Resource recommendation method, feature extraction model training method, device and equipment

Abstract

The invention provides a resource recommendation method, a feature extraction model training method, devices and equipment, and relates to the technical field of artificial intelligence. The resource recommendation method comprises the following steps: in response to a resource recommendation request, obtaining a historical behavior sequence and candidate resources of a target object contained in the resource recommendation request; inputting the historical interaction resources and the candidate resources into a feature extraction model for feature extraction to obtain respective meta-embedding sets of the historical interaction resources and the candidate resources; determining a target fusion strategy corresponding to a cooperative relationship type from preset fusion strategies according to the cooperative relationship type between the historical interaction resources and the candidate resources; performing fusion processing on the plurality of meta-inserts of the historical interaction resources and the plurality of meta-inserts of the candidate resources to obtain historical resource features of the historical interaction resources and candidate resource features of the candidate resources; and according to the similarity between the historical resource features and the candidate resource features, generating a resource recommendation result for the target object.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the field of resource recommendation technology, and more specifically, to a resource recommendation method, a training method for a feature extraction model, an apparatus, and a device. Background Technology

[0002] E-commerce platforms typically use recommendation systems to analyze users' historical interactions with products (such as likes, favorites, and purchases) to predict user preferences. For example, by assigning unique identifiers to users and items, the association between these unique identifiers is used to characterize historical interactions and predict user preferences.

[0003] In realizing the concept disclosed herein, the inventors discovered at least the following problems in the related technologies: there is relatively little interaction data between some products and users, making it difficult to accurately determine the association between the unique identifiers of users and items, resulting in low accuracy of model recommendations. Summary of the Invention

[0004] In view of this, this disclosure provides a resource recommendation method, a training method for a feature extraction model, an apparatus, an electronic device, a storage medium, and a program product.

[0005] One aspect of this disclosure provides a resource recommendation method, comprising: responding to a resource recommendation request, acquiring a historical behavior sequence of a target object and candidate resources contained in the resource recommendation request, wherein the historical behavior sequence includes multiple historical interaction resources; inputting the historical interaction resources and candidate resources into a feature extraction model for feature extraction to obtain a set of meta-embeddings for each of the historical interaction resources and candidate resources, wherein the set of meta-embeddings includes meta-embeddings under multiple preset semantic dimensions; determining a target fusion strategy corresponding to the type of collaborative relationship between the historical interaction resources and candidate resources from a preset fusion strategy; using the target fusion strategy to perform fusion processing on the multiple meta-embeddings of the historical interaction resources and the multiple meta-embeddings of the candidate resources respectively to obtain historical resource features of the historical interaction resources and candidate resource features of the candidate resources; and performing resource recommendation processing based on the similarity between the historical resource features and the candidate resource features to generate a resource recommendation result for the target object.

[0006] Another aspect of this disclosure provides a method for training a feature extraction model, comprising: acquiring multiple sample resource pairs, wherein the sample resource pairs are labeled with a type of collaborative relationship representing the collaborative relationship between sample resources in the sample resource pair, and the multiple sample resource pairs contain at least two types of collaborative relationship; inputting the multiple sample resource pairs into a feature extraction model for feature extraction, generating a set of meta-embeddings for each of the multiple sample resources, wherein the set of meta-embeddings includes meta-embeddings of the sample resources under multiple preset semantic dimensions; for each sample resource, performing fusion processing on the meta-embeddings of the sample resource under multiple preset semantic dimensions according to an initial fusion strategy corresponding to the collaborative relationship type of the sample resource pair to which the sample resource belongs, to obtain sample resource features of the sample resource; determining the similarity between sample resources based on the sample resource features, and adjusting the network parameters and initial fusion strategy of the feature extraction model according to the loss value determined by the similarity, to obtain a trained feature extraction model.

[0007] Another aspect of this disclosure provides a resource recommendation apparatus, comprising: a first acquisition module, configured to, in response to a resource recommendation request, acquire a historical behavior sequence and candidate resources of a target object contained in the resource recommendation request, wherein the historical behavior sequence includes multiple historical interaction resources; a first feature extraction module, configured to input the historical interaction resources and candidate resources into a feature extraction model for feature extraction, obtaining a meta-embedding set for each of the historical interaction resources and candidate resources, wherein the meta-embedding set includes meta-embeddings under multiple preset semantic dimensions; a first determination module, configured to, based on the type of collaborative relationship between the historical interaction resources and candidate resources, determine a target fusion strategy corresponding to the type of collaborative relationship from a preset fusion strategy; a first fusion processing module, configured to, using the target fusion strategy, perform fusion processing on the multiple meta-embeddings of the historical interaction resources and the multiple meta-embeddings of the candidate resources respectively, to obtain historical resource features of the historical interaction resources and candidate resource features of the candidate resources; and a resource recommendation module, configured to, based on the similarity between the historical resource features and the candidate resource features, perform resource recommendation processing to generate a resource recommendation result for the target object.

[0008] Another aspect of this disclosure provides a training apparatus for a feature extraction model, comprising: a second acquisition module for acquiring multiple sample resource pairs, wherein the sample resource pairs are labeled with a collaborative relationship type representing the collaborative relationship between sample resources in the sample resource pair, and the multiple sample resource pairs include at least two collaborative relationship types; a second feature extraction module for inputting the multiple sample resource pairs into a feature extraction model for feature extraction, generating a set of meta-embedded features for each of the multiple sample resources, wherein the set of meta-embedded features includes meta-embedded features of the sample resources under multiple preset semantic dimensions; a second fusion processing module for fusing the meta-embedded features of the sample resources under multiple preset semantic dimensions for each sample resource according to an initial fusion strategy corresponding to the collaborative relationship type of the sample resource pair to which the sample resource belongs, thereby obtaining sample resource features; and an adjustment module for determining the similarity between sample resources based on the sample resource features, and adjusting the network parameters and the initial fusion strategy of the feature extraction model according to the loss value determined by the similarity, thereby obtaining a trained feature extraction model.

[0009] This disclosure also provides an electronic device, including one or more processors; and a memory for storing one or more programs that, when executed by the one or more processors, cause the one or more processors to perform the method described above.

[0010] Another aspect of this disclosure provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the method described above.

[0011] Another aspect of this disclosure provides a computer program product including computer-executable instructions that, when executed, implement the method described above. Attached Figure Description

[0012] The above and other objects, features and advantages of this disclosure will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0013] Figure 1 A schematic diagram illustrating the relevant recommendation methods is provided.

[0014] Figure 2 The illustration schematically shows an exemplary system architecture for which the application resource recommendation method, feature extraction model training method, apparatus and device of this disclosure can be applied;

[0015] Figure 3 A flowchart illustrating an application resource recommendation method according to an embodiment of this disclosure is shown schematically.

[0016] Figure 4 A schematic diagram of a feature extraction model according to an embodiment of the present disclosure is shown.

[0017] Figure 5 A flowchart illustrating a training method for a feature extraction model according to an embodiment of the present disclosure is shown schematically.

[0018] Figure 6 A schematic diagram of a feature extraction model according to another embodiment of the present disclosure is shown;

[0019] Figure 7 A schematic diagram illustrating comparative learning according to an embodiment of the present disclosure is shown.

[0020] Figure 8 A block diagram of a resource recommendation apparatus according to an embodiment of the present disclosure is shown schematically;

[0021] Figure 9 A block diagram of a training apparatus for a feature extraction model according to an embodiment of the present disclosure is shown schematically.

[0022] Figure 10 A block diagram of an electronic device suitable for implementing the methods described above, according to embodiments of the present disclosure, is illustrated schematically. Detailed Implementation

[0023] The embodiments of the present disclosure will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the disclosure. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of the present disclosure for ease of explanation. However, it will be apparent that one or more embodiments may be practiced without these specific details. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concepts of the present disclosure.

[0024] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0025] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0026] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0027] In the embodiments disclosed herein, the collection, updating, analysis, processing, use, transmission, provision, disclosure, and storage of data (e.g., including but not limited to user personal information) comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. In particular, necessary measures have been taken to prevent unauthorized access to user personal information data and to safeguard user personal information security and network security.

[0028] In the embodiments disclosed herein, user authorization or consent is obtained before acquiring or collecting user personal information.

[0029] For recommendation systems, on the one hand, unique identifiers are typically used to represent different users and items. The association between these unique identifiers characterizes historical interaction behavior, and user preferences are predicted based on this association. For example, user A is assigned a user ID of 1001, item A (e.g., a mobile phone) is assigned an item ID of 2001, and item B (e.g., a phone case) is assigned an item ID of 2002. By statistically analyzing the association between user IDs and item IDs, the recommendation system discovers that user A (1001) purchased item 2001 (the mobile phone), and many users who purchased item 2001 (the mobile phone) also purchased item 2002 (the phone case). Therefore, item 2002 (the phone case) can be recommended to user A (1001).

[0030] However, the sparsity of unique identifier features limits the accuracy of recommendation models, especially in long-tail recommendation scenarios. Specifically, for niche, low-volume, but numerous long-tail products, there is insufficient user interaction data with these products. This makes it difficult to fully optimize the embedding of unique identifiers for these products through backpropagation, and the recommendation model struggles to accurately capture the relationship between these products and users, resulting in low product recommendation accuracy.

[0031] On the other hand, historical interaction behaviors include types such as co-click, co-order, and co-like. Recommendation systems may exhibit drastically different collaborative behaviors in historical interaction behaviors of different interaction types.

[0032] Figure 1 A schematic diagram illustrating the relevant recommendation methods is provided.

[0033] like Figure 1 As shown, "co-click" reflects users comparing competing products within the same product category. For example, if a user searches for "mobile phone" and finds two mobile phones, and the user "co-clicks" on both phones, it indicates that the user may be comparing the price or color of the phones, and these two phones are competing products.

[0034] Co-purchasing reflects the complementary relationship between products across different categories. For example, if users "co-purchase" mobile phones and headphones, it means that users usually buy mobile phones and headphones together, and mobile phones and headphones are complementary products.

[0035] Shared preferences reflect multiple products that a user might like simultaneously. For example, if a user "shares" headphones and sports shoes, or headphones and rackets, it indicates that the user is likely a sports enthusiast who likes rackets, headphones, and sports shoes.

[0036] The differences in collaborative behaviors corresponding to different interaction types indicate that collaborative semantics essentially depends on the context of the interaction type. Furthermore, methods that train recommendation models based solely on a single interaction type can only capture a partial and biased view of product relationships, failing to fully represent the semantic richness of multimodal data. For example, such methods can only partially capture competitive relationships between products, ignoring other important relationships such as complementary relationships. Moreover, recommendation models may infinitely amplify the association weight of a single interaction type, treating the association under that interaction type as the only association between products, leading to misjudgments of the true relationships between products.

[0037] In view of this, embodiments of the present disclosure provide a resource recommendation method, comprising: responding to a resource recommendation request, obtaining a historical behavior sequence and candidate resources of a target object contained in the resource recommendation request, wherein the historical behavior sequence includes multiple historical interaction resources; inputting the historical interaction resources and candidate resources into a feature extraction model for feature extraction to obtain a meta-embedding set for each of the historical interaction resources and candidate resources, wherein the meta-embedding set includes meta-embeddings under multiple preset semantic dimensions; determining a target fusion strategy corresponding to the collaborative relationship type from preset fusion strategies based on the collaborative relationship type between the historical interaction resources and candidate resources; using the target fusion strategy to perform fusion processing on the multiple meta-embeddings of the historical interaction resources and the multiple meta-embeddings of the candidate resources respectively to obtain historical resource features of the historical interaction resources and candidate resource features of the candidate resources; and performing resource recommendation processing based on the similarity between the historical resource features and the candidate resource features to generate a resource recommendation result for the target object.

[0038] Figure 2 An exemplary system architecture 200 is illustrated, in which the application resource recommendation method, feature extraction model training method, apparatus, and device of this disclosure can be applied. It should be noted that... Figure 2 The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, in order to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other devices, systems, environments or scenarios.

[0039] like Figure 2 As shown, the system architecture 200 according to this embodiment may include a first terminal device 201, a second terminal device 202, a third terminal device 203, a network 204, and a server 205. The network 204 serves as a medium for providing communication links between the first terminal device 201, the second terminal device 202, the third terminal device 203, and the server 205. The network 204 may include various connection types, such as wired and / or wireless communication links, etc.

[0040] Users can use the first terminal device 201, the second terminal device 202, and the third terminal device 203 to interact with the server 205 via the network 204 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 201, the second terminal device 202, and the third terminal device 203, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, and / or social media platform software, etc. (for example only).

[0041] The first terminal device 201, the second terminal device 202, and the third terminal device 203 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0042] Server 205 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 201, the second terminal device 202, and the third terminal device 203 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0043] For example, a user can initiate a resource recommendation request through any one of the first terminal device 201, the second terminal device 202, and the third terminal device 203. In response to the resource recommendation request, the server 205 can obtain the historical behavior sequence and candidate resources of the target object contained in the resource recommendation request. The historical behavior sequence includes multiple historical interaction resources. The historical interaction resources and candidate resources are input into a feature extraction model for feature extraction to obtain the meta-embedding sets of the historical interaction resources and candidate resources, where the meta-embedding sets include meta-embeddings under multiple preset semantic dimensions. Based on the type of collaborative relationship between the historical interaction resources and candidate resources, a target fusion strategy corresponding to the collaborative relationship type is determined from preset fusion strategies. Using the target fusion strategy, the multiple meta-embeddings of the historical interaction resources and the multiple meta-embeddings of the candidate resources are fused to obtain the historical resource features of the historical interaction resources and the candidate resource features of the candidate resources. Based on the similarity between the historical resource features and the candidate resource features, resource recommendation processing is performed to generate resource recommendation results for the target object.

[0044] It should be noted that the resource recommendation method and feature extraction model training method provided in this embodiment can generally be executed by server 205. Correspondingly, the system provided in this embodiment can generally be set up in server 205. The resource recommendation method and feature extraction model training method provided in this embodiment can also be executed by a server or server cluster that is different from server 205 and capable of communicating with the first terminal device 201, the second terminal device 202, the third terminal device 203, and / or server 205. Correspondingly, the system provided in this embodiment can also be set up in a server or server cluster that is different from server 205 and capable of communicating with the first terminal device 201, the second terminal device 202, the third terminal device 203, and / or server 205. Alternatively, the resource recommendation method and feature extraction model training method provided in this embodiment can also be executed by the first terminal device 201, the second terminal device 202, or the third terminal device 203, or by other terminal devices different from the first terminal device 201, the second terminal device 202, or the third terminal device 203. Accordingly, the system provided in this embodiment may also be installed in the first terminal device 201, the second terminal device 202 or the third terminal device 203, or in other terminal devices different from the first terminal device 201, the second terminal device 202 or the third terminal device 203.

[0045] It should be understood that Figure 2 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0046] Figure 3 A flowchart illustrating a resource recommendation method according to an embodiment of this disclosure is shown schematically.

[0047] like Figure 3 As shown, the method includes operations S301 to S305.

[0048] In operation S301, in response to the resource recommendation request, the historical behavior sequence of the target object and candidate resources contained in the resource recommendation request are obtained, wherein the historical behavior sequence includes multiple historical interaction resources.

[0049] It should be noted that in the technical solutions disclosed herein, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, they do not violate public order and good morals, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0050] For example, a resource recommendation request can be a demand instruction submitted to the recommendation system by a target object through system interaction, requesting the acquisition of resources. The target object can include users, merchants, etc., and the system interaction can include clicking on a recommendation entry point, initiating a search, browsing pages, etc., within the recommendation system (such as an e-commerce platform). Resources can include goods, services, information, etc. For example, a user can search for the product resource "mobile phone" on an e-commerce platform.

[0051] Historical behavior sequences can include multiple historical interaction resources corresponding to various historical interaction behaviors of a target object within a predetermined historical period. The predetermined historical period can be set according to actual needs, and historical interaction behaviors include, for example, clicks and purchases. For instance, if a user clicked on phone A, purchased phone B, and purchased headphones on an e-commerce platform within the past month, the historical behavior sequence would include: phone A, phone B, and headphones. Historical interaction resources can include resources corresponding to the historical interaction behaviors, such as phone A, phone B, and headphones.

[0052] Candidate resources can include: a set of all resources that may be suitable for the target user's needs, selected from a resource repository by the recommendation system in response to a resource recommendation request. The resource repository can store all resources of the recommendation platform. For example, after a user searches for the product "mobile phone" on an e-commerce platform, candidate resources may include a selection of multiple mobile phones, headphones, phone cases, and other similar products.

[0053] In operation S302, the historical interaction resources and candidate resources are input into the feature extraction model for feature extraction, resulting in a meta-embedding set for each of the historical interaction resources and candidate resources. The meta-embedding set includes meta-embeddings under multiple preset semantic dimensions.

[0054] For example, the feature extraction model may include a Multimodal Large Language Model (MLLM). Meta-embeddings may include feature vectors corresponding to key attributes of a resource under a predetermined semantic dimension. Predetermined semantic dimensions may include, for example, the resource's functional dimension or applicable scenario dimension. The set of meta-embeddings may specifically include multiple meta-embeddings under multiple predetermined semantic dimensions.

[0055] For example, candidate resources include mobile phone D, whose functions include photography and applicable scenarios include outdoor photography. Key attributes corresponding to the photography function dimension may include focusing speed, resolution, and background blur naturalness, while key attributes corresponding to outdoor photography scenarios may include light adaptability and image stabilization. The above information from mobile phone D can be input into a feature extraction model to obtain a set of meta-embeddings corresponding to mobile phone D. This set of meta-embeddings may include: feature vectors obtained by feature extraction of the key attributes corresponding to the photography function dimension, and feature vectors obtained by feature extraction of the key attributes corresponding to the outdoor photography scenario.

[0056] For example, historical interaction resources such as mobile phone A, mobile phone B, and headphones, and candidate resources of mobile phone D can be input into the feature extraction model to obtain the meta-embedding set corresponding to mobile phone A, the meta-embedding set corresponding to mobile phone B, the meta-embedding set corresponding to headphones, and the meta-embedding set corresponding to mobile phone D.

[0057] In operation S303, based on the type of collaborative relationship between historical interactive resources and candidate resources, a target fusion strategy corresponding to the type of collaborative relationship is determined from the preset fusion strategies.

[0058] For example, the types of collaborative relationships may include competitive relationships, complementary product relationships, etc., and the types of collaborative relationships may be determined based on the category of resources, the function of resources, etc.

[0059] For example, historical interaction resources include phone A, and candidate resources include phone D. Both phone A and phone D belong to the category of mobile phones, and users may choose to purchase one of these two phones. Therefore, the collaborative relationship type between phone A and phone D includes a competitive relationship.

[0060] For example, historical interaction resources include mobile phone A, and candidate resources include mobile phone cases. Since the function of a mobile phone case is to ensure the normal operation of the mobile phone, its function is complementary to that of the mobile phone. Therefore, the collaborative relationship type between mobile phone A and mobile phone cases includes a complementary product relationship.

[0061] For example, the type of collaborative relationship between each historical interactive resource and each candidate resource can be determined. For instance, if the historical interactive resources may include mobile phone A and mobile phone B, and the candidate resources include mobile phone C and a certain mobile phone case, the type of collaborative relationship between mobile phone A and mobile phone C, mobile phone A and the mobile phone case, mobile phone B and mobile phone C, and mobile phone B and the mobile phone case can be determined.

[0062] The type of collaborative relationship can affect the resource recommendation results. For example, if many users like both item X and item Y, then item X and item Y have a collaborative relationship. If a user likes item X, the system can recommend item Y, which is collaborative with item X, to that user.

[0063] For example, a preset fusion strategy can be used to fuse multiple meta-embedded elements in a set of meta-embedded elements to obtain a final feature vector. The preset fusion strategy may include the weights of each meta-embedded element in the set of meta-embedded elements.

[0064] A type of collaborative relationship can correspond to a target fusion strategy. For example, when the collaborative relationship type includes a competitor relationship, the target fusion strategy may include: increasing the weight of meta-embedding in the category dimension; when the collaborative relationship type includes a complementary product relationship, the target fusion strategy may include: increasing the weight of meta-embedding in the function dimension and the applicable scenario dimension.

[0065] In operation S304, using the target fusion strategy, multiple meta-embedded elements of historical interaction resources and multiple meta-embedded elements of candidate resources are fused to obtain the historical resource features of historical interaction resources and the candidate resource features of candidate resources.

[0066] For example, the fusion process can be a weighted summation of multiple meta-embedded elements. By fusing the multiple meta-embedded elements of each historical interaction resource, historical resource features of each historical interaction resource are obtained. Similarly, the multiple meta-embedded elements of each candidate resource can be fused to obtain candidate resource features of each candidate resource. For instance, if historical interaction resources include phone A and phone B, and candidate resources include phone C and a certain phone case, then historical resource feature 1 for phone A, historical resource feature 2 for phone B, candidate resource feature 1 for phone C, and candidate resource feature 2 for the phone case can be obtained.

[0067] In operation S305, resource recommendation processing is performed based on the similarity between historical resource features and candidate resource features to generate resource recommendation results for the target object.

[0068] For example, similarity calculation methods include cosine similarity, Euclidean distance, etc., and other similarity calculation methods can also be used. There is no limitation on the method used to calculate similarity. The similarity between each historical resource feature and each candidate resource feature can be calculated, such as the similarity between historical resource feature 1 and candidate resource feature 1, the similarity between historical resource feature 1 and candidate resource feature 2, the similarity between historical resource feature 2 and candidate resource feature 1, and the similarity between historical resource feature 2 and candidate resource feature 2.

[0069] Resource recommendation processing may include: determining the target historical resource features and the target candidate resource features, wherein the similarity between the target historical resource features and the target candidate resource features is greater than a preset similarity threshold, and the resource recommendation result may include recommending the candidate resources corresponding to the target candidate resource features to the target object. The preset similarity threshold can be set according to actual needs and is not limited here.

[0070] For example, if the similarity between historical resource feature 2 and candidate resource feature 2 is greater than a preset similarity threshold, then the phone case corresponding to candidate resource feature 2 can be recommended to the target object.

[0071] By acquiring the meta-embedded sets of historical interaction resources and candidate resources, and including meta-embedded sets under multiple preset semantic dimensions, and determining the target fusion strategy corresponding to the collaborative relationship type between historical interaction resources and candidate resources, and fusing multiple meta-embedded sets, the dependence of recommendation results on user-product interaction data can be reduced. Even when interaction data is insufficient, accurate recommendations can still be made based on meta-embedded sets and collaborative relationship types. This avoids the cold start problem of recommendation systems being unable to provide effective personalized recommendations for newly added users or items due to a lack of sufficient historical interaction data. It also effectively solves the problem in related recommendation methods where popular items receive more and more exposure and recommendations, becoming more popular, while unpopular or unpopular items become less popular due to a lack of exposure opportunities, forming a cycle of "the strong get stronger and the weak get weaker" (i.e., the Matthew effect).

[0072] According to embodiments of this disclosure, the preset semantic dimension includes at least one of the following: resource function, resource category, resource compensation value, and resource applicable scenario.

[0073] For example, a preset semantic dimension can be used to determine the type of collaborative relationship.

[0074] Resource functions can include the core uses of the resource. For example, if the resource includes a mobile phone, its resource functions may include taking photos and entertainment. For instance, the resource functions of a mobile phone and the functions of a phone case are complementary, and users often purchase them together; therefore, the type of synergistic relationship between the two includes a complementary product relationship.

[0075] Resource categories can include the product classification to which the resource belongs. For example, mobile phone A and mobile phone C both belong to the category of terminal devices. Users usually choose one to buy. Therefore, the collaborative relationship between the two can include a competitive relationship.

[0076] Resource compensation value can include the cost that a user needs to pay to acquire or use the resource, such as the price of the resource. For example, if phone A costs XXXX yuan and phone C costs XXXXX yuan, users will usually compare the prices of the two phones and choose one to buy. Therefore, the collaborative relationship between the two can be classified as a competitive relationship.

[0077] The applicable scenarios for resources can include the usage context, target audience, or application scenarios of the resource. For example, if the primary usage scenario for a certain mobile phone is outdoor photography, and the usage scenario for a selfie stick is also outdoor photography, then the type of synergistic relationship between the two includes a complementary product relationship.

[0078] By pre-setting semantic dimensions, the feature extraction model can capture the multi-dimensional features of historical interaction resources and candidate resources more precisely, thereby improving the accuracy of resource recommendation results.

[0079] According to embodiments of this disclosure, the feature extraction model includes query vectors for multiple preset semantic dimensions.

[0080] For example, query vectors are used to accurately capture the key attributes of a resource under the corresponding preset semantic dimension. For instance, query vectors for the preset semantic dimension of resource function are used to capture the key attributes of a resource in the resource function dimension, such as the focusing speed and resolution of the resource; query vectors for the preset semantic dimension of resource category are used to capture the key attributes of a resource in the resource category dimension, such as the major category to which the resource belongs (e.g., terminal device category) and the sub-category (e.g., mobile phone category or tablet category).

[0081] According to embodiments of this disclosure, inputting historical interaction resources and candidate resources into a feature extraction model for feature extraction to obtain meta-embedded sets for each of the historical interaction resources and candidate resources includes: extracting image features and text features of the historical interaction resources and candidate resources respectively to obtain historical resource multimodal features of the historical interaction resources and candidate resource multimodal features of the candidate resources; using query vectors of multiple preset semantic dimensions, performing cross-attention processing on the historical resource multimodal features and candidate resource multimodal features respectively to obtain meta-embedded sets for each of the historical interaction resources and candidate resources.

[0082] For example, image features can reflect the visual attributes of a resource, such as its shape, appearance, structure, color, and identifier, while text features can reflect textual information such as the resource's name, parameters, functional description, and attribute description.

[0083] By extracting multimodal features such as image and text features, a comprehensive analysis of resources can be provided, improving recommendation accuracy. Furthermore, since text and image features are determined solely based on the objective attributes of the resource itself (such as shape and name) and are unrelated to the resource's popularity (e.g., unaffected by the number of user interactions with the resource), they can effectively alleviate the cold start problem in recommendation scenarios (i.e., the problem where recommendation systems cannot provide effective personalized recommendations for newly added users or items due to a lack of sufficient historical interaction data). Therefore, by extracting multimodal features such as image and text features from historically interacted resources and candidate resources, reliance on historical interaction data between users and resources can be reduced, allowing for accurate recommendations even for long-tail products.

[0084] For example, cross-attention can be used to: accurately filter out semantically relevant information in image features and text features that is related to the query vector, and filter out irrelevant information, thereby determining the accurate features under the preset semantic dimension and obtaining the meta-embedding under the preset semantic dimension.

[0085] For example, for phone A, its historical resource multimodal features include information points such as: black, matte texture, photography, and price of XXX yuan. For the query vector of the preset semantic dimension of resource function, the query vector will traverse each information point in the above historical resource multimodal features, calculating the similarity between the query vector and each information point. Higher attention weights can be assigned to information points with high similarity. For example, the information point "photography" reflects the function of phone A and has a high similarity to the query vector, so it is assigned a higher attention weight, while "black" has a low similarity to the query vector and is assigned a lower attention weight. The information points can be weighted and summed according to their attention weights to obtain the meta-embeddedness under the preset semantic dimension of resource function. The above cross-attention processing can be performed on the query vectors of each preset semantic dimension to obtain the meta-embeddedness set.

[0086] By extracting image features and text features respectively, we obtain historical resource multimodal features and candidate resource multimodal features. We then perform cross-attention processing on the historical resource multimodal features and candidate resource multimodal features respectively. Since the text, image and other modal features of resources are not related to the popularity of resources, this method can effectively alleviate the cold start problem in recommendation scenarios, which is unable to provide effective personalized recommendations for newly added users or items due to the lack of sufficient historical interaction data.

[0087] According to embodiments of this disclosure, the method further includes: performing self-attention processing on the query vectors of each of the multiple preset semantic dimensions to obtain the processed query vectors of each of the multiple preset semantic dimensions.

[0088] According to embodiments of this disclosure, cross-attention processing is performed on historical resource multimodal features and candidate resource multimodal features using query vectors for multiple preset semantic dimensions, including: using multiple processed query vectors to perform cross-attention processing on historical resource multimodal features and candidate resource multimodal features respectively.

[0089] For example, multiple preset semantic dimensions are not entirely independent. For instance, there is a correlation between resource functions and resource application scenarios, such as outdoor photography functions typically corresponding to outdoor scenarios. If cross-attention processing is performed directly using the original query vector, the meta-embedding may ignore this inherent correlation, such as extracting only the photography function feature while ignoring the outdoor scenario feature.

[0090] For example, the query vectors for resource function, resource category, resource compensation value, and resource applicable scenario can each include: Q1, Q2, Q3, and Q4. The relevance between each query vector and other query vectors can be calculated, and weights can be assigned based on the degree of relevance (represented by relevance scores). For instance, higher weights can be assigned to query vectors with high relevance. Multiple query vectors can be weighted and fused to obtain the processed query vector. For example: calculate the relevance scores between Q1 and Q2, Q3, and Q4 respectively, normalize the relevance scores, determine the weights for Q2, Q3, and Q4, and then perform a weighted sum of Q1, Q2, Q3, and Q4 based on these weights to obtain the processed query vector Q1'.

[0091] Through self-attention processing, the query vector corresponding to the resource function can not only retain the core semantics of the resource function, but also reflect strongly related information such as the applicable scenario of the resource, thereby enabling accurate extraction of resource features.

[0092] For example, multiple processed query vectors can be used for cross-attention processing. For instance, for historical resource multimodal features or candidate resource multimodal features, the processed query vector can traverse each information point in the multimodal features, calculate the similarity between the processed query vector and each information point one by one, determine the magnitude of the attention weight based on the similarity, and sum the information points according to the attention weight.

[0093] By performing self-attention processing on the query vectors of multiple preset semantic dimensions, and using the multiple processed query vectors to perform cross-attention processing on the multimodal features of historical resources and the multimodal features of candidate resources, the model can simultaneously capture the correlation within a single modality and the correlation between multiple modalities. This allows the model to consider both the semantics of a single modality and the correlation between different modalities when generating resource recommendation results.

[0094] According to embodiments of this disclosure, the target fusion strategy includes fusion weights for each of multiple preset semantic dimensions.

[0095] For example, multiple preset semantic dimensions can correspond one-to-one with multiple meta-embeds, and each preset semantic dimension has its own fusion weight, which is used to perform weighted processing on the multiple meta-embeds.

[0096] According to embodiments of this disclosure, the fusion processing of multiple meta-embedded elements of historical interactive resources and multiple meta-embedded elements of candidate resources using a target fusion strategy includes: weighting multiple meta-embedded elements of historical interactive resources and multiple meta-embedded elements of candidate resources according to the fusion weights of multiple preset semantic dimensions.

[0097] For example, for each historical interaction resource, multiple meta-embedded elements in its meta-embedded set can be weighted. Similarly, for each candidate resource, multiple meta-embedded elements in its meta-embedded set can be weighted. Weighting methods include, for example, linear weighted summation, polynomial weighting, etc. The weighting method can be set according to actual needs and is not limited here.

[0098] According to embodiments of this disclosure, the method further includes: mapping the location information of historical interaction resources in the historical behavior sequence to a location vector; mapping the collaborative relationship type between historical interaction resources and candidate resources to a collaborative relationship vector; and fusing the similarity vector, location vector, and collaborative relationship vector to obtain a fused feature vector.

[0099] For example, the position information of historical interaction resources in the historical behavior sequence can reflect the time when the historical interaction occurred. For instance, if a user interacts with a certain historical interaction resource earlier, its position in the historical behavior sequence will be earlier. The time when the historical interaction occurred can affect the accuracy of the recommendation results. For example, for historical interaction that occurred earlier, the user's interests may have changed, so its reference value is lower. On the other hand, historical interaction that occurred recently can better reflect the user's current needs, so its reference value is higher.

[0100] The location information of historical interaction resources in the historical behavior sequence can be mapped to a location vector. The method of mapping to a location vector can be set according to actual needs. For example, it can be mapped by linear mapping or by embedding layer mapping, etc., which is not limited here.

[0101] For example, the type of collaborative relationship can also affect the recommendation results. For instance, when two products are complementary, both products can be recommended to the user. The method for mapping the collaborative relationship types between historical interaction resources and candidate resources to collaborative relationship vectors can be set according to actual needs. For example, each collaborative relationship can be directly assigned a fixed, differentiated vector, or each collaborative relationship type can be assigned a unique discrete ID, and then a trainable collaborative relationship embedding layer can be used to map the discrete IDs to a dense vector of fixed length. There is no limitation here.

[0102] For example, the fusion process may include weighted fusion of the similarity vector, position vector, and collaborative relationship vector to obtain a fused feature vector. The similarity vector, position vector, and collaborative relationship vector can be set according to actual needs.

[0103] According to embodiments of this disclosure, resource recommendation processing includes: performing resource recommendation processing based on fused feature vectors.

[0104] For example, the fusion feature vectors corresponding to each candidate resource can be weighted and summed to obtain the final score of a single candidate resource. Resource recommendations can then be made based on the final score. For instance, candidate resources with a final score greater than a preset score threshold can be recommended to the user. The preset score threshold can be set according to actual needs and is not limited here.

[0105] By mapping the location information of historical interaction resources in the historical behavior sequence to location vectors, the temporal information of historical interaction behavior can be captured; by mapping the collaborative relationship type between historical interaction resources and candidate resources to collaborative relationship vectors, the relationships between resources, such as competitive relationships and complementary product relationships, can be clarified; by fusing similarity vectors, location vectors and collaborative relationship vectors, the multi-faceted characteristics of resources can be comprehensively evaluated, resulting in more accurate resource recommendation results.

[0106] According to embodiments of this disclosure, the method further includes: dividing the historical behavior sequence into multiple sub-sequences according to a preset partitioning method, wherein the sub-sequences include at least one historical interaction resource arranged in sequence, and the historical interaction resource corresponds to a meta-embedding set; performing local attention processing on the multiple sub-sequences respectively to obtain local conversation features of each of the multiple sub-sequences; and performing global attention processing on the meta-embedding set of the candidate resource and the local conversation features of each of the multiple sub-sequences to obtain global features.

[0107] For example, the preset division method may include division according to time windows. For instance, the historical interaction resources corresponding to the three most recent historical interaction behaviors can be divided into one subsequence, and the historical interaction resources corresponding to the three historical interaction behaviors before that can be divided into another subsequence.

[0108] Preset categorization methods can also include categorization by resource theme. For example, outdoor mobile phones and outdoor tripods can be categorized as an outdoor theme subsequence, while mobile phones, headphones, and power banks can be categorized as a digital accessories theme subsequence.

[0109] For example, a subsequence may include at least one historical interaction resource arranged in chronological order, such as sorting the historical interaction resources in order of the time when the historical interaction behavior was generated from early to late.

[0110] According to embodiments of this disclosure, historical behavior sequences may be too long, causing long-term interests to interfere with short-term interests. For example, a user's purchase of a mobile phone six months ago might influence current outdoor gear recommendations. By dividing the historical behavior sequence into multiple sub-sequences, the user's primary interests at each stage can be accurately extracted.

[0111] For example, local attention processing can be performed on multiple subsequences separately. Local attention processing may include, for example, extracting the meta-embedded set of each historical interaction resource within each subsequence; calculating the correlation between each meta-embedded set and other meta-embedded sets; assigning attention weights to the meta-embedded sets based on the correlation; and using the attention weights to perform a weighted summation of each meta-embedded set to obtain local session features.

[0112] According to embodiments of this disclosure, each subsequence can correspond to a user's interests at a certain stage (such as a time window, a resource topic, etc.). Through local attention processing, the recommendation model can focus on the core interests within each subsequence, ignoring interference from irrelevant resources and accurately identifying local interest features.

[0113] Figure 4 A schematic diagram of a feature extraction model according to an embodiment of the present disclosure is shown.

[0114] Candidate resources can include items to be recommended, and historical interaction resources can include items that the user has interacted with in the past, such as items that have been clicked, purchased, or favorited. Candidate resources and historical interaction resources can be input into the embedding layer (4M Embedding Layer) of the feature extraction model. For example, the embedding layer can handle four feature dimensions.

[0115] Feature extraction models can include Session Window Transformer (SWT). For example... Figure 4 As shown, the session window converter can include a local session converter (Transformer).

[0116] For example, the introduction of meta-embeddings under multiple predefined semantic dimensions can lead to excessively long historical behavior sequences, such as extending the sequence length to K times its original length (where K is the number of meta-embeddings). If the standard Transformer is directly applied to the extended historical behavior sequence of length K×L (where L is the original sequence length), the computational complexity will be as high as O((KL)). 2 When K or L is large, the model is difficult to withstand.

[0117] For a local session Transformer, the expanded historical behavior sequence of length K×L can be divided into multiple non-overlapping subsequences. Each subsequence can correspond to a session window (SessionTransformer) of fixed length W, thus resulting in P=[KL / S] subsequences. Each session subsequence can be processed independently by the local session Transformer to capture fine-grained in-session behavior patterns within each session window.

[0118] For example, global attention processing may include: calculating the correlation between the meta-embedding set of candidate resources and the local session features of each of the multiple sub-sequences, determining the weight of each local session feature based on the correlation, and performing weighted fusion of the local session features of each of the multiple sub-sequences based on the weight to obtain global features.

[0119] According to embodiments of this disclosure, by performing global attention processing on the meta-embedding set of candidate resources and the local session features of each of the multiple sub-sequences, global interest patterns can be mined, highlighting the interest points most relevant to the candidate resources and reflecting the user's complete interest preferences.

[0120] like Figure 4 As shown, the session window converter can also include a global converter.

[0121] For example, the output of the local session Transformer (i.e., local session features) can be fed into the global Transformer along with the set of meta-embeddings of candidate resources. The global Transformer can then integrate the local session features and model the interaction between the entire expanded sequence of historical behaviors and candidate resources, generating global features.

[0122] According to the embodiments of this disclosure, the user's historical behavior sequence is first split to extract local interest features; then, the local features and candidate resource information are integrated to explore global interest patterns, so that resource recommendations not only match the user's short-term behavioral preferences, but also take into account the overall picture of long-term interests, thereby further improving the accuracy of recommendations.

[0123] According to embodiments of this disclosure, resource recommendation processing based on the similarity between historical resource features and candidate resource features includes: performing resource recommendation processing based on the similarity between historical resource features and candidate resource features, as well as global features.

[0124] For example, a combined score of similarity and global features can be calculated, and candidate resources corresponding to candidate resource features with a combined score greater than a preset value can be recommended to the user.

[0125] For example, the weights of similarity and global features can be determined according to actual needs, and the weighted sum of similarity and global features can be obtained to obtain a comprehensive score.

[0126] According to embodiments of this disclosure, the similarity between historical resource features and candidate resource features reflects the direct matching degree between them. The higher the similarity, the higher the matching degree between historical resource features and candidate resource features. Global features reflect the user's multi-stage interests and the interest points most relevant to the candidate resources. By comprehensively processing resource recommendations based on both the similarity between historical resource features and candidate resource features and global features, the recommendation accuracy can be improved.

[0127] According to embodiments of this disclosure, the method further includes: performing information enhancement on any subsequence based on the position information of historical interaction resources in the historical behavior sequence and the meta-embedding identifier corresponding to the meta-embedding set of historical interaction resources, to obtain an information-enhanced subsequence.

[0128] like Figure 4 As shown, for a local session Transformer, position information and meta-embedded identifiers (Meta ID & Position Embedding) can be added to each meta-embedded set.

[0129] For example, the position information of historical interaction resources in the historical behavior sequence (such as the historical interaction resource being ranked 3rd in the historical behavior sequence) can reflect the time of the historical interaction behavior corresponding to the historical interaction resource. For example, the historical interaction resource corresponding to the user's recent historical interaction behavior is ranked relatively late in the historical behavior sequence.

[0130] For example, meta-embedding identifiers are used to uniquely identify each set of meta-embeddings, enabling the feature extraction model to distinguish the source information of different meta-embeddings. Through meta-embedding identifiers, recommendation models can differentiate the source information of different meta-embeddings.

[0131] For example, location information and meta-embedding identifiers can be directly appended to each meta-embedding set to form an enhanced meta-embedding set. For instance, each enhanced meta-embedding set may include: the original meta-embedding set + location information + meta-embedding identifier. The information enhancement sub-sequence may include at least one historical interaction resource arranged in sequence, with each historical interaction resource corresponding to an enhanced meta-embedding set.

[0132] According to embodiments of this disclosure, feature extraction is performed on multiple sub-sequences to obtain local conversation features for each sub-sequence, including: performing local attention processing on multiple information-enhanced sub-sequences to obtain local conversation features for each information-enhanced sub-sequence.

[0133] For example, performing local attention processing on multiple information enhancement subsequences may include: for each information enhancement subsequence, extracting the set of enhancement meta-embeddings for each historical interaction resource within the information enhancement subsequence; calculating the correlation between each set of enhancement meta-embeddings and other sets of enhancement meta-embeddings; assigning attention weights to the sets of enhancement meta-embeddings based on the correlation; and using the attention weights to perform a weighted summation of the sets of enhancement meta-embeddings to obtain local conversation features.

[0134] According to embodiments of this disclosure, by providing the location information and meta-embedding identifier information of each historical interaction resource in the sub-sequence, the recommendation model can simultaneously focus on the interaction time between the user and the historical interaction resource and the characteristics of the historical interaction resource itself, so that the finally extracted local session features are more in line with the user's real interests.

[0135] According to embodiments of this disclosure, the method further includes: discretizing a preset similarity interval to obtain a preset number of sub-intervals; and mapping the interval identifier information of the sub-intervals to which the similarity between historical resource features and candidate resource features belongs to a similarity vector.

[0136] like Figure 4 As shown, the feature extraction model can also include a multi-collaboration similarity module (MSM). For example, the multi-collaboration similarity module learns the cross features of candidate merchant resource user behavior sequences under multiple collaborative relationships. For example, it can explicitly calculate the similarity between candidate resources and historical interaction resources corresponding to each collaborative behavior type, and utilize the association between capturing the global features of candidate products and user behavior sequences.

[0137] The multi-cooperative similarity module may include a cosine similarity discretization (CSD) module, which can be used to calculate the cosine similarity between each historical resource feature and the candidate resource feature.

[0138] For example, the preset similarity interval may include a preset range of similarity values, such as [-1, 1].

[0139] Data noise (such as slight differences in images and text descriptions) and computational precision errors can cause minor variations in similarity scores. Changing the recommendation ranking of candidate resources based solely on these minor fluctuations may affect the accuracy of the recommendations. For example, prioritizing a candidate resource with a similarity score of 0.83 over one with a similarity score of 0.82, even though their actual fit is not substantially different, leads to unstable recommendations and negatively impacts user experience. Therefore, the preset similarity interval can be discretized to obtain a preset number of sub-intervals. For instance, the preset similarity interval can be discretized into 20 equally wide intervals. The preset number can be set according to actual needs and is not limited here.

[0140] For example, each sub-interval corresponds to an interval identifier, which may include the interval ID. The interval identifier is used to uniquely identify a sub-interval. For example, ID0 corresponds to the sub-interval [-1, -0.9]...ID=19 corresponds to the sub-interval [0.9, 1.0].

[0141] It can determine the sub-interval to which the similarity between each historical resource feature and candidate resource feature belongs, and determine the interval identifier information of the sub-interval. For example, if the similarity between historical resource feature 1 and candidate resource feature 1 is -0.82, then the interval identifier information of the sub-interval to which it belongs is ID2.

[0142] For example, since interval identifiers consist only of isolated letters and / or numbers, recommendation models cannot directly understand the semantic relationships between interval identifiers. For instance, a recommendation model cannot understand that the sub-intervals corresponding to ID1 and ID2 are relatively similar. The interval identifiers of the sub-intervals to which the similarity between historical resource features and candidate resource features belongs can be mapped to similarity vectors. For example, the interval identifiers can be converted into a dense vector of fixed length, so that the similarity vectors corresponding to semantically similar IDs are also very similar.

[0143] For example, the interval identifiers corresponding to the similarity between historical resource feature 1 and candidate resource feature 1, the similarity between historical resource feature 1 and candidate resource feature 2, the similarity between historical resource feature 2 and candidate resource feature 1, and the similarity between historical resource feature 2 and candidate resource feature 2 are ID1, ID3, ID8, and ID9, respectively. The above interval identifiers can be mapped to similarity vectors to obtain vector [1,3,8,9].

[0144] like Figure 4As shown, for CSD, the meta-embedded element most relevant to the current collaborative relationship (e.g., competitor / complementary) can be selected from multiple meta-embedded elements; this is called collaborative embedding. For example, when calculating the complementary relationship between a phone and a phone case, the selector can prioritize meta-embedded elements related to resource functions and applicable scenarios, filtering out irrelevant ones. For the meta-embedded elements filtered by the selector, the cosine similarity between historical resource features and candidate resource features is calculated. The preset similarity interval can be discretized to obtain sub-intervals.

[0145] According to embodiments of this disclosure, resource recommendation processing based on the similarity between historical resource features and candidate resource features includes: performing resource recommendation processing based on a similarity vector.

[0146] For example, for each candidate resource, the interval identifier information corresponding to its similarity with all historical resources can be summarized, and the average or maximum value can be calculated to obtain the comprehensive fit of each candidate resource feature. For example, for candidate resource feature 1, its comprehensive fit can be (1+8) / 2=4.5.

[0147] Resource recommendation can be performed based on the overall fit of each candidate resource feature. For example, candidate resources corresponding to candidate resource features with an overall fit greater than the fit threshold can be recommended to customers. The fit threshold can be set according to actual needs.

[0148] like Figure 4 As shown, the feature extraction module can also include multi-task layering.

[0149] Multi-task layering can be the final output module of a feature extraction model, which may include a multi-expert network. Its input may include the output features of MSM, SWT, and other modules. Multi-task layering can adaptively activate different expert layers (Expert1, Expert2, Expert3, Expert4…ExpertN) based on different collaborative relationship types (such as Click, Order, Like, Add to Cart, etc., M collaborative relationship types), thereby achieving joint modeling of multiple tasks and reducing the seesaw effect between tasks. Furthermore, multi-task layering can also handle “Other Features” (other auxiliary features) through feature concatenation (…). Figure 4The ⊕) in the model is further integrated with historical resource features and candidate resource features. Other auxiliary features may include user profile features, resource attribute features, etc.

[0150] Among related methods, the most direct approach to obtaining modal embeddings typically involves using publicly available sentence embedding models or employing pooled sequence embeddings from large language models as text representations. However, these pre-trained models suffer from two key limitations in industrial deployments: domain inconsistency between general training and specific application scenarios, and unrealistic embedding dimensions (such as 768-2048 dimensions) leading to excessive computational resource consumption.

[0151] In view of this, another aspect of this disclosure provides a method for training a feature extraction model, comprising: acquiring multiple sample resource pairs, wherein the sample resource pairs are labeled with a type of collaborative relationship representing the collaborative relationship between sample resources in the sample resource pair, and the multiple sample resource pairs contain at least two types of collaborative relationship; inputting the multiple sample resource pairs into a feature extraction model for feature extraction, generating a set of meta-embedded features for each of the multiple sample resources, wherein the set of meta-embedded features includes meta-embedded features of the sample resources under multiple preset semantic dimensions; for each sample resource, performing fusion processing on the meta-embedded features of the sample resource under multiple preset semantic dimensions according to an initial fusion strategy corresponding to the collaborative relationship type of the sample resource pair to which the sample resource belongs, to obtain sample resource features of the sample resource; and determining the similarity between sample resources according to the sample resource features, and adjusting the network parameters and the initial fusion strategy of the feature extraction model according to the loss value determined by the similarity, to obtain a trained feature extraction model.

[0152] Figure 5 A flowchart illustrating a training method for a feature extraction model according to an embodiment of the present disclosure is shown schematically.

[0153] like Figure 5 As shown, the method includes operations S501 to S504.

[0154] In operation S501, multiple sample resource pairs are acquired. Each sample resource pair is labeled with a collaborative relationship type that characterizes the collaborative relationship between the sample resources in the sample resource pair. The multiple sample resource pairs contain at least two collaborative relationship types.

[0155] For example, a sample resource pair may include at least two sample resources with a collaborative relationship. The type of collaborative relationship may include competitive relationship, complementary product relationship, etc. The collaborative relationship type of multiple sample resource pairs may include at least competitive relationship and complementary product relationship, but is not limited to these. It may also include shared preference type, for example, when a user likes or favorites at least two products at the same time, the collaborative relationship type of the above-mentioned at least two products may be shared preference type.

[0156] Multiple sample resource pairs can be obtained through public datasets, business databases, etc. For example, a public dataset matching the recommendation scenario can be obtained, sample resource pairs can be extracted from it, and the type of collaborative relationship can be labeled. Alternatively, sample resources associated with historical interaction behaviors can be extracted from the recommendation platform's own business database, and sample resource pairs can be determined based on the collaborative relationship between the sample resources, and the type of collaborative relationship can be labeled.

[0157] For example, sample resource pair 1 extracted from public datasets in e-commerce scenarios includes: mobile phones and mobile phone cases, where the collaborative relationship type is complementary product relationship; sample resource pair 2 extracted from the business database of recommendation platforms (such as e-commerce platforms) includes: mobile phone A and mobile phone B, where the collaborative relationship type is complementary product relationship.

[0158] In operation S502, multiple sample resources are used to extract features from the input feature extraction model to generate a set of meta-embedded elements for each sample resource. The set of meta-embedded elements includes the meta-embedded elements of the sample resources under multiple preset semantic dimensions.

[0159] Figure 6 A schematic diagram of a feature extraction model according to another embodiment of the present disclosure is shown.

[0160] like Figure 6 As shown, feature extraction models include, for example, the Large Language Model (LLM). A single sample resource pair among multiple sample resource pairs may include two sample resources: plush toys and baby clothing. Since most users who buy baby clothing also buy plush toys, plush toys and baby clothing are complementary goods.

[0161] Sample resource pairs can also include irrelevant sample resources. Irrelevant sample resources can include those that are not significantly related to other sample resources in core dimensions such as user needs, product attributes, and usage scenarios. Correspondingly, collaborative relationship types can also include irrelevant product relationships. For example, a sample resource pair could include wine, plush toys, and baby clothing that are not significantly related. Users typically do not compare plush toys or baby clothing with wine and choose one to buy, nor do they typically buy wine when buying plush toys or baby clothing.

[0162] The above sample resources can be used to extract features from the input large language model.

[0163] For example, the feature extraction model may include a lightweight fine-tuning module (LoRA Adaptor). The lightweight fine-tuning module can insert a small number of low-rank matrix parameters into the key layers of the feature extraction model without modifying the original massive parameters. During training, only these small number of parameters need to be trained. Compared with the original parameter scale of the model, the memory usage and training cost are greatly reduced, avoiding the high cost of full training.

[0164] Through feature extraction, for each sample resource pair, a set of meta-embedded elements corresponding to each sample resource in that sample resource pair can be obtained. The set of meta-embedded elements can include multiple meta-embedded elements, and a meta-embedded element can include the feature vector corresponding to the key attributes of the sample resource under a certain preset semantic dimension. The preset semantic dimension includes at least one of the following: resource function, resource category, resource compensation value, and resource applicable scenario.

[0165] For example, a sample resource pair may include mobile phone A. The meta-embedding set of mobile phone A may include: resource category meta-embedding (digital product) + resource function meta-embedding (image stabilization photography) + resource usage scenario meta-embedding (outdoor scenario).

[0166] In operation S503, for each sample resource, according to the initial fusion strategy corresponding to the collaborative relationship type of the sample resource pair to which the sample resource belongs, the meta-embedding of the sample resource under multiple preset semantic dimensions is fused to obtain the sample resource features of the sample resource.

[0167] The initial fusion strategy may include: increasing the weight of the meta-embedding of the resource category dimension when the collaborative relationship type includes competitive product relationships; and increasing the weight of the meta-embedding of the resource function dimension and the resource applicable scenario dimension when the collaborative relationship type includes complementary product relationships. For each sample resource, the fusion processing may include weighting multiple meta-embeddings to obtain the sample resource features of the sample resource.

[0168] For example, if mobile phone A and mobile phone case in a sample resource pair are complementary products, the weight of the resource function embedding (e.g., set the weight to 0.5) and the resource applicable scenario dimension (e.g., set the weight to 0.3) can be increased, while the weight of the resource category dimension embedding (e.g., set the weight to 0.2) can be decreased.

[0169] In operation S504, the similarity between sample resources is determined based on the characteristics of the sample resources, and the network parameters and initial fusion strategy of the feature extraction model are adjusted based on the loss value determined by the similarity to obtain the trained feature extraction model.

[0170] For example, the loss value can characterize the difference between the actual similarity predicted by the model and the preset theoretical similarity. The smaller the difference, the lower the loss value. Different types of collaborative relationships can correspond to different theoretical similarities. For example, since the sample resources of competitive relationships mostly correspond to the same functions, the same usage scenarios, and the same categories, a high theoretical similarity (e.g., 0.95~1) can be set for competitive relationships. Since complementary products are related in terms of applicable scenarios and functions, a moderately high theoretical similarity (e.g., 0.85~0.95) can be set for complementary product relationships. Since unrelated products have no obvious correlation in core dimensions such as user needs, product attributes, and usage scenarios, a low theoretical similarity (e.g., 0) can be set for unrelated product relationships. For example, the relationship between red wine and plush toys, and red wine and baby clothes are unrelated product relationships, and their corresponding similarities are low.

[0171] For example, the network parameters and initial fusion strategy of the feature extraction model can be adjusted according to the loss value to achieve a smaller loss value, thus obtaining a trained feature extraction model.

[0172] By training a feature extraction model with samples containing multiple types of collaborative relationships, the model learns fine-grained meta-embeddings for each product across multiple preset semantic dimensions. It then adaptively selects fusion strategies based on different collaborative relationship types, dynamically combining these meta-embeddings into comprehensive product features. This process enables the ultimately learned product features to comprehensively and unbiasedly represent the complex and diverse relationships between products, fully utilizing the rich semantic information of multimodal data, thereby effectively improving the accuracy of product relationship modeling and the performance of the recommendation system.

[0173] According to embodiments of this disclosure, the feature extraction model includes initial query vectors for multiple preset semantic dimensions.

[0174] For example, the initial query vector can be used to capture the key attributes of the sample resource under the corresponding preset semantic dimension. For instance, the initial query vector of the resource function dimension can be used to capture the key attributes of the sample resource under the resource function dimension.

[0175] For example, if mobile phone A's resource functions include photography, then the initial query vector of the resource function dimension can capture the key attributes of mobile phone A under the resource function dimension, such as the focus speed and resolution of the resource.

[0176] According to embodiments of this disclosure, feature extraction is performed on multiple sample resources in an input feature extraction model to generate a set of meta-embedded features for each sample resource. This includes: extracting image features and text features for each sample resource to obtain multimodal features of the sample resource; and using initial query vectors for multiple preset semantic dimensions to perform cross-attention processing on the multimodal features to obtain the meta-embedded features of the sample resource under multiple preset semantic dimensions.

[0177] like Figure 6 As shown, feature extraction can be performed on plush toys, baby clothing, and red wine. Large language models can include vision transformers, which can handle image-based visual tasks, such as... Figure 5 Feature extraction was performed on images of plush toys, baby clothes, and red wine to obtain their respective image features.

[0178] Large language models can also include a word embedding module, which can extract text information, such as... Figure 5 Textual features were obtained from the textual descriptions of plush toys, baby clothes, and red wine. Examples of these descriptions include: "Plump toy: Pink; Teddy bear plush toy," "Baby clothes: Beige; Size XX," and "Red wine: Origin XXX; Vintage XXXX," etc.

[0179] like Figure 6 As shown, a large language model can include a query transformer (QFormer) layer. The QFormer layer can include, for example, a self-attention layer, a cross-attention layer, and a feedforward network (FFN). The initial query vectors (Learned Queries) can be processed through the QFormer layer.

[0180] In the self-attention layer, the initial query vector can be processed using formula (1).

[0181]

[0182] in, E Q Let E be the initial query vector, K represent the number of query embeddings (such as the initial query vector), and H be the hidden layer dimension. Q H serves as a learnable cue parameter, controlling the method of feature extraction. Q This is the output of the self-attention layer, for example, the processed initial query vector. In the self-attention layer, E can be...Q It encodes itself to obtain HQ.

[0183] In the Cross Attention layer, the processed initial query vector can be subjected to cross attention processing using formula (2).

[0184]

[0185] Among them, H CA H is the output of the cross-attention layer. MLLM This is the output of the last layer of the feature extraction model. For example, it includes a feature set containing all information after processing image and text features. This feature set might include information such as the category, function, color, and user reviews of the sample resources. In the cross-attention layer, it can be based on H... Q Extract H MLLM The collaborative information in different spaces yields H. CA In the feedforward network (FFN) layer, it can be achieved through formula (3).

[0186]

[0187] in, For meta-embedding, H CA After encoding by the feedforward network layer, the final representation H of the meta-embedding can be obtained. Meta .

[0188] For example, after processing by the QFormer layer, K meta-embeddings can be extracted. Figure 5 As shown, the large language model can also include a compression layer, which is used to reduce the dimensionality of the K meta-embeddeds output by the QFormer layer to obtain the final set of meta-embeddeds.

[0189] like Figure 6 As shown, the feature extraction model can also include a gating layer, specifically, the gating layer can include a selector as a functional component. The gating layer can be used to dynamically fuse the above K-element embeddings.

[0190] In the gating layer, K element embeddings can be fused using formula (4).

[0191]

[0192] (6)

[0193] For each type of collaboration relationship, a unique collaboration ID can be assigned, i.e. . After querying the embedding layer, the gate weights are obtained through the Softmax function. Each gate weight dynamically adjusts the K co-embedding elements. Different combinations yield collaborative spaces Item embedding representation .

[0194] In related methods, for product embedding generation, these methods typically perform aggregation operations (such as average / max / min pooling) on ​​the hidden states of the final layer, or use the hidden state of the last token as the overall representation, and fine-tune the model through parameter-efficient fine-tuning (PEFT, such as LoRA). However, this framework inherently limits the model to expressing only one specific feature extraction preference it has learned. Furthermore, different collaborative relationships require different extraction strategies, i.e., they focus on different aspects or tokens in the multimodal input. Therefore, a single fixed embedding cannot adaptively capture the semantic nuances of these dependencies, ultimately limiting the model's expressive power and making it difficult to capture the multifaceted interaction patterns inherent in recommendation systems. According to embodiments of this disclosure, by using QFormer layers for self-attention and cross-attention processing, diverse collaborative relationship types can be better captured, improving the accuracy of recommendation results.

[0195] By applying a QFormer layer to the MLLM sequence output, K meta-cooperative embeddings are extracted. Each meta-embedding is guided by a learnable query vector, capturing specific semantics for the potential collaborative scenario. By introducing a lightweight Session Window Transformer, it can be seamlessly integrated into industrial systems. The Session Window Transformer module dynamically aggregates K meta-embeddings based on recent user behavior sequences, enabling the final resource representation to adapt to the current collaborative scenario while reducing computational complexity and achieving a significant accuracy improvement with extremely low computational overhead.

[0196] According to embodiments of this disclosure, adjusting the network parameters and initial fusion strategy of the feature extraction model based on the loss value determined by similarity includes: adjusting the network parameters of the feature extraction model, the initial fusion strategy, and the initial query vectors of multiple preset semantic dimensions based on the loss value determined by similarity.

[0197] For example, network parameters may include learnable numerical variables in the feature extraction model, such as self-attention layer parameters, feedforward network parameters, word embedding parameters, etc. Adjusting the network parameters of the feature extraction model based on the loss value determined by similarity may include: adjusting the parameter values ​​of the above network parameters according to the loss value; the higher the loss value, the more significantly the feature extraction model will adjust the parameter values ​​of the above network parameters.

[0198] Adjusting the initial fusion strategy based on the loss value can include adjusting the weights of meta-embeddings. For example, for competitor relationships, the weight of the meta-embedding corresponding to the resource category can be increased.

[0199] Adjusting the initial query vector based on the loss value can include adjusting the key attributes of the sample resources to be captured under the corresponding preset semantic dimension. For example, in the initial state, the initial query vector for the resource category dimension can only query information in the broad dimension of "mobile phone", while the adjusted initial query vector can query more granular information such as "system type" and "chip type", which can reflect the essential differences between the sample resource and its competitors.

[0200] According to embodiments of this disclosure, by adjusting network parameters, initial fusion strategies, and initial query vectors for multiple preset semantic dimensions based on loss values, the features extracted by the feature extraction model can better fit different collaborative relationship types, thereby improving the accuracy of resource recommendation. For example... Figure 6 As shown, the feature extraction model may include a loss optimization layer (such as...) Figure 5 As shown in the Loss section, in the loss optimization layer, the loss value is determined by the following operations: based on the characteristics of the sample resources, the first similarity between sample resources in the sample resource pair and the second similarity between sample resources belonging to different sample resource pairs are determined; the first similarity and the second similarity are processed using the contrastive loss function to obtain the loss value.

[0201] For example, the loss optimization layer can be used to teach the feature extraction model to distinguish between target sample resources and irrelevant sample items through contrastive language-image pre-training. For instance, the loss optimization layer can distinguish between target sample resources and irrelevant sample items using a contrastive loss function.

[0202] For example, the loss optimization layer can compare the similarity and difference between different sample resources, bringing similar sample resources closer together and pushing dissimilar sample resources further apart, so that the feature extraction model can better understand the essential features of the data.

[0203] like Figure 6 As shown, a trigger can be one of the sample resources in a sample resource pair, a target can be another sample resource in the sample resource pair, and a negative sample can be a sample resource that has no synergistic relationship with the trigger.

[0204] Using contrastive loss functions to handle first and second similarities can include: using contrastive loss functions to reduce the feature distance between trigger items and target items, such as increasing the first similarity; and increasing the feature distance between trigger items and negative sample items, such as decreasing the second similarity.

[0205] For example, the trigger could be a plush toy, the target could be baby clothing, and the negative sample could be red wine. You can pull the distance between the plush toy and the baby clothing, or push the distance between the plush toy and the red wine.

[0206] Figure 7 A schematic diagram illustrating comparative learning according to an embodiment of the present disclosure is shown.

[0207] like Figure 7 As shown, the input information can include text input and image input. Text input can include N sets of descriptive text, and image input can include N images corresponding to the descriptive text. The text encoder can process each piece of text (such as...) Figure 7 The input (e.g., a black and white puppy) is converted into a corresponding text feature vector, such as... An image encoder can convert each image into a corresponding image feature vector, such as... Pairwise similarity calculations can be performed on all text and image features to obtain an N×N similarity matrix. This matrix can include multiple similarity levels, such as T1, I1, etc. The diagonal positions in the similarity matrix correspond to the similarity of positive sample pairs; for example, for the first similarity level, a contrastive loss function can be used to maximize this similarity. The off-diagonal positions correspond to the similarity of negative sample pairs; a contrastive loss function can be used to minimize this similarity. Thus, feature vectors describing the same thing in text and images will be spatially close, while feature vectors describing different things will be spatially far apart, enabling the feature extraction model to achieve cross-modal semantic understanding.

[0208] For example, the contrastive learning in this disclosure does not introduce a learnable temperature parameter τ of the relevant method to scale the cosine similarity. This is because, in implementing this disclosure, the inventors discovered that the model will continuously converge to a very small value of τ during training. Although this scaling operation can amplify the variance of the output similarity (which may improve the discriminative ability), it also suppresses the difference in the embedding itself. This side effect ultimately weakens the performance of the model in downstream tasks because the compressed embedding space limits the expressive power of subsequent tasks.

[0209] Figure 8A block diagram of a resource recommendation apparatus according to an embodiment of the present disclosure is shown schematically.

[0210] like Figure 8 As shown, the resource recommendation device 800 includes a first acquisition module 810, a first determination module 820, a first fusion processing module 830, and a resource recommendation 880.

[0211] The first acquisition module 810 is used to respond to a resource recommendation request by acquiring the historical behavior sequence and candidate resources of the target object contained in the resource recommendation request, wherein the historical behavior sequence includes multiple historical interaction resources; in one embodiment, the first acquisition module 810 can be used to perform the operation S301 described above, which will not be repeated here.

[0212] The first feature extraction module 820 is used to input historical interaction resources and candidate resources into the feature extraction model for feature extraction, and obtain the meta-embedding sets of the historical interaction resources and candidate resources respectively. The meta-embedding sets include meta-embeddings under multiple preset semantic dimensions. In one embodiment, the first feature extraction module 820 can be used to perform the operation S302 described above, which will not be repeated here.

[0213] The first determining module 830 is used to determine the target fusion strategy corresponding to the collaborative relationship type from the preset fusion strategies based on the collaborative relationship type between historical interactive resources and candidate resources. In one embodiment, the first determining module 830 can be used to perform the operation S303 described above, which will not be repeated here.

[0214] The first fusion processing module 840 is used to perform fusion processing on multiple meta-embedded elements of historical interaction resources and multiple meta-embedded elements of candidate resources using a target fusion strategy, so as to obtain historical resource features of historical interaction resources and candidate resource features of candidate resources. In one embodiment, the first fusion processing module 840 can be used to perform the operation S304 described above, which will not be repeated here.

[0215] The resource recommendation module 850 is used to perform resource recommendation processing based on the similarity between historical resource features and candidate resource features, and generate resource recommendation results for the target object; in one embodiment, the resource recommendation module 850 can be used to perform the operation S305 described above, which will not be repeated here.

[0216] The feature extraction model includes query vectors for multiple preset semantic dimensions. The first feature extraction module includes a first extraction submodule and a first cross-attention processing module. The first extraction submodule is used to extract image features and text features of historical interaction resources and candidate resources, respectively, to obtain historical resource multimodal features of historical interaction resources and candidate resource multimodal features of candidate resources. The first cross-attention processing module is used to perform cross-attention processing on the historical resource multimodal features and candidate resource multimodal features using the query vectors for multiple preset semantic dimensions, respectively, to obtain meta-embedding sets for historical interaction resources and candidate resources.

[0217] The aforementioned first feature extraction module further includes a self-attention processing submodule. This submodule performs self-attention processing on the query vectors for each of the multiple preset semantic dimensions to obtain the processed query vectors for each of the preset semantic dimensions. The first cross-attention processing module further utilizes the multiple processed query vectors to perform cross-attention processing on the historical resource multimodal features and the candidate resource multimodal features, respectively. The target fusion strategy includes fusion weights for each of the multiple preset semantic dimensions.

[0218] The first fusion processing module includes a weighted submodule. The weighted submodule is used to perform weighted processing on multiple meta-embedded elements of historical interactive resources and multiple meta-embedded elements of candidate resources according to the fusion weights of multiple preset semantic dimensions.

[0219] The aforementioned device also includes: a discretization processing module and a first mapping module.

[0220] The discretization module discretizes a preset similarity interval to obtain a preset number of sub-intervals. The first mapping module maps the interval identifier information of the sub-intervals to which the similarity between historical resource features and candidate resource features belongs to a similarity vector. The resource recommendation module also performs resource recommendation processing based on the similarity vector.

[0221] The aforementioned apparatus further includes: a second mapping module, a third mapping module, and a third fusion processing module. The second mapping module maps the position information of historical interaction resources in the historical behavior sequence into position vectors. The third mapping module maps the collaborative relationship type between historical interaction resources and candidate resources into collaborative relationship vectors. The third fusion module fuses the similarity vector, position vector, and collaborative relationship vector to obtain a fused feature vector.

[0222] The resource recommendation module is also used to perform resource recommendation processing based on the fused feature vector.

[0223] The aforementioned device further includes: a segmentation module, a local attention processing module, and a global attention processing module. The segmentation module is used to divide the historical behavior sequence into multiple sub-sequences according to a preset segmentation method, wherein each sub-sequence includes at least one historical interaction resource arranged in sequence, and the historical interaction resource corresponds to a meta-embedding set; the local attention processing module is used to perform local attention processing on each of the multiple sub-sequences to obtain local conversation features for each sub-sequence; the global attention processing module performs global attention processing on the meta-embedding set of the candidate resources and the local conversation features of each of the multiple sub-sequences to obtain global features.

[0224] According to embodiments of this disclosure, the recommendation module is further configured to perform resource recommendation processing based on the similarity between historical resource features and candidate resource features, as well as global features.

[0225] The aforementioned apparatus further includes an information enhancement module. The information enhancement module is used to enhance the information of any sub-sequence based on the position information of historical interaction resources in the historical behavior sequence and the meta-embedding identifier corresponding to the meta-embedding set of the historical interaction resources, thereby obtaining an information-enhanced sub-sequence.

[0226] The local attention processing module is also used to perform local self-attention processing on multiple information-enhanced sub-sequences respectively, to obtain the local conversation features of each information-enhanced sub-sequence. The preset semantic dimensions include at least one of the following: resource function, resource category, resource compensation value, and resource applicable scenario.

[0227] For example, any plurality of the first acquisition module 810, the first determination module 820, the first fusion processing module 830, and the resource recommendation module 880 can be combined into one module / unit / subunit, or any one of these modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least part of the functionality of one or more of these modules / units / subunits can be combined with at least part of the functionality of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of this disclosure, at least one of the first acquisition module 810, the first determination module 820, the first fusion processing module 830, and the resource recommendation module 880 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the first acquisition module 810, the first determination module 820, the first fusion processing module 830, and the resource recommendation module 880 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0228] It should be noted that the resource recommendation system part in the embodiments of this disclosure corresponds to the data processing method part in the embodiments of this disclosure. For a detailed description of the resource recommendation system part, please refer to the resource recommendation method part, which will not be repeated here.

[0229] Figure 9 A block diagram of a training apparatus for a feature extraction model according to an embodiment of the present disclosure is shown schematically.

[0230] like Figure 9 As shown, the training device 900 for the feature extraction model includes a second acquisition module 910, a second feature extraction module 920, a second fusion processing module 930, and an adjustment module 990.

[0231] The second acquisition module 910 is used to acquire multiple sample resource pairs, wherein the sample resource pairs are labeled with a collaborative relationship type that characterizes the collaborative relationship between the sample resources in the sample resource pair, and the multiple sample resource pairs contain at least two collaborative relationship types; in one embodiment, the second acquisition module 910 can be used to perform the operation S501 described above, which will not be repeated here.

[0232] The second feature extraction module 920 is used to extract features from multiple sample resources in the input feature extraction model and generate a set of meta-embedded elements for each of the multiple sample resources. The set of meta-embedded elements includes the meta-embedded elements of the sample resources under multiple preset semantic dimensions. In one embodiment, the second feature extraction module 920 can be used to perform the operation S502 described above, which will not be repeated here.

[0233] The second fusion processing module 930 is used to perform fusion processing on the meta-embedding of the sample resource under multiple preset semantic dimensions for each sample resource according to the initial fusion strategy corresponding to the collaborative relationship type of the sample resource pair to which the sample resource belongs, so as to obtain the sample resource features of the sample resource. In one embodiment, the second fusion processing module 930 can be used to execute the operation S503 described above, which will not be repeated here.

[0234] The adjustment module 940 is used to determine the similarity between sample resources based on the characteristics of the sample resources, and adjust the network parameters and initial fusion strategy of the feature extraction model according to the loss value determined by the similarity, so as to obtain the trained feature extraction model. In one embodiment, the adjustment module 940 can be used to perform the operation S504 described above, which will not be repeated here.

[0235] The feature extraction model includes initial query vectors for multiple preset semantic dimensions. The second feature extraction module includes a second extraction submodule and a second cross-attention processing submodule. The second extraction submodule extracts image and text features for each sample resource, obtaining multimodal features. The second cross-attention processing submodule performs cross-attention processing on the multimodal features using the initial query vectors for each of the multiple preset semantic dimensions, obtaining meta-embeddings of the sample resource across the multiple preset semantic dimensions. The adjustment module further adjusts the network parameters of the feature extraction model, the initial fusion strategy, and the initial query vectors for each of the multiple preset semantic dimensions based on the loss value determined by similarity.

[0236] The training device further includes a second determination module and a loss value determination module. The second determination module is used to determine, based on the characteristics of the sample resources, the first similarity between sample resources in a sample resource pair, and the second similarity between sample resources belonging to different sample resource pairs. The loss value determination module is used to process the first and second similarities using a contrastive loss function to obtain a loss value.

[0237] For example, any plurality of the second acquisition module 910, the second feature extraction module 920, the second fusion processing module 930, and the adjustment module 990 can be combined into one module / unit / subunit, or any one of these modules / units / subunits can be split into multiple modules / units / subunits. Alternatively, at least part of the functionality of one or more of these modules / units / subunits can be combined with at least part of the functionality of other modules / units / subunits and implemented in one module / unit / subunit. According to embodiments of the present disclosure, at least one of the second acquisition module 910, the second feature extraction module 920, the second fusion processing module 930, and the adjustment module 940 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the second acquisition module 910, the second feature extraction module 920, the second fusion processing module 930, and the adjustment module 990 may be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0238] Figure 10 A block diagram of an electronic device suitable for implementing the methods described above, according to embodiments of the present disclosure, is illustrated schematically. Figure 10 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0239] like Figure 10 As shown, an electronic device 1000 according to an embodiment of the present disclosure includes a processor 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage portion 1008 into a random access memory (RAM) 1003. The processor 1001 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 1001 may also include onboard memory for caching purposes. The processor 1001 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0240] RAM 1003 stores various programs and data required for the operation of electronic device 1000. Processor 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. Processor 1001 performs various operations of the method flow according to embodiments of the present disclosure by executing programs in ROM 1002 and / or RAM 1003. It should be noted that programs may also be stored in one or more memories other than ROM 1002 and RAM 1003. Processor 1001 may also perform various operations of the method flow according to embodiments of the present disclosure by executing programs stored in one or more memories.

[0241] According to embodiments of this disclosure, the electronic device 1000 may further include an input / output (I / O) interface 1005, which is also connected to a bus 1004. The electronic device 1000 may also include one or more of the following components connected to the input / output (I / O) interface 1005: an input section 1006 including a keyboard, mouse, etc.; an output section 1007 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the input / output (I / O) interface 1005 as needed. A removable medium 1011, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1010 as needed so that computer programs read from it can be installed into the storage section 1008 as needed.

[0242] The method flow according to embodiments of this disclosure can be implemented as a computer software program. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowchart. The computer program can be downloaded and installed from a network via communication section 1009, and / or installed from removable medium 1011. When the computer program is executed by processor 1001, it performs the functions defined in the system of embodiments of this disclosure. The systems, devices, apparatuses, modules, units, etc., described above can be implemented using computer program modules.

[0243] This disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs that, when executed, implement the method according to the embodiments of this disclosure.

[0244] Computer-readable storage media can be non-volatile computer-readable storage media. Examples include, but are not limited to: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. Computer-readable storage media can include ROM 1002 and / or RAM 1003 and / or one or more memories other than ROM 1002 and RAM 1003 described above.

[0245] Embodiments of this disclosure also include a computer program product comprising a computer program containing program code for performing the methods provided in the embodiments of this disclosure. When the computer program product is run on an electronic device, the program code is used to enable the electronic device to implement the methods provided in the embodiments of this disclosure.

[0246] When the computer program is executed by the processor 1001, it performs the functions defined in the system / apparatus of this disclosure embodiments. According to embodiments of this disclosure, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0247] The computer program can rely on tangible storage media such as optical storage devices or magnetic storage devices. In another embodiment, the computer program can also be transmitted and distributed in the form of signals over a network medium, and can be downloaded and installed via communication section 1009, and / or installed from removable medium 1011. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0248] Program code for executing the computer programs provided in the embodiments of this disclosure can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0249] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions. Those skilled in the art will understand that the features described in the various embodiments of the present disclosure can be combined and / or combined in various ways, even if such combinations are not explicitly described in the present disclosure. In particular, the features described in the various embodiments of this disclosure can be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure. All such combinations and / or combinations fall within the scope of this disclosure. Embodiments of this disclosure have been described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of this disclosure. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Various substitutions and modifications can be made by those skilled in the art without departing from the scope of this disclosure, and all such substitutions and modifications should fall within the scope of this disclosure.

Claims

1. A resource recommendation method, comprising: In response to a resource recommendation request, the historical behavior sequence and candidate resources of the target object contained in the resource recommendation request are obtained, wherein the historical behavior sequence includes multiple historical interaction resources; The historical interaction resources and the candidate resources are input into the feature extraction model for feature extraction to obtain the meta-embedding sets of the historical interaction resources and the candidate resources respectively. The meta-embedding sets include meta-embeddings under multiple preset semantic dimensions. Based on the type of collaborative relationship between the historical interaction resources and the candidate resources, a target fusion strategy corresponding to the type of collaborative relationship is determined from the preset fusion strategies; Using the target fusion strategy, multiple meta-embeddings of the historical interaction resources and multiple meta-embeddings of the candidate resources are fused to obtain the historical resource features of the historical interaction resources and the candidate resource features of the candidate resources. Based on the similarity between the historical resource features and the candidate resource features, resource recommendation processing is performed to generate resource recommendation results for the target object.

2. The method according to claim 1, wherein, The feature extraction model includes query vectors for multiple preset semantic dimensions; The step of inputting the historical interaction resources and the candidate resources into the feature extraction model for feature extraction to obtain the meta-embedding sets of the historical interaction resources and the candidate resources respectively includes: The image features and text features of the historical interaction resources and the candidate resources are extracted respectively to obtain the historical resource multimodal features of the historical interaction resources and the candidate resource multimodal features of the candidate resources; Using the query vectors of the multiple preset semantic dimensions, cross-attention processing is performed on the multimodal features of historical resources and the multimodal features of candidate resources to obtain the meta-embedding sets of the historical interactive resources and the candidate resources respectively.

3. The method according to claim 2, further comprising: Self-attention processing is performed on the query vectors of each of the multiple preset semantic dimensions to obtain the processed query vectors of each of the multiple preset semantic dimensions. The step of using the query vectors of the multiple preset semantic dimensions to perform cross-attention processing on the historical resource multimodal features and candidate resource multimodal features includes: Using multiple processed query vectors, cross-attention processing is performed on the historical resource multimodal features and the candidate resource multimodal features, respectively.

4. The method according to claim 1, wherein, The target fusion strategy includes fusion weights for multiple preset semantic dimensions; The step of using the target fusion strategy to perform fusion processing on multiple meta-embedded elements of the historical interaction resources and multiple meta-embedded elements of the candidate resources includes: Based on the fusion weights of the multiple preset semantic dimensions, the multiple meta-embedded elements of the historical interaction resources and the multiple meta-embedded elements of the candidate resources are weighted.

5. The method according to any one of claims 1 to 4, further comprising: Discretize the preset similarity interval to obtain a preset number of sub-intervals; Map the interval identifier information of the sub-intervals to which the similarity between the historical resource features and the candidate resource features belongs to a similarity vector; The resource recommendation process based on the similarity between the historical resource features and the candidate resource features includes: Resource recommendation processing is performed based on the similarity vector.

6. The method according to claim 5, further comprising: Map the location information of the historical interaction resources in the historical behavior sequence into a location vector; Map the collaborative relationship type between the historical interaction resources and the candidate resources into a collaborative relationship vector; The similarity vector, the position vector, and the collaborative relationship vector are fused to obtain a fused feature vector; The resource recommendation process includes: Resource recommendation processing is performed based on the fused feature vector.

7. The method according to any one of claims 1 to 6, further comprising: The historical behavior sequence is divided into multiple subsequences according to a preset division method, wherein each subsequence includes at least one historical interaction resource arranged in sequence, and the historical interaction resource corresponds to a meta-embedding set. Local attention processing is applied to each of the multiple sub-sequences to obtain the local conversation features of each of the multiple sub-sequences; Global attention processing is performed on the meta-embedding set of the candidate resources and the local session features of each of the multiple sub-sequences to obtain global features; The resource recommendation process based on the similarity between the historical resource features and the candidate resource features includes: Resource recommendation is performed based on the similarity between the historical resource features and the candidate resource features, as well as the global features.

8. The method according to claim 7, further comprising: Based on the position information of the historical interaction resources in the historical behavior sequence in any of the sub-sequences, and the meta-embedding identifier corresponding to the meta-embedding set of the historical interaction resources, information enhancement is performed on any of the sub-sequences to obtain an information-enhanced sub-sequence; The step of extracting features from multiple sub-sequences to obtain the local session features of each sub-sequence includes: Local attention processing is applied to each of the multiple information-enhanced subsequences to obtain the local conversation features of each subsequence.

9. A method for training a feature extraction model, comprising: Multiple sample resource pairs are obtained, wherein each sample resource pair is labeled with a collaborative relationship type that characterizes the collaborative relationship between sample resources in the sample resource pair, and the multiple sample resource pairs contain at least two collaborative relationship types; The multiple sample resources are used to extract features from the input feature extraction model to generate multiple meta-embedding sets for each of the sample resources, wherein the meta-embedding sets include the meta-embeddings of the sample resources under multiple preset semantic dimensions. For each sample resource, based on the initial fusion strategy corresponding to the collaborative relationship type of the sample resource pair to which the sample resource belongs, the meta-embeddings of the sample resource under the multiple preset semantic dimensions are fused to obtain the sample resource features of the sample resource; and The similarity between sample resources is determined based on the characteristics of the sample resources, and the network parameters of the feature extraction model and the initial fusion strategy are adjusted based on the loss value determined by the similarity to obtain the trained feature extraction model.

10. The method according to claim 9, wherein, The feature extraction model includes initial query vectors for multiple preset semantic dimensions; The step of extracting features from the multiple sample resources into the input feature extraction model to generate a set of meta-embedded features for each of the multiple sample resources includes: For each sample resource, extract the image features and text features of the sample resource to obtain the multimodal features of the sample resource; Using the initial query vectors of each of the multiple preset semantic dimensions, cross-attention processing is performed on the multimodal features to obtain the meta-embedding of the sample resources under the multiple preset semantic dimensions.

11. The method according to claim 10, wherein, Adjusting the network parameters of the feature extraction model and the initial fusion strategy based on the loss value determined by the similarity includes: The network parameters of the feature extraction model, the initial fusion strategy, and the initial query vectors of the multiple preset semantic dimensions are adjusted based on the loss value determined by the similarity.

12. The method according to claim 9, wherein, The loss value is determined using the following operation: Based on the characteristics of the sample resources, determine the first similarity between sample resources in the sample resource pair, and the second similarity between sample resources belonging to different sample resource pairs; The first similarity and the second similarity are processed using a contrastive loss function to obtain the loss value.

13. A resource recommendation device, comprising: The first acquisition module is used to respond to a resource recommendation request by acquiring the historical behavior sequence and candidate resources of the target object contained in the resource recommendation request, wherein the historical behavior sequence includes multiple historical interaction resources; The first feature extraction module is used to input the historical interaction resources and the candidate resources into the feature extraction model for feature extraction, and obtain the meta-embedding sets of the historical interaction resources and the candidate resources respectively, wherein the meta-embedding sets include meta-embeddings under multiple preset semantic dimensions. The first determining module is used to determine a target fusion strategy corresponding to the collaborative relationship type from a preset fusion strategy based on the collaborative relationship type between the historical interaction resources and the candidate resources. The first fusion processing module is used to perform fusion processing on multiple meta-embeddings of the historical interaction resources and multiple meta-embeddings of the candidate resources using the target fusion strategy, so as to obtain the historical resource features of the historical interaction resources and the candidate resource features of the candidate resources. The resource recommendation module is used to perform resource recommendation processing based on the similarity between the historical resource features and the candidate resource features, and generate resource recommendation results for the target object.

14. A training device for a feature extraction model, comprising: The second acquisition module is used to acquire multiple sample resource pairs, wherein the sample resource pairs are labeled with a collaborative relationship type that characterizes the collaborative relationship between sample resources in the sample resource pair, and the multiple sample resource pairs contain at least two collaborative relationship types; The second feature extraction module is used to extract features from the input feature extraction model from the multiple sample resources and generate multiple meta-embedding sets for each of the sample resources, wherein the meta-embedding sets include meta-embeddings of the sample resources under multiple preset semantic dimensions. The second fusion processing module is used to perform fusion processing on the meta-embeddings of each sample resource under the multiple preset semantic dimensions, based on the initial fusion strategy corresponding to the collaborative relationship type of the sample resource pair to which the sample resource belongs, to obtain the sample resource features of the sample resource; and An adjustment module is used to determine the similarity between sample resources based on the characteristics of the sample resources, and adjust the network parameters of the feature extraction model and the initial fusion strategy based on the loss value determined by the similarity, so as to obtain a trained feature extraction model.

15. An electronic device comprising: One or more processors; Memory, used to store one or more programs. Wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 12.

16. A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method of any one of claims 1 to 12.

17. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 12.

Images