Hybrid multi-tower model training method and device, hybrid multi-tower model recommendation method and device, medium and equipment

Abstract

The invention relates to the technical field of computers, and discloses a hybrid multi-tower model training method, a recommendation method, a hybrid multi-tower model training device, a recommendation device, a computer readable medium and electronic equipment. The hybrid multi-tower model trained by the method can realize cross-scene-domain multi-domain training and analysis of heterogeneous contents through different content types corresponding to a plurality of sub-towers of an object tower and different scene types included in content types corresponding to private towers under the sub-towers, avoids feature interference among contents of different types and scenes, and improves the training efficiency. And content characteristics are analyzed more sufficiently, so that new shallow contents and long-tail contents are mined more deeply and comprehensively. On the basis, the purpose of fitting user consumption content samples by dynamically covering private towers under multi-scene-type training features is achieved, the problem of negative migration possibly caused by different user distribution and different sample sources in cross-domain training can be avoided, feature information of various services is fully mined, and the user experience is improved. And the model recall accuracy is improved.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of computer technology. More specifically, the embodiments of the present disclosure relate to a hybrid multi-tower model training method, a recommendation method, a hybrid multi-tower model training device, a recommendation device, a computer-readable medium and an electronic device. Background Art

[0002] This section is intended to provide a background or context to the embodiments of the disclosure that are recited in the claims, and no description herein is admitted to be prior art by inclusion in this section.

[0003] In the recommendation process, recall refers to the rapid screening of a small number of candidate items from a large candidate set for further processing in the subsequent rough sorting and fine sorting stages. Therefore, the recall efficiency will affect the subsequent fine sorting effect and the efficiency of the overall recommendation process to a certain extent.

[0004] At present, the recall is usually done by statistical analysis of the user's historical behavior data. There is a lack of sufficient historical behavior data for new and shallow content, so it is impossible to accurately recall it. In addition, it is greatly affected by the popularity effect, and it is also impossible to specifically recall long-tail candidate items. Furthermore, for businesses with multiple content types and cross-scenario, it is usually impossible to fully utilize the characteristics of heterogeneous content and cross-domain scenarios of the business, resulting in insufficient in-depth information extraction and understanding, resulting in inaccurate recall results. Summary of the invention

[0005] In this context, embodiments of the present disclosure are intended to provide a hybrid multi-tower model training method, a recommendation method, a hybrid multi-tower model training device, a recommendation device, a computer-readable medium, and an electronic device.

[0006] According to a first aspect of an embodiment of the present disclosure, a hybrid multi-tower model training method is provided, wherein the hybrid multi-tower model includes a user tower and an item tower, wherein the item tower includes at least one sub-tower, and the sub-tower includes a public feature tower and a private tower. The method may include: extracting training features based on training samples; the training samples include user behavior samples and content samples; the content samples include at least one content type, and each content type includes at least one scene type; the scene types are divided based on the purpose of the user consuming the content samples; the training features are input into the hybrid multi-tower model to be trained; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type; the private tower is dynamically masked based on the scene type, and training is performed in a direction where the similar distance between the outputs of the user tower and the item tower conforms to the direction of the training sample to obtain the hybrid multi-tower model.

[0007] Optionally, training features are extracted based on training samples, including: extracting initial features from training samples based on feature fields through a large language model; constructing feature fields based on business characteristics; constructing hub units based on the initial features of user behavior samples, constructing heterogeneous units corresponding to different content types based on the initial features of content samples, and constructing edges between hub units and heterogeneous units based on the behavioral relationships between user behavior samples and content samples to obtain a heterogeneous graph model; performing meta-path sampling in the heterogeneous graph model, adjusting the initial features based on the sampling results to obtain sampling features; initializing the sampling features to construct training features corresponding to the training samples.

[0008] Optionally, the construction conditions of the meta-path include at least one of the following: training samples of different content types appear symmetrically alternately in the meta-path; training samples whose creation time is within the historical time limit or before the historical time limit and training samples whose creation time is after the historical time limit appear symmetrically alternately in the meta-path; training samples whose accumulated user behavior data is less than or equal to a preset number and training samples whose accumulated user behavior data is greater than a preset number appear symmetrically alternately in the meta-path.

[0009] Optionally, initialization is performed based on the sampling features to construct training features corresponding to the training samples, including: when the training samples are user behavior samples, initialization is performed based on the sampling features, and user attribute features are added to obtain training features; when the training samples are content samples, initialization is performed based on the sampling features, and label features of the content samples and statistical features of the content samples under the corresponding scene types are added to obtain training features.

[0010] Optionally, the training features are input into the hybrid multi-tower model to be trained, including: in the hybrid multi-tower model to be trained, the initialized sampling features and user attribute features are input into the user tower; and the initialized sampling features and label features are input into the public feature tower according to the content type corresponding to the sub-tower, and the statistical features are input into the private tower corresponding to the scene type.

[0011] Optionally, the training features do not include identity identification number ID features.

[0012] Optionally, the content type includes audio content and playlist content.

[0013] Optionally, private towers are dynamically masked based on the scene type, and training is performed in a direction where the similarity distance between the outputs of the user tower and the item tower matches the direction of the training samples to obtain a hybrid multi-tower model, including: dynamically masking private towers based on the scene type, and using the output of the sub-tower corresponding to the playlist content as the input of the sub-tower corresponding to the sound content to obtain the output of the user tower and the item tower; training is performed in a direction where the similarity distance between the outputs of the user tower and the item tower matches the direction of the training samples to obtain a hybrid multi-tower model.

[0014] Optionally, the scenario type includes at least one of a recommendation scenario, a revisit scenario, and a search scenario.

[0015] According to a second aspect of an embodiment of the present disclosure, a recommendation method is provided, which is implemented based on a hybrid multi-tower model trained by the hybrid multi-tower model training method disclosed in the first aspect above, and the method may include: obtaining behavior data and content data of a target user, and extracting predictive features based on the behavior data and the content data; the content data includes at least one content type, and each content type includes at least one scene type; inputting the predictive features into the hybrid multi-tower model; and determining the content data to be provided to the target user based on the similarity distance between the outputs of the user tower and the item tower in the hybrid multi-tower model.

[0016] According to a third aspect of an embodiment of the present disclosure, a hybrid multi-tower model training device is provided, wherein the hybrid multi-tower model includes a user tower and an item tower, wherein the item tower includes at least one sub-tower, and the sub-tower includes a public feature tower and a private tower. The device may include: a feature extraction module, used to extract training features based on training samples; the training samples include user behavior samples and content samples; the content samples include at least one content type, and each content type includes at least one scene type; the scene types are divided based on the purpose of the user consuming the content samples; a model training module, used to input the training features into the hybrid multi-tower model to be trained; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type; the model training module is also used to dynamically mask the private tower based on the scene type, and train in a direction where the similar distance between the outputs of the user tower and the item tower conforms to the direction of the training sample to obtain the hybrid multi-tower model.

[0017] Optionally, a feature extraction module is specifically used to extract initial features from training samples based on feature fields through a large language model; feature fields are constructed based on business characteristics; a heterogeneous graph model is constructed with user behavior samples as hub units and training samples of different content types as heterogeneous units, and each training sample in the graph model corresponds to an initial feature; meta-path sampling is performed in the heterogeneous graph model, and the initial features are adjusted based on the sampling results to obtain sampling features; the sampling features are initialized, and training features corresponding to the training samples are constructed.

[0018] Optionally, the construction conditions of the meta-path include at least one of the following: training samples of different content types appear symmetrically alternately in the meta-path; training samples whose creation time is within the historical time limit or before the historical time limit and training samples whose creation time is after the historical time limit appear symmetrically alternately in the meta-path; training samples whose accumulated user behavior data is less than or equal to a preset number and training samples whose accumulated user behavior data is greater than a preset number appear symmetrically alternately in the meta-path.

[0019] Optionally, the feature extraction module is specifically used to initialize based on the sampling features when the training samples are user behavior samples, and add user attribute features to obtain training features; when the training samples are content samples, it is initialized based on the sampling features, and label features of the content samples and statistical features of the content samples under the corresponding scene types are added to obtain training features.

[0020] Optionally, the model training module is specifically used to input the initialized sampling features and user attribute features into the user tower in the hybrid multi-tower model to be trained; and input the initialized sampling features and label features into the public feature tower according to the content type corresponding to the sub-tower, and input the statistical features into the private tower corresponding to the scene type.

[0021] Optionally, the training features do not include identity identification number ID features.

[0022] Optionally, the content type includes audio content and playlist content.

[0023] Optionally, the model training module is specifically used to dynamically mask the private tower based on the scene type, and use the output of the sub-tower corresponding to the playlist content as the input of the sub-tower corresponding to the sound content, to obtain the output of the user tower and the item tower; and train with the similar distance between the outputs of the user tower and the item tower in the direction of the training sample to obtain a hybrid multi-tower model.

[0024] Optionally, the scenario type includes at least one of a recommendation scenario, a revisit scenario, and a search scenario.

[0025] According to a fourth aspect of an embodiment of the present disclosure, a recommendation device is provided, which is implemented based on a hybrid multi-tower model trained by the model training device disclosed in the third aspect. The device may include: a feature extraction module, used to obtain behavior data and content data of a target user, and extract predictive features based on the behavior data and the content data; the content data includes at least one content type, and each content type includes at least one scene type; a model input module, used to input predictive features into the hybrid multi-tower model; a content recommendation module, used to determine the content data provided to the target user based on the similarity distance between the outputs of the user tower and the item tower in the hybrid multi-tower model.

[0026] According to a fifth aspect of an embodiment of the present disclosure, a computer-readable medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the hybrid multi-tower model training method disclosed in the first aspect or the recommendation method disclosed in the second aspect is implemented.

[0027] According to the sixth aspect of the embodiment of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the hybrid multi-tower model training method disclosed in the first aspect, or the recommendation method disclosed in the second aspect, by executing the executable instructions.

[0028] According to the hybrid multi-tower model training method of the embodiment of the present disclosure, the hybrid multi-tower model includes a user tower and an item tower, the item tower includes at least one sub-tower, and the sub-tower includes a public feature tower and a private tower. The method extracts training features from training samples, the training samples include user behavior samples and content samples, the content samples include at least one content type, and each content type includes at least one scene type, and the scene type is divided based on the purpose of the user consuming the content sample; on this basis, the training features extracted based on the training samples are input into the hybrid multi-tower model to be trained, and the private tower is dynamically masked based on the scene type, and the similar distance between the output of the user tower and the item tower is in line with the direction of the training sample to obtain the hybrid multi-tower model; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type. The hybrid multi-tower model trained by this method can achieve multi-domain training and analysis of heterogeneous content across scene domains through multiple sub-towers of the item tower corresponding to different content types, and private towers under the sub-towers corresponding to different scene types of content types, avoiding feature interference between content of different types and scenes, and more fully analyzing content characteristics, so as to more deeply and comprehensively mine new shallow content and long-tail content. On this basis, by dynamically masking private towers under the training characteristics of multiple scene types to fit the purpose of user consumption content samples, it can also avoid negative transfer problems caused by different user distributions and different sample sources in cross-domain training, fully mine the feature information of diverse businesses, and thus improve the accuracy of model recall. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present disclosure are shown in an exemplary and non-limiting manner, in which:

[0030] Figure 1 A structural schematic diagram of a hybrid multi-tower model according to an embodiment of the present disclosure is shown.

[0031] Figure 2 One of the step flow charts of the hybrid multi-tower model training method according to an embodiment of the present disclosure is shown.

[0032] Figure 3 The second step flow chart of the hybrid multi-tower model training method according to an embodiment of the present disclosure is shown.

[0033] Figure 4 A schematic diagram of a heterogeneous graph model according to an embodiment of the present disclosure is shown.

[0034] Figure 5 A schematic diagram of a hybrid multi-tower model applied to podcast business according to an embodiment of the present disclosure is shown.

[0035] Figure 6 The third step flow chart of the hybrid multi-tower model training method of the exemplary embodiment of the present disclosure is shown.

[0036] Figure 7 A flowchart showing the steps of a recommended method according to an embodiment of the present disclosure is shown.

[0037] Figure 8 A schematic diagram of a hybrid multi-tower model training device according to an embodiment of the present disclosure is shown.

[0038] Fig. 9 A schematic diagram showing a recommended arrangement of an embodiment of the present disclosure is shown.

[0039] Fig.10 A schematic diagram of a computer readable medium illustrating an embodiment of the present disclosure is shown.

[0040] Fig.11 A schematic diagram of an electronic device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0041] The principles and spirit of the present disclosure will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present disclosure, and are not intended to limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0042] Those skilled in the art will appreciate that the embodiments of the present disclosure may be implemented as a system, device, apparatus, method or computer program product. Therefore, the present disclosure may be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0043] According to an embodiment of the present disclosure, a hybrid multi-tower model training method, a hybrid multi-tower model training device, a computer-readable medium and an electronic device are provided.

[0044] In this document, any number of elements in the drawings is used for illustration rather than limitation, and any naming is used only for distinction and does not have any limiting meaning.

[0045] The principle and spirit of the present disclosure are explained in detail below with reference to several representative embodiments of the present disclosure.

[0046] At present, the recall scheme based on user behavior is often used, which can be achieved through data statistics or model calculation. There are some problems when this recall scheme is applied in actual scenarios. For example, the recall scheme based on user behavior lacks sufficient data to support the relevance calculation for new and shallow content, which causes the new and shallow content to sink to the bottom and the recall results to be assimilated; there is also a hot effect caused by the concentrated distribution of positive samples, which causes the recall results to be biased towards known hot content, while there is less exploration of user interests in long-tail items; and heterogeneous content is usually modeled separately for different content types, or heterogeneous content is treated as the same type; and it cannot distinguish and explore the characteristics of business scenarios, resulting in certain negative transfer problems.

[0047] To address the above issues, content-based and collaborative filtering-based recommendations can be combined for training. User feedback needs to be introduced to adjust the content-based recommendation model to improve the recall effect of the model and enhance its performance. This solution is more complicated in terms of model adjustment and training, and has limited improvement in the recall of new shallow content and long-tail items. It also cannot understand and process business characteristics across types and scenarios.

[0048] In the embodiments of the present disclosure, a training method for a hybrid multi-tower model is provided. The hybrid multi-tower model adopts the structure of a user tower and an item tower. The item tower includes sub-towers, and each sub-tower includes a public feature tower and a private tower. On this basis, training features are extracted based on training samples. The training samples include user writing samples and content samples. Content samples can be divided based on content types. Content samples of each content type can also be divided based on scene types. The scene types are defined based on the purpose of users consuming content samples. The sub-towers in the item tower correspond to the content types, and the private towers correspond to the scene types included in the content types. Therefore, after extracting cross-content and cross-scene domain training features based on training samples and inputting them into the hybrid multi-tower model, each sub-tower and private tower can fully process content features of different types and scenes. Dynamic masking is used during the training process to avoid negative transfer problems and improve the model recall performance.

[0049] After introducing the basic principles of the present invention, various non-limiting embodiments of the present invention are described in detail below.

[0050] Exemplary application scenarios

[0051] It should be noted that the following application scenarios are only shown to facilitate understanding of the spirit and principle of the present invention, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario.

[0052] The hybrid multi-tower model training method of the disclosed embodiment can be applied to a variety of application scenarios involving recommendation and search in communication networks.

[0053] In one application scenario, a hybrid multi-tower model related to sounds, playlists, etc. corresponding to the blog business in a music platform can be trained, where user behavior samples can be behavioral data generated by the consumption of content such as sounds and playlists, such as playing, commenting, sharing, and collecting sounds and playlists, and content types are divided in the form of sounds and playlists. The scene type can be determined based on the purpose or intention of the user when consuming sounds and playlists, such as the search scene type for purposeful acquisition, the recommendation scene type for purposeless exploration, and so on.

[0054] In another application scenario, it may involve hybrid multi-tower model training related to videos, video collections, video favorites, etc. in video platforms; it may also involve hybrid multi-tower model training related to articles, article collections, article favorites, etc. in information platforms; it may also involve hybrid multi-tower model training related to products, shopping carts, etc. in e-commerce platforms, and the present disclosed embodiments do not impose specific restrictions on this.

[0055] Exemplary Methods

[0056] In combination with the above application scenarios, Figure 1 The hybrid multi-tower model training method according to an exemplary embodiment of the present disclosure is described. The hybrid multi-tower model includes a user tower and an item tower, the item tower includes at least one sub-tower, and the sub-tower includes a public feature tower and a private tower.

[0057] Figure 1 shows a schematic diagram of the structure of a hybrid multi-tower model of an exemplary embodiment of the present disclosure. It includes a user tower and an item tower. The user tower can process user-related features and output user prediction results, and the item tower can process features related to the recommended content and output item prediction results. On this basis, the hybrid multi-tower model can output the correlation calculation result between the user prediction result and the item prediction result to determine the recommended content for the user. Among them, the item tower also includes multiple sub-towers, each sub-tower includes a public feature tower and a private tower, so that more diverse features of the recommended content can be discovered and processed. In addition, the number of sub-towers and private towers can also be set and expanded according to actual business characteristics, so as to meet the analysis needs of different business scenarios.

[0058] like Figure 2 One of the step flow charts of the hybrid multi-tower model training method according to the exemplary embodiment of the present disclosure may include the following steps 201 to 203:

[0059] Step 201, extracting training features based on training samples; the training samples include user behavior samples and content samples; the content samples include at least one content type, and each content type includes at least one scene type; the scene types are divided based on the purpose of the user consuming the content sample.

[0060] In the disclosed embodiment, a training feature that can be recognized, analyzed and processed by a hybrid multi-tower model can be constructed based on training samples. Among them, the training samples may include user behavior samples and content samples. The user behavior sample is the behavior data generated by the consumption behavior of the sample user on the content sample, and the consumption behavior may include browsing, playing, liking, commenting, sharing, forwarding, collecting, etc.; the content sample includes the consumable content provided to the sample user.

[0061] On this basis, content samples can be divided based on content type. Content samples of each content type can also be divided into corresponding scene types according to the purpose of the user consuming the content sample. For example, if the sample user plays sound type content 1 through search, a user behavior sample is generated to indicate that the user played content 1. A content sample includes that the content type of content 1 is sound, and it was played once during search; if the sample user plays sound type content 1 through recommendation, a user behavior sample is generated to indicate that the user played content 1. A content sample includes that the content type of content 1 is sound, and it was played once during recommendation. The same can be said for other content samples. Therefore, training features related to user behavior can be extracted based on training samples, and training features corresponding to different content types can be extracted based on content samples, and the training features can further correspond to different scene types.

[0062] Step 202: input the training features into the hybrid multi-tower model to be trained; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type.

[0063] In the disclosed embodiment, the training features can be input into the hybrid multi-tower model to be trained so that the hybrid multi-tower model can fully understand and process them and output the corresponding training prediction results. Wherein, the sub-tower corresponds to the content type, and the training features of different content types can be processed respectively. Each sub-tower also includes a public feature tower and a private tower. The private tower corresponds to the scene type included in the content type. For example, if a content sample is content type 1, the training features related to the content sample are processed by the sub-tower 1 corresponding to the content type 1. On this basis, the private tower 1 of the sub-tower 1 can process the training features related to the scene type 1 extracted based on the content sample, the private tower 2 can process the training features related to the scene type 2 extracted based on the content sample, and the private tower 3 can process the training features related to the scene type 3 extracted based on the content sample. The disclosed embodiment does not impose specific restrictions on this.

[0064] Step 203: dynamically mask the private tower based on the scene type, and train in the direction that the similar distance between the output of the user tower and the item tower conforms to the training sample to obtain a hybrid multi-tower model.

[0065] In the disclosed embodiment, masking can indicate the parts to be paid attention to or ignored during model training. By dynamically masking private towers based on scene types, partial ignoring can be indicated between private towers during model training to avoid mutual interference or inhibition between private towers, which results in the inability to fully understand and process the features related to each scene type in the training features of the multi-scene domain, and the inability to accurately identify and simulate the purpose of the user's consumption of content samples, thereby affecting the recall performance of the trained model.

[0066] On this basis, the outputs of the user tower and the item tower are obtained respectively, and the similarity distance of their outputs is calculated. Under normal circumstances, the recommendation result can be determined based on the similarity distance. For example, when the similarity distance of the outputs of the user tower and the item tower is greater than or equal to a certain threshold, it can be considered to meet the recommendation conditions. Among them, meeting the training sample refers to meeting the behavior record of the sample user on the content sample. When the similarity distance output by the user tower and the item tower indicates that it does not meet the recommendation conditions, but the training sample includes the consumption behavior of the sample user on the content sample, the hybrid multi-tower model can be adjusted in this direction so that the similarity distance of the outputs of the user tower and the item tower after adjustment is updated in the direction that meets the recommendation conditions. The hybrid multi-tower model is obtained based on the above-mentioned model training method.

[0067] like Figure 3 The second step flow chart of the hybrid multi-tower model training method according to the exemplary embodiment of the present disclosure may include the following steps 301 to 306:

[0068] Step 301: Extract initial features from training samples based on feature fields through a large language model; feature fields are constructed based on business characteristics. Training samples include user behavior samples and content samples; content samples include at least one content type, and each content type includes at least one scene type; scene types are divided based on the purpose of users consuming content samples.

[0069] In the disclosed embodiment, a large language model (LLM) can extract initial features in vector form from data such as text, pictures, and audio to obtain voice information that is convenient for model analysis and processing. The large language model can adapt to different sample data and feature construction requirements according to the source, structure, and function. For example, the Baichuan large language model can be used to extract the initial features of text, and the ViT (Vision Transformer) large language model can be used to extract the initial features of pictures. Those skilled in the art can specifically select the large language model to be used for feature extraction.

[0070] The feature field is used to represent the multimodal features in the training samples. The multimodal features can be selected according to the business characteristics by evaluating the influence of the features on the user's consumption behavior, the degree of influence and other factors, and then selected. For example, when distributing audio content in the podcast business, the consumption behaviors such as clicking and playing the audio content are evaluated to be highly influenced by factors such as the audio title, introduction and cover, and the consumption behaviors such as pausing, exiting, finishing, collecting, sharing and commenting on the audio content are highly influenced by factors such as content information and audio quality; or, when distributing the playlist of the podcast business, the consumption behaviors such as clicking on the playlist content are evaluated to be highly influenced by factors such as the audio title and cover, and whether to play a specific sound on the playlist page is highly influenced by the introduction and the title of the front row sound, etc. On this basis, the multimodal features including title, introduction, cover, etc. can be determined based on the business characteristics, and the feature field pointing to the multimodal features can be constructed to avoid the large language model from being able to extract the corresponding initial features based on the feature field.

[0071] Step 302: construct a hub unit based on the initial features of the user behavior samples, construct heterogeneous units corresponding to different content types based on the initial features of the content samples, and construct edges between the hub unit and the heterogeneous units based on the behavioral relationship between the user behavior samples and the content samples to obtain a heterogeneous graph model.

[0072] In the disclosed embodiment, different content samples of different content types may have different consumption behaviors in the business, and there may also be inclusion relationships between content samples, such as consumption behaviors such as clicking, playing, commenting, and liking audio content, and clicking and collecting playlist content, and audio content may belong to at least one playlist content, thereby constructing relationships between different consumption behaviors. On this basis, the hub unit of the heterogeneous graph model can be constructed with the initial features of the user behavior samples, and the heterogeneous units corresponding to different content types can be constructed with the initial features of the content samples, and the edges between the hub unit and the heterogeneous unit can be constructed with the behavioral relationship between the user behavior samples of the sample users and the content samples, so as to construct a heterogeneous graph model.

[0073] like Figure 4A schematic diagram of a heterogeneous graph model of an exemplary embodiment of the present disclosure is shown. Among them, three user behavior samples are respectively extracted through a large language model to obtain corresponding initial features 1, five content samples of content type 1 are respectively extracted through a large language model to obtain corresponding initial features 2, and four content samples of content type 2 are respectively extracted through a large language model to obtain corresponding initial features 3. Based on the initial feature 1, hub units U1, U2 and U3 corresponding to the hub are constructed, based on the initial feature 2, heterogeneous units C11, C12, C13, C14 and C15 corresponding to content type 1 are constructed, and based on the initial feature 3, heterogeneous units C21, C22, C23 and C24 corresponding to content type 2 are constructed, and further edges between hub units and heterogeneous units are constructed based on the behavioral relationship between user behavior samples of sample users and content samples, to obtain a heterogeneous graph model such as Figure 4 shown.

[0074] Step 303: perform meta-path sampling in the heterogeneous graph model, adjust the initial features based on the sampling results, and obtain sampling features.

[0075] In the embodiments of the present disclosure, on the basis of obtaining a heterogeneous graph model, a meta-path can be constructed for sampling, the semantic and structural relationships between different units can be extracted, and the user behavior information and behavior relationship chain information extracted by the large language model can be fitted, so that the sampling features can fully combine the characteristics of the behavior and the sample, and deeply explore the upper and lower connections of the behavior chain. Based on the sampling results of the meta-path sampling, the initial features extracted by the aforementioned large language model are adjusted and the behavior is calibrated to obtain the sampling features, so that the sampling features are more consistent with the action representation in the business scenario, and the accuracy of the consumer behavior representation is improved, thereby improving the recall accuracy of the trained model. Among them, the construction method of the meta-path can be set according to the implementation conditions and recall requirements. For example, the meta-path can be constructed using the sampling method of metapath2vec. The embodiments of the present disclosure do not impose specific restrictions on this.

[0076] In an optional method embodiment of the present disclosure, the construction condition of the meta-path in step 303 includes that training samples of different content types appear symmetrically and alternately in the meta-path.

[0077] In the embodiment of the present disclosure, the training samples of different content types in the constructed conditional meta-path appear symmetrically and alternately, which can be expressed as a distribution of "content type 1-content type 2-content type 1-content type 2". Figure 4 The meta-path constructed in the heterogeneous graph model shown can be C11-C23-C13-C24, or C21-C13-C23, etc., so that the heterogeneous units corresponding to content type 1 and content type 2 in the meta-path are symmetrically sampled alternately.

[0078] In an optional method embodiment of the present disclosure, the construction condition of the meta-path in step 303 includes training samples whose creation time is within the historical time limit or before the historical time limit, and training samples whose creation time is after the historical time limit appearing symmetrically in the meta-path.

[0079] In an optional method embodiment of the present disclosure, the construction condition of the meta-path in step 303 includes that the accumulated user behavior data is less than or equal to a preset number of training samples and the accumulated user behavior data is greater than the preset number of training samples appear symmetrically in the meta-path.

[0080] In the embodiment of the present disclosure, the construction of the meta-path can also be set for the heterogeneous units corresponding to the new shallow content samples. Among them, it can be considered that the samples created within the historical time limit or before the historical time limit are old samples (such as Figure 4 N in the figure), a new sample whose creation time is after the historical time limit; or a new sample whose accumulated user behavior data is less than or equal to the preset amount, and an old sample whose accumulated user behavior data is greater than the preset amount (such as Figure 4 In the meta-path, the heterogeneous units corresponding to the new and old samples can appear alternately and symmetrically, such as "old sample-new sample-old sample-new sample", so that the sampling features can fully mine the new and shallow content in the feature construction stage, reduce the influence of the hot effect, and improve the diversity of the recall results of the trained model.

[0081] In the embodiments of the present disclosure, the settings of indicators such as historical time limits and preset quantities can be customized according to the distribution of training samples, data processing conditions, and actual recall requirements. For example, if the creation time of the training samples is distributed within 10 days, the zero o'clock of the 10th day can be set as the historical time limit, or if the accumulated user behavior data is distributed within the range of 0 to 100, 10 items can be set as the preset number. The embodiments of the present disclosure do not impose specific restrictions on this.

[0082] Step 304: Initialize the sampling features and construct training features corresponding to the training samples.

[0083] In the disclosed embodiment, the obtained sampling features may be initialized to facilitate subsequent analysis and understanding of the model and distance calculation, thereby obtaining training features corresponding to the training samples.

[0084] In an optional method embodiment of the present disclosure, the training features do not include identity identification number ID features.

[0085] Generally, ID features are easy to collect and analyze, and therefore are widely used in model training and applications, such as using ID to represent the content when representing the user's consumption behavior of the content in the user behavior sequence feature, or using ID to represent the content in the content feature. However, the ID feature may cause the model to pay too much attention to the relevant information of the ID, while weakening the importance of the generalized features and deepening the popularity effect. In the disclosed embodiment, the ID is removed from the training features, and the ID feature is not added. The initialized sampled features are used as the main features for content representation, thereby strengthening the model's attention to the generalized features of the content, which can effectively reduce the popularity effect caused by the focus on the old content ID.

[0086] In an optional method embodiment of the present disclosure, step 304 may include the following steps A1 to A2:

[0087] Step A1: When the training sample is a user behavior sample, it is initialized based on the sampling features, and the user attribute features are added to obtain the training features.

[0088] In the embodiment of the present disclosure, for the user behavior samples in the training samples, the sampling features can be initialized to construct the behavior sequence features of the sample users, and basic user attribute features can be added, such as gender, age, etc. Among them, the behavior sequence features of the sample users represent the content with the sampling features corresponding to the content when representing the user's consumption behavior of the content, rather than representing the content with the ID features.

[0089] Step A2: When the training sample is a content sample, it is initialized based on the sampling feature, and the label feature of the content sample and the statistical feature of the content sample under the corresponding scene type are added to obtain the training feature.

[0090] In the embodiment of the present disclosure, in feature selection, for the content samples in the training samples, the sampling features can be initialized as the main features of the content samples. On this basis, the label features of the content samples can be added, such as content labels, primary classification labels, secondary classification labels, etc. Similarly, the label features of the sound content can include its content labels, such as children's songs, love songs, pure music, etc., the primary classification labels can include pure music, and the secondary classification labels can include electronic music, classical music, etc., and the present disclosure does not make specific restrictions on this.

[0091] Furthermore, we can also extract statistical features corresponding to content samples of different scene types. The statistical features can be the total number, frequency, and number of people of different consumption behaviors. Since users have different purposes for consuming content in different scene types, the distribution of the same content sample in different scene types will be quite different. Therefore, the feature analysis across scene domains may have negative transfer problems that mislead the model direction. We can extract statistical features of different scene types as training features.

[0092] In an optional method embodiment of the present disclosure, the scenario type includes at least one of a recommendation scenario, a revisit scenario, and a search scenario.

[0093] In the embodiments of the present disclosure, the scene types may include recommendation scenes, that is, recommending content to users without providing specific information, and the users then consume the recommended content; they may also include search scenes, in which relevant content based on the key information search is returned to the users when the users provide key information, and the users then consume the relevant content of the key information; they may also include revisit scenes, that is, the content that the users have not consumed in other scenes is consumed by repeated visits. The above scene types are only used for examples, and those skilled in the art may determine the scene types based on actual sample distribution, training conditions, and recall requirements, etc., and the embodiments of the present disclosure do not impose specific restrictions on this.

[0094] Step 305: input the training features into the hybrid multi-tower model to be trained; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type.

[0095] In the embodiment of the present disclosure, step 305 may correspond to the relevant description of the aforementioned step 202, and will not be described again here to avoid repetition.

[0096] In an optional method embodiment of the present disclosure, on the basis of step A1 to step A2, step 305 may specifically include the following step A3:

[0097] Step A3: In the hybrid multi-tower model to be trained, the initialized sampling features and user attribute features are input into the user tower; and the initialized sampling features and label features are input into the public feature tower according to the content type corresponding to the sub-tower, and the statistical features are input into the private tower corresponding to the scene type.

[0098] In the disclosed embodiment, the training features are input into the hybrid multi-tower model to be trained, the initialized sampling features and user attribute features corresponding to the user behavior samples are input into the user tower, and the user tower uses the initialized sampling features as the behavior sequence features for end-to-end training; sub-towers are constructed in the item tower for the heterogeneous characteristics of different content types, and private towers of sub-towers are constructed based on the scene domain characteristics. Among them, the initialized sampling features and label features corresponding to the content samples are input into the public feature tower, and the statistical features are input into the private tower corresponding to the scene type, and the comprehensive heterogeneous characteristics biased towards the content features and the cross-scene domain model training are carried out to improve the recall effect of the model.

[0099] Step 306: Dynamically mask the private tower based on the scene type, and train in the direction that the similar distance between the output of the user tower and the item tower conforms to the training sample to obtain a hybrid multi-tower model.

[0100] In the embodiment of the present disclosure, step 306 may correspond to the relevant description of the aforementioned step 203, and will not be described again here to avoid repetition.

[0101] In an optional method embodiment of the present disclosure, the content type may include audio content and playlist content.

[0102] In the podcast business, content types may include audio content and playlist content. Figure 5 The schematic diagram of the hybrid multi-tower model applied to the podcast service according to the exemplary embodiment of the present disclosure is shown, wherein the item tower may include a sub-tower corresponding to the sound content, and may include a sub-tower corresponding to the playlist content. Figure 6 The third step flow chart of the hybrid multi-tower model training method according to the exemplary embodiment of the present disclosure may include the following steps 601 to 604:

[0103] Step 601, extracting training features based on training samples; the training samples include user behavior samples and content samples; the content samples include at least one content type, and each content type includes at least one scene type; the scene types are divided based on the purpose of the user consuming the content sample.

[0104] In the embodiment of the present disclosure, step 601 may correspond to the relevant description of the aforementioned step 201, or steps 301 to 304, which will not be repeated here to avoid repetition.

[0105] Step 602: Input the training features into the hybrid multi-tower model to be trained; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type.

[0106] In the embodiment of the present disclosure, step 602 may correspond to the relevant description of the aforementioned step 202 or step 305, and will not be described again here to avoid repetition.

[0107] like Figure 5 As shown, in the hybrid multi-tower model to be trained, the user tower performs end-to-end processing, and its input includes basic user attribute features and behavior sequence features. In the behavior sequence features, the initialized sampling features represent the sound content or playlist content consumed by the sample user; in the sub-tower of the sound content, the shared feature tower can process the sampling features corresponding to the initialized sound content, and the private tower corresponding to the scene type processes the statistical features under the scene type; and, in the sub-tower of the playlist content, the public feature tower can process the sampling features corresponding to the initialized playlist content, and the private tower corresponding to the scene type processes the statistical features under the scene type;

[0108] Step 603: dynamically mask the private tower based on the scene type, and use the output of the sub-tower corresponding to the playlist content as the input of the sub-tower corresponding to the sound content, to obtain the output of the user tower and the item tower.

[0109] In the disclosed embodiment, based on the characteristics of the podcast service, the sound content is usually included in at least one playlist. Therefore, the output of the sub-tower corresponding to the playlist content can also be used as the input of the sub-tower corresponding to the sound content. Through joint training between the sub-towers, it is possible to avoid misleading the training of the sound content due to the popularity effect of the playlist content itself. For example, if a certain sound content is included in two playlist contents, and one of the playlist contents is a popular playlist, a large number of users will try to consume the sound content, while the other playlist content has fewer consumption records. At this time, there may be a problem of attributing a large number of consumption behaviors brought by the popular playlist to the sound content itself, thereby misleading the model analysis. Through joint training, in-depth analysis can be performed and the correct attribution can be made to the playlist attributes, thereby further improving the model recall performance.

[0110] like Figure 5 As shown, the statistical features under the scene type can also be input into the user tower for processing to characterize the impact of the scene type on the user behavior sequence characteristics, and the statistical features of the scene type are also calculated together when calculating the similarity distance output by the user tower and the item tower to further characterize the impact of the scene type on the degree of relevance.

[0111] Step 604: Perform training in a direction where the similar distance between the outputs of the user tower and the item tower matches the training sample, and obtain a hybrid multi-tower model.

[0112] In the embodiment of the present disclosure, step 604 may correspond to the relevant description of the aforementioned step 203 or step 306, and will not be described again here to avoid repetition.

[0113] According to the hybrid multi-tower model training method of the embodiment of the present disclosure, the training samples include user behavior samples and content samples, the content samples include at least one content type, and each content type includes at least one scene type, and the scene type is divided based on the purpose of the user consuming the content sample; on this basis, the training features extracted based on the training samples are input into the hybrid multi-tower model to be trained, and the private tower is dynamically masked based on the scene type, and the training is carried out in the direction of the training sample in which the similar distance between the output of the user tower and the item tower conforms to the training sample, so as to obtain a hybrid multi-tower model; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type. The hybrid multi-tower model trained by this method can realize multi-domain training and analysis of heterogeneous content across scene domains through multiple sub-towers of the item tower corresponding to different content types, and private towers under the sub-towers corresponding to different scene types included in the content type, so as to avoid feature interference between content of different types and scenes, and more fully analyze the content characteristics, so as to more deeply and comprehensively mine new shallow content and long-tail content. On this basis, by dynamically masking private towers under the training features of multiple scene types to fit the user consumption content samples, it is also possible to avoid the negative transfer problem caused by different user distribution and different sample sources in cross-domain training, fully explore the characteristic information of diverse businesses, and thus improve the accuracy of model recall.

[0114] like Figure 7 A flowchart of the steps of a recommended method of an exemplary embodiment of the present disclosure is shown, the method is based on Figure 2 , 3 And the hybrid multi-tower model obtained by training the hybrid multi-tower model training method described in any one of 6 is realized, and the hybrid multi-tower model is as follows Figure 1 , 5 As shown, the method may include the following steps 701 to 703:

[0115] Step 701: Obtain behavior data and content data of a target user, and extract prediction features based on the behavior data and the content data; the content data includes at least one content type, and each content type includes at least one scene type.

[0116] In the embodiment of the present disclosure, when the hybrid multi-tower model obtained through training is applied, the behavior data of the target user can be obtained, describing the historical consumption behavior of the target user on the consumed content, and the content data that can be provided to the target user. On this basis, the prediction features are constructed based on the behavior data and the content data, which can correspond to the relevant description of extracting the training features based on the training samples in the aforementioned step 201, or steps 301 to 304. To avoid repetition, it will not be repeated here.

[0117] Step 702: Input the prediction features into the hybrid multi-tower model.

[0118] In the embodiment of the present disclosure, the prediction features are input into the hybrid multi-tower model, which can correspond to the relevant description of the aforementioned step 202 or step 305, and will not be described again here to avoid repetition.

[0119] Step 703: Determine content data to be provided to the target user based on the similarity distance between the outputs of the user tower and the item tower in the hybrid multi-tower model.

[0120] In the embodiment of the present disclosure, the hybrid multi-tower model can further determine the similarity distance based on the outputs of the user tower and the item tower. When the similarity distance meets the recommendation condition, the corresponding content data is provided to the target user.

[0121] According to the recommendation method of the embodiment of the present disclosure, a hybrid multi-tower model trained by a hybrid multi-tower model training method is implemented. The training samples include user behavior samples and content samples, the content samples include at least one content type, and each content type includes at least one scene type, and the scene type is divided based on the purpose of the user consuming the content sample; on this basis, the training features extracted based on the training samples are input into the hybrid multi-tower model to be trained, and the private tower is dynamically masked based on the scene type, and the training is performed in the direction of the training sample in which the similar distance between the output of the user tower and the item tower conforms to the training sample, so as to obtain a hybrid multi-tower model; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type. The hybrid multi-tower model trained by this method can realize multi-domain training and analysis of heterogeneous content across scene domains through multiple sub-towers of the item tower corresponding to different content types, and the private towers under the sub-towers corresponding to different scene types included in the content type, so as to avoid feature interference between content of different types and scenes, and more fully analyze the content characteristics, so as to more deeply and comprehensively mine new shallow content and long-tail content. On this basis, by dynamically masking private towers under the training features of multiple scene types to fit the user consumption content samples, it is also possible to avoid the negative transfer problem caused by different user distribution and different sample sources in cross-domain training, fully explore the characteristic information of diverse businesses, and thus improve the accuracy of model recall.

[0122] Exemplary Devices

[0123] After introducing the hybrid multi-tower model training method and recommendation method of the exemplary embodiment of the present disclosure, next, refer to Figure 8 The hybrid multi-tower model training device of the exemplary embodiment of the present disclosure is described, and reference is made to Fig. 9 A recommended device according to an exemplary embodiment of the present disclosure is described.

[0124] It should be noted that other specific details of the various functional modules of the hybrid multi-tower model training device and recommendation device of the embodiment of the present disclosure have been described in detail in the implementation of the above-mentioned hybrid multi-tower model training method and recommendation method, and will not be repeated here.

[0125] Figure 8 A hybrid multi-tower model training device 800 according to an exemplary embodiment of the present disclosure is shown. The hybrid multi-tower model includes a user tower and an item tower. The item tower includes at least one sub-tower. The sub-tower includes a public feature tower and a private tower. The device may include: a feature extraction module 801, used to extract training features based on training samples; the training samples include user behavior samples and content samples; the content samples include at least one content type, and each content type includes at least one scene type; the scene types are divided based on the purpose of the user consuming the content samples; a model training module 802, used to input the training features into the hybrid multi-tower model to be trained; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type; the model training module 802 is also used to dynamically mask the private tower based on the scene type, and train in a direction that matches the training sample with a similar distance between the outputs of the user tower and the item tower to obtain a hybrid multi-tower model.

[0126] In an optional device embodiment of the present disclosure, the feature extraction module 801 is specifically used to extract initial features from training samples based on feature fields through a large language model; the feature fields are constructed based on business characteristics; a heterogeneous graph model is constructed with user behavior samples as hub units and training samples of different content types as heterogeneous units, and each training sample in the graph model corresponds to an initial feature; meta-path sampling is performed in the heterogeneous graph model, and the initial features are adjusted based on the sampling results to obtain sampled features; the sampled features are initialized to construct training features corresponding to the training samples.

[0127] In an optional device embodiment of the present disclosure, the construction conditions of the meta-path include at least one of the following: training samples of different content types appear symmetrically and alternately in the meta-path; training samples whose creation time is within the historical time limit or before the historical time limit and training samples whose creation time is after the historical time limit appear symmetrically and alternately in the meta-path; training samples whose accumulated user behavior data is less than or equal to a preset number and training samples whose accumulated user behavior data is greater than a preset number appear symmetrically and alternately in the meta-path.

[0128] In an optional device embodiment of the present disclosure, the feature extraction module 801 is specifically used to initialize based on the sampling features and add user attribute features to obtain training features when the training samples are user behavior samples; and to initialize based on the sampling features and add label features of the content samples and statistical features of the content samples under the corresponding scene types to obtain training features when the training samples are content samples.

[0129] In an optional device embodiment of the present disclosure, the model training module 802 is specifically used to input the initialized sampling features and user attribute features into the user tower in the hybrid multi-tower model to be trained; and input the initialized sampling features and label features into the public feature tower according to the content type corresponding to the sub-tower, and input the statistical features into the private tower corresponding to the scene type.

[0130] In an optional device embodiment of the present disclosure, the training features do not include identity identification number ID features.

[0131] In an optional device embodiment of the present disclosure, the content type includes audio content and playlist content.

[0132] In an optional device embodiment of the present disclosure, the model training module 802 is specifically used to dynamically mask the private tower based on the scene type, and use the output of the sub-tower corresponding to the playlist content as the input of the sub-tower corresponding to the sound content to obtain the output of the user tower and the item tower; train with the similar distance between the outputs of the user tower and the item tower in the direction of the training sample to obtain a hybrid multi-tower model.

[0133] In an optional device embodiment of the present disclosure, the scenario type includes at least one of a recommendation scenario, a revisit scenario, and a search scenario.

[0134] According to the hybrid multi-tower model training device of the embodiment of the present disclosure, the hybrid multi-tower model includes a user tower and an item tower, the item tower includes at least one sub-tower, and the sub-tower includes a public feature tower and a private tower. The method extracts training features from training samples, the training samples include user behavior samples and content samples, the content samples include at least one content type, and each content type includes at least one scene type, and the scene type is divided based on the purpose of the user consuming the content sample; on this basis, the training features extracted based on the training samples are input into the hybrid multi-tower model to be trained, and the private tower is dynamically masked based on the scene type, and the training is performed in the direction that the similar distance between the output of the user tower and the item tower conforms to the training sample, so as to obtain a hybrid multi-tower model; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type. The hybrid multi-tower model trained by this method can achieve multi-domain training and analysis of heterogeneous content across scene domains through multiple sub-towers of the item tower corresponding to different content types, and private towers under the sub-towers corresponding to different scene types of content types, avoiding feature interference between content of different types and scenes, and more fully analyzing content characteristics, so as to more deeply and comprehensively mine new shallow content and long-tail content. On this basis, by dynamically masking private towers under the training characteristics of multiple scene types to fit the purpose of user consumption content samples, it can also avoid negative transfer problems caused by different user distributions and different sample sources in cross-domain training, fully mine the feature information of diverse businesses, and thus improve the accuracy of model recall.

[0135] Fig. 9 The recommendation device 900 according to an exemplary embodiment of the present disclosure is shown. Figure 8 The hybrid multi-tower model obtained by training the disclosed model training device 800 is implemented, and the device 900 may include: a feature extraction module 901, used to obtain the behavior data and content data of the target user, and extract predictive features based on the behavior data and the content data; the content data includes at least one content type, and each content type includes at least one scene type; a model input module 902, used to input the predictive features into the hybrid multi-tower model; a content recommendation module 903, used to determine the content data provided to the target user based on the similarity distance between the outputs of the user tower and the item tower in the hybrid multi-tower model.

[0136] According to the recommendation device of the embodiment of the present disclosure, the hybrid multi-tower model trained by the hybrid multi-tower model training device is implemented. The training samples include user behavior samples and content samples, the content samples include at least one content type, and each content type includes at least one scene type, and the scene type is divided based on the purpose of the user consuming the content sample; on this basis, the training features extracted based on the training samples are input into the hybrid multi-tower model to be trained, and the private tower is dynamically masked based on the scene type, and the training is carried out in the direction of the training sample in which the similar distance between the output of the user tower and the item tower conforms to the training sample, so as to obtain a hybrid multi-tower model; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type. The hybrid multi-tower model trained by this method can realize multi-domain training and analysis of heterogeneous content across scene domains through multiple sub-towers of the item tower corresponding to different content types, and the private towers under the sub-towers corresponding to different scene types included in the content type, so as to avoid feature interference between content of different types and scenes, and more fully analyze the content characteristics, so as to more deeply and comprehensively mine new shallow content and long-tail content. On this basis, by dynamically masking private towers under the training features of multiple scene types to fit the user consumption content samples, it is also possible to avoid the negative transfer problem caused by different user distribution and different sample sources in cross-domain training, fully explore the characteristic information of diverse businesses, and thus improve the accuracy of model recall.

[0137] It should be noted that although several modules or units of the hybrid multi-tower model training device and the recommendation device are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be concretized in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into multiple modules or units to be concretized.

[0138] Exemplary computer readable media

[0139] The following is a description of a computer-readable medium according to an exemplary embodiment of the present disclosure.

[0140] In this exemplary embodiment, reference Fig.10 As shown, a program product 1000 for implementing the above method according to an exemplary embodiment of the present disclosure is described, such as a portable compact disk read-only memory (CD-ROM) and including program code, and can be run on a device, such as a personal computer. However, the program product of the present disclosure is not limited thereto. In this document, a computer-readable medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus, or a device.

[0141] The program product 1000 may use any combination of one or more readable media. The readable medium may be a readable signal medium or a computer readable medium. The computer readable 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 (non-exhaustive list) of computer readable media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0142] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, wherein readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than computer readable media, which may send, propagate, or transmit programs for use by or in conjunction with an instruction execution system, apparatus, or device.

[0143] The program code contained on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RE, etc., or any suitable combination of the foregoing.

[0144] Program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user computing device, partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (FAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., using an Internet service provider to connect through the Internet).

[0145] Exemplary Electronic Devices

[0146] refer to Fig.11 An electronic device according to an exemplary embodiment of the present disclosure is described.

[0147] Fig.11 The electronic device 1100 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0148] like Fig.11 As shown, the electronic device 1100 is in the form of a general computing device. The components of the electronic device 1100 may include, but are not limited to: at least one processing unit 1110, at least one storage unit 1120, a bus 1130 connecting different system components (including the storage unit 1120 and the processing unit 1110), and a display unit 1140.

[0149] The storage unit stores program codes, which can be executed by the processing unit 1110, so that the processing unit 1110 executes the steps according to various exemplary embodiments of the present disclosure described in the above “Exemplary Method” section of this specification. For example, the processing unit 1110 can execute the following steps: Figure 2 , 3 and the method steps shown in 6, etc.

[0150] The storage unit 1120 may include a volatile storage unit, such as a random access storage unit (RAM) 1121 and / or a cache storage unit 1122 , and may further include a read-only storage unit (ROM) 1123 .

[0151] The storage unit 1120 may also include a program / utility 1124 having a set (at least one) of program modules 1125, such program modules 1125 including but not limited to: an operating system, one or more server applications, other program modules and program data, each of which or some combination may include an implementation of a network environment.

[0152] The bus 1130 may include a data bus, an address bus, and a control bus.

[0153] The electronic device 1100 may also communicate with one or more external devices 1200 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), and such communication may be performed through an input / output (I / O) interface 1150. The electronic device 1100 also includes a display unit 1140, which is connected to the input / output (I / O) interface 1150 for display. In addition, the electronic device 1100 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) through a network adapter 1160. As shown, the network adapter 1160 communicates with other modules of the electronic device 1100 through a bus 1130. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 1100, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0154] It should be noted that, although several modules or submodules of the device are mentioned in the above detailed description, such division is merely exemplary and not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided into multiple units / modules to be embodied.

[0155] In addition, although the operations of the disclosed method are described in a specific order in the drawings, this does not require or imply that the operations must be performed in this specific order, or that all the operations shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0156] Although the spirit and principle of the present disclosure have been described with reference to several specific embodiments, it should be understood that the present disclosure is not limited to the disclosed specific embodiments, and the division of various aspects does not mean that the features in these aspects cannot be combined to benefit, and such division is only for the convenience of expression. The present disclosure is intended to cover various modifications and equivalent arrangements included in the spirit and scope of the attached claims.

Claims

1. A hybrid multi-tower model training method, characterized in that: The hybrid multi-tower model includes a user tower and an item tower, the item tower includes at least one sub-tower, the sub-tower includes a public feature tower and a private tower, and the method includes: Extracting training features based on training samples; the training samples include user behavior samples and content samples; the content samples include at least one content type, and each of the content types includes at least one scene type; the scene types are divided based on the purpose of the user consuming the content sample; Inputting the training features into a hybrid multi-tower model to be trained; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type; The private tower is dynamically masked based on the scene type, and training is performed in a direction that matches the training sample with a similar distance between the outputs of the user tower and the item tower to obtain the hybrid multi-tower model.

2. The method according to claim 1, characterized in that The extracting of training features based on training samples includes: Extracting initial features from the training sample based on feature fields through a large language model; the feature fields are constructed based on business characteristics; Constructing a hub unit based on the initial features of the user behavior sample, constructing heterogeneous units corresponding to different content types based on the initial features of the content sample, and constructing edges between the hub unit and the heterogeneous units based on the behavior relationship between the user behavior sample and the content sample to obtain a heterogeneous graph model; Performing meta-path sampling in the heterogeneous graph model, and adjusting the initial features based on the sampling results to obtain sampling features; The sampling features are initialized to construct training features corresponding to the training samples.

3. The method according to claim 2, characterized in that The construction conditions of the meta-path include at least one of the following: The training samples of different content types appear alternately and symmetrically in the meta-path; The training samples whose creation time is within the historical time limit or before the historical time limit and the training samples whose creation time is after the historical time limit appear alternately and symmetrically in the meta-path; The training samples whose accumulated user behavior data is less than or equal to a preset number and the training samples whose accumulated user behavior data is greater than the preset number appear alternately and symmetrically in the meta-path.

4. The method according to claim 2, characterized in that: The initialization based on the sampling features to construct the training features corresponding to the training samples includes: In the case where the training sample is the user behavior sample, initialization is performed based on the sampling feature, and user attribute features are added to obtain the training feature; In the case where the training sample is the content sample, initialization is performed based on the sampling feature, and label features of the content sample and statistical features of the content sample corresponding to the scene type are added to obtain the training feature.

5. The method according to claim 4, characterized in that The step of inputting the training features into the hybrid multi-tower model to be trained comprises: In the hybrid multi-tower model to be trained, the initialized sampling features and the user attribute features are input into the user tower; and the initialized sampling features and the label features are input into the public feature tower according to the content type corresponding to the sub-tower, and the statistical features are input into the private tower corresponding to the scene type.

6. A recommendation method, characterized in that: The method is implemented based on a hybrid multi-tower model trained by the hybrid multi-tower model training method according to any one of claims 1 to 5, and the method comprises: Obtaining behavior data and content data of a target user, and extracting prediction features based on the behavior data and the content data; the content data includes at least one content type, and each of the content types includes at least one scene type; inputting the predicted features into the hybrid multi-tower model; The content data provided to the target user is determined based on the similarity distance between the outputs of the user tower and the item tower in the hybrid multi-tower model.

7. A hybrid multi-tower model training device, characterized in that: The hybrid multi-tower model includes a user tower and an item tower, the item tower includes at least one sub-tower, the sub-tower includes a public feature tower and a private tower, and the device includes: A feature extraction module, configured to extract training features based on training samples; the training samples include user behavior samples and content samples; the content samples include at least one content type, and each of the content types includes at least one scene type; the scene types are divided based on the purpose of the user consuming the content sample; A model training module, used for inputting the training features into a hybrid multi-tower model to be trained; wherein the sub-tower corresponds to the content type, and the private tower corresponds to the scene type included in the content type; The model training module is also used to dynamically mask the private tower based on the scene type, and to perform training in a direction that matches the training sample with a similar distance between the outputs of the user tower and the item tower to obtain the hybrid multi-tower model.

8. A recommendation device, characterized in that: The device is implemented based on the hybrid multi-tower model trained by the hybrid multi-tower model training device according to claim 7, and the device comprises: A feature extraction module, used to obtain behavior data and content data of a target user, and extract prediction features based on the behavior data and the content data; the content data includes at least one content type, and each of the content types includes at least one scene type; A model input module, used for inputting the prediction features into the hybrid multi-tower model; The content recommendation module is used to determine the content data provided to the target user based on the similarity distance between the outputs of the user tower and the item tower in the hybrid multi-tower model.

9. A computer-readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the hybrid multi-tower model training method according to any one of claims 1 to 5, or the recommendation method according to claim 6 is implemented.

10. An electronic device, characterized in that: include: processor; as well as A memory, configured to store executable instructions of the processor; Wherein, the processor is configured to execute the hybrid multi-tower model training method described in any one of claims 1 to 5, or the recommendation method described in claim 6 by executing the executable instructions.

Images