to generate portrait feature information based on behavior perception and behavior feature information based on portrait perception, as the above-mentioned portrait feature information and the above-mentioned behavior feature information, respectively.

[0075] Optionally, the generating of the behavior-perception-based portrait feature information and the portrait-perception-based behavior feature information according to the second initial behavior feature information and the second initial portrait feature information may include the following steps:

[0076] The first step is to obtain a historical interactive item information sequence and an item attribute information sequence corresponding to the historical interactive item information sequence, wherein the item attribute information includes: at least one item category and item value level information. Among them, the historical interactive item information sequence may be an information sequence of items with interactive behaviors within a historical time period. There is a one-to-one correspondence between the historical interactive item information in the historical interactive item information sequence and the item attribute information in the item attribute information sequence. The at least one item category may include: a primary category, a corresponding secondary category, and a corresponding tertiary category corresponding to the historical interactive item information. The item value level information may include: category value level information under the corresponding primary category, category value level information under the corresponding secondary category, and category value level information under the corresponding tertiary category. The item value level information may also include: a price level corresponding to the interactive item information.

[0077] The second step is to generate third embedded feature information for the item attribute information sequence and the historical interaction item information sequence. The specific implementation method can refer to the generation of the first embedded feature information.

[0078] The third step is to generate fused embedded feature information based on the second initial behavior feature information, the second initial portrait feature information and the third embedded feature information. The fused embedded feature information can represent the fused feature semantics between the second initial behavior feature information, the second initial portrait feature information and the third embedded feature information.

[0079] As an example, first, the execution subject may splice the second initial screen feature information with the second initial behavior feature information to obtain first spliced ​​feature information. Then, the second initial screen feature information is spliced ​​with the third embedded feature information to obtain second spliced ​​feature information. Next, the first spliced ​​feature information and the second spliced ​​feature information are added to obtain fused embedded feature information.

[0080] The fourth step is to use the self-attention mechanism model to generate the image feature information and behavior feature information for the above fused embedded feature information. The self-attention mechanism model can be a Tranformer structure model.

[0081] As an example, the execution subject can directly input the fused embedded feature information into the self-attention mechanism model to generate an output result. Then, the output result is split into feature elements at predetermined positions (split operation) to generate image feature information and behavior feature information.

[0082] Step 203, using the above-mentioned initial value level information output layer, execute the following first generation step:

[0083] Step 2031: Generate at least one interest feature information for at least one user interest dimension based on the above-mentioned portrait feature information and the above-mentioned behavior feature information.

[0084] In some embodiments, the execution subject may generate at least one interest feature information for at least one user interest dimension based on the portrait feature information and the behavior feature information. Among them, the user interest dimension may be an interest dimension corresponding to the user. At least one user interest dimension may be a description of user interests from multiple dimensions. In practice, at least one user interest dimension may include but is not limited to at least one of the following: category interest dimension, primary category interest dimension, and tertiary category interest dimension. The category interest dimension may characterize the category in which the user is interested. The primary category interest dimension may characterize the primary category in which the user is interested. The secondary category interest dimension may characterize the secondary category in which the user is interested. There is a one-to-one correspondence between the user interest dimension in at least one user interest dimension and the interest feature information in at least one interest feature information. The interest feature information may characterize the characteristic performance information of the user under the corresponding user interest dimension. For example, at least one interest feature information may include: interest feature information under the category interest dimension and interest feature information under the primary category interest dimension.

[0085] As an example, the execution entity may input the portrait feature information and the behavior feature information into an interest feature information generation layer to generate at least one interest feature information for at least one user interest dimension.

[0086] Step 2032: Generate category interest feature information and category value level feature information for the category tag based on the at least one interest feature information.

[0087] In some embodiments, the execution subject may generate category interest feature information and category value level feature information for the category tag based on the at least one interest feature information. The category interest feature information may represent feature information of the first target user's interest in the category tag. The category value level feature information may represent level feature information of the user's category value level under the category tag. The category value level under the category tag may be the value level of the first target user under the category tag. That is, the first target user's frequent commodity purchasing power under the category tag.

[0088] As an example, the above-mentioned execution entity can use the category interest feature information generation layer and the category value level feature information generation layer to generate category interest feature information and category value level feature information for the above-mentioned category tag according to the above-mentioned at least one interest feature information.

[0089] In some optional implementations of some embodiments, the generating of the category interest feature information and the category value level feature information for the category tag according to the at least one interest feature information may include the following steps:

[0090] The first step is to obtain the category representation vector and the category value level vector corresponding to the above category tag. In practice, the category representation vector can represent the semantic information of the category tag. The category value level vector can represent the level semantic information of each value level corresponding to the category tag. That is, the category value level vector can be the level semantic information for all value levels.

[0091] As an example, first, the above-mentioned execution entity can determine the first-level category information, second-level category information, and third-level category information corresponding to the category tag. Then, determine the first-level category vector corresponding to the above-mentioned first-level category information, the second-level category vector corresponding to the second-level category information, and the third-level category vector corresponding to the third-level category information. Next, average the corresponding elements of the above-mentioned first-level category vector, second-level category vector, and third-level category vector to generate an average vector. Next, multiply the average vector by the fifth parameter to generate a third multiplication result. Finally, add the third multiplication result to the sixth parameter to obtain a category representation vector. The fifth parameter and the sixth parameter are parameters in the initial value gear information output layer. As the model parameters of the initial value gear information output layer are updated, the corresponding parameter values ​​also change.

[0092] As another example, first, determine the 10 gear representations corresponding to the category label. Then, embed each gear representation to generate a gear embedding representation vector, and obtain a gear embedding representation vector set. Finally, concatenate the gear embedding representation vectors in the gear embedding representation vector set to generate a category value gear vector.

[0093] In the second step, the above-mentioned crowd-level interest feature information, the above-mentioned target-level category interest feature information and the above-mentioned user-level hidden interest feature information are spliced ​​to generate spliced ​​feature information. Among them, the crowd-level interest feature information can be interest feature information based on the crowd level. The target-level category interest feature information can be interest feature information based on the target-level category level. The user-level hidden interest feature information can be interest feature information based on the user level. The interest level granularity corresponding to the crowd-level interest feature information is greater than the interest level granularity corresponding to the target-level category interest feature information. The interest level granularity corresponding to the target-level category interest feature information is greater than the interest level granularity corresponding to the user-level hidden interest feature information. The target-level category level can be the category level corresponding to the secondary category. The crowd-level interest feature information, the above-mentioned target-level category interest feature information and the above-mentioned user-level hidden interest feature information can all be feature information in the form of vectors. The target-level category interest feature information can characterize the interest feature information of the first target user under each secondary category.

[0094] The third step is to determine the category interest feature information for the category label whose similarity between the concatenated feature information and the category representation vector satisfies a preset similarity condition, wherein the preset similarity condition may be that the vector distance between the corresponding vectors is the shortest.

[0095] The fourth step is to generate, based on the above-mentioned category interest feature information and the above-mentioned category value level vector, the category value level feature information representing the above-mentioned first target user's signature in the above-mentioned category.

[0096] As an example, the execution subject may perform a dot product on the category interest feature information and the category value level vector to generate a dot product result, and then input the dot product result into a softmax function to generate the category value level feature information.

[0097] As another example, the execution entity may input the category interest feature information and the category value level vector into a fully connected layer to generate category value level feature information.

[0098] Step 204, using the above loss function layer, generates model loss information for the above category interest feature information and the above category value level feature information.

[0099] In some embodiments, the execution subject may use the loss function layer to generate model loss information for the category interest feature information and the category value level feature information. The loss function layer includes at least one loss function. The model loss information represents the information difference between the predicted category value level information and the actual category value level information.

[0100] In some optional implementations of some embodiments, the training data further includes: actual category value level information, wherein the actual category value level information may be the actual category value level signed by the first target user in the category target.

[0101] Optionally, the above-mentioned use of the above-mentioned loss function layer to generate the model loss information for the above-mentioned category interest feature information and the above-mentioned category value level feature information may include the following steps:

[0102] In the first step, using the above loss function layer, the following third generation step is performed:

[0103] Sub-step 1: Generate category loss information between the above-mentioned category representation vector and the above-mentioned category interest feature information. The category loss information may represent the category difference between the category corresponding to the category representation vector and the category corresponding to the category interest feature information.

[0104] As an example, the execution entity may generate category loss information between the category representation vector and the category interest feature information using a noise contrast loss function.

[0105] Sub-step 2: Generate value level loss information between the above-mentioned category value level characteristic information and the above-mentioned actual category value level information. The value level loss information can represent the level difference between the value level corresponding to the category value level characteristic information and the value level corresponding to the actual category value level information.

[0106] As an example, the execution entity may utilize a cross entropy loss function to generate value level loss information between the category value level feature information and the actual category value level information.

[0107] Sub-step 3: determining the predetermined level category corresponding to the above category tag as predetermined level category information. The predetermined level category may be a secondary category corresponding to the category tag. The predetermined level category information may be category information corresponding to the secondary category.

[0108] Sub-step 4: determining predetermined level category characteristic information corresponding to the predetermined level category information from the splicing characteristic information.

[0109] Sub-step 5: Generate interest loss information for the above-mentioned category interest characteristic information and the above-mentioned predetermined level category characteristic information. The interest loss information may represent the difference between the category interest content corresponding to the category interest characteristic information and the category interest content corresponding to the predetermined level category characteristic information.

[0110] As an example, the execution subject may generate interest loss information for the above-mentioned category interest feature information and the above-mentioned predetermined level category feature information using a regularized loss function based on label-aware attention. The regularized loss function based on label-aware attention includes a function that characterizes whether the secondary category corresponding to the category interest feature information exists in the secondary category set corresponding to the user's historical value behavior.

[0111] The second step is to generate the above-mentioned model loss information based on the above-mentioned category loss information, the above-mentioned target level category feature information and the above-mentioned interest loss information.

[0112] As an example, the execution entity may perform weighted summation on the category loss information, the target-level category feature information, and the interest loss information to generate weighted sum loss information as model loss information.

[0113] Step 205, in response to determining that the training of the above-mentioned model loss information characterization model is completed, the above-mentioned initial category value level information generation model is determined as the category value level information generation model.

[0114] In some embodiments, in response to determining that the training of the above-mentioned model loss information representation model is completed, the above-mentioned execution subject can determine the above-mentioned initial category value level information generation model as the category value level information generation model. Wherein, when the model loss information corresponds to each loss information in the loss information set, the representation model training is completed. The loss information set includes: model loss information and a subset of adjacent loss information corresponding to the model loss information.

[0115] The above-mentioned embodiments of the present disclosure have the following beneficial effects: through the model training method of some embodiments of the present disclosure, using the category value level information generation model, the corresponding category value level information can be accurately generated under the category where the user has no corresponding value behavior data. Specifically, the reason why the relevant category value level information is not accurate enough is that for users who do not have a value circulation data set or have a small amount of corresponding data in the value circulation data set, the conventional neural network model cannot effectively extract features and often lacks category generalization ability, resulting in the generated category value level information being not accurate enough. Based on this, the model training method of some embodiments of the present disclosure first obtains training data. Among them, the above-mentioned training data includes: a first user portrait for a first target user, a first user value behavior information sequence, and a category tag. Here, through the first user portrait, the initial category value level information generation model can subsequently learn more user features in the first user portrait that are associated with the first target user. Through the first user value behavior information sequence, the initial category value level information generation model can subsequently learn more value behavior features corresponding to the first user value behavior information sequence. The obtained category tag is used for the disadvantage of subsequent loss of information. Then, the initial model input layer is used to generate the portrait feature information corresponding to the first user portrait and the behavior feature information corresponding to the first user value behavior information sequence. Among them, the initial category value level information generation model includes: the initial model input layer, the initial value level information output layer and the loss function layer. The obtained portrait feature information can comprehensively characterize the main feature information of the portrait of the first target user in the first user portrait. The behavior feature information can comprehensively characterize the main feature information of the value behavior of the first target user in the first user value behavior information sequence. Here, the obtained portrait feature information and behavior feature information can allow the initial value level information output layer to learn more important and targeted feature information, not learn useless related feature information, and can generate more accurate category interest feature information and category value level feature information. Then, using the above-mentioned initial value level information output layer, the following first generation step is performed: the first step, according to the above-mentioned portrait feature information and the above-mentioned behavior feature information, the main feature information of the user and the main value behavior feature information of the user in the learned portrait can be used to accurately generate at least one interest feature information for at least one user interest dimension. Here, the corresponding interest feature information is determined from multiple user interest dimensions, which can fully reflect the user's corresponding category interest information from various fine-grained levels. In the second step, based on the above at least one interest feature information, the category interest feature information and category value level feature information for the above category tag can be accurately generated. The category interest feature information can represent the category interest feature information of the first target user under the category tag. The category value level feature information can represent the value level information of the first target user under the category tag.Furthermore, by using the above-mentioned loss function layer, the model loss information for the above-mentioned category interest feature information and the above-mentioned category value level feature information can be accurately generated, so as to be used for determining whether the subsequent initial category value level information generation model has been trained. Finally, in response to determining that the above-mentioned model loss information represents that the model training is completed, the above-mentioned initial category value level information generation model is determined as the category value level information generation model. In summary, through the initial model input layer, the portrait feature information and behavior feature information can be accurately generated, so that the subsequent initial value level information output layer can learn more refined user important feature information. Then, through the initial value level information output layer, under the premise that at least one interest feature information corresponding to at least one user interest dimension can be accurately generated, the category interest feature information and the category value level feature information are accurately generated by matching with the category tag. Based on this, the loss function layer is finally used to generate the corresponding model loss information to determine whether the model training is completed. Therefore, the trained category value level information generation model can be used to accurately generate the corresponding category value level information under the category where the user has no corresponding value behavior data.

[0116] Further references Figure 3 , shows a process 300 of other embodiments of the model training method according to the present disclosure. The model training method comprises the following steps:

[0117] Step 301, obtaining training data.

[0118] Step 302, using the initial model input layer, generates the portrait feature information corresponding to the first user portrait and the behavior feature information corresponding to the first user value behavior information sequence.

[0119] Step 303, using the above-mentioned initial value level information output layer, execute the following first generation step:

[0120] Step 3031, performing a linear transformation on the second initial portrait feature information and the portrait feature information to generate the population-level interest feature information corresponding to the population interest dimension as interest feature information.

[0121] In some embodiments, an execution entity (e.g. Figure 1 The electronic device 101 shown) can perform a linear transformation on the second initial portrait feature information and the portrait feature information to generate the crowd-level interest feature information corresponding to the crowd interest dimension as interest feature information. The at least one user interest dimension includes: crowd interest dimension. The crowd interest dimension can be an interest dimension of the user at the crowd level. The crowd-level interest feature information can be in vector form, characterizing the interest features of the first target user at the crowd interest dimension.

[0122] As an example, first, the above-mentioned execution entity can add the second initial portrait feature information and the portrait feature information to generate added feature information. Then, the added feature information is input into the RELU function to obtain the function output result. Next, the function output result is multiplied by the seventh parameter to obtain the fourth multiplication result. Finally, the fourth multiplication result is added to the eighth parameter to obtain the crowd-level interest feature information. The seventh parameter and the eighth parameter are parameter values ​​updated as the model is trained.

[0123] Step 3032, obtaining a category information set corresponding to the target level category.

[0124] In some embodiments, the execution subject may obtain a category information set corresponding to a target level category. The target level category may be a pre-set level category. For example, the target level category may be a secondary category. Each category information in the category information set may be a full category information set corresponding to the secondary category. That is, the category information set may be each secondary category information. The at least one user interest dimension further includes: a target level category interest dimension.

[0125] Step 3033: Generate a category vector corresponding to each category information in the above category information set to obtain a category vector set.

[0126] In some embodiments, the execution entity may generate a category vector corresponding to each category information in the category information set to obtain a category vector set.

[0127] Step 3034: Generate target-level category interest feature information corresponding to the target-level category interest dimension as interest feature information based on the category vector set and the behavior feature information.

[0128] In some embodiments, the execution subject may generate target-level category interest characteristic information corresponding to the target-level category interest dimension as interest characteristic information based on the category vector set and the behavior characteristic information.

[0129] As an example, first, the execution subject can perform dot multiplication processing on the category matrix corresponding to the category vector set and the matrix corresponding to the behavior feature information to generate dot product information. Then, the dot product information is multiplied by the ninth parameter to generate the fifth multiplication result. Then, the fifth multiplication result is added to the tenth parameter to obtain the target level category interest feature information. The ninth parameter and the tenth parameter are parameter values ​​updated as the model is trained.

[0130] In some optional implementations of some embodiments, the at least one user interest dimension further includes: a user-level implicit interest dimension.

[0131] Optionally, the execution subject may generate user-level implicit interest feature information for the behavior feature information as interest feature information using a capsule network based on the portrait feature information. The user-level implicit interest feature information is feature information corresponding to the user-level interest dimension, and the user-level implicit interest feature information includes: implicit interest feature information under at least one interest representation. The at least one interest representation corresponding to the capsule network may be pre-set. The interest representation may be an information representation that the user may be interested in. Each interest representation has corresponding implicit interest feature information.

[0132] As an example, the above-mentioned execution entity can input the portrait feature information and the behavior feature information into the capsule network to generate user-level implicit interest feature information.

[0133] Step 3035: Generate category interest feature information and category value level feature information for the category tag based on the at least one interest feature information.

[0134] Step 304: using the loss function layer, generate model loss information for the category interest feature information and the category value level feature information.

[0135] Step 305, in response to determining that the training of the above-mentioned model loss information characterization model is completed, the above-mentioned initial category value level information generation model is determined as the category value level information generation model.

[0136] In some embodiments, the specific implementation of steps 301, 3035, 304-305 and the technical effects thereof can be referred to in Figure 2 Steps 201, 2022, 204-205 in the corresponding embodiments are not repeated here.

[0137] from Figure 3 It can be seen that Figure 2 Compared with the description of some corresponding embodiments, Figure 3 The process 300 of the model training method in some corresponding embodiments can comprehensively obtain the interest information of the first target user at various granularities by generating interest feature information corresponding to the target-level category interest dimension and interest feature information corresponding to the user-level latent interest dimension, so that subsequent models can learn more feature information and make model training more effective.

[0138] Further references Figure 4 , shows a process 400 of some embodiments of the information generation method according to the present disclosure. The information generation method comprises the following steps:

[0139] Step 401, obtain target category information, a second user portrait for a second target user, and a second user value behavior information sequence.

[0140] In some embodiments, the execution subject (for example, an electronic device) may obtain target category information, a second user portrait for a second target user, and a second user value behavior information sequence. Among them, the second target user may be the same user as the first target user, or a different user. The target category information may be predetermined category information. The category level corresponding to the target category information may be preset. For example, the target category information may be "tops". For specific explanations of the second user portrait and the second user value behavior information sequence, please refer to the relevant explanations of the first user portrait and the first user value behavior information sequence.

[0141] Step 402, input the above-mentioned target category information, the above-mentioned second user portrait and the above-mentioned second user value behavior information sequence into a pre-trained category value level information generation model to generate category value level information.

[0142] In some embodiments, the execution subject may input the target category information, the second user portrait and the second user value behavior information sequence into a pre-trained category value level information generation model to generate category value level information. The category value level information generation model is generated based on a model training method.

[0143] The above-mentioned embodiments of the present disclosure have the following beneficial effects: through the information generation method of some embodiments of the present disclosure, using the trained category value rank information generation model, the category value rank information of the first target user under the target category information can be accurately generated.

[0144] Further references Figure 5 As an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of a model training device, and these device embodiments are Figure 2 Corresponding to the method embodiments shown, the model training device can be specifically applied to various electronic devices.

[0145] like Figure 5As shown, a model training device 500 includes: a first acquisition unit 501, a first generation unit 502, an execution unit 503, a second generation unit 504 and a determination unit 505. The first acquisition unit 501 is configured to acquire training data, wherein the training data includes: a first user portrait for a first target user, a first user value behavior information sequence and a category label; the first generation unit 502 is configured to use an initial model input layer to generate the portrait feature information corresponding to the first user portrait and the behavior feature information corresponding to the first user value behavior information sequence, wherein the initial category value level information generation model includes: the initial model input layer, the initial value level information output layer and the loss function layer; the execution unit 503 is configured to use the initial value level information output layer to perform the following first Generation step: based on the above-mentioned portrait feature information and the above-mentioned behavior feature information, generate at least one interest feature information for at least one user interest dimension; based on the above-mentioned at least one interest feature information, generate category interest feature information and category value level feature information for the above-mentioned category tag; the second generation unit 504 is configured to use the above-mentioned loss function layer to generate model loss information for the above-mentioned category interest feature information and the above-mentioned category value level feature information; the determination unit 505 is configured to determine the above-mentioned initial category value level information generation model as the category value level information generation model in response to determining that the above-mentioned model loss information is less than a preset value.

[0146] In some optional implementations of some embodiments, the above-mentioned initial model input layer includes: an initial compression excitation layer; and the first generation unit 502 can be further configured to: use the above-mentioned initial model input layer to perform the following second generation step: generate the first initial portrait feature information for the above-mentioned first user portrait, and generate the first initial behavior feature information for the above-mentioned first user value behavior information sequence; perform low-dimensional embedding processing on the above-mentioned first initial portrait feature information and the above-mentioned first initial behavior feature information respectively to generate first embedded feature information and second embedded feature information; input the above-mentioned first embedded feature information into the above-mentioned initial compression excitation layer to generate the main feature information of the portrait; generate the above-mentioned portrait feature information and the above-mentioned behavior feature information according to the above-mentioned main feature information of the portrait and the above-mentioned second embedded feature information.

[0147] In some optional implementations of some embodiments, the first generation unit 502 can be further configured to: generate second initial behavior feature information based on the above-mentioned second embedded feature information; generate second initial portrait feature information based on the above-mentioned portrait main feature information and the above-mentioned first embedded feature information; generate portrait feature information based on behavior perception and behavior feature information based on portrait perception based on the above-mentioned second initial behavior feature information and the above-mentioned second initial portrait feature information, as the above-mentioned portrait feature information and the above-mentioned behavior feature information, respectively.

[0148] In some optional implementations of some embodiments, the first generation unit 502 can be further configured to: obtain a historical interactive item information sequence and an item attribute information sequence corresponding to the historical interactive item information sequence, the item attribute information including: at least one item category and item value level information; generate third embedded feature information for the above-mentioned item attribute information sequence and the above-mentioned historical interactive item information sequence; generate fused embedded feature information based on the above-mentioned second initial behavior feature information, the above-mentioned second initial portrait feature information and the above-mentioned third embedded feature information; and use a self-attention mechanism model to generate portrait feature information and behavior feature information for the above-mentioned fused embedded feature information.

[0149] In some optional implementations of some embodiments, the at least one user interest dimension includes: a population interest dimension; and the execution unit 503 can be further configured to: perform a linear transformation on the second initial portrait feature information and the portrait feature information to generate population-level interest feature information corresponding to the population interest dimension as interest feature information.

[0150] In some optional implementations of some embodiments, the at least one user interest dimension includes: a target level category interest dimension; and the execution unit 503 can be further configured to: obtain a category information set corresponding to the target level category; generate a category vector corresponding to each category information in the above category information set to obtain a category vector set; and generate target level category interest feature information corresponding to the above target level category interest dimension as interest feature information based on the above category vector set and the above behavioral feature information.

[0151] In some optional implementations of some embodiments, the at least one user interest dimension includes: a user-level latent interest dimension; and the execution unit 503 can be further configured to: generate user-level latent interest feature information for the above-mentioned behavioral feature information according to the above-mentioned portrait feature information using a capsule network as interest feature information, wherein the user-level latent interest feature information is feature information corresponding to the above-mentioned user-level interest dimension, and the user-level latent interest feature information includes: latent interest feature information under at least one interest representation.

[0152] In some optional implementations of some embodiments, the execution unit 503 can be further configured to: obtain the category representation vector and the category value level vector corresponding to the above-mentioned category tag; perform feature information splicing on the above-mentioned population-level interest feature information, the above-mentioned target-level category interest feature information and the above-mentioned user-level latent interest feature information to generate spliced ​​feature information; determine the category interest feature information for the above-mentioned category tag whose similarity with the above-mentioned category representation vector in the above-mentioned spliced ​​feature information meets the preset similarity condition; generate the category value level feature information representing the above-mentioned first target user under the above-mentioned category tag based on the above-mentioned category interest feature information and the above-mentioned category value level vector.

[0153] In some optional implementations of some embodiments, the above-mentioned training data also includes: actual category value level information; and the second generation unit 504 can be further configured to: use the above-mentioned loss function layer to perform the following third generation step: generate category loss information between the above-mentioned category representation vector and the above-mentioned category interest feature information; generate value level loss information between the above-mentioned category value level feature information and the above-mentioned actual category value level information; determine the predetermined level category corresponding to the above-mentioned category tag as the predetermined level category information; determine the predetermined level category feature information corresponding to the above-mentioned predetermined level category information from the above-mentioned splicing feature information; generate interest loss information for the above-mentioned category interest feature information and the above-mentioned predetermined level category feature information; generate the above-mentioned model loss information based on the above-mentioned category loss information, the above-mentioned target level category feature information and the above-mentioned interest loss information.

[0154] It is understandable that the units recorded in the model training device 500 are similar to the reference Figure 2 Therefore, the operations, features and beneficial effects described above for the method are also applicable to the model training device 500 and the units contained therein, and will not be repeated here.

[0155] Further references Figure 6 As an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of an information generating device, and these device embodiments are Figure 4 Corresponding to the method embodiments shown, the information generating device can be specifically applied to various electronic devices.

[0156] like Figure 6As shown, an information generating device 600 includes: a second acquisition unit 601 and a third generation unit 602. The second acquisition unit 601 is configured to acquire target category information, a second user portrait for a second target user, and a second user value behavior information sequence, wherein the target category information is the category information of the category under the user's non-value behavior; the third generation unit 602 is configured to input the target category information, the second user portrait, and the second user value behavior information sequence into a pre-trained category value level information generation model to generate category value level information, wherein the category value level information represents the value level status of the second target user under the target category information, and the category value level information generation model is generated based on a model training method.

[0157] It is understandable that the units recorded in the information generating device 600 are similar to those in the reference Figure 4 Therefore, the operations, features and beneficial effects described above for the method are also applicable to the information generating device 600 and the units included therein, and will not be described in detail here.

[0158] Reference below Figure 7 , which shows an electronic device (eg, Figure 1 Schematic diagram of the structure of the electronic device 101)700. Figure 7 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0159] like Figure 7 As shown, the electronic device 700 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory 702 or a program loaded from a storage device 708 into a random access memory 703. Various programs and data required for the operation of the electronic device 700 are also stored in the random access memory 703. The processing device 701, the read-only memory 702, and the random access memory 703 are connected to each other via a bus 704. An input / output interface 705 is also connected to the bus 704.

[0160] Typically, the following devices may be connected to the input / output interface 705: an input device 706 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 707 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 708 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 709. The communication device 709 may allow the electronic device 700 to communicate with other devices wirelessly or by wire to exchange data. Although Figure 7 The electronic device 700 is shown with various devices, but it should be understood that it is not required to implement or possess all the devices shown. More or fewer devices may be implemented or possessed instead. Figure 7 Each block shown in the figure may represent one device, or may represent multiple devices as required.

[0161] In particular, according to some embodiments of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, some embodiments of the present disclosure include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In some such embodiments, the computer program can be downloaded and installed from a network through a communication device 709, or installed from a storage device 708, or installed from a read-only memory 702. When the computer program is executed by the processing device 701, the above-mentioned functions defined in the method of some embodiments of the present disclosure are executed.

[0162] It should be noted that the computer-readable medium in some embodiments of the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the two. The computer-readable storage medium may 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 of computer-readable storage media may include, but are not limited to: 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 some embodiments of the present disclosure, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In some embodiments of the present disclosure, the computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The computer readable signal medium 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. The program code contained on the computer readable medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (radio frequency), etc., or any suitable combination of the above.

[0163] In some embodiments, the client and the server may communicate using any currently known or future developed network protocol such as HTTP (HyperText Transfer Protocol), and may be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), an internet (e.g., the Internet), and a peer-to-peer network (e.g., an ad hoc peer-to-peer network), as well as any currently known or future developed network.

[0164] The computer-readable medium may be included in the electronic device, or may exist independently without being incorporated into the electronic device. The computer-readable medium carries one or more programs. When the one or more programs are executed by the electronic device, the electronic device: obtains training data, wherein the training data includes: a first user portrait for a first target user, a first user value behavior information sequence and a category tag; uses an initial model input layer to generate portrait feature information corresponding to the first user portrait and behavior feature information corresponding to the first user value behavior information sequence, wherein the initial category value level information generation model includes: the initial model input layer, the initial value level information output layer and the loss function layer; uses the initial value level information output layer to perform the following first generation step: based on the portrait feature information and the behavior feature information, generate at least one interest feature information for at least one user interest dimension; based on the at least one interest feature information, generate category interest feature information and category value level feature information for the category tag; uses the loss function layer to generate model loss information for the category interest feature information and the category value level feature information; in response to determining that the model loss information indicates that model training is complete, the initial category value level information generation model is determined as the category value level information generation model. Obtain target category information, a second user portrait for a second target user, and a second user value behavior information sequence, wherein the target category information is the category information of a category under a user's valueless behavior; input the target category information, the second user portrait, and the second user value behavior information sequence into a pre-trained category value level information generation model to generate category value level information, wherein the category value level information represents the value level status of the second target user under the target category information, and the category value level information generation model is generated based on a model training method.

[0165] Computer program code for performing the operations of some embodiments of the present disclosure 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 cases involving a remote computer, the remote computer may be connected to the user's computer via 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).

[0166] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some implementations as replacements, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0167] The units described in some embodiments of the present disclosure may be implemented by software or by hardware. The described units may also be provided in a processor, for example, may be described as: a processor includes a first acquisition unit, a first generation unit, an execution unit, a second generation unit, and a determination unit. The names of these units do not, in some cases, constitute limitations on the units themselves, for example, the first acquisition unit may also be described as a "unit for acquiring training data".

[0168] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), and the like.

[0169] Some embodiments of the present disclosure also provide a computer program product, including a computer program, which implements any of the above-mentioned model training methods or information generation methods when executed by a processor.

[0170] The above descriptions are only some preferred embodiments of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) and the technical solutions formed.

Claims

1. A model training method, comprising: Acquire training data, wherein the training data includes: a first user portrait for a first target user, a first user value behavior information sequence, and a category label; Using the initial model input layer, generating the portrait feature information corresponding to the first user portrait and the behavior feature information corresponding to the first user value behavior information sequence, wherein the initial category value level information generation model includes: the initial model input layer, the initial value level information output layer and the loss function layer; Using the initial value level information output layer, the following first generation step is performed: Generate at least one interest feature information for at least one user interest dimension according to the portrait feature information and the behavior feature information; Generating category interest feature information and category value level feature information for the category tag according to the at least one interest feature information; Using the loss function layer, generating model loss information for the category interest feature information and the category value level feature information; In response to determining that the model loss information characterization model training is completed, the initial category value level information generation model is determined as the category value level information generation model.

2. The method according to claim 1, wherein: The initial model input layer includes: an initial compression excitation layer; and The using the initial model input layer to generate the portrait feature information corresponding to the first user portrait and the behavior feature information corresponding to the first user value behavior information sequence includes: Using the initial model input layer, the following second generation step is performed: Generate first initial portrait feature information for the first user portrait, and generate first initial behavior feature information for the first user value behavior information sequence; Performing low-dimensional embedding processing on the first initial portrait feature information and the first initial behavior feature information respectively to generate first embedded feature information and second embedded feature information; Inputting the first embedded feature information into the initial compression excitation layer to generate main feature information of the image; The portrait feature information and the behavior feature information are generated according to the portrait main feature information and the second embedded feature information.

3. The method according to claim 2, wherein: The step of generating the portrait feature information and the behavior feature information according to the portrait main feature information and the second embedded feature information includes: generating second initial behavior feature information according to the second embedded feature information; Generating second initial portrait feature information according to the portrait main feature information and the first embedded feature information; Based on the second initial behavior feature information and the second initial portrait feature information, the portrait feature information based on behavior perception and the behavior feature information based on portrait perception are generated as the portrait feature information and the behavior feature information, respectively.

4. The method according to claim 3, wherein: The generating, according to the second initial behavior feature information and the second initial portrait feature information, the portrait feature information based on behavior perception and the behavior feature information based on portrait perception includes: Acquire a historical interactive item information sequence and an item attribute information sequence corresponding to the historical interactive item information sequence, wherein the item attribute information includes: at least one item category and item value level information; generating third embedded feature information for the item attribute information sequence and the historical interaction item information sequence; generating fused embedded feature information according to the second initial behavior feature information, the second initial portrait feature information and the third embedded feature information; The self-attention mechanism model is used to generate portrait feature information and behavioral feature information for the fused embedded feature information.

5. The method according to claim 3, wherein: The at least one user interest dimension comprises: a crowd interest dimension; and The generating, according to the portrait feature information and the behavior feature information, at least one interest feature information for at least one user interest dimension comprises: A linear transformation is performed on the second initial portrait feature information and the portrait feature information to generate crowd-level interest feature information corresponding to the crowd interest dimension as interest feature information.

6. The method according to claim 5, wherein: The at least one user interest dimension further includes: a target level category interest dimension; and The generating at least one interest feature information for at least one user interest dimension according to the portrait feature information and the behavior feature information further includes: Get the category information set corresponding to the target level category; Generate a category vector corresponding to each category information in the category information set to obtain a category vector set; According to the category vector set and the behavior characteristic information, target-level category interest characteristic information corresponding to the target-level category interest dimension is generated as interest characteristic information.

7. The method according to claim 6, wherein: The at least one user interest dimension further includes: a user-level implicit interest dimension; and The generating at least one interest feature information for at least one user interest dimension according to the portrait feature information and the behavior feature information further includes: According to the portrait feature information, a capsule network is used to generate user-level implicit interest feature information for the behavioral feature information as interest feature information, wherein the user-level implicit interest feature information is feature information corresponding to the user-level interest dimension, and the user-level implicit interest feature information includes: implicit interest feature information under at least one interest representation.

8. The method according to claim 7, wherein: The step of generating the category interest feature information and the category value level feature information for the category tag according to the at least one interest feature information includes: Obtaining a category representation vector and a category value level vector corresponding to the category tag; Performing feature information splicing on the crowd-level interest feature information, the target-level category interest feature information, and the user-level latent interest feature information to generate spliced ​​feature information; Determine the category interest feature information for the category tag, the similarity between which in the spliced ​​feature information and the category representation vector satisfies a preset similarity condition; According to the category interest feature information and the category value level vector, category value level feature information representing the first target user's signature in the category is generated.

9. The method according to claim 8, wherein: The training data also includes: actual category value level information; and The using the loss function layer to generate model loss information for the category interest feature information and the category value level feature information includes: Using the loss function layer, the following third generation step is performed: Generating category loss information between the category representation vector and the category interest feature information; Generating value level loss information between the category value level characteristic information and the actual category value level information; Determining a predetermined level category corresponding to the category tag as predetermined level category information; Determining predetermined level category characteristic information corresponding to the predetermined level category information from the splicing characteristic information; generating interest loss information for the category interest feature information and the predetermined level category feature information; The model loss information is generated according to the category loss information, the target level category feature information and the interest loss information.

10. A method for generating information, comprising: Obtain target category information, a second user portrait for a second target user, and a second user value behavior information sequence, wherein the target category information is category information of a category under a user's valueless behavior; The target category information, the second user portrait and the second user value behavior information sequence are input into a pre-trained category value level information generation model to generate category value level information, wherein the category value level information represents the value level status of the second target user under the target category information, and the category value level information generation model is generated based on the method described in one of claims 1-9.

11. A model training device, comprising: A first acquisition unit is configured to acquire training data, wherein the training data includes: a first user portrait for a first target user, a first user value behavior information sequence, and a category label; A first generating unit is configured to generate, by using an initial model input layer, portrait feature information corresponding to the first user portrait and behavior feature information corresponding to the first user value behavior information sequence, wherein the initial category value level information generating model comprises: the initial model input layer, the initial value level information output layer and the loss function layer; The execution unit is configured to use the initial value level information output layer to perform the following first generation step: generating at least one interest feature information for at least one user interest dimension according to the portrait feature information and the behavior feature information; generating category interest feature information and category value level feature information for the category tag according to the at least one interest feature information; A second generating unit is configured to generate model loss information for the category interest feature information and the category value level feature information using the loss function layer; The determination unit is configured to determine the initial category value level information generation model as the category value level information generation model in response to determining that the model loss information characterization model training is completed.

12. An information generating device, comprising: A second acquisition unit is configured to acquire target category information, a second user portrait for a second target user, and a second user value behavior information sequence, wherein the target category information is category information of a category under a user's valueless behavior; The third generating unit is configured to input the target category information, the second user portrait and the second user value behavior information sequence into a pre-trained category value level information generating model to generate category value level information, wherein the category value level information represents the value level status of the second target user under the target category information, and the category value level information generating model is generated based on the method described in one of claims 1-9.

13. An electronic device comprising: one or more processors; a storage device having one or more programs stored thereon, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 10.

14. A computer readable medium having a computer program stored thereon, wherein: When the computer program is executed by a processor, the method according to any one of claims 1 to 10 is implemented.

15. A computer program product comprising a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 10.

Images