Resource cold start method, apparatus, equipment and storage medium based on large model

Abstract

This disclosure provides a method, apparatus, device, and storage medium for resource cold start based on a large model. This disclosure relates to the field of artificial intelligence technology, particularly to large models, natural language processing, deep learning, and intelligent recommendation, and can be used in application scenarios such as recommendation systems, advertising, content distribution, e-commerce platforms, and search engine optimization. The specific implementation scheme is as follows: feature data of the target object is parsed from online requests; the target object's vector is determined based on the feature data, and a first recall result is obtained from a multimodal vector index based on the target object's vector; a first target resource set is determined based on the feature data of the target object, and a second recall result is obtained from the multimodal vector index based on the multimodal vectors corresponding to the target resources in the first target resource set; the first and second recall results are integrated, and the integrated result is sent to the recommendation system.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the fields of large models, natural language processing, deep learning and intelligent recommendation, and can be used in application scenarios such as recommendation systems, advertising, content distribution, e-commerce platforms, and search engine optimization. Specifically, it relates to a resource cold start method, apparatus, device and storage medium based on a large model. Background Technology

[0002] Recommendation systems are an indispensable part of modern internet services, and personalized recommendations provide users with great convenience. However, in cold start scenarios, because new resources have no interactive information within the recommendation system, effective recommendations are difficult to make, and the problem of poor recommendation results still exists. Summary of the Invention

[0003] This disclosure provides a resource cold start method, apparatus, device, and storage medium based on a large model.

[0004] According to one aspect of this disclosure, a resource cold start method based on a large model is provided, the method comprising:

[0005] Extract the characteristic data of the target object from the online request;

[0006] The vector of the target object is determined based on the feature data of the target object, and the first recall result is obtained from the multimodal vector index library based on the vector of the target object. The multimodal vector index library stores the multimodal vectors of new resources, and the multimodal vectors of new resources are extracted from the fine-tuned multimodal large model.

[0007] A first set of target resources is determined based on the feature data of the target objects, and a second recall result is obtained from the multimodal vector index library based on the multimodal vectors corresponding to the target resources in the first set of target resources.

[0008] The results of the first and second recalls are integrated and then sent to the recommendation system, which determines new resources to be recommended based on the integration results.

[0009] According to another aspect of this disclosure, a resource cold start apparatus based on a large model is provided, the apparatus comprising:

[0010] The parsing module is used to extract the feature data of the target object from online requests;

[0011] The first determination module is used to determine the vector of the target object based on the feature data of the target object, and to obtain the first recall result from the multimodal vector index library based on the vector of the target object. The multimodal vector index library stores the multimodal vectors of new resources, and the multimodal vectors of new resources are extracted from the fine-tuned multimodal large model.

[0012] The second determination module is used to determine a first set of target resources based on the feature data of the target object, and to obtain a second recall result from the multimodal vector index library based on the multimodal vectors corresponding to the target resources in the first set of target resources.

[0013] The integration and sending module is used to integrate the results based on the first recall result and the second recall result, and send the integration result to the recommendation system so that the recommendation system can determine the new resources to be recommended based on the integration result.

[0014] According to another aspect of this disclosure, an electronic device is provided, comprising:

[0015] At least one processor; and

[0016] The memory is communicatively connected to the at least one processor; wherein,

[0017] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the methods described in the present disclosure.

[0018] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform any of the methods according to embodiments of this disclosure.

[0019] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements any of the methods according to embodiments of this disclosure.

[0020] By combining the feature data of the target object with the multimodal vector of the new resource, personalized recommendations are achieved in the absence of historical data support. This not only improves the accuracy and diversity of recommendations, but also accelerates the exposure and distribution of new resources and promotes their effective utilization.

[0021] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0022] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0023] Figure 1 This is a flowchart illustrating the resource cold start method based on a large model according to an embodiment of the present disclosure. Figure 1 ;

[0024] Figure 2 This is a flowchart illustrating the resource cold start method based on a large model according to an embodiment of the present disclosure. Figure 2 ;

[0025] Figure 3 This is an architecture diagram of resource cold start based on a large model according to an embodiment of the present disclosure;

[0026] Figure 4 This is a schematic diagram of the structure of a resource cold start device based on a large model according to an embodiment of the present disclosure;

[0027] Figure 5 This is a schematic diagram of a scenario for a resource cold start method based on a large model according to an embodiment of the present disclosure;

[0028] Figure 6 This is a schematic diagram of the structure of an electronic device used to implement the resource cold start method based on a large model in the embodiments of this disclosure. Detailed Implementation

[0029] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0030] In this document, the term "and / or" merely describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. The term "at least one" in this document indicates any combination of at least two of a plurality of elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C. The terms "first" and "second" in this document refer to and distinguish between multiple similar technical terms, not to restrict the order or to limit there to only two. For example, "first feature" and "second feature" refer to two categories / two features; the first feature can be one or more, and the second feature can also be one or more.

[0031] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0032] Before introducing the technical solutions of the embodiments of this disclosure, the technical terms that may be used in this disclosure will be further explained:

[0033] Resource cold start: This refers to the situation where, when new resources (such as new products or content) are first introduced into the system, personalized recommendations are difficult to make directly due to a lack of user interaction data or historical behavior data. In this case, specific strategies and methods are needed to accelerate the exposure and distribution of resources and improve the accuracy and diversity of recommendations.

[0034] Multimodal data refers to data that contains multiple types of information, such as text, images, video, and audio. This data can describe the characteristics and information of an object from different angles and levels.

[0035] Multimodal vectors are high-dimensional numerical representations generated by comprehensively analyzing and processing various types of data (such as text, images, and videos) through a large multimodal model. They comprehensively reflect the characteristics and information of the data and are an important basis for recommendation systems.

[0036] Multimodal large models refer to deep learning models that can process multiple types of data simultaneously. They can capture the inherent relationships and interactions between different types of data, thereby gaining a more comprehensive understanding of the characteristics and information of objects.

[0037] A multimodal vector index is a database that stores vector representations of various types of resources (such as text, images, and videos). These vectors are extracted and integrated using advanced algorithms to capture multiple features of the resources, and can be used for tasks such as similarity calculation and cluster analysis. In resource cold start scenarios, the multimodal vector index can provide strong support for recommending new resources.

[0038] Offline logs: These are log files that record user interactions with the system. These log files contain a large amount of user behavior data, which can be used to analyze user preferences and optimize recommendation algorithms.

[0039] Resource cold start, as a startup or recommendation strategy, is primarily used to address the problem of insufficient data support for new resources or new users in the initial stage. Most recommendation systems face issues such as data sparsity and cold start.

[0040] 1. Lack of user feedback for new resources: Newly listed resources have not yet accumulated enough user reviews, ratings or purchase records since they have just been launched, making it difficult for the system to assess their popularity and relevance to users, thus making it difficult to make effective recommendations.

[0041] 2. Data sparsity: Even for users and resources with some historical data, data sparsity may still exist. That is, some users may have only interacted with a very small number of resources, while some resources may have only been followed by a very small number of users, which increases the difficulty and uncertainty of recommendations.

[0042] In related technologies, resource cold start methods mainly include the following categories:

[0043] 1. Content-based recommendation: This method utilizes the inherent attributes of resources (such as product descriptions, categories, and tags) for recommendations. Even without user interaction data, it can recommend similar items based on the user's initial preferences.

[0044] 2. Graph-based recommendation: Using graph structures to model the relationship between users and resources, such as using graph neural networks to learn the internal representation of resources and making recommendations based on the similarity between resources.

[0045] 3. Based on active learning, taking the exploration-exploitation (EE) strategy as an example, when the popularity of an item is unknown, it is first tentatively exposed to a group of seed users to observe their feedback. If the feedback is positive, the item is recommended to more users; otherwise, the number of recommendations is reduced. This method can help the recommendation system quickly discover users' interests and optimize recommendation performance.

[0046] However, all of the above cold start methods have obvious drawbacks.

[0047] For example, content-based cold start methods often cannot be customized according to users' individual needs, which may result in recommended content that does not match users' interests.

[0048] For example, other cold-start methods based on personalized models do not fully utilize resource attribute information, lack multimodal information fusion, and have limited ability to extract resource-related information. Taking e-commerce recommendations as an example, product image information is generally extracted using traditional algorithms such as Scale-Invariant Feature Transform (SIFT) and Speeded UpRobust Features (SURF) to extract key points from images, while deep learning models such as Convolutional Neural Networks (CNN) extract high-level feature representations from images. Text content information such as product titles and descriptions is extracted using bag-of-words models such as Term Frequency-Inverse Document Frequency (TF-IDF) and word embedding methods to convert words in the text into fixed-length vector representations. Separate extraction and modeling of image and text information, lacking the fusion of different modalities, cannot fully characterize the attribute information of a product.

[0049] Therefore, during the cold start phase, the interaction data between new resources and users is sparse. The model trained based on user feedback signals lacks sufficient input, and the model training cannot converge to the confidence result. Instead, the recommendation system relies more on the attribute information of the new resource itself. Due to the lack of fusion of information from different modalities, it is impossible to effectively characterize the attribute features of a new resource, which exacerbates the cold start problem.

[0050] To at least partially address one or more of the aforementioned problems and other potential issues, this disclosure proposes a resource cold-start method based on a large model. By combining the feature data of the target object with the multimodal vectors of new resources, it achieves personalized recommendations even in the absence of historical data support. This method not only improves the accuracy and diversity of recommendations but also accelerates the exposure and distribution of new resources, promoting the effective utilization of new resources and enhancing the satisfaction of the target object.

[0051] This disclosure provides a resource cold start method based on a large model. Figure 1This is a flowchart illustrating a resource cold start method based on a large model according to an embodiment of this disclosure. This method can be applied to a resource cold start device based on a large model. The large model-based resource cold start device is located in an electronic device. This electronic device includes, but is not limited to, fixed devices and / or mobile devices. For example, fixed devices include, but are not limited to, servers, which can be cloud servers or ordinary servers. Mobile devices include, but are not limited to, mobile phones and tablets. In some possible implementations, the resource cold start method based on a large model can also be implemented by a processor calling computer-readable instructions stored in memory. Figure 1 As shown, this resource cold start method based on a large model includes:

[0052] S101. Parse the feature data of the target object from the online request;

[0053] S102. Determine the vector of the target object based on the feature data of the target object, and obtain the first recall result from the multimodal vector index library based on the vector of the target object. The multimodal vector index library stores the multimodal vector of the new resource, and the multimodal vector of the new resource is extracted from the fine-tuned multimodal large model.

[0054] S103. Determine a first set of target resources based on the feature data of the target object, and obtain a second recall result from the multimodal vector index library based on the multimodal vectors corresponding to the target resources in the first set of target resources.

[0055] S104. Based on the first recall result and the second recall result, perform integration processing and send the integration result to the recommendation system so that the recommendation system can determine the new resources to be recommended based on the integration result.

[0056] In this embodiment of the disclosure, the target object refers to the subject that needs resource recommendations or services. In a personalized recommendation system, the target object can be understood as those users who need resources to be recommended.

[0057] In this embodiment of the disclosure, an online request refers to a request initiated by a target object to the system through some means (such as web browsing, application usage, search query, etc.). This online request contains some basic information, behavioral data, or a specific query requirement of the target object. By parsing these online requests, the system can understand the target object's current state, needs, and preferences, thereby providing personalized recommendation services to the target object.

[0058] In this embodiment of the disclosure, the feature data of the target object may include basic information that the target object has authorized to be disclosed (such as nickname, geographical location), interests, hobbies, historical behavior records, etc. By parsing this feature data, the needs and preferences of the target object can be initially understood, and then subsequent steps such as vector generation, similarity calculation, and recommendation can be performed.

[0059] In this embodiment of the disclosure, the multimodal vector index library stores the multimodal vectors of new resources. The multimodal vectors are vectors that extract and integrate multiple features of the new resources.

[0060] In some implementations, a vector representation of the target object is generated using machine learning or deep learning algorithms based on the object's feature data. This vector captures the key features of the target object and is used for similarity calculation in subsequent steps. Then, this vector is compared with new resource vectors in a multimodal vector index, and the top K (K is an integer greater than 1) new resources with the highest similarity are retrieved as the first recall result.

[0061] In some implementations, the first target resource set is a set of target resources (e.g., those with a posterior metric greater than a first threshold) extracted from the target object's historical consumption resource set. Based on the multimodal vectors corresponding to the target resources in this set, new resources with the highest similarity ranking (M is an integer greater than 1) are retrieved from a multimodal vector index and used as the second recall result. Here, the values ​​of M and K can be the same or different.

[0062] In some implementations, the first and second recall results are integrated. This integration process may include deduplication, sorting, and other operations to ensure that the final recommended resource set is both rich and accurate. The integrated result is then sent to the recommendation system, which determines the final new resources to be recommended based on the integration result and the user's specific needs.

[0063] For example, suppose an e-commerce platform encounters a resource cold start problem when promoting a newly launched smartwatch. Using this method, the e-commerce platform can parse user feature data such as nicknames and shopping preferences from online user requests and generate a user vector representation. Then, it uses smartwatch multimodal vectors from a multimodal vector index to calculate similarity, obtaining the top K smartwatches with the highest similarity as the first recall result. Simultaneously, based on the user's shopping history, a first target resource set is determined (such as previously purchased electronic products), and new resources with the highest similarity to this set, ranked Mth, are obtained from the multimodal vector index as the second recall result. Finally, the two recall results are integrated and recommended to the user, improving the user experience.

[0064] It should be noted that this disclosure does not restrict the execution order of S102 and S103. They can be executed simultaneously, and S102 can be executed before or after S103.

[0065] The technical solution of this disclosure combines the feature data of the target object with the multimodal vector of the new resource to achieve personalized recommendation in the absence of historical data support. This not only improves the accuracy and diversity of recommendations, but also accelerates the exposure and distribution of new resources and promotes the effective utilization of new resources.

[0066] In some embodiments, the large-model-based resource cold start method further includes:

[0067] Extract multiple modal data from multiple datasets corresponding to each published resource;

[0068] Input the multimodal data corresponding to each published resource into the multimodal large model to obtain the multimodal vector of each published resource;

[0069] By using the interaction data of published resources and objects mined from the first offline log, the multimodal large model is fine-tuned to obtain the fine-tuned multimodal large model.

[0070] In this embodiment of the disclosure, the dataset may include types such as text descriptions, images, videos, and audio, covering different aspects and characteristics of the resources.

[0071] In some implementations, multimodal data corresponding to each published resource is input into a multimodal large model. This multimodal large model can process multiple types of data simultaneously and capture the intrinsic relationships and interactions between them. Through the model's calculations, a multimodal vector for each published resource is obtained. This vector is a high-dimensional numerical representation that comprehensively reflects the characteristics and information of the resource.

[0072] In some implementations, to further improve the model's accuracy and adaptability, the multimodal large model is fine-tuned using interaction data between published resources and objects mined from the first offline log. This interaction data records actual interaction behaviors between objects and resources, such as clicks, browsing, and purchases, reflecting the objects' preferences and needs for resources. Through the fine-tuning process, the model can learn more object behavior patterns and resource characteristics, thereby more accurately predicting objects' interests and preferences for new resources.

[0073] For example, suppose an e-commerce platform is promoting a newly launched smart speaker and uses the resource cold start method described above. First, it extracts multimodal data, such as text descriptions, product images, and usage videos, from multiple datasets related to the smart speaker. Then, this data is input into a large multimodal model to obtain the smart speaker's multimodal vectors. Next, the model is fine-tuned using user interaction data with smart speaker-like products from offline logs. Finally, during the resource cold start phase, the fine-tuned large multimodal model and multimodal vector index library are used to quickly identify target users interested in the smart speaker and recommend the product to them.

[0074] Thus, by extracting multimodal data and utilizing a large multimodal model, we can more comprehensively capture the features and information of resources, thereby improving the accuracy of recommendations. Fine-tuning the model using interaction data from offline logs allows it to better adapt to changes in user behavior and needs in real-world scenarios. During the resource cold start phase, by leveraging a multimodal vector index library and the fine-tuned large multimodal model, we can quickly find similar target objects for new resources, thereby accelerating resource exposure and distribution.

[0075] In some embodiments, the large-model-based resource cold start method further includes:

[0076] Extract multimodal vectors of new resources using a fine-tuned multimodal large model;

[0077] A multimodal vector index library is constructed based on the multimodal vectors of new resources.

[0078] In some implementations, after fine-tuning the multimodal large model, the model is used to extract multimodal vectors from new resources. These new resources also contain multiple modalities of data, such as text descriptions, images, and videos. This data, after preprocessing, is input into the fine-tuned multimodal large model. Based on its learned knowledge, the model comprehensively analyzes the multiple modalities of the new resources and generates a multimodal vector that fully reflects the characteristics and information of the new resources. This vector will serve as an important basis for recommending new resources in subsequent processes.

[0079] In some implementations, after obtaining the multimodal vectors of new resources, a multimodal vector index needs to be built to store these vectors. The index is a key component of the recommendation system, allowing the system to quickly retrieve and match similar resources or users. When building the index, efficient indexing algorithms and data structures can be employed to optimize query performance. Simultaneously, to ensure the timeliness and accuracy of the index, the vector information in the index needs to be updated regularly to reflect the latest status of new resources.

[0080] For example, suppose a video platform launches a new documentary with rich text descriptions, a high-quality trailer, and detailed information about the production team. During the resource cold start phase, the platform can use a fine-tuned multimodal model to extract the documentary's multimodal vectors and add them to its multimodal vector index. Subsequently, when a user visits the platform, the recommendation system can query the index to find other documentaries or related content similar to the documentary and recommend them to the user. This not only increases the exposure of the new documentary but also enhances the user's viewing experience and satisfaction.

[0081] Thus, by extracting multimodal vectors of new resources in real time and building an index library, the recommendation system can quickly incorporate new resources into its recommendations, improving the real-time performance of recommendations. Simultaneously, the introduction of multimodal vectors makes the recommendation process more comprehensive and accurate, better meeting users' personalized needs. During the resource cold start phase, traditional recommendation algorithms often fail due to a lack of user interaction data. However, by using a finely tuned multimodal large model to extract multimodal vectors of new resources and building an index library for similarity matching, the resource cold start problem can be effectively alleviated, increasing the exposure and click-through rate of new resources. With the continuous release of new resources and the constant changes in user needs, the recommendation system requires good scalability and flexibility. By building a multimodal vector index library and continuously optimizing and updating the indexing algorithm and data structure, the recommendation system can adapt to these changes and maintain efficient and stable operation.

[0082] In some embodiments, fine-tuning of a multimodal large model is performed using interaction data of published resources and objects mined from a first offline log, including:

[0083] A content-based collaborative filtering (CB2CF) model is adopted to fine-tune the multimodal large model by using interaction data of published resources and objects mined from the first offline log.

[0084] In this embodiment, CB2CF is a recommendation algorithm that combines content filtering and collaborative filtering. It uses the content features of resources to predict user preferences and incorporates the user's historical interaction behavior to optimize the recommendation results. The CB2CF model can simultaneously consider the intrinsic attributes of resources and the extrinsic behavioral characteristics of users, thereby providing a more comprehensive and accurate recommendation service.

[0085] In some implementations, interaction data between published resources and objects is mined from a first offline log. This data may include user clicks, browsing, comments, and other behavioral records, as well as multimodal information related to the resources, such as text descriptions, images, and videos. The mined data is then cleaned and preprocessed to remove noise and outliers, ensuring data quality and consistency.

[0086] In some implementations, the CB2CF model is used as the basis for fine-tuning the multimodal large model. The CB2CF model combines the advantages of content filtering and collaborative filtering, and can optimize recommendation performance by leveraging the content features of resources and user interaction behavior.

[0087] In some implementations, preprocessed interaction data is used as a training set and input into a multimodal large model. During training, the model learns the correlation between the multimodal features of resources and user behavior. Utilizing the mechanism of the CB2CF model, it predicts user preferences for resources based on content features (such as text descriptions and image features), and combines this with collaborative filtering to optimize these predictions. Specifically, the model attempts to find other resources similar to those in the user's historical interactions and recommends these resources to the user. During fine-tuning, the model's parameters and training strategies, such as learning rate, batch size, and training epochs, can be adjusted according to actual conditions to achieve better recommendation results.

[0088] In some implementations, appropriate evaluation metrics (such as precision, recall, etc.) are used to assess the performance of the fine-tuned multimodal large model. Based on the evaluation results, the model is further optimized and adjusted to improve recommendation effectiveness and user satisfaction.

[0089] For example, suppose an e-commerce platform launches a new smartwatch with various functions and features, providing detailed text descriptions and high-resolution images on its product details page. By mining user interaction data with similar smartwatches from offline logs (such as purchase history and browsing duration), and fine-tuning a multimodal model using a CB2CF model, the model learns the correlation between the smartwatch's text descriptions and image features and user purchasing behavior. When a new user visits the platform and views the smartwatch, the model can predict the user's preference for the watch based on its multimodal features and the user's potential needs, and recommend it to the user, thus improving the user experience.

[0090] Thus, by combining content features and user behavior data, the fine-tuned multimodal model can more accurately predict user preferences and needs, thereby providing more personalized recommendation services. For newly released resources or new users, traditional recommendation algorithms often fail due to a lack of sufficient interaction data. Content-based collaborative filtering models, however, can utilize the content features of resources for recommendations, thus alleviating the cold start problem to some extent. As new resources are continuously released and user behavior changes, the fine-tuned multimodal model can continuously learn and optimize its recommendation capabilities to adapt to these changes and maintain efficient recommendation performance.

[0091] In some embodiments, determining the vector of a target object based on the feature data of the target object includes: using a collaborative filtering (CF) model employing the CB2CF model to determine the vector of the target object based on the feature data of the target object.

[0092] Here, the CF model is the CF model of the CB2CF model used when fine-tuning a large multimodal model.

[0093] In this embodiment of the disclosure, the CF model is a recommendation algorithm that analyzes user-item interaction data (such as ratings, purchase records, etc.) to discover similarities between user interests or items and makes recommendations accordingly.

[0094] In some implementations, a suitable collaborative filtering (CF) model is selected based on the application scenario and data characteristics. Common CF models include user-based collaborative filtering (User-based CF) and item-based collaborative filtering (Item-based CF). User-based collaborative filtering focuses on finding other users with similar interests to the target user and recommending items that these users like; while item-based collaborative filtering focuses on finding other items similar to the target item and recommending them to users who like these similar items.

[0095] In some implementations, interaction data related to the target object is extracted, including user ratings, purchase records, browsing history, etc. The data is preprocessed, including noise removal, missing value imputation, and normalization, to ensure data quality.

[0096] In some implementations, within the framework of the CB2CF model, the output of the CF model can be used as part or all of the target object vector. For item-based CF, the target object vector can be directly obtained by weighted averaging of the feature vectors of its similar items. For user-based CF, if it is necessary to incorporate user features into the target object vector, the target object vector can be constructed by calculating the similarity between the target user and other users and combining the feature vectors of these users.

[0097] In some implementations, considering that the CB2CF model also incorporates content-based (CB) features, the target object vector can be further optimized by fusing content features and collaborative filtering results. For example, a deep learning model (such as a neural network) can be used as input to the content features and collaborative filtering results to train the model and generate a more accurate target object vector.

[0098] For example, suppose an e-commerce platform is using a CB2CF model for product recommendation. For a newly listed product A, we can determine its vector as follows: Use an item-based collaborative filtering model to find other products similar to product A (such as product B, product C, etc.). Extract feature vectors from these similar products (such as price, brand, category, user ratings, etc.). Based on the similarity between product A and these similar products, perform a weighted average of these feature vectors to obtain a preliminary vector for product A. Combine the content features of product A (such as product description, images, videos, etc.) and use a deep learning model to optimize and adjust the preliminary vector to obtain the final vector for product A.

[0099] Thus, by combining collaborative filtering and content features, user interests and item attributes can be captured more comprehensively, thereby improving recommendation accuracy. For new users or new items, even in the absence of interaction data, their vectors can be constructed using content features, allowing for initial recommendations. Collaborative filtering models can explain the reasons behind recommendation results (such as similar users or similar items), thereby improving the model's interpretability and user trust.

[0100] In some embodiments, determining a first set of target resources based on feature data of the target object includes:

[0101] Based on the characteristic data of the target object, obtain the list of published resources consumed by the target object from the second offline log;

[0102] The first target resource set is formed by filtering out published resources whose posterior metrics are greater than the first threshold from the list of published resources.

[0103] In some implementations, obtaining the list of consumed resources from a second offline log includes: retrieving a list of published resources consumed by the target object from the second offline log using the target object's unique identifier (such as a user identifier, ID). The second offline log is typically a large database storing user behavior data (such as clicks, views, purchases, etc.), which is used for subsequent analysis and recommendations.

[0104] In some implementations, filtering resources with posterior metrics greater than a first threshold includes: for a list of published resources obtained from logs, calculating or extracting the posterior metric for each resource. The posterior metric is a quantitative standard for measuring resource quality or user satisfaction, and may include user ratings, number of comments, conversion rate, repurchase rate, etc. Published resources with posterior metrics greater than the first threshold are then selected and included in the first target resource set.

[0105] In this embodiment of the disclosure, the first threshold is a judgment standard set in advance based on factors such as business needs, data distribution, and actual results.

[0106] For example, suppose an e-commerce platform wants to recommend high-quality products to its users. First, the platform extracts feature data from the user's profile and purchase history. Then, it searches for a list of products the user has purchased in offline logs. Next, the platform calculates posterior metrics for each product, such as user ratings, number of reviews, and conversion rate. Let's assume the first threshold is set at 4.5 out of 5. Finally, the platform filters out all products with user ratings greater than 4.5, forming the first target resource set.

[0107] Thus, by selecting resources with excellent posterior metrics, we can ensure that the content recommended to users better matches their needs and interests, thereby improving the accuracy of recommendations and user satisfaction. With limited resources, allocating more exposure opportunities and resources to verified, high-quality resources helps improve overall content quality and user experience. Automated filtering and recommendation processes can reduce the cost of resource cold starts.

[0108] In some embodiments, the large-model-based resource cold start method further includes:

[0109] Based on the first target resource set, a multimodal vector storage system is requested to obtain the multimodal vector corresponding to the target resource in the first target resource set. The multimodal vector storage system stores the multimodal vectors of published resources and the multimodal vectors of new resources.

[0110] In this embodiment of the disclosure, the multimodal vector storage system is a system capable of storing and retrieving multimodal vectors of published and new resources. This system typically possesses efficient indexing and query mechanisms to quickly respond to requests from the recommendation system.

[0111] In the resource cold start method, requesting the multimodal vector storage system based on the first target resource set to obtain the multimodal vector corresponding to the target resource is an important step to improve the recommendation system's ability to understand and process new resources.

[0112] In some implementations, a request to the multimodal vector storage system is constructed based on the resource ID or unique identifier in the first target resource set. The type of multimodal vector to be retrieved is determined, such as text vectors, image vectors, audio vectors, etc., and any additional parameters that may be required (such as vector dimension, similarity measurement method, etc.). The prepared request is then sent to the multimodal vector storage system. This system should already have pre-stored multimodal vectors of published and new resources.

[0113] In some implementations, a response is received from a multimodal vector storage system, containing multimodal vectors of the resource specified in the request. The received multimodal vectors are then processed as necessary, such as format conversion and normalization, for subsequent use in the recommendation system. The processed multimodal vectors are integrated into relevant modules of the recommendation system, such as feature extraction and similarity calculation modules. These multimodal vectors are then used to enhance the recommendation system's understanding and recommendation capabilities for new resources.

[0114] For example, suppose a video recommendation platform wants to recommend newly released movies to its users. The platform first determines a primary target resource set based on users' historical behavior data—that is, movie genres or directors the user might be interested in. Then, the platform sends a request to a multimodal vector storage system to retrieve the multimodal vectors (including vectors corresponding to text descriptions, poster images, trailer videos, etc.) of these movies and the newly released movies. Finally, the platform uses these multimodal vectors to calculate the similarity between the user and the new movies, and recommends movies with high similarity to the user, thus improving the user experience.

[0115] Thus, by introducing multimodal vectors, recommender systems can gain a more comprehensive understanding of resource features and content, thereby more accurately matching user interests and needs. For newly released resources, even without user interaction data, their multimodal vectors can be used for initial display and promotion within the recommender system, accelerating their cold start process. The introduction of a multimodal vector storage system enables recommender systems to easily handle resources of different types and formats, enhancing the system's scalability and flexibility.

[0116] In some embodiments, such as Figure 2 As shown, this resource cold start method based on a large model also includes:

[0117] S201: Calculate the similarity between the multimodal vector of the new resource and the multimodal vector of the published resource to identify a second set of target resources whose similarity to the new resource is higher than the second threshold;

[0118] S202: Utilize historical data of published resources in the second target resource set to generate recommendation information for new resources;

[0119] S203: Push new resources to the target audience in the form of recommendation information.

[0120] In this embodiment of the disclosure, similarity calculation refers to the process of measuring the degree of similarity between two or more objects. In recommendation systems, similarity calculation is often used to discover the inherent connections between user interests or resources, or the inherent connections between resources.

[0121] In this embodiment of the disclosure, recommendation information refers to information used to recommend relevant resources or services to users. Recommendation information typically includes a description of the resource, its highlights, and the reasons for the recommendation.

[0122] In some implementations, the multimodal vector library of published resources is traversed, and the similarity between the new resource and each published resource is calculated one by one. Similarity calculation can employ various metrics such as cosine similarity and Euclidean distance, with the specific choice depending on data characteristics and business requirements.

[0123] In some implementations, a second threshold is set to determine the similarity level. Only published resources with a similarity level higher than this threshold are included in the second target resource set.

[0124] In some implementations, for each published resource in the second target resource set, historical data such as user ratings, reviews, page views, and conversion rates are extracted. This historical data is analyzed to understand user preferences and reactions to similar resources.

[0125] In some implementations, recommendations are generated based on the analysis results for new resources. These recommendations may include the highlights of the new resource, its similarities and differences with similar resources, and reasons why users might be interested in it.

[0126] In some implementations, the generated recommendation information is pushed to the target audience in an appropriate format (such as pop-ups, recommendation lists, email notifications, etc.). The effectiveness of the push is monitored, and feedback data from the target audience on the new resources is collected for subsequent optimization and improvement.

[0127] For example, suppose an e-commerce platform launches a new smartwatch with multiple functions and innovations. The platform first extracts the smartwatch's multimodal vectors, including product descriptions, images, and videos. Then, it calculates the similarity between these vectors and the multimodal vectors of already released smartwatches, identifying those with similarity scores above a set threshold as references. Next, the platform analyzes the historical sales data and user reviews of these reference smartwatches, discovering that users are particularly interested in the health monitoring and activity tracking functions. Based on this information, the platform generates recommendations for the new smartwatch, emphasizing its health management and activity assistance features, and pushes these recommendations to potentially interested users.

[0128] In this way, by utilizing historical data and user feedback from similar resources, recommendation information for new resources can be quickly generated and pushed to target users, effectively shortening the cold start cycle of new resources. Similarity calculation based on multimodal vectors can more accurately capture the inherent connections between resources, thereby generating recommendation information that better matches user interests and needs, improving recommendation accuracy. Users can promptly access new resources relevant to their interests, improving the user experience.

[0129] In some embodiments, the large-model-based resource cold start method further includes:

[0130] The target group is determined based on the objects corresponding to the second target resource set. The target group is the object actively pushed with new resources.

[0131] Push new resources to the terminals corresponding to the target group;

[0132] Collect feedback from the target group regarding the new resources;

[0133] Based on feedback information, the multimodal large model is dynamically adjusted and optimized.

[0134] The resource cold start method further includes determining the target group based on the objects corresponding to the second target resource set, pushing new resources to these target groups, and collecting feedback information to dynamically adjust and optimize the multimodal large model.

[0135] In some implementations, identifying the target group includes analyzing the user characteristics of each published resource in the second target resource set, including geographic location, interests, and preferences. Using this characteristic information, techniques such as cluster analysis and user profiling are employed to identify user groups with high similarity to the new resource as the target group. The determination of the target group should comprehensively consider users' historical behavior data, real-time behavior data, and potential needs to improve the accuracy of the push notifications.

[0136] In some implementations, new resources are pushed to the corresponding terminals (such as mobile phones, computers, tablets, etc.) of a defined target group. Push methods can include various forms such as push notifications, email notifications, and in-app pop-ups. When pushing content, emphasis should be placed on personalization and customization, providing differentiated recommended content based on different users' interests, preferences, and needs.

[0137] In some implementations, feedback from the target group regarding the new resource is collected through user behavior data (such as clicks, browsing, purchases, and reviews) and proactive feedback (such as surveys and user ratings). This feedback is then organized and analyzed to extract useful information, such as user satisfaction with the new resource and suggestions for improvement.

[0138] In some implementations, dynamically adjusting and optimizing the multimodal large model includes: dynamically adjusting and optimizing the multimodal large model based on collected feedback information. This includes adjusting the model's parameters, updating the model's training data, and improving the model's algorithm. Through continuous learning and optimization, the multimodal large model's ability to understand and recommend new resources is improved, making it more in line with users' actual needs and interests.

[0139] For example, suppose a music platform releases a new original song. The platform first uses multimodal vector technology to calculate the similarity between this song and other released songs, identifying the user groups corresponding to songs with high similarity as the target audience. Then, the platform pushes the new song to these target users via in-app pop-ups. For a period after the push, the platform collects user feedback information such as clicks, plays, shares, and comments on the song, and finds that users highly rate the melody and lyrics. Based on this feedback, the platform optimizes and adjusts the multimodal vector model, improving its ability to recommend similar new songs.

[0140] In this way, by accurately identifying the target audience and providing personalized new resources, the accuracy of recommendations and user satisfaction are improved. By rapidly collecting user feedback and optimizing the model, the cold start process for new resources is accelerated, allowing them to be accepted and recognized by users more quickly. Through continuous optimization of the recommendation algorithm and model, users are provided with recommended content that better matches their interests and needs, enhancing the user experience.

[0141] Figure 3 The diagram illustrates an architecture for resource cold start based on a large model, such as... Figure 3 As shown, the resource cold start based on the large model is divided into two parts: offline and online. The offline part mainly involves fine-tuning the multimodal large model, while the online part includes the application of two recall methods: User to Item (U2I) and Resource to Item (I2I).

[0142] The offline process includes: extracting modal data for each resource from the resource image / text dataset, resource text dataset, and resource audio dataset, respectively, and using a multimodal fusion model to obtain the multimodal fusion feature vector for each resource. To improve the accuracy of personalized recommendations, interaction data between historically distributed resources and users is mined from offline logs, and the CB2CF model is fine-tuned to better adapt to the data distribution and feature representation of specific recommendation scenarios. The fine-tuned multimodal model is then used to calculate the multimodal vector of newly released resources. The multimodal vectors of all resources are stored in a key-value store (Redis), and a multimodal vector index is built using the new resource vectors for real-time access by online requests.

[0143] The online processes include the U2I recall process and the I2I recall process.

[0144] The U2I recall process includes:

[0145] Parse user-related feature data from online real-time requests;

[0146] The CF model used for fine-tuning the CB2CF model is used to estimate the user CF vector corresponding to the current request.

[0147] Using user CF vectors, the multimodal vector index library built in the offline part is indexed, and the resources with the top K similarity are taken as the U2I recall results (i.e., the first recall results).

[0148] The I2I recall process includes:

[0149] Based on the parsed user data and combined with offline logs, the list of resources consumed by the current user is parsed, and the user's consumed resources are filtered using post-hoc metrics such as click-through rate to obtain a set of resources that satisfies the user (i.e., the first target resource set).

[0150] Using a set of resources that satisfies the user, request the Redis in the offline part to obtain the multimodal vector corresponding to the target resource;

[0151] Using the target resource multimodal vector, the multimodal vector index library built in the offline part is indexed, and the resources with the top M similarity are taken as the I2I recall result (i.e., the second recall result).

[0152] The U2I recall result list and I2I recall result list are deduplicated and returned to the downstream ranking model.

[0153] Compared to common resource cold start methods for recommendation systems, the resource cold start method disclosed in this paper has the following two significant advantages:

[0154] First, multimodal fusion can significantly improve the model's ability to learn and understand resource attributes, and more comprehensively and accurately characterize the resource's own attribute features, which is especially important in the cold start stage when resource and user interaction behaviors are sparse.

[0155] Second, the multimodal large model used in this disclosure has a large number of parameters, enabling it to learn more complex data patterns and knowledge representations. Its ability to understand image or text modalities is also stronger than that of the single-modal model. The knowledge learned by the large model during training is more general, and it can maintain high accuracy and stability when processing data from different domains.

[0156] The proposed solution can effectively improve the cold start efficiency of new resources in recommendation systems, with a particularly significant improvement for long-tail new resources.

[0157] This disclosure provides a resource cold start device based on a large model, such as... Figure 4 As shown, the device may include: a parsing module 401, used to parse feature data of a target object from an online request; a first determining module 402, used to determine the vector of the target object based on the feature data of the target object, and obtain a first recall result from a multimodal vector index library based on the vector of the target object, wherein the multimodal vector index library stores the multimodal vectors of new resources, and the multimodal vectors of new resources are extracted by a fine-tuned multimodal large model; a second determining module 403, used to determine a first set of target resources based on the feature data of the target object, and obtain a second recall result from the multimodal vector index library based on the multimodal vectors corresponding to the target resources in the first set of target resources; and an integration and sending module 404, used to integrate the first recall result and the second recall result, and send the integration result to a recommendation system so that the recommendation system can determine new resources to be recommended based on the integration result.

[0158] In some embodiments, the apparatus further includes: a first extraction module, configured to extract multiple modal data from multiple datasets corresponding to each published resource; an acquisition module, configured to input the multiple modal data corresponding to each published resource into a multimodal large model to obtain a multimodal vector for each published resource; and a fine-tuning module, configured to fine-tune the multimodal large model using interaction data between published resources and objects mined from a first offline log to obtain a fine-tuned multimodal large model.

[0159] In some embodiments, the apparatus further includes: a second extraction module for extracting multimodal vectors of new resources using a fine-tuned multimodal large model; and a construction module for constructing a multimodal vector index library based on the multimodal vectors of the new resources.

[0160] In some embodiments, the fine-tuning module is configured to: fine-tune a multimodal large model using interaction data of published resources and objects mined from a first offline log, employing a CB2CF-based model.

[0161] In some embodiments, the first determining module 402 includes: a first determining submodule, used to determine the vector of the target object based on the CF model used when fine-tuning a multimodal large model based on the CB2CF model.

[0162] In some embodiments, the second determining module 403 includes: an acquisition submodule, configured to acquire a list of published resources consumed by the target object from a second offline log based on the feature data of the target object; and a filtering submodule, configured to filter published resources from the list of published resources whose posterior metrics are greater than a first threshold, to form a first target resource set.

[0163] In some embodiments, the second determining module 403 further includes: a second determining submodule, configured to request a multimodal vector storage system based on a first target resource set to obtain multimodal vectors corresponding to target resources in the first target resource set, wherein the multimodal vector storage system stores multimodal vectors of published resources and multimodal vectors of new resources.

[0164] In some embodiments, the apparatus further includes: a third determining module, configured to perform similarity calculation between the multimodal vector of the new resource and the multimodal vector of the published resource, so as to identify a second set of target resources with a similarity higher than a second threshold to the new resource; a generating module, configured to generate recommendation information for the new resource using historical data of the published resources in the second set of target resources; and a first pushing module, configured to push the new resource to the target object in the form of recommendation information.

[0165] In some embodiments, the device further includes: a fourth determining module, configured to determine a target group based on objects corresponding to the second target resource set, wherein the target group is the active target of the new resource; a second pushing module, configured to push the new resource to the terminal corresponding to the target group; a collection module, configured to collect feedback information from the target group regarding the new resource; and an adjustment and optimization module, configured to dynamically adjust and optimize the multimodal large model based on the feedback information.

[0166] The specific functions and examples of each module and submodule of the apparatus in this disclosure can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0167] The resource cold start device based on a large model disclosed in this embodiment achieves personalized recommendations in the absence of historical data support by combining the feature data of the target object and the multimodal vector of the new resource. This not only improves the accuracy and diversity of recommendations, but also accelerates the exposure and distribution of new resources and promotes the effective utilization of new resources.

[0168] This disclosure provides a scenario illustration of a resource cold start method based on a large model, such as... Figure 5 As shown.

[0169] As previously described, the resource cold start method based on a large model provided in this disclosure is applied to electronic devices. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers. Electronic devices can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices.

[0170] Specifically, the electronic device may perform the following operations:

[0171] Extract the characteristic data of the target object from the online request;

[0172] The vector of the target object is determined based on the feature data of the target object, and the first recall result is obtained from the multimodal vector index library based on the vector of the target object. The multimodal vector index library stores the multimodal vectors of new resources, and the multimodal vectors of new resources are extracted from the fine-tuned multimodal large model.

[0173] A first set of target resources is determined based on the feature data of the target objects, and a second recall result is obtained from the multimodal vector index library based on the multimodal vectors corresponding to the target resources in the first set of target resources.

[0174] The results of the first and second recalls are integrated and then sent to the recommendation system, which determines new resources to be recommended based on the integration results.

[0175] It should be understood that Figure 5 The scene diagrams shown are merely illustrative and not restrictive; those skilled in the art can interpret them based on... Figure 5 Even with various obvious changes and / or substitutions to the examples, the resulting technical solutions still fall within the scope of this disclosure.

[0176] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0177] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0178] Figure 6 A schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0179] like Figure 6As shown, device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 602 or a computer program loaded from storage unit 608 into random access memory (RAM) 603. RAM 603 may also store various programs and data required for the operation of device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0180] Multiple components in device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0181] The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processors, controllers, microcontrollers, etc. The computing unit 601 performs the various methods and processes described above, such as the large model-based resource cold start method. For example, in some embodiments, the large model-based resource cold start method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of the large model-based resource cold start method described above can be performed. Alternatively, in other embodiments, computing unit 601 may be configured by any other suitable means (e.g., by means of firmware) to perform a resource cold start method based on a large model.

[0182] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0183] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0184] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory (EPROM), flash memory, optical fiber, compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0185] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a cathode ray tube (CRT) or liquid crystal display (LCD) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0186] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0187] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0188] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0189] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A resource cold start method based on a large model, comprising: Extract the characteristic data of the target object from the online request; The vector of the target object is determined based on the feature data of the target object, and a first recall result is obtained from the multimodal vector index library based on the vector of the target object. The multimodal vector index library stores the multimodal vectors of new resources, and the multimodal vectors of the new resources are extracted by a fine-tuned multimodal large model. The first recall result includes the top K new resources in the multimodal vector index library with the top K vector similarity to the target object, where K is greater than 1 and K is a positive integer. A first target resource set is determined based on the feature data of the target object, and a second recall result is obtained from the multimodal vector index based on the multimodal vectors corresponding to the target resources in the first target resource set. The first target resource set includes the historical consumption resources of the target object. The multimodal vectors corresponding to the target resources in the first target resource set are obtained by requesting a multimodal vector storage system based on the first target resource set. The multimodal vector storage system stores the multimodal vectors of published resources and the multimodal vectors of new resources. The second recall result includes the top M new resources in the first target resource set with the highest similarity to the multimodal vectors corresponding to the target resources, where M is greater than 1 and M is a positive integer. The first recall result and the second recall result are integrated and processed, and the integrated result is sent to the recommendation system so that the recommendation system can determine new resources to be recommended based on the integrated result; The method further includes: Extract multimodal vectors of new resources using a fine-tuned multimodal large model; A multimodal vector index library is constructed based on the multimodal vectors of the new resources.

2. The method according to claim 1, further comprising: Extract multiple modal data from multiple datasets corresponding to each published resource; Input the multimodal data corresponding to each published resource into the multimodal large model to obtain the multimodal vector of each published resource; The multimodal large model is fine-tuned using the interaction data of published resources and objects mined from the first offline log, resulting in a fine-tuned multimodal large model.

3. The method according to claim 2, wherein, The multimodal large model is fine-tuned using interaction data of published resources and objects mined from the first offline log, including: A content-based collaborative filtering (CB2CF) model is adopted, and the interaction data of published resources and objects mined from the first offline log is used to fine-tune the multimodal large model.

4. The method according to claim 3, wherein, Determining the vector of the target object based on its feature data includes: The collaborative filtering CF model using the CB2CF model determines the vector of the target object based on the feature data of the target object.

5. The method according to claim 1, wherein, Determining the first target resource set based on the feature data of the target object includes: Based on the feature data of the target object, obtain the list of published resources consumed by the target object from the second offline log; The first target resource set is formed by selecting published resources whose posterior metrics are greater than the first threshold from the published resource list.

6. The method according to claim 1, further comprising: The multimodal vectors of new resources are compared with the multimodal vectors of published resources to identify a second set of target resources whose similarity to the new resources is higher than a second threshold. Using historical data of published resources in the second target resource set, recommendation information is generated for the new resource; The new resources are pushed to the target object in the form of the recommended information.

7. The method according to claim 6, further comprising: The target group is determined based on the objects corresponding to the second target resource set, and the target group is the object actively pushed by the new resource; The new resources are pushed to the terminals corresponding to the target group. Collect feedback from the target group regarding the new resource; Based on the feedback information, the multimodal large model is dynamically adjusted and optimized.

8. A resource cold start device based on a large model, comprising: The parsing module is used to extract the feature data of the target object from online requests; The first determining module is used to determine the vector of the target object based on the feature data of the target object, and to obtain a first recall result from the multimodal vector index library based on the vector of the target object. The multimodal vector index library stores the multimodal vectors of new resources, and the multimodal vectors of the new resources are extracted by a fine-tuned multimodal large model. The first recall result includes the top K new resources in the multimodal vector index library with the top K vector similarity to the target object, where K is greater than 1 and K is a positive integer. The second determining module is used to determine a first target resource set based on the feature data of the target object, and to obtain a second recall result from the multimodal vector index library based on the multimodal vectors corresponding to the target resources in the first target resource set; the first target resource set includes the historical consumption resources of the target object; the multimodal vectors corresponding to the target resources in the first target resource set are obtained by requesting a multimodal vector storage system based on the first target resource set, and the multimodal vector storage system stores the multimodal vectors of published resources and the multimodal vectors of new resources; the second recall result includes the top M new resources in the first target resource set with the highest similarity to the multimodal vectors corresponding to the target resources, where M is greater than 1 and M is a positive integer; An integration and sending module is used to integrate the first recall result and the second recall result, and send the integration result to the recommendation system so that the recommendation system can determine new resources to be recommended based on the integration result; The device further includes: The second extraction module is used to extract the multimodal vectors of new resources using the fine-tuned multimodal large model; A construction module is used to construct a multimodal vector index library based on the multimodal vectors of the new resource.

9. The apparatus according to claim 8, further comprising: The first extraction module is used to extract multiple modal data from multiple datasets corresponding to each published resource; The acquisition module is used to input the multimodal data corresponding to each published resource into the multimodal large model to obtain the multimodal vector of each published resource; The fine-tuning module is used to fine-tune the multimodal large model by using the interaction data of published resources and objects mined from the first offline log, so as to obtain the fine-tuned multimodal large model.

10. The apparatus according to claim 9, wherein, The fine-tuning module is used for: A content-based collaborative filtering (CB2CF) model is adopted, and the interaction data of published resources and objects mined from the first offline log is used to fine-tune the multimodal large model.

11. The apparatus according to claim 10, wherein, The first determining module includes: The first determining submodule is used to determine the vector of the target object based on the feature data of the target object by employing the collaborative filtering CF model of the CB2CF model.

12. The apparatus according to claim 8, wherein, The second determining module includes: The acquisition submodule is used to obtain a list of published resources consumed by the target object from the second offline log based on the feature data of the target object; The filtering submodule is used to filter out published resources from the published resource list whose posterior indicators are greater than a first threshold, forming the first target resource set.

13. The apparatus according to claim 8, further comprising: The third determination module is used to calculate the similarity between the multimodal vector of the new resource and the multimodal vector of the published resource, so as to identify the second set of target resources with a similarity higher than the second threshold with the new resource. The generation module is used to generate recommendation information for the new resource by utilizing historical data of published resources in the second target resource set; The first push module is used to push the new resources to the target object in the form of the recommendation information.

14. The apparatus of claim 13, further comprising: The fourth determining module is used to determine the target group based on the objects corresponding to the second target resource set, wherein the target group is the active push object of the new resource; The second push module is used to push the new resources to the terminals corresponding to the target group; A collection module is used to collect feedback information from the target group regarding the new resource; The adjustment and optimization module is used to dynamically adjust and optimize the multimodal large model based on the feedback information.

15. An electronic device comprising: At least one processor; as well as A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor to enable the at least one processor to perform the method of any one of claims 1-7.

16. A non-transitory computer-readable storage medium storing computer instructions, wherein, Computer instructions are used to cause a computer to perform the method according to any one of claims 1-7.

17. A computer program product comprising a computer program stored on a storage medium, wherein the computer program, when executed by a processor, implements the method according to any one of claims 1-7.

Images