Article recommendation method and device in target scene

Abstract

The embodiment of the invention discloses an article recommendation method and device in a target scene. A specific embodiment of the method comprises the steps of encoding an article image and an article text in a target scene, and generating to-be-recommended article features in the target scene; encoding the historical data of the user in the target scene to generate personal attribute features of the user in the target scene; the to-be-recommended article features and the user personal attribute features in the target scene are input into a recommendation model in the target scene, recommended article information is generated, and the recommendation model in the target scene takes the simulated to-be-recommended article features and the user personal attribute features in the target scene as input and generates recommended article information; and taking the simulated recommended item information in the target scene as target output, and performing training. According to the embodiment, the recommendation model in the target scene is utilized, and the item recommendation accuracy in the target scene is improved.

Description

Technical Field

[0001] Embodiments of the present disclosure relate to the field of computer technology, and in particular to a method and device for recommending items in a target scene. Background Art

[0002] The main task of the recommendation model is to distribute and filter information according to user interests to solve the problem of information overload. At present, the mainstream recommendation models in the industry will divide their recommendation processes into modules such as retrieval recall, rough sorting, and fine sorting. Among them, the retrieval recall module mainly completes the initial screening of the target set and limits the recommended set to a certain range. The sorting module mainly completes the refined sorting of the recall results to output a more accurate final result.

[0003] The function of the sorting module is to estimate the click rate of items. In the sorting module, transfer learning technology and multi-task learning can be used to improve the recommendation effect. Transfer learning refers to the transfer of knowledge from the source scene to the target scene to improve the performance of the recommendation model in the target scene. Such methods usually achieve cross-scene migration by sharing model parameters or mapping feature spaces. However, the generalization ability of transfer learning methods is affected by the difference in data distribution between the source scene and the target scene, and the effect is often limited in large-scale and diverse scenes.

[0004] Multi-task learning improves the generalization ability of the model in different tasks or scenarios by learning multiple related tasks at the same time. In the recommendation model, multi-task learning can achieve cross-scenario generalization by sharing feature representations or jointly optimizing recommendation tasks in multiple scenarios. However, this method may cause interference between tasks and reduce the recommendation effect when there are large differences between scenarios. Summary of the invention

[0005] The embodiments of the present disclosure provide a method and device for recommending items in a target scenario.

[0006] In a first aspect, an embodiment of the present disclosure provides a cross-generator training method, comprising: encoding an object image and an object text in a source scene to generate features of an object to be recommended in the source scene; encoding user history data in the source scene to generate user personal attribute features in the source scene; performing feature decoupling of the user personal attribute features in the source scene and features of the object to be recommended to generate a user personal preference feature encoding and an object information encoding in the source scene; inputting the user personal preference feature encoding and the object information encoding in the source scene into a cross-generator to generate a simulated object image in the source scene; calculating a similarity loss based on the simulated object image and the object image in the source scene; and adjusting parameters of the cross-generator based on the similarity loss until the similarity loss meets a preset condition.

[0007] In a second aspect, an embodiment of the present disclosure provides a method for generating a simulated image in a target scene, including: generating vocabulary in the target scene; encoding the vocabulary in the target scene to generate features of items to be recommended in the target scene; encoding user history data in the target scene to generate user personal attribute features in the target scene; feature decoupling the user personal attribute features in the target scene and features of items to be recommended to generate user personal preference feature codes and item information codes in the target scene; inputting the user personal preference feature codes and item information codes in the target scene into a cross generator to generate a simulated object image in the target scene.

[0008] In a third aspect, an embodiment of the present disclosure provides a training method for a recommendation model in a target scenario, including: encoding simulated object images and simulated object texts in the target scenario to generate simulated features of items to be recommended in the target scenario; encoding user history data in the target scenario to generate user personal attribute features in the target scenario; inputting the simulated features of items to be recommended in the target scenario and the user personal attribute features into the recommendation model in the target scenario to generate first predicted recommended item information; calculating a first loss based on the first predicted recommended item information and the simulated recommended item information; and adjusting parameters of the recommendation model in the target scenario based on the first loss until the first loss meets a first preset condition.

[0009] In a fourth aspect, an embodiment of the present disclosure provides an item recommendation method in a target scene, comprising: encoding item images and item texts in the target scene to generate features of items to be recommended in the target scene; encoding user history data in the target scene to generate user personal attribute features in the target scene; inputting the features of items to be recommended in the target scene and the user personal attribute features into a recommendation model in the target scene to generate recommended item information, wherein the recommendation model in the target scene is trained using simulated features of items to be recommended and user personal attribute features in the target scene as input, and using simulated recommended item information in the target scene as target output.

[0010] In a fifth aspect, an embodiment of the present disclosure provides a cross-generator training device, comprising: a first encoding module, configured to encode object images and object texts in a source scene, and generate features of objects to be recommended in the source scene; a second encoding module, configured to encode user history data in the source scene, and generate user personal attribute features in the source scene; a decoupling module, configured to perform feature decoupling on user personal attribute features in the source scene and features of objects to be recommended, and generate user personal preference feature codes and object information codes in the source scene; a generating module, configured to input the user personal preference feature codes and object information codes in the source scene into a cross-generator, and generate a simulated object image in the source scene; a calculating module, configured to calculate the similarity loss based on the simulated object image and the object image in the source scene; and an adjusting module, configured to adjust the parameters of the cross-generator based on the similarity loss until the similarity loss meets a preset condition.

[0011] In a sixth aspect, an embodiment of the present disclosure provides a device for generating a simulated image in a target scene, comprising: a first generating module, configured to generate vocabulary in the target scene; a first encoding module, configured to encode the vocabulary in the target scene, and generate features of items to be recommended in the target scene; a second encoding module, configured to encode user history data in the target scene, and generate user personal attribute features in the target scene; a decoupling module, configured to perform feature decoupling on user personal attribute features in the target scene and features of items to be recommended, and generate user personal preference feature codes and item information codes in the target scene; a second generating module, configured to input the user personal preference feature codes and item information codes in the target scene into a cross generator, and generate a simulated object image in the target scene.

[0012] In the seventh aspect, an embodiment of the present disclosure provides a training device for a recommendation model in a target scenario, comprising: a first encoding module, configured to encode simulated object images and simulated object texts in the target scenario, and generate simulated features of objects to be recommended in the target scenario; a second encoding module, configured to encode user history data in the target scenario, and generate user personal attribute features in the target scenario; a first recommendation module, configured to input the simulated features of objects to be recommended in the target scenario and the user personal attribute features into the recommendation model in the target scenario, and generate first predicted recommended object information; a first calculation module, configured to calculate a first loss based on the first predicted recommended object information and the simulated recommended object information; a first adjustment module, configured to adjust the parameters of the recommendation model in the target scenario based on the first loss, until the first loss meets a first preset condition.

[0013] In an eighth aspect, an embodiment of the present disclosure provides an item recommendation device in a target scene, comprising: a first encoding module, configured to encode item images and item texts in the target scene, and generate features of items to be recommended in the target scene; a second encoding module, configured to encode user history data in the target scene, and generate features of user personal attributes in the target scene; a recommendation module, configured to input features of items to be recommended in the target scene and features of user personal attributes into a recommendation model in the target scene, and generate recommended item information, wherein the recommendation model in the target scene is trained using simulated features of items to be recommended and features of user personal attributes in the target scene as input, and using simulated recommended item information in the target scene as target output.

[0014] In the ninth aspect, an embodiment of the present disclosure provides an electronic device, comprising: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by one or more processors, the one or more processors implement the method described in any implementation method of the first to fourth aspects.

[0015] In a tenth aspect, an embodiment of the present disclosure provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any one of the implementation modes of the first to fourth aspects.

[0016] In an eleventh aspect, an embodiment of the present disclosure proposes a computer program product, including a computer program, which implements the method described in any implementation manner of the first to fourth aspects when executed by a processor.

[0017] This disclosure systematically solves a series of challenges in cross-scenario migration of recommendation models through a series of technical steps, including training of target scenario data generator, generation of target scenario pseudo-label data, and training of target scenario real data based on large model distillation. Through efficient data generation and feature processing, combined with knowledge distillation technology of large models, this disclosure not only ensures the efficiency and accuracy of recommendation models in target scenarios, but also greatly improves the adaptability of recommendation models in different scenarios, providing strong technical support for personalized recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Other features, objects and advantages of the present disclosure will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0019] Figure 1 is a flow chart of an embodiment of a cross generator training method according to the present disclosure;

[0020] Figure 2It is the training block diagram of the cross generator in the target scenario;

[0021] Figure 3 is a flow chart of an embodiment of a method for generating a simulated image under a target scene according to the present disclosure;

[0022] Figure 4 is a flowchart of an embodiment of a training method for a recommendation model in a target scenario according to the present disclosure;

[0023] Figure 5 It is the training framework diagram of the recommendation model in the target scenario;

[0024] Figure 6 is a flowchart of an embodiment of a method for recommending items in a target scenario according to the present disclosure;

[0025] Figure 7 is a structural schematic diagram of an embodiment of a cross generator training device according to the present disclosure;

[0026] Figure 8 is a structural schematic diagram of an embodiment of a device for generating a simulated image under a target scene according to the present disclosure;

[0027] Fig. 9 is a structural schematic diagram of an embodiment of a training device for a recommendation model in a target scenario according to the present disclosure;

[0028] Fig.10 is a structural schematic diagram of an embodiment of an item recommendation device in a target scenario according to the present disclosure;

[0029] Fig.11 It is a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present disclosure. DETAILED DESCRIPTION

[0030] The present disclosure is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It is also necessary to explain that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0031] It should be noted that in the technical solution of the present invention, the collection / collection, update, analysis, use, transmission, storage and other aspects of user personal information involved are in compliance with the provisions of relevant laws and regulations, are used for legal and reasonable purposes, are not shared, leaked or sold outside of these legal uses, and are subject to supervision and management by national regulatory authorities. Necessary measures should be taken for user personal information to selectively block the use or access to personal information data to prevent illegal access to such personal information data, ensure that persons who have access to personal information data comply with the provisions of relevant laws and regulations, and ensure the security of user personal information. In addition, once these user personal information data are no longer needed, the risks should be minimized by limiting or even prohibiting data collection and / or deleting data.

[0032] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present disclosure may be combined with each other. The present disclosure will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0033] This paper aims to solve many technical challenges in existing recommendation models in cross-scenario applications, especially to optimize the following issues:

[0034] 1. Decouple scene dependencies and improve generalization: Existing recommendation models perform well in specific scenarios, but due to their strong dependence on scenarios, it is difficult to maintain consistent recommendation effects in target scenarios or dynamically changing scenarios. The goal of this disclosure is to decouple scene dependencies through technical means and enhance the generalization ability of the recommendation model, so that it can provide stable, high-quality recommendations in a variety of scenarios.

[0035] 2. Enhance the dynamic adaptability of user behavior: Since user behavior varies in different scenarios, existing recommendation models often fail to effectively capture these differences, resulting in a decrease in the relevance of recommendation results in the target scenario. This disclosure aims to improve the adaptability of recommendation models to dynamic changes in user behavior by improving algorithms, thereby maintaining high relevance and user satisfaction in different scenarios.

[0036] 3. Improve the efficiency of feature migration: The efficiency of cross-scenario feature migration is one of the key factors affecting the performance of the recommendation model. Existing methods have low migration efficiency in complex and diverse scenarios and cannot fully utilize the knowledge of existing scenarios. This paper optimizes the feature migration technology to improve the efficiency of feature transfer between different scenarios, ensuring that the recommendation model can quickly and efficiently make recommendations in the target scenario.

[0037] 4. Reduce model training and reasoning costs: Although complex deep learning models can improve the cross-scenario generalization ability of recommendation models, their high computational costs limit their application in resource-constrained environments. This disclosure is committed to reducing the computational costs of training and reasoning by optimizing the model structure and training process, so that the recommendation model can run efficiently in scenarios with high real-time requirements or limited resources.

[0038] 5. Optimize the application of pre-trained models: Although pre-trained models show good generalization of feature space, their application in recommendation models still has certain deficiencies, especially in matching visual features with scene requirements. This disclosure further optimizes the application method of pre-trained models to improve the alignment effect of feature space and the accuracy of recommendation models.

[0039] By solving the above technical problems, the present invention will significantly improve the cross-scenario application capability of the recommendation model and provide more personalized and accurate recommendation services. At the same time, it will optimize the model performance and computing cost to meet the needs of a wide range of application scenarios.

[0040] This paper proposes a domain migration technology solution designed specifically for cross-scenario recommendation models. Through target scenario data generator training, target scenario pseudo-label data generation, and target scenario real data training based on large model distillation, the challenges faced by recommendation models in scenario migration are comprehensively addressed. This solution not only focuses on data generation and processing, but also deeply considers feature encoding and decoupling to ensure that the recommendation model can efficiently and accurately provide personalized recommendation services in the target scenario.

[0041] When a recommendation model is migrated from one source scene (such as entertainment) to another target scene (such as education), the biggest challenge is the lack of sufficient training data for the target scene. This data scarcity not only limits the generalization ability of the recommendation model in the target scene, but also may cause the recommendation model to fail to correctly capture the characteristics of the target scene, thereby affecting the recommendation effect. Therefore, the present disclosure generates simulated data related to the target scene through an efficient target scene data generator, thereby providing sufficient data support for the training of the recommendation model in the target scene.

[0042] Figure 1 A process 100 of an embodiment of a cross generator training method according to the present disclosure is shown. The cross generator training method comprises the following steps:

[0043] Step 101 : Encode the object image and the object text in the source scene to generate features of the object to be recommended in the source scene.

[0044] In this embodiment, the execution subject of the cross generator training method can encode the object image and the object text in the source scene to generate the features of the object to be recommended in the source scene.

[0045] In some embodiments, the item image and item text in the source scene are encoded respectively to generate item image features and item text features in the source scene. For example, the item image in the source scene is encoded using a large image model to generate item image features in the source scene. The item text in the source scene is encoded using a large text model to generate item text features in the source scene.

[0046] Among them, large models refer to deep learning models that are pre-trained on large-scale data sets, such as CLIP (Contrastive Language-Image Pre-training), GPT (Generative Pre-Trained Transformer), LLaMA (Large Language Model MetaAI), etc. These models usually contain hundreds of millions or even billions of parameters, have powerful feature extraction and generalization capabilities, and can handle complex tasks and diverse input data.

[0047] The image large model can be, for example, a CLIP model. CLIP is a multimodal pre-trained model developed by OpenAI that can align images and text in the same feature space. Through contrastive learning of large-scale image-text pairs, CLIP can understand the content of an image without explicit annotations and generate text descriptions related to it. CLIP performs well in multimodal tasks and is widely used for feature extraction and alignment of images and text.

[0048] The text model can be, for example, an LLaMA model. LLaMA is a large-scale language model developed by Meta AI that aims to generate high-quality text. LLaMA has powerful natural language understanding and generation capabilities by pre-training on large-scale text data. LLaMA is widely used in tasks such as text generation, translation, and dialogue systems.

[0049] Source data feature encoding is the foundation of the entire technical solution. For the recommendation model data in the source scene, including item images and item texts, the CLIP model and LLaMA2 model pre-trained with large-scale data are used for feature encoding. These two models have been fully trained on large-scale cross-modal datasets and can map diverse image features and text features into a unified universal feature space. This feature space has a high degree of generalization ability and can support the recommendation model to effectively process data features in different scenarios.

[0050] F img= CLIP(Image);

[0051] F text =LLaMA2(Text).

[0052] Among them, Image represents the object image, F img Indicates the image features of the item. Text indicates the text of the item. text Represents the text features of items. For item text, the LLaMA2 model is used for encoding. LLaMA2 is pre-trained on a wide range of text datasets and can effectively capture the semantic and structural characteristics of text. Through this feature encoding, the recommendation model can obtain a highly consistent feature representation. Whether it is processing user viewing history in entertainment scenarios or recommending learning content in educational scenarios, the recommendation model can use this common feature space for efficient feature extraction and processing.

[0053] It should be noted that this embodiment can perform multimodal feature alignment through CLIP and LLaMA2 to ensure the consistency of image and text features in the same feature space. This embodiment can also use a general self-supervised pre-trained model (such as ResNet (Residual Networks) or BERT (Bidirectional Encoder Representation from Transformers)) for feature alignment. These models have been pre-trained on large-scale data sets, have good feature extraction capabilities, and can achieve feature alignment across scenarios to a certain extent.

[0054] In some embodiments, knowledge of the image model can be transferred to the image encoder through knowledge distillation. The image of the object in the source scene is input to the image encoder to generate the image features of the object in the source scene. Similarly, knowledge of the text model can be transferred to the text encoder through knowledge distillation. The text of the object in the source scene is input to the text encoder to generate the text features of the object in the source scene.

[0055] Among them, distillation learning is a technique that transfers the knowledge of a large pre-trained model (teacher model) to a smaller model (student model). Through the distillation process, the small model is able to learn and replicate the feature representation and decision logic of the large model, thereby retaining the performance of the large model as much as possible while maintaining computational efficiency. This technique is often used to deploy efficient models in scenarios with limited computing resources.

[0056] By distilling the features of a large image model into a smaller image encoder, computational overhead can be reduced. Similarly, by distilling the knowledge of a large text model into a smaller text encoder, efficient text encoding can be achieved.

[0057] It should be noted that this embodiment can use knowledge distillation technology to distill the features of a large model into a smaller model, reduce computational overhead, and maintain efficient feature alignment capabilities. This embodiment can also use transfer learning to migrate some parameters of the source scene model to the target scene and perform fine-tuning on the real data of the target scene.

[0058] In some embodiments, the image features and text features of the items in the source scene are encoded to generate the features of the items to be recommended in the source scene. For example, the image features and text features of the items in the source scene are input into the item to be recommended encoder to generate the features of the items to be recommended in the source scene.

[0059] After obtaining the item image features and item text features in the source scene, the feature encoder of the recommendation model can be trained. Among them, the item encoder to be recommended is a feature encoder of the recommendation model. By using the item encoder to encode the item image features and item text features in the source scene, the item features to be recommended in the source scene can be obtained.

[0060] F item =ItemEncoder(F img ,F text ).

[0061] Among them, F img represents the image features of the item, F text Indicates the item text feature. item Represents the features of the item to be recommended. ItemEncoder represents the item encoder to be recommended. It combines the image and text feature vectors to generate a comprehensive item feature representation to match the user's personalized needs.

[0062] Step 102: encode the user's historical data in the source scenario to generate the user's personal attribute characteristics in the source scenario.

[0063] In this embodiment, the execution subject may encode the user's historical data in the source scenario to generate the user's personal attribute characteristics in the source scenario.

[0064] In some embodiments, the user historical data in the source scenario is input into the user portrait encoder to generate the user personal attribute characteristics in the source scenario.

[0065] Among them, the user profile encoder is another feature encoder of the recommendation model. By encoding the user's historical data with the user profile encoder, the user's personal attribute features in the source scenario can be obtained. These features describe the user's personalized preferences and are represented in the general feature space.

[0066] F user =UserProfileEncoder(UserHistory).

[0067] Among them, F user Represents the user's personal attribute characteristics. UserHistory represents the user's historical data, including the user's previous behavior and preferences. UserProfileEncoder represents the user profile encoder that generates an encoding feature that describes the user's personalized attributes and preferences based on the user's historical data.

[0068] Through the user portrait encoder and the recommended item encoder, the recommendation model can match the characteristics of the recommended item with the user's personal characteristics to achieve accurate personalized recommendations. This not only enhances the performance of the recommendation model in the source scenario, but also lays a solid foundation for subsequent target scenario migration.

[0069] Step 103 , feature decoupling is performed on the user's personal attribute features and the features of the items to be recommended in the source scene, and the user's personal preference feature code and the item information code in the source scene are generated.

[0070] In this embodiment, the execution subject may perform feature decoupling on the user's personal attribute features and the features of the items to be recommended in the source scene, and generate the user's personal preference feature code and the item information code in the source scene.

[0071] Among them, feature decoupling technology can separate the user's personal preference attributes from the inherent attributes of the item. The user's personal preference feature encoding can represent the user's preference for a certain type of item, which is the user's hidden attribute. The item information encoding can represent the inherent attributes of the item, such as appearance, color, etc.

[0072] In some embodiments, the user's personal attribute features and the features of the items to be recommended in the source scene are input into the feature decoupler, and the user's personal preference feature coding and the item information coding in the source scene can be generated.

[0073] In the cross-scenario migration of recommendation models, it is crucial to handle the relationship between users' personalized preferences and the inherent characteristics of items. In order to improve the adaptability of recommendation models in target scenarios, feature decoupling technology is used to separate users' personal preference attributes from the inherent attributes of items.

[0074] F user_pref ,F item_info =FeatureDisentangler(F item ,F user ).

[0075] Among them, F user_prefIt is the user's personal preference feature encoding, which indicates the user's preference for a certain type of item. This is the user's hidden attribute. item_info It is an item information code that indicates the inherent properties of the item, such as appearance, color, etc. item Indicates the features of the items to be recommended. user Represents the user's personal attribute characteristics.

[0076] Specifically, by constructing feature pairs for different types of items for the same user, the item ID with the highest CTR (Click-Through-Rate) is selected for each type of item, and the features of the item are input into the feature decoupler. The feature decoupler generates two independent features: one is the personal preference feature encoding, which represents the hidden attributes of users who like this type of item; the other is the item information encoding, which represents the inherent attributes of the item such as appearance and color. This process not only ensures that the recommendation model can flexibly handle the personalized needs of different users in the target scenario, but also provides a more robust foundation for feature migration in the target scenario.

[0077] Through this feature decoupling, the recommendation model can better understand the personalized needs of users in cross-scenario migration, avoiding the problem of decreased recommendation accuracy caused by feature confusion in traditional methods.

[0078] It should be noted that this embodiment can optimize the cross-scenario migration performance of the recommendation model through feature decoupling and knowledge distillation technology. This embodiment can also introduce reinforcement learning technology to gradually adjust the recommendation strategy through a reward mechanism, especially in the target scenario. Reinforcement learning can dynamically optimize the recommendation strategy through continuous interaction with the environment and has strong adaptability. Among them, reinforcement learning is a machine learning method that gradually optimizes decision-making strategies through continuous interaction with the environment. The model adjusts its behavior by receiving feedback (rewards or penalties) from the environment to maximize the cumulative reward. Reinforcement learning performs well in strategy optimization and adaptive tasks in dynamic environments, and is often used in robot control, game AI (Artificial Intelligence) and other fields.

[0079] Step 104 , inputting the user's personal preference feature code and the object information code in the source scene into a cross generator to generate a simulated object image in the source scene.

[0080] In this embodiment, the execution entity may input the user's personal preference feature code and the object information code in the source scene into the cross generator to generate a simulated object image in the source scene.

[0081] In order to further improve the accuracy of feature decoupling and the adaptability of the recommendation model in the target scenario, a cross generator is introduced for auxiliary training. The cross generator generates simulated item images by inputting the user's personal preference feature encoding and item information encoding. The generated simulated item images should be as close to the original item images as possible. This not only verifies the effectiveness of feature decoupling, but also ensures that appropriate simulated data can be generated in the target scenario.

[0082] By inputting the user's personal preference feature encoding and the item information encoding into the cross generator, a simulated item image can be generated.

[0083] I gen =CrossGenerator(F user_pref ,F item_info ).

[0084] Among them, I gen represents the simulated object image, F user_pref Represents the user's personal preference feature encoding, F item_info Indicates the item information code.

[0085] Through this process, the recommendation model can generate accurate simulation data based on user preferences and item attributes in the target scenario, thereby supporting the recommendation service of the recommendation model in the target scenario.

[0086] Step 105 , calculating the similarity loss based on the simulated object image and the object image in the source scene.

[0087] In this embodiment, the execution subject may calculate the similarity loss based on the simulated object image and the object image in the source scene, wherein the similarity loss may be used to characterize the similarity between the simulated object image and the original object image.

[0088] Step 106: Adjust the parameters of the cross generator based on the similarity loss until the similarity loss meets a preset condition.

[0089] In this embodiment, the execution subject may adjust the parameters of the cross generator based on the similarity loss until the similarity loss meets a preset condition, wherein the preset condition may be that the similarity between the simulated object image and the original object image is greater than a preset similarity threshold.

[0090] In some embodiments, based on the similarity loss, the parameters of the user portrait encoder, the recommended item encoder, the feature decoupler, and the cross generator may be adjusted until the similarity loss meets a preset condition.

[0091] Figure 2 A training block diagram of the cross generator in the target scenario is shown.

[0092] In the first step, the image 201 of the object A in the source scene is input into the image encoder 202 to obtain the image features 203 of the object A. The knowledge of the large image model 204 can be transferred to the image encoder 202 through knowledge distillation.

[0093] In the second step, the text of item A 205 in the source scene is input into the text encoder 206 to obtain the text feature 207 of item A. The user history data 208 is input into the text encoder 206 to obtain the user history data feature 209. Among them, the knowledge of the text large model 210 can be transferred to the text encoder 206 through knowledge distillation.

[0094] In the third step, the image feature 203 of item A and the text feature 207 of item A are input into the recommended item encoder 211 to obtain the recommended item features. The user history data feature 209 is input into the user portrait encoder 212 to obtain the user personal attribute features.

[0095] In the fourth step, the features of the item to be recommended and the features of the user's personal attributes are input into the feature decoupler 213 to obtain the item A information code 214 and the user's personal preference feature code 215.

[0096] In the fifth step, the item A information code 214 and the user's personal preference feature code 215 are input to the cross generator 218 to obtain a simulated item A image 219. Based on the simulated item A image 219 and the item A image 201, the similarity loss is calculated to adjust the parameters of the recommended item encoder 211, the user portrait encoder 212, the feature decoupler 213 and the cross generator 218.

[0097] In the sixth step, the item B information code 217 of the item B image 216 is input to the cross generator 218 to obtain a simulated item B image 220. Based on the simulated item B image 220 and the item B image 216, the similarity loss is calculated to adjust the parameters of the recommended item encoder 211, the user portrait encoder 212, the feature decoupler 213 and the cross generator 218.

[0098] In addition, the source domain recommendation model 221 can also be trained using the item A information encoding 214 and the user personal preference feature encoding 215.

[0099] In the process of cross-scenario migration, the recommendation model needs to process a large amount of unlabeled data in the target scenario. Since real data in the target scenario is scarce, the generation of pseudo-labeled data becomes a key link. Pseudo-labeled data can not only serve as auxiliary data in the training process, but also help the system quickly adapt to the target scenario during the migration process.

[0100] Figure 3A process 300 of an embodiment of a method for generating a simulated image in a target scene according to the present disclosure is shown. The method for generating a simulated image in a target scene comprises the following steps:

[0101] Step 301, generating vocabulary in a target scenario.

[0102] In this embodiment, the execution subject of the method for generating a simulated image in a target scene may generate vocabulary in the target scene.

[0103] In some embodiments, the prompt words in the target scenario are input into the text macro model to generate the vocabulary in the target scenario.

[0104] In order to generate high-quality pseudo-label data for target scenes, the conversational mode of the LLaMA2 model can be used to guide the generation of a large number of candidate target scene vocabularies through prompts. These vocabularies include typical elements and feature descriptions in the target scene. For example, in an educational scene, descriptions such as "scenes of recommended math learning videos" and "photos of educational tools used by students in class" may be generated. This process ensures that the feature vocabulary in the target scene is closely integrated with the actual needs of the scene, laying the foundation for subsequent data generation.

[0105] T Edu =LLaMA2(Edu Prompt).

[0106] Among them, Edu Prompt is a specific prompt designed for educational scenarios and is used to generate education-related vocabulary. Edu It is an educational scenario-specific feature vocabulary generated using the LLaMA2 model.

[0107] Step 302: Encode the vocabulary in the target scene to generate features of the items to be recommended in the target scene.

[0108] In this embodiment, the execution subject may encode the vocabulary in the target scene to generate features of the items to be recommended in the target scene.

[0109] In some embodiments, the vocabulary in the target scene is input into a text encoder to generate vocabulary features in the target scene. The vocabulary features in the target scene are input into a to-be-recommended item encoder to generate to-be-recommended item features in the target scene.

[0110] After obtaining the characteristic words in the target scene, the LLaMA2 model can be used to feature encode these words, and they can be further encoded using the previously trained encoder for recommended items. This process generates the encoding of item attributes related to the target scene.

[0111] F edu_item=ItemEncoder(T Edu ).

[0112] Among them, T Edu It is a feature vocabulary specific to educational scenarios generated using the LLaMA2 model. edu_item It is the attribute encoding of education-related items in the target scene. ItemEncoder further encodes the education feature vocabulary generated by the LLaMA2 model.

[0113] It should be noted that the text encoder and the recommended item encoder can be Figure 1 The embodiment shown is obtained by training and will not be described in detail here.

[0114] Step 303: encode the user's historical data in the target scenario to generate the user's personal attribute characteristics in the target scenario.

[0115] In this embodiment, the execution subject may encode the user's historical data in the target scenario to generate the user's personal attribute characteristics in the target scenario.

[0116] In some embodiments, the user's historical data in the target scenario is input into the user portrait encoder to generate the user's personal attribute characteristics in the target scenario.

[0117] It should be noted that the user profile encoder can be used Figure 1 The embodiment shown is obtained by training and will not be described in detail here.

[0118] Step 304 , decoupling the user's personal attribute features and the features of the items to be recommended in the target scenario, and generating the user's personal preference feature code and the item information code in the target scenario.

[0119] In this embodiment, the execution subject may perform feature decoupling on the user's personal attribute features and the features of the items to be recommended in the target scenario, and generate the user's personal preference feature code and the item information code in the target scenario.

[0120] In some embodiments, the user's personal attribute features and the features of the items to be recommended in the target scene are input into a feature decoupler, and the user's personal preference feature coding and the item information coding in the target scene can be generated.

[0121] By passing the user's personal attribute characteristics and the characteristics of the recommended items in the target scene into the feature decoupler of personal preference attributes and item's own attributes, an item attribute information library of the target scene can be generated.

[0122] It should be noted that the feature decoupler can be used to Figure 1 The embodiment shown is obtained by training and will not be described in detail here.

[0123] Step 305 , inputting the user's personal preference feature code and the object information code in the target scene into a cross generator to generate a simulated object image in the target scene.

[0124] In this embodiment, the execution subject may input the user's personal preference feature code and the object information code in the target scene into the cross generator to generate a simulated object image in the target scene.

[0125] Through the cross generator of personal preference and item attribute information, synthetic data of the target scene can be generated. These synthetic data will be used to train the recommendation model under the target scene to improve its adaptability to the target scene.

[0126] I edu_gen =CrossGenerator(F user_pref ,F edu_item ).

[0127] Among them, F user_pref Indicates the user’s personal preference feature encoding. edu_item Represents the attribute encoding of education-related items in the target scene. edu_gen Represents a simulated image of an educational scenario generated by a cross generator. CrossGenerator represents the generation of simulated data by combining the user's personal preferences and the attributes of the target scene items. This process ensures that the recommendation model can be effectively trained using the generated data when facing the target scene, avoiding the performance degradation caused by the scarcity of real data.

[0128] It should be noted that the cross generator can be used Figure 1 The embodiment shown is obtained by training and will not be described in detail here.

[0129] Figure 4 A process 400 of an embodiment of a method for training a recommendation model in a target scenario according to the present disclosure is shown. The method for training a recommendation model in a target scenario includes the following steps:

[0130] Step 401 : Encode the simulated object image and the simulated object text in the target scene to generate the features of the simulated object to be recommended in the target scene.

[0131] In this embodiment, the execution subject of the training method of the recommendation model in the target scene can encode the simulated object image and the simulated object text in the target scene to generate the simulated recommended object features in the target scene.

[0132] In some embodiments, the simulated object image and the simulated object text in the target scene are encoded respectively, and the simulated object image features and the simulated object text features in the target scene can be generated.

[0133] Optionally, the image of the simulated object in the target scene is encoded using the large image model to generate the image features of the simulated object in the target scene. For example, knowledge of the large image model can be transferred to the image encoder through knowledge distillation. The image of the simulated object in the target scene is input to the image encoder to generate the image features of the simulated object in the target scene.

[0134] Optionally, the simulated object text in the target scene is encoded using the large text model to generate the simulated object text features in the target scene. For example, through knowledge distillation, the knowledge of the large text model can be transferred to the text encoder. The simulated object text in the target scene is input into the text encoder to generate the simulated object text features in the target scene.

[0135] Although pseudo-labeled data can help the recommendation model quickly adapt to the target scenario, in order to ensure the final recommendation effect, the recommendation model needs to be further optimized and trained based on the real data of the target scenario. However, directly using a large model for feature encoding will bring high computational overhead in practical applications. Therefore, knowledge distillation becomes a key optimization method.

[0136] During the knowledge distillation process, the initial encoding of all image features can be kept through CLIP for feature alignment. This process ensures the consistency of image features in the common feature space and provides a robust feature foundation for subsequent recommendation tasks. At the same time, when generating images for feature decoupling, alignment is also performed through CLIP to ensure the feature consistency of the generated image and the source image.

[0137] In order to improve the efficiency of the recommendation model, CLIP will not be used directly for encoding in the final model. Instead, the features of CLIP will be distilled into a smaller image encoder to reduce computational overhead. Similarly, for text data processing, the knowledge of the LLaMA2 model will also be distilled into a smaller text encoder to achieve efficient text encoding.

[0138]

[0139] Among them, L distill Represents the distillation loss, which is used to transfer the knowledge of a large model (CLIP) to a smaller model (SmallEncoder) that is more computationally efficient. SmallEncoder represents a smaller model that is trained to replicate the output of the CLIP model, reducing computational resource requirements. gen) represents the output of the CLIP model when the generated image is input, providing high-quality feature representation. gen Represents the generated image used as input in the distillation process and is also the input of SmallEncoder.

[0140] Through knowledge distillation, the recommendation model can maintain efficient operation while inheriting the advantages of the large model and ensuring the universality of the feature space. In some embodiments, the simulated item image features and simulated item text features in the target scene are encoded to generate the simulated item features to be recommended in the target scene. For example, the simulated item image features and simulated item text features in the target scene are encoded and input into the item to be recommended encoder to generate the simulated item features in the target scene.

[0141] It should be noted that the image encoder, the text encoder and the recommended item encoder can be Figure 1 The simulated object image can be obtained by training the embodiment shown in FIG. Figure 3 The embodiment shown is generated and will not be described in detail here.

[0142] Step 402: Encode the user's historical data in the target scenario to generate the user's personal attribute characteristics in the target scenario.

[0143] In this embodiment, the execution subject may encode the user's historical data in the target scenario to generate the user's personal attribute characteristics in the target scenario.

[0144] In some embodiments, the user's historical data in the target scenario is input into the user portrait encoder to generate the user's personal attribute characteristics in the target scenario.

[0145] It should be noted that the user profile encoder can be used Figure 1 The embodiment shown is obtained by training and will not be described in detail here.

[0146] Step 403: Input the simulated recommended item features and the user's personal attribute features in the target scenario into the recommendation model in the target scenario to generate first predicted recommended item information.

[0147] In this embodiment, the above execution can mainly input the simulated recommended item features and the user's personal attribute features in the target scenario into the recommendation model in the target scenario to generate the first predicted recommended item information.

[0148] In some embodiments, the parameters of the recommendation model in the target scenario can be initialized by migrating the parameters of the recommendation model in the source scenario. Among them, transfer learning is a technology that transfers the knowledge learned by the model in the source task or source domain to the target task or target domain. Through transfer learning, the model can quickly adapt to the target scenario or task, reducing the dependence on a large amount of labeled data. Transfer learning is particularly effective in cross-domain tasks, such as migrating from one recommendation model to another recommendation model with different user behavior patterns.

[0149] Step 404: Calculate a first loss based on the first predicted recommended item information and the simulated recommended item information.

[0150] In this embodiment, the execution entity may calculate the first loss based on the first predicted recommended item information and the simulated recommended item information.

[0151] use Figure 2 The simulation data generated by the embodiment shown can further optimize the recommendation model of the target scene. In this process, the recommendation model gradually adjusts its parameters to better adapt to the characteristics of the target scene and user needs.

[0152]

[0153] Among them, L pseudo_train represents the training loss calculated using pseudo-labeled data, which is essential for training the recommendation model in the target scenario. M represents the total number of pseudo-labeled data points used for training. Represents the loss function used to measure the difference between the prediction result and the pseudo label. edu,j I represents the pseudo label of the jth data point, which is specific to the educational scenario and represents the expected output of the recommendation model. edu_gen,j θ represents the input features generated by the jth data point, created by the cross generator using the user preference and item attribute encoding. f represents the trained function or model that accepts the generated input features and outputs a prediction. r Indicates the recommendation model parameters being optimized. Through pseudo-label training, the recommendation model can complete the migration from the source scene to the target scene in a short time, ensuring the recommendation effect in the target scene.

[0154] Step 405: Based on the first loss, adjust the parameters of the recommendation model in the target scenario until the first loss meets the first preset condition.

[0155] In this embodiment, the execution entity may adjust the parameters of the recommendation model in the target scenario based on the first loss until the first loss satisfies a first preset condition, wherein the first preset condition may be that the first loss is less than a first preset threshold.

[0156] Although pseudo-labeled data can help the recommendation model quickly adapt to the target scenario, in order to ensure the final recommendation effect, the recommendation model needs to be further optimized and trained based on the real data of the target scenario. However, directly using a large model for feature encoding will bring high computational overhead in practical applications. Therefore, knowledge distillation becomes a key optimization method.

[0157] During the knowledge distillation process, the initial encoding of all image features can be kept through CLIP for feature alignment. This process ensures the consistency of image features in the common feature space and provides a robust feature foundation for subsequent recommendation tasks. At the same time, when generating images for feature decoupling, alignment is also performed through CLIP to ensure the feature consistency of the generated image and the source image.

[0158] In some embodiments, the recommendation model can also use the real data of the target scene for final training. This process follows the traditional recommendation model training process to ensure that the recommendation model can make full use of the data of the target scene for optimization, further improving the accuracy and relevance of the recommendation. For example, the real object image and the real object text in the target scene are encoded to generate the real features of the recommended items in the target scene; the real features of the recommended items in the target scene and the user's personal attribute features are input into the recommendation model in the target scene to generate the second predicted recommended item information; based on the second predicted recommended item information and the real recommended item information, the second loss is calculated; based on the second loss, the parameters of the recommendation model in the target scene are adjusted until the second loss meets the second preset condition. Among them, the second preset condition can be that the second loss is less than the second preset threshold.

[0159]

[0160] Among them, L real_train represents the training loss using real data in the target scenario, which is crucial for final tuning and optimization of the model. N represents the total number of real data points. i I represents the true label or target result of the i-th true data point. real,i θ represents the true input feature of the i-th data point, which comes from the target scene. r Represents the recommendation model parameters optimized based on real data. By combining training with real data, the recommendation model can provide more accurate and personalized recommendation services in the target scenario while maintaining efficient computing performance.

[0161] Figure 5 This is a training diagram of the recommendation model in the target scenario. The training of the recommendation model in the target scenario includes the generation and training of pseudo-label data of the target scenario and the training of real scene data of the target scenario.

[0162] The generation and training of pseudo-label data for target scenes include the following steps:

[0163] The first step is to input the prompt words 501 in the target scene into the text model 502 to obtain the target scene vocabulary 503.

[0164] In the second step, the target scene vocabulary 503 is input into the text encoder 504 to obtain vocabulary features 505. The user history data 506 is input into the text encoder 504 to obtain user history data features 507.

[0165] In the third step, the vocabulary feature 505 is input to the recommended item encoder 508 to obtain the recommended item feature. The user history data feature 507 is input to the user portrait encoder 509 to obtain the user personal attribute feature.

[0166] In the fourth step, the features of the item to be recommended and the features of the user's personal attributes are input into the feature decoupler 510 to obtain the item information code 511 and the user's personal preference feature code 512.

[0167] In the fifth step, the object information code 511 and the user personal preference feature code 512 are input into the cross generator 513 to obtain a simulated object image 514.

[0168] The sixth step is to train the target domain recommendation model 515 based on the simulated object image 514.

[0169] The target scene real scene data training includes the following steps:

[0170] In the first step, the image 516 of the object C in the target scene is input into the image encoder 517 to obtain the image features 518 of the object C. The knowledge of the large image model 519 can be transferred to the image encoder 517 through knowledge distillation.

[0171] In the second step, the object C text 520 in the target scene is input into the text encoder 521 to obtain the object C text feature 522. The user history data 523 is input into the text encoder 521 to obtain the user history data feature 524. Among them, through knowledge distillation, the knowledge of the text large model 525 can be transferred to the text encoder 521.

[0172] In the third step, the image feature 518 of item C and the text feature 522 of item C are input into the recommended item encoder 526 to obtain the recommended item features. The user history data feature 524 is input into the user portrait encoder 527 to obtain the user personal attribute features.

[0173] In the fourth step, the features of the items to be recommended and the features of the user's personal attributes are input into the target domain recommendation model 515 for training.

[0174] This disclosure systematically solves a series of challenges in cross-scenario migration of recommendation models through a series of technical steps, including training of target scenario data generator, generation of target scenario pseudo-label data, and training of target scenario real data based on large model distillation. Through efficient data generation and feature processing, combined with knowledge distillation technology of large models, this disclosure not only ensures the efficiency and accuracy of recommendation models in target scenarios, but also greatly improves the adaptability of recommendation models in different scenarios, providing strong technical support for personalized recommendations.

[0175] This paper proposes a series of innovative technologies in the field of cross-scenario recommendation models, aiming to solve many challenges encountered by traditional recommendation models during scenario migration.

[0176] 1. Feature decoupling and cross-generation technology of target scene data generator

[0177] This disclosure proposes a data generation method based on feature decoupling and cross generation to deal with the data scarcity problem in cross-scenario migration. By decoupling the user's personal preference features from the inherent attribute features of the item and recombining these features through a cross generator, this disclosure can generate simulated data that adapts to the needs of the target scenario. This innovation ensures that the recommendation model can use these generated data for effective pre-training when facing the target scenario, thereby improving the generalization ability of the system.

[0178] Through feature decoupling, the user's preference features are processed separately from the item attributes, thus avoiding feature confusion. Subsequently, through cross-generation technology, different features are combined to generate new data, which can reflect the typical features in the target scene and provide rich training samples for the recommendation model.

[0179] 2. Knowledge distillation technology based on large models

[0180] This disclosure uses knowledge distillation technology based on large models (such as CLIP and LLaMA2) to achieve efficient feature alignment and migration. In the cross-scenario recommendation model, directly using a large model for feature encoding can ensure high accuracy, but the computational overhead is large. This disclosure distills the knowledge of the large model into a smaller encoder, which not only retains the feature alignment capability of the large model, but also greatly reduces the computational complexity, so that the recommendation model can run efficiently in the target scenario.

[0181] Knowledge distillation technology allows small models to learn the output (feature representation) of large models, so that small models can inherit the superior characteristics of large models while maintaining high efficiency. This process ensures that the features in the target scene are aligned with the features in the source scene in the same feature space, thereby improving the cross-scene migration capability of the recommendation model.

[0182] 3. Target scene pseudo-label generation and training optimization technology

[0183] The present disclosure introduces a pseudo-label-based data generation and training optimization technology for quickly adapting recommendation tasks in target scenarios. By generating pseudo-label data related to the target scenario and using these data to train and optimize the target scenario recommendation model, the present disclosure can adapt the recommendation model to the characteristics of the target scenario and user behavior in a short period of time, thereby improving the accuracy of the recommendation.

[0184] Pseudo-label generation technology generates simulated data in the target scenario and assigns pseudo-labels to these data, so that the recommendation model can be trained using these pseudo-labeled data in the absence of real data. This method not only speeds up the adaptation of the recommendation model, but also further optimizes the performance of the recommendation model through transfer learning.

[0185] 4. Decoupling and generation methods of personal preferences and item attributes

[0186] The present disclosure introduces a method for decoupling and generating personal preferences and item attributes in the feature processing of the recommendation model. This method not only enhances the recommendation model's personalized understanding of different users, but also allows the system to generate recommendation data that meets the needs of the target scenario through feature recombination in the target scenario.

[0187] By separating the user's personal preferences from the inherent attributes of items, the system can more flexibly handle different users' preferences for different items. This decoupling enables the recommendation model to recombine features in the target scenario, generate new item attributes and user preferences, and thus meet the diverse recommendation needs in scenario migration.

[0188] Figure 6 This is a process 600 of an embodiment of a method for recommending items in a target scenario according to the present disclosure. The method for recommending items in a target scenario includes the following steps:

[0189] Step 601: Encode the object image and the object text in the target scene to generate features of the object to be recommended in the target scene.

[0190] In this embodiment, the execution subject of the method for recommending items in a target scene may encode the image and text of the items in the target scene to generate features of the items to be recommended in the target scene.

[0191] In some embodiments, the object image and the object text in the target scene are encoded respectively, and the object image features and the object text features in the target scene can be generated.

[0192] Optionally, the image of the object in the target scene is encoded using the large image model to generate the image features of the object in the target scene. For example, through knowledge distillation, the knowledge of the large image model can be transferred to the image encoder. The image of the object in the target scene is input to the image encoder to generate the image features of the object in the target scene.

[0193] Optionally, the object text in the target scene is encoded using the text big model to generate the object text features in the target scene. For example, through knowledge distillation, the knowledge of the text big model can be transferred to the text encoder. The object text in the target scene is input into the text encoder to generate the object text features in the target scene.

[0194] In some embodiments, the image features and text features of the items in the target scene are encoded to generate the features of the items to be recommended in the target scene. For example, the image features and text features of the items in the target scene are encoded and input into the encoder of the items to be recommended to generate the features of the items to be recommended in the target scene.

[0195] It should be noted that the image encoder, the text encoder and the recommended item encoder can be Figure 1 The embodiment shown is obtained by training and will not be described in detail here.

[0196] Step 602: Encode the user's historical data in the target scenario to generate the user's personal attribute characteristics in the target scenario.

[0197] In this embodiment, the execution subject may encode the user's historical data in the target scenario to generate the user's personal attribute characteristics in the target scenario.

[0198] In some embodiments, the user's historical data in the target scenario is input into the user portrait encoder to generate the user's personal attribute characteristics in the target scenario.

[0199] It should be noted that the user profile encoder can be used Figure 1 The embodiment shown is obtained by training and will not be described in detail here.

[0200] Step 603: Input the features of the items to be recommended in the target scenario and the features of the user's personal attributes into the recommendation model in the target scenario to generate recommended item information.

[0201] In this embodiment, the execution subject may input the features of the items to be recommended in the target scenario and the features of the user's personal attributes into the recommendation model in the target scenario to generate recommended item information.

[0202] It should be noted that the recommendation model in the target scenario can be Figure 4The embodiment shown is obtained by training and will not be described in detail here.

[0203] Further references Figure 7 As an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a cross generator training device, which is similar to Figure 1 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic equipment.

[0204] like Figure 7 As shown, the cross generator training device 700 provided in this embodiment includes a first encoding module 701, a second encoding module 702, a decoupling module 703, a generating module 704, a calculating module 705 and an adjusting module 706. Among them, the first encoding module 701 is configured to encode the object image and object text in the source scene to generate the features of the to-be-recommended object in the source scene; the second encoding module 702 is configured to encode the user history data in the source scene to generate the user personal attribute features in the source scene; the decoupling module 703 is configured to perform feature decoupling on the user personal attribute features in the source scene and the features of the to-be-recommended object, to generate the user personal preference feature encoding and the object information encoding in the source scene; the generating module 704 is configured to input the user personal preference feature encoding and the object information encoding in the source scene into the cross generator to generate the simulated object image in the source scene; the calculating module 705 is configured to calculate the similarity loss based on the simulated object image and the object image in the source scene; the adjusting module 706 is configured to adjust the parameters of the cross generator based on the similarity loss until the similarity loss meets the preset conditions.

[0205] In this embodiment, in the cross generator training device 700, the specific processing of the first encoding module 701, the second encoding module 702, the decoupling module 703, the generating module 704, the calculating module 705 and the adjusting module 706 and the technical effects thereof can be referred to respectively. Figure 1 The relevant descriptions of steps 101 - 106 in the corresponding embodiment are not repeated here.

[0206] In some optional implementations of this embodiment, the first encoding module 701 is further configured to: encode the object image and the object text in the source scene respectively to generate the object image features and the object text features in the source scene; encode the object image features and the object text features in the source scene to generate the features of the items to be recommended in the source scene.

[0207] In some optional implementations of this embodiment, the first encoding module 701 is further configured to: use the image macro model to encode the object image in the source scene to generate the object image features in the source scene; use the text macro model to encode the object text in the source scene to generate the object text features in the source scene.

[0208] In some optional implementations of this embodiment, the first encoding module 701 is further configured to: transfer the knowledge of the large image model to the image encoder through knowledge distillation; input the object image in the source scene into the image encoder to generate the object image features in the source scene; transfer the knowledge of the large text model to the text encoder through knowledge distillation; input the object text in the source scene into the text encoder to generate the object text features in the source scene.

[0209] In some optional implementations of this embodiment, the first encoding module 701 is further configured to: input the image features and text features of the items in the source scene into the to-be-recommended item encoder to generate the to-be-recommended item features in the source scene.

[0210] In some optional implementations of this embodiment, the second encoding module 702 is further configured to: input the user historical data in the source scenario into the user portrait encoder to generate the user personal attribute characteristics in the source scenario.

[0211] In some optional implementations of this embodiment, the decoupling module 703 is further configured to: input the user's personal attribute characteristics and the characteristics of the recommended items in the source scene into the feature decoupler to generate the user's personal preference feature code and the item information code in the source scene.

[0212] In some optional implementations of this embodiment, the adjustment module 706 is further configured to: adjust the parameters of the user portrait encoder, the recommended item encoder, the feature decoupler and the cross generator based on the similarity loss.

[0213] Further references Figure 8 As an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a device for generating a simulated image under a target scene. Figure 3 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic equipment.

[0214] like Figure 8As shown, the device 800 for generating a simulated image in a target scene provided by this embodiment includes a first generating module 801, a first encoding module 802, a second encoding module 803, a decoupling module 804 and a second generating module 805. Among them, the first generating module 801 is configured to generate vocabulary in the target scene; the first encoding module 802 is configured to encode the vocabulary in the target scene to generate the features of the items to be recommended in the target scene; the second encoding module 803 is configured to encode the user history data in the target scene to generate the user's personal attribute features in the target scene; the decoupling module 804 is configured to perform feature decoupling on the user's personal attribute features in the target scene and the features of the items to be recommended to generate the user's personal preference feature code and the item information code in the target scene; the second generating module 805 is configured to input the user's personal preference feature code and the item information code in the target scene into the cross generator to generate the simulated item image in the target scene.

[0215] In the device 800 for generating a simulated image under a target scene, the specific processing of the first generating module 801, the first encoding module 802, the second encoding module 803, the decoupling module 804 and the second generating module 805 and the technical effects thereof can be referred to in detail. Figure 3 The relevant descriptions of steps 301 - 305 in the corresponding embodiment are not repeated here.

[0216] In some optional implementations of this embodiment, the first generating module 801 is further configured to: input the prompt words in the target scenario into the text macro model to generate the vocabulary in the target scenario.

[0217] In some optional implementations of this embodiment, the first encoding module 802 is further configured to: input the vocabulary in the target scene into the text encoder to generate vocabulary features in the target scene; input the vocabulary features in the target scene into the to-be-recommended item encoder to generate features of the to-be-recommended items in the target scene.

[0218] In some optional implementations of this embodiment, the second encoding module 803 is further configured to: input the user historical data in the target scenario into the user portrait encoder to generate the user personal attribute characteristics in the target scenario.

[0219] In some optional implementations of this embodiment, the decoupling module 804 is further configured to: input the user's personal attribute characteristics and the characteristics of the items to be recommended in the target scene into the feature decoupler to generate the user's personal preference feature coding and the item information coding in the target scene.

[0220] Further references Fig. 9As an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of a training device for a recommendation model in a target scenario. Figure 4 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic equipment.

[0221] like Fig. 9 As shown, the training device 900 of the recommendation model in the target scene provided by this embodiment includes a first encoding module 901, a second encoding module 902, a first recommendation module 903, a first calculation module 904 and a first adjustment module 905. Among them, the first encoding module 901 is configured to encode the simulated item image and the simulated item text in the target scene to generate the simulated to-be-recommended item features in the target scene; the second encoding module 902 is configured to encode the user history data in the target scene to generate the user personal attribute features in the target scene; the first recommendation module 903 is configured to input the simulated to-be-recommended item features in the target scene and the user personal attribute features into the recommendation model in the target scene to generate the first predicted recommended item information; the first calculation module 904 is configured to calculate the first loss based on the first predicted recommended item information and the simulated recommended item information; the first adjustment module 905 is configured to adjust the parameters of the recommendation model in the target scene based on the first loss until the first loss meets the first preset condition.

[0222] In the training device 900 for the recommendation model in the target scenario, the specific processing of the first encoding module 901, the second encoding module 902, the first recommendation module 903, the first calculation module 904 and the first adjustment module 905 and the technical effects thereof can be referred to respectively. Figure 4 The relevant descriptions of steps 401-405 in the corresponding embodiment are not repeated here.

[0223] In some optional implementations of the present embodiment, the first encoding module 901 is further configured to: respectively encode the simulated object image and the simulated object text in the target scene to generate the simulated object image features and the simulated object text features in the target scene; encode the simulated object image features and the simulated object text features in the target scene to generate the simulated recommended object features in the target scene.

[0224] In some optional implementations of this embodiment, the first encoding module 901 is further configured to: use the image macro model to encode the image of the simulated object in the target scene to generate the image features of the simulated object in the target scene; use the text macro model to encode the text of the simulated object in the target scene to generate the text features of the simulated object in the target scene.

[0225] In some optional implementations of this embodiment, the first encoding module 901 is further configured to: transfer the knowledge of the large image model to the image encoder through knowledge distillation; input the image of the simulated object in the target scene into the image encoder to generate the image features of the simulated object in the target scene; transfer the knowledge of the large text model to the text encoder through knowledge distillation; input the text of the simulated object in the target scene into the text encoder to generate the text features of the simulated object in the target scene.

[0226] In some optional implementations of this embodiment, the first encoding module 901 is further configured to: encode the image features and text features of the simulated objects in the target scene and input them into the recommended object encoder to generate the features of the simulated objects in the target scene.

[0227] In some optional implementations of this embodiment, the second encoding module 902 is further configured to: input the user historical data in the target scenario into the user portrait encoder to generate the user personal attribute characteristics in the target scenario.

[0228] In some optional implementations of the present embodiment, the training device 900 for the recommendation model in the target scenario also includes: a third encoding module, configured to encode the real object image and the real object text in the target scene, and generate the real features of the to-be-recommended objects in the target scene; a fourth encoding module, configured to input the real features of the to-be-recommended objects in the target scene and the personal attribute features of the user into the recommendation model in the target scene, and generate second predicted recommended object information; a second calculation module, configured to calculate the second loss based on the second predicted recommended object information and the real recommended object information; and a second adjustment module, configured to adjust the parameters of the recommendation model in the target scenario based on the second loss until the second loss meets the second preset condition.

[0229] In some optional implementations of this embodiment, the parameters of the recommendation model in the target scenario are initialized by migrating the parameters of the recommendation model in the source scenario.

[0230] Further references Fig.10 As an implementation of the methods shown in the above figures, the present disclosure provides an embodiment of an item recommendation device in a target scene. Figure 6 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic equipment.

[0231] like Fig.10As shown, the device 1000 for recommending items in a target scene provided by this embodiment includes a first encoding module 1001, a second encoding module 1002 and a recommendation module 1003. The first encoding module 1001 is configured to encode the item image and item text in the target scene to generate the features of the items to be recommended in the target scene; the second encoding module 1002 is configured to encode the user history data in the target scene to generate the user personal attribute features in the target scene; the recommendation module 1003 is configured to input the features of the items to be recommended in the target scene and the user personal attribute features into the recommendation model in the target scene to generate the recommended item information, wherein the recommendation model in the target scene is trained by taking the simulated features of the items to be recommended in the target scene and the user personal attribute features as inputs and taking the simulated recommended item information in the target scene as target output.

[0232] In the device for recommending items in the target scene 1000, the specific processing of the first encoding module 1001, the second encoding module 1002 and the recommendation module 1003 and the technical effects thereof can be referred to in Figure 6 The relevant descriptions of steps 601-603 in the corresponding embodiment are not repeated here.

[0233] In some optional implementations of this embodiment, the first encoding module 1001 is further configured to: encode the object image and the object text in the target scene respectively to generate the object image features and the object text features in the target scene; encode the object image features and the object text features in the target scene to generate the features of the items to be recommended in the target scene.

[0234] In some optional implementations of this embodiment, the first encoding module 1001 is further configured to: use the image macro model to encode the object image in the target scene to generate the object image features in the target scene; use the text macro model to encode the object text in the target scene to generate the object text features in the target scene.

[0235] In some optional implementations of this embodiment, the first encoding module 1001 is further configured to: transfer the knowledge of the large image model to the image encoder through knowledge distillation; input the object image in the target scene into the image encoder to generate the object image features in the target scene; transfer the knowledge of the large text model to the text encoder through knowledge distillation; input the object text in the target scene into the text encoder to generate the object text features in the target scene.

[0236] In some optional implementations of this embodiment, the first encoding module 1001 is further configured to: encode the image features and text features of the items in the target scene and input them into the to-be-recommended item encoder to generate the features of the items in the target scene.

[0237] In some optional implementations of this embodiment, the second encoding module 1002 is further configured to: input the user historical data in the target scenario into the user portrait encoder to generate the user personal attribute characteristics in the target scenario.

[0238] Reference below Fig.11 , which shows a schematic diagram of the structure of an electronic device 1100 suitable for implementing the embodiments of the present disclosure. The terminal device in the embodiments of the present disclosure may include but is not limited to mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), vehicle terminals (such as vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Fig.11 The server 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.

[0239] like Fig.11 As shown, the electronic device 1100 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 1101, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1102 or a program loaded from a storage device 1108 to a random access memory (RAM) 1103. In the RAM 1103, various programs and data required for the operation of the electronic device 1100 are also stored. The processing device 1101, the ROM 1102, and the RAM 1103 are connected to each other via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0240] Typically, the following devices may be connected to the I / O interface 1105: an input device 1106 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 1107 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 1108 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1109. The communication device 1109 may allow the electronic device 1100 to communicate with other devices wirelessly or by wire to exchange data. Although Fig.11 The electronic device 1100 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. Fig.11Each block shown in the figure may represent one device, or may represent multiple devices as required.

[0241] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device 1109, or installed from a storage device 1108, or installed from a ROM 1102. When the computer program is executed by the processing device 1101, the above-mentioned functions defined in the method of the embodiment of the present disclosure are executed.

[0242] It should be noted that the computer-readable medium described in the embodiments of the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above 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 the embodiments of the present disclosure, the computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, device or device. In the 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.

[0243] The computer-readable medium may be included in the electronic device; or it may exist independently without being installed in 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: encodes the item image and item text in the target scene to generate the features of the items to be recommended in the target scene; encodes the user history data in the target scene to generate the user's personal attribute features in the target scene; inputs the features of the items to be recommended in the target scene and the user's personal attribute features into the recommendation model in the target scene to generate recommended item information.

[0244] Computer program code for performing the operations of 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 a user's computer, partially on a user's computer, as a separate software package, partially on a 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).

[0245] 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.

[0246] The modules involved in the embodiments described in the present disclosure may be implemented by software or by hardware. The modules described may also be set in a processor. For example, they may be described as: a processor includes a first encoding module, a second encoding module, and a recommendation module. The names of these modules do not constitute a limitation on the modules themselves in some cases. For example, the first encoding module may also be described as "a module for encoding the object images and object texts in the target scene and generating the features of the items to be recommended in the target scene".

[0247] The above description is only a preferred embodiment 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) to form a technical solution.

Claims

1. A cross generator training method, comprising: Encoding the object image and the object text in the source scene to generate features of the object to be recommended in the source scene; Encoding the user historical data in the source scenario to generate the user personal attribute characteristics in the source scenario; Decoupling the user's personal attribute features and the features of the items to be recommended in the source scene to generate the user's personal preference feature code and the item information code in the source scene; Inputting the user's personal preference feature code and the item information code in the source scene into a cross generator to generate a simulated item image in the source scene; Calculating a similarity loss based on the simulated object image and the object image in the source scene; Based on the similarity loss, the parameters of the cross generator are adjusted until the similarity loss meets a preset condition.

2. The method according to claim 1, wherein: The encoding of the item image and the item text in the source scene to generate features of the item to be recommended in the source scene includes: Encoding the object image and the object text in the source scene respectively to generate the object image feature and the object text feature in the source scene; The image features and text features of the objects in the source scene are encoded to generate features of the objects to be recommended in the source scene.

3. The method according to claim 2, wherein: The encoding of the object image and the object text in the source scene respectively to generate the object image features and the object text features in the source scene includes: Encoding the image of the object in the source scene by using the image macro model to generate image features of the object in the source scene; The object text in the source scene is encoded using a large text model to generate object text features in the source scene.

4. The method according to claim 3, wherein: The step of encoding the object image in the source scene by using the image macro model to generate the image features of the object in the source scene includes: Transferring the knowledge of the large image model to the image encoder through knowledge distillation; Inputting the object image in the source scene into the image encoder to generate the image features of the object in the source scene; and The step of encoding the object text in the source scene by using the large text model to generate the object text features in the source scene includes: By knowledge distillation, the knowledge of the large text model is transferred to the text encoder; The object text in the source scene is input into the text encoder to generate the object text features in the source scene.

5. The method according to claim 2, wherein: The encoding of the image features and text features of the items in the source scene to generate features of the items to be recommended in the source scene includes: The image features and text features of the items in the source scene are input into a to-be-recommended item encoder to generate the to-be-recommended item features in the source scene.

6. The method according to claim 5, wherein: The encoding of the user historical data in the source scenario to generate the user personal attribute characteristics in the source scenario includes: The user historical data in the source scenario is input into the user portrait encoder to generate the user personal attribute characteristics in the source scenario.

7. The method according to claim 6, wherein: The step of decoupling the user's personal attribute features and the features of the items to be recommended in the source scene to generate the user's personal preference feature code and the item information code in the source scene includes: The user's personal attribute features and the features of the items to be recommended in the source scene are input into the feature decoupler to generate the user's personal preference feature code and the item information code in the source scene.

8. The method according to claim 7, wherein: The adjusting the parameters of the cross generator based on the similarity loss includes: Based on the similarity loss, the parameters of the user portrait encoder, the recommended item encoder, the feature decoupler and the cross generator are adjusted.

9. A method for generating a simulated image in a target scene, comprising: Generate vocabulary in target contexts; Encoding the vocabulary in the target scenario to generate features of items to be recommended in the target scenario; Encoding the user's historical data in the target scenario to generate the user's personal attribute characteristics in the target scenario; Decoupling the user's personal attribute features and the features of the items to be recommended in the target scenario to generate the user's personal preference feature code and the item information code in the target scenario; The user's personal preference feature code and the object information code in the target scene are input into a cross generator to generate a simulated object image in the target scene.

10. The method according to claim 9, wherein: The generating of the vocabulary in the target scenario includes: The prompt words in the target scenario are input into the text macro model to generate the vocabulary in the target scenario.

11. The method according to claim 9, wherein: The encoding of the vocabulary in the target scene to generate features of items to be recommended in the target scene includes: Inputting the vocabulary in the target scenario into a text encoder to generate vocabulary features in the target scenario; The vocabulary features in the target scenario are input into a to-be-recommended item encoder to generate the to-be-recommended item features in the target scenario.

12. The method according to claim 9, wherein: The encoding of the user historical data in the target scenario to generate the user personal attribute characteristics in the target scenario includes: The user historical data in the target scenario is input into the user portrait encoder to generate the user personal attribute characteristics in the target scenario.

13. The method according to claim 9, wherein: The step of decoupling the user's personal attribute features and the features of the items to be recommended in the target scenario to generate the user's personal preference feature code and the item information code in the target scenario includes: The user's personal attribute features and the features of the items to be recommended in the target scenario are input into the feature decoupler to generate the user's personal preference feature code and the item information code in the target scenario.

14. A method for training a recommendation model in a target scenario, comprising: Encoding the simulated object image and the simulated object text in the target scene to generate features of the simulated object to be recommended in the target scene; Encoding the user's historical data in the target scenario to generate the user's personal attribute characteristics in the target scenario; Inputting the simulated recommended item features and the user's personal attribute features in the target scenario into the recommendation model in the target scenario to generate first predicted recommended item information; Calculating a first loss based on the first predicted recommended item information and the simulated recommended item information; Based on the first loss, parameters of the recommendation model in the target scenario are adjusted until the first loss meets a first preset condition.

15. The method according to claim 14, wherein: The step of encoding the simulated object image and the simulated object text in the target scene to generate the features of the simulated object to be recommended in the target scene includes: Encoding the simulated object image and the simulated object text in the target scene respectively to generate simulated object image features and simulated object text features in the target scene; The simulated object image features and the simulated object text features in the target scene are encoded to generate simulated object features to be recommended in the target scene.

16. The method according to claim 15, wherein: The step of encoding the simulated object image and the simulated object text in the target scene respectively to generate the simulated object image features and the simulated object text features in the target scene includes: Encoding the simulated object image in the target scene by using the large image model to generate the simulated object image features in the target scene; The simulated object text in the target scene is encoded using a large text model to generate simulated object text features in the target scene.

17. The method according to claim 16, wherein: The step of encoding the simulated object image in the target scene by using the large image model to generate the simulated object image features in the target scene includes: Transferring the knowledge of the large image model to the image encoder through knowledge distillation; Inputting the simulated object image in the target scene into the image encoder to generate simulated object image features in the target scene; and The step of encoding the simulated object text in the target scene by using the large text model to generate the simulated object text features in the target scene includes: By knowledge distillation, the knowledge of the large text model is transferred to the text encoder; The simulated object text in the target scene is input into the text encoder to generate the simulated object text features in the target scene.

18. The method according to claim 15, wherein: The step of encoding the simulated object image features and the simulated object text features in the target scene to generate the simulated object features to be recommended in the target scene includes: The simulated object image features and the simulated object text features in the target scene are encoded and input into a to-be-recommended object encoder to generate simulated to-be-recommended object features in the target scene.

19. The method according to claim 14, wherein: The encoding of the user historical data in the target scenario to generate the user personal attribute characteristics in the target scenario includes: The user historical data in the target scenario is input into the user portrait encoder to generate the user personal attribute characteristics in the target scenario.

20. The method according to claim 14, wherein: The method further comprises: Encoding the real object image and the real object text in the target scene to generate the real features of the recommended object in the target scene; Inputting the real to-be-recommended item features and the user's personal attribute features in the target scenario into the recommendation model in the target scenario to generate second predicted recommended item information; Calculating a second loss based on the second predicted recommended item information and the actual recommended item information; Based on the second loss, adjust the parameters of the recommendation model in the target scenario until the second loss meets a second preset condition.

21. The method according to any one of claims 14 to 20, wherein: The parameters of the recommendation model in the target scenario are initialized by migrating the parameters of the recommendation model in the source scenario.

22. A method for recommending items in a target scene, comprising: Encoding the object images and object texts in the target scene to generate features of the objects to be recommended in the target scene; Encoding the user's historical data in the target scenario to generate the user's personal attribute characteristics in the target scenario; The features of the items to be recommended in the target scenario and the personal attribute features of the user are input into the recommendation model in the target scenario to generate recommended item information, wherein the recommendation model in the target scenario is trained using the simulated features of the items to be recommended in the target scenario and the personal attribute features of the user as input and the simulated recommended item information in the target scenario as target output.

23. The method according to claim 22, wherein: The encoding of the object image and the object text in the target scene to generate features of the to-be-recommended object in the target scene includes: Encoding the object image and the object text in the target scene respectively to generate the object image features and the object text features in the target scene; The image features and text features of the objects in the target scene are encoded to generate features of the objects to be recommended in the target scene.

24. The method according to claim 23, wherein: The encoding of the object image and the object text in the target scene respectively to generate the object image features and the object text features in the target scene includes: Encoding the image of the object in the target scene by using the large image model to generate image features of the object in the target scene; The object text in the target scene is encoded using a large text model to generate object text features in the target scene.

25. The method according to claim 24, wherein: The step of encoding the object image in the target scene by using the large image model to generate the image features of the object in the target scene includes: Transferring the knowledge of the large image model to the image encoder through knowledge distillation; Inputting the object image in the target scene into the image encoder to generate the image features of the object in the target scene; and The method of encoding the object text in the target scene by using the large text model to generate the object text features in the target scene includes: By knowledge distillation, the knowledge of the large text model is transferred to the text encoder; The object text in the target scene is input into the text encoder to generate the object text features in the target scene.

26. The method of claim 23, wherein: The encoding of the image features and text features of the objects in the target scene to generate features of the objects to be recommended in the target scene includes: The image features and text features of the objects in the target scene are encoded and input into the to-be-recommended object encoder to generate the features of the to-be-recommended objects in the target scene.

27. The method of claim 22, wherein: The encoding of the user historical data in the target scenario to generate the user personal attribute characteristics in the target scenario includes: The user historical data in the target scenario is input into the user portrait encoder to generate the user personal attribute characteristics in the target scenario.

28. A cross generator training device, comprising: A first encoding module is configured to encode the object image and the object text in the source scene to generate features of the object to be recommended in the source scene; A second encoding module is configured to encode the user historical data in the source scenario to generate the user personal attribute characteristics in the source scenario; A decoupling module is configured to perform feature decoupling on the user's personal attribute features and the features of the items to be recommended in the source scene, and generate the user's personal preference feature code and the item information code in the source scene; A generating module, configured to input the user's personal preference feature code and the item information code in the source scene into a cross generator to generate a simulated item image in the source scene; A calculation module configured to calculate a similarity loss based on the simulated object image and the object image in the source scene; The adjustment module is configured to adjust the parameters of the cross generator based on the similarity loss until the similarity loss meets a preset condition.

29. A device for generating a simulated image in a target scene, comprising: A first generation module is configured to generate vocabulary in a target scenario; A first encoding module is configured to encode the vocabulary in the target scene to generate features of items to be recommended in the target scene; A second encoding module is configured to encode the user historical data in the target scenario to generate the user personal attribute characteristics in the target scenario; A decoupling module is configured to perform feature decoupling on the user's personal attribute features and the features of the items to be recommended in the target scene, and generate a user's personal preference feature code and an item information code in the target scene; The second generating module is configured to input the user's personal preference feature code and the object information code in the target scene into the cross generator to generate a simulated object image in the target scene.

30. A training device for a recommendation model in a target scenario, comprising: A first encoding module is configured to encode the simulated object image and the simulated object text in the target scene to generate features of the simulated object to be recommended in the target scene; A second encoding module is configured to encode the user historical data in the target scenario to generate the user personal attribute characteristics in the target scenario; A first recommendation module is configured to input the simulated recommended item features and the user's personal attribute features in the target scenario into the recommendation model in the target scenario to generate first predicted recommended item information; A first calculation module is configured to calculate a first loss based on the first predicted recommended item information and the simulated recommended item information; The first adjustment module is configured to adjust the parameters of the recommendation model in the target scenario based on the first loss until the first loss meets a first preset condition.

31. An item recommendation device in a target scene, comprising: A first encoding module is configured to encode the object image and the object text in the target scene to generate features of the object to be recommended in the target scene; A second encoding module is configured to encode the user historical data in the target scenario to generate the user personal attribute characteristics in the target scenario; The recommendation module is configured to input the characteristics of the items to be recommended in the target scenario and the personal attribute characteristics of the user into the recommendation model in the target scenario to generate recommended item information, wherein the recommendation model in the target scenario is trained with the characteristics of the simulated items to be recommended in the target scenario and the personal attribute characteristics of the user as input and the simulated recommended item information in the target scenario as target output.

32. 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 27.

33. A computer readable medium having a computer program stored thereon, wherein: When the program is executed by a processor, the method according to any one of claims 1 to 27 is implemented.

34. 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 27.

Images