Data processing method and device, equipment and medium

Abstract

The embodiment of the invention discloses a data processing method and device, equipment and a medium, and is applied to the technical field of computers. The method comprises the following steps: acquiring a first interaction data sequence of a plurality of first objects, and obtaining first object vectors of the plurality of first objects based on the first interaction data sequence of the plurality of first objects; prototype vectors of N interest prototypes in the interest prototype space are constructed through the first object vectors of the multiple first objects; n is a positive integer greater than 1; when a second interaction data sequence of a second object is obtained, mapping the second interaction data sequence to the interest prototype space to obtain weight distribution used for representing the second object under the N interest prototypes; obtaining a second object vector used for representing a second object through the weight distribution and prototype vectors of the N interest prototypes; the second object vector is used for carrying out resource recommendation on the second object. By adopting the embodiment of the invention, the accuracy of the user vector determined for the new user can be improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a data processing method, apparatus, device and medium. Background Technology

[0002] Currently, recommendation systems have become a core technology for various platforms to improve user experience. Traditional recommendation systems mainly rely on users' historical behavior data for modeling. By analyzing the user's past behavior sequences such as clicks and purchases, a user vector is determined (the user vector is determined by the sequence of items the user has clicked) to learn user preferences and then predict the user's level of interest in new items.

[0003] However, the inventors found in practice that this method heavily relies on the amount of historical behavioral data of users. For new users, due to insufficient behavioral data, it is difficult to determine a user vector that can accurately represent the new user, which in turn affects the subsequent application effect. Summary of the Invention

[0004] This application provides a data processing method, apparatus, device, and medium that can improve the accuracy of user vectors determined for new users, thereby improving the application effect of user vectors.

[0005] On one hand, embodiments of this application provide a data processing method, the method comprising:

[0006] Obtain the first interaction data sequence of multiple first objects, and based on the first interaction data sequence of multiple first objects, obtain the first object vector of multiple first objects;

[0007] The prototype vectors of N interest prototypes are constructed in the interest prototype space using the first object vectors of multiple first objects; N is a positive integer greater than 1.

[0008] When the second interaction data sequence of the second object is obtained, the second interaction data sequence is mapped to the interest prototype space to obtain the weight distribution used to represent the second object under N interest prototypes; the interaction data volume of the first interaction data sequences of multiple first objects is greater than the interaction data volume of the second interaction data sequence of the second object;

[0009] By using the weight distribution and prototype vectors of N interest prototypes, a second object vector is obtained to represent the second object; the second object vector is used to recommend resources for the second object.

[0010] On one hand, embodiments of this application provide a data processing apparatus, the apparatus comprising:

[0011] The sequence acquisition module is used to acquire the first interaction data sequence of multiple first objects, and based on the first interaction data sequence of multiple first objects, to obtain the first object vector of multiple first objects;

[0012] The sequence acquisition module is also used to construct prototype vectors of N interest prototypes in the interest prototype space through the first object vectors of multiple first objects; N is a positive integer greater than 1.

[0013] The sequence mapping module is used to map the second interaction data sequence of the second object to the interest prototype space when the second interaction data sequence of the second object is obtained, so as to obtain the weight distribution used to represent the second object under N interest prototypes; the interaction data volume of the first interaction data sequences of multiple first objects is greater than the interaction data volume of the second interaction data sequence of the second object;

[0014] The sequence mapping module is also used to obtain a second object vector to represent the second object through the weight distribution and the prototype vectors of N interest prototypes; the second object vector is used to make resource recommendations for the second object.

[0015] On one hand, embodiments of this application provide an electronic device including a processor and a memory, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute some or all of the steps in the above method.

[0016] On one hand, embodiments of this application provide a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, are used to perform some or all of the steps in the above-described method.

[0017] Accordingly, according to one aspect of this application, a computer program product or computer program is provided, which includes computer instructions that, when executed by a processor, can implement some or all of the steps in the above-described method.

[0018] In this embodiment, multiple first objects (i.e., old users, such as users with a large amount of interaction data) can be pre-acquired with their first interaction data sequences. N interest prototypes representing the object interests of the first objects are then obtained by clustering the first object vectors of these first objects. At this point, a second interaction data sequence of a second object (i.e., a new user, such as a user with a small amount of interaction data, where the amount of interaction data in the first interaction data sequence is greater than the amount of interaction data in the second interaction data sequence) can be obtained to determine the weight distribution of the second object under the N interest prototypes in the interest prototype space. Then, based on the weight distribution and the prototype vectors of the N interest prototypes, the second object vector of the second object is determined. This allows for the prediction of the weight distribution of new users under the interest prototypes based on the interest prototypes representing user interests abstracted from old users. The new user is represented by the prototype vector combining the weight distribution and the interest prototypes. This effectively utilizes the general interest patterns contained in the behavioral data of old users, enabling new users to quickly obtain accurate second object vectors with minimal information. Subsequent downstream tasks can be performed using the second object vectors to improve application performance, such as resource recommendations for the first object, achieving high-quality personalized recommendations. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A schematic diagram of a network architecture provided for an embodiment of this application;

[0021] Figure 2 A schematic diagram illustrating a data processing scenario provided in an embodiment of this application;

[0022] Figure 3 A flowchart illustrating a data processing method provided in this application embodiment. Figure 1 ;

[0023] Figure 4 This is a schematic diagram illustrating a resource recommendation scenario provided in an embodiment of this application.

[0024] Figure 5 A flowchart illustrating a data processing method provided in this application embodiment. Figure 2 ;

[0025] Figures 6a-6b A schematic diagram of a model training scenario provided in an embodiment of this application;

[0026] Figure 7This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0027] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] The data processing method proposed in this application is implemented in an electronic device, which can be a server or a terminal. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet computer, laptop computer, desktop computer, etc., but is not limited to these.

[0029] In the description of the embodiments of this application, the resources involved can be of any type, such as goods, published content, activities, etc. No limitation is made here. Published content can refer to notes (including text and images), short videos, medium-length videos, etc., posted by users on content platforms. Alternatively, it can also be live streams, snapshots (a type of short content that is instantly shared and exists), points of interest (such as group chats, live streams, etc.), routes (such as cycling routes, travel guides, etc.). The specific content information included in the published content applicable to different scenarios can be adjusted accordingly, and the content information in the published content can be determined by the publisher (such as the target audience). No limitation is made here on the specific type of published content.

[0030] The interactive data involved is obtained through interactive operations on resources. Each piece of interactive data indicates the resource on which the interactive operation was performed (i.e., the resource the user is interested in), such as a resource identifier. The sequence of interactive data can be a sequence of resource identifiers from multiple resources. The interactive operation can be a click operation, a purchase operation, a comment operation, etc., on the resource. The corresponding interactive operation varies depending on the resource. For example, if the resource is a product, it could be a purchase operation; if the resource is published content, it could be a click operation (in this case, clicking the published content is equivalent to performing a click operation, or clicking the published content and viewing it for a specified time is equivalent to performing a click operation), and so on. No further limitations are imposed here.

[0031] The objects involved can be divided into a second object and a first object. The second object can be considered as a new user on the platform (i.e., a user with insufficient interaction data, such as a user who has only performed 3-5 interaction operations, meaning there are 3-5 interaction data points for this new user; the interaction data sequence of this new user consists of the interaction data from these 3-5 interaction operations). The first object can be considered as an old user on the platform (i.e., a user with sufficient interaction data, such as a user who has performed multiple interaction operations; the interaction data sequence of this old user can consist of the interaction data from these multiple interaction operations). This application aims to utilize the interaction data sequence of old users to participate in the determination of the user representation (i.e., object vector) of new users, thereby ensuring the accuracy of the user representation.

[0032] The interest prototypes involved refer to the prototypes obtained by clustering the first interaction data sequence of the first object. These interest prototypes can be used to indicate a type of interest pattern (e.g., clustering multiple first objects to abstract multiple interest patterns, where an interest pattern represents a type of resource that a user is interested in extracted from multiple first objects) or a class of first objects (e.g., dividing multiple first objects into multiple classes, where a class includes a group of first objects with similar interests). In other words, they can represent the typical interest direction of a user group. For example, if the resource is published content, an interest prototype could represent a preference for travel-related published content (or an object that likes this type of published content), or an interest prototype could represent a preference for food-related published content (or an object that likes this type of published content), and so on. Therefore, for the second object, since there is less interaction data, the prototype vector of the second object can be determined using the prototype vector of the interest prototypes determined by the first object.

[0033] One of the network architecture diagrams proposed by the data processing method is as follows: Figure 1 As shown, the network architecture may include server 100 (the number of servers is not limited) and a cluster of terminal devices (the number of terminal devices is not limited, such as terminal device 200a, terminal device 200b, ..., terminal device 200n), wherein communication connections may exist between the servers. Simultaneously, a server may have a communication connection with any terminal device, so that the server can interact with the terminal device through this communication connection. The aforementioned communication connection is not limited in method; it can be directly or indirectly connected via wired communication, or directly or indirectly connected via wireless communication, or other methods, which are not limited herein. Furthermore, it is understood that the electronic devices involved in the embodiments of this application may be... Figure 1 The terminal device shown can also be Figure 1 The server shown.

[0034] It should be understood that, such as Figure 1 Each terminal device in the terminal device cluster shown can have an application client installed for resource recommendation. This application client can be of any type, such as a social networking client, instant messaging client (e.g., a conferencing client), entertainment client (e.g., a live streaming client), multimedia client (e.g., a video client), information client (e.g., a news client), shopping client, or any other client capable of displaying text, images, audio, and video data. No specific type of application client is limited here.

[0035] For example, an application client refers to a client that can send and receive internet messages instantly and has information functions. Specifically, the target account's terminal device (such as...) Figure 1 When the terminal device 200a) shown in the figure starts the application client and enters the application homepage, it can request the published content to be displayed on the application homepage from the server. At this time, if the server determines that the target account belongs to the second object, it can obtain the second interaction data sequence of the second object to determine the second object vector of the second object, and then determine the published content recommended to the target account through the second object vector, and return it to the terminal device for display.

[0036] Optionally, the aforementioned terminal devices and servers can be logically separated. Therefore, when referring to terminal devices and servers below, they may be physically the same device or different devices.

[0037] For further information, please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic diagram of a data processing scenario provided in an embodiment of this application. In this scenario, multiple first interaction data sequences of first objects (such as interaction data sequence 22a of object 21a, interaction data sequence 22b of object 21b, ..., interaction data sequence 22k of object 21k) can be pre-acquired. Then, the first interaction data sequences of the multiple first objects can be clustered to obtain a prototype vector for each of the N interest prototypes.

[0038] Specifically, when recommending resources to the second object, the second interaction data sequence of the second object can be obtained. This means that when recommending resources to a target account, it can be determined whether the target account belongs to the second object or the first object.

[0039] For example, the target account can be identified based on its existing interaction data volume (the interaction data volume of multiple first-object first interaction data sequences is greater than the interaction data volume of the second-object second interaction data sequence). For instance, if the target account's interaction data volume is greater than a preset threshold (e.g., 10), it is considered a first-object, and the corresponding first interaction data sequence can be obtained to determine the first-object vector for resource recommendation. Conversely, if the target account's interaction data volume is less than or equal to a preset threshold, it is considered a second-object, and the object's second interaction data sequence can be obtained, combined with N interest prototypes to determine the second-object vector for resource recommendation.

[0040] The weight distribution can be obtained by the first business model based on the second interaction data sequence. Therefore, when the target object is the second object, the corresponding object vector can be determined by inputting the second interaction data sequence into the first business model, which then maps the second interaction data sequence to the interest prototype space to obtain the weight distribution used to represent the second object under N interest prototypes in the interest prototype space (where one interest prototype corresponds to one interest weight).

[0041] Among them, the second object vector used to represent the second object can be obtained through the weight distribution and the prototype vectors of N interest prototypes (such as by weighted summation).

[0042] It's understandable that the second object vector can be used to recommend resources to a second object. For example, given a resource vector for a resource (such as a product), resources can be selected for recommendation to the second object based on the similarity between the resource vector and the second object vector.

[0043] Optionally, in some embodiments, the electronic device can execute the data processing method to achieve resource recommendation for new users according to actual business needs. For example, it can use a small amount of behavioral data from new users and an interest prototype representing the interest direction of the user group abstracted from a large amount of behavioral data from old users to represent new users as objects, so as to obtain accurate object vectors, thereby improving the application effect of object vectors in downstream tasks and the user experience.

[0044] For example, product recommendations can be made. For a target account, the corresponding object vector (second object vector or first object vector) can be determined according to the vector determination method extracted in this application, and the product vector can be determined (e.g., the corresponding resource vector can be determined based on multimodal data such as the product's title, image, video, sales volume, and type). Then, the products recommended to the target account can be determined based on the similarity between the product vector and the object vector. Similarly, content recommendations can be made. For a target account, the corresponding object vector (second object vector or first object vector) can be determined according to the vector determination method extracted in this application, and the content vector of the published content can be determined (e.g., the corresponding resource vector can be determined based on multimodal data such as the product's title, body text, image, and video). Then, the published content recommended to the target account can be determined based on the similarity between the content vector and the object vector. The application of downstream tasks is not limited here.

[0045] Optionally, the data involved in this application, such as published content, may be stored in a database or in a blockchain, such as through a blockchain distributed system. This application does not impose any restrictions on this.

[0046] It should be noted that in specific embodiments of this application, when scenarios involving the acquisition of user information and related data, such as acquiring user-published content, user permission or consent is required. That is, when these embodiments are applied to specific products or technologies, the collection, use, and processing of relevant user data comply with the relevant laws, regulations, and standards of the relevant regions. For example, a prompt message can be issued through an interactive interface to indicate what data will be collected or acquired. Specifically, the types and content of this data can be presented to the user through a list or similar method. Further data collection and processing will only proceed after a confirmation or instruction to allow data collection is received on the interactive interface.

[0047] It is understood that the above scenarios are merely examples and do not constitute a limitation on the application scenarios of the technical solutions provided in the embodiments of this application. The technical solutions of this application can also be applied to other scenarios. For example, as those skilled in the art will know, with the evolution of system architecture and the emergence of new business scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0048] Based on the foregoing description, this application proposes a data processing method that can be executed by the aforementioned electronic device. Please refer to... Figure 3 , Figure 3 A flowchart illustrating a data processing method provided in this application embodiment. Figure 1 Specifically, it can be Figure 1 The server shown. (As shown) Figure 3 As shown, the data processing method of this application embodiment may include the following flow:

[0049] S101. Obtain the first interaction data sequence of multiple first objects, and based on the first interaction data sequence of multiple first objects, obtain the first object vector of multiple first objects.

[0050] Specifically, N interest prototypes representing the group interests of user groups can be constructed in advance based on the first interaction data sequence of the first object. At this time, the prototype parameters (such as prototype vectors) of the interest prototypes can be obtained, and the second object can be represented based on the prototype parameters of the interest prototypes.

[0051] Among them, N interest prototypes are constructed through the first interaction data sequences of multiple first objects, where N is a positive integer greater than 1, and the amount of interaction data of the first interaction data sequences of multiple first objects is greater than the amount of interaction data of the second interaction data sequences of second objects.

[0052] Here, the first object refers to an object whose interaction data volume exceeds a preset threshold (i.e., it can be considered a returning user, a user who has performed many interactive operations, and whose interests in resources can be predicted). Therefore, the first interaction data sequence can be a sequence consisting of a specified number of interaction data of the first object (e.g., [item1, item2, ..., itemN], where itemi can represent the interaction data obtained from the i-th interactive operation, i.e., the identifier of the resource corresponding to the i-th interactive operation). For example, it can be a sequence consisting of the content identifiers of the 20 most recently clicked published content, or it can be a sequence of the content of all clicked published content.

[0053] Specifically, N interest prototypes can be obtained by clustering the first interaction data sequences of multiple first objects. The first interaction data sequences of multiple first objects are input into a third business model (such as a feature-extraction neural network like a transformer encoder). The third business model processes the first interaction data sequences of multiple first objects to obtain first object vectors for clustering, thus obtaining the abstracted interest prototypes.

[0054] For example, in the third business model, the first interaction data sequence can be sequence encoded, that is, deep feature extraction can be performed on each first interaction data sequence. Specifically, the first interaction data sequence [item1, item2, ..., itemN] can be converted into an embedding vector, and then the dependencies within the embedding vector can be captured through a self-attention mechanism. Finally, a fixed-dimensional representation vector can be obtained through pooling operations to serve as the first object vector of the first object. The training of the third business model can be found in the relevant description of the following embodiments.

[0055] Optionally, when inputting the first interaction data sequence into the third business model, the object information of the first object (such as the age, gender, and geographical location of the first object) can also be input simultaneously. The third business model then performs deep feature extraction on the first interaction data sequence and the object information to obtain the corresponding first object vector. In this way, the first object vector contains not only the relevant behavioral features of the first object but also the object information features. No restrictions are placed on the feature data used to determine the first object vector.

[0056] S102. Construct prototype vectors of N interest prototypes in the interest prototype space using the first object vectors of multiple first objects.

[0057] Optionally, multiple first objects can be periodically acquired and clustered to update the prototype parameters (including prototype vectors) of N interest prototypes. This allows for an accurate representation of the second object based on the latest interest prototypes. For example, a certain number of first objects can be acquired at fixed times each day, and N interest prototypes can be obtained based on the current first object vector of each first object.

[0058] The clustering of multiple first objects can be done using any clustering method, such as k-means clustering, variational Gaussian mixture clustering, etc., without limitation. Here, we will use variational Gaussian mixture clustering as the clustering method to illustrate the construction of N interest prototypes.

[0059] In this model, each interest prototype corresponds to a Gaussian distribution, meaning that the prototype parameters of an interest prototype are parameters related to the Gaussian distribution. These parameters may include the mean μ (the center position of the Gaussian distribution, which can serve as a prototype vector for an interest prototype, representing a typical interest feature of a user group), the covariance matrix Σ (describing the shape and orientation of each Gaussian distribution, reflecting the diversity and distribution range of user interests), and the mixing coefficient π (representing the weight or prior probability of each Gaussian distribution in the mixture model, reflecting the relative importance of each cluster component in the overall data; the sum of the mixing coefficients of all Gaussian distributions is 1).

[0060] Therefore, N initial Gaussian distributions can be constructed, each initial Gaussian distribution including the relevant initialization parameters. The mean μ can be a vector with the same dimension as the first object vector.

[0061] Specifically, the posterior probability of each first object belonging to each initial Gaussian distribution can be determined based on the first object vector of each first object, and the N initial Gaussian distributions can be adjusted based on the posterior probability to obtain N target Gaussian distributions.

[0062] Specifically, N prototype vectors for interest prototypes can be obtained based on N target Gaussian distributions. For example, an initial Gaussian distribution includes an initial center vector μ with the same dimension as the first object vector of the first object, and a target Gaussian distribution includes a target center vector μ obtained by adjusting the initial center vector. A target center vector can serve as a prototype vector for an interest prototype. That is, a mean μ of a target Gaussian distribution can be used as a prototype vector for an interest prototype.

[0063] The adjustment of N initial Gaussian distributions based on posterior probabilities can be achieved by adjusting the distribution parameters of the k-th initial Gaussian distribution based on the posterior probability that each first object belongs to the k-th initial Gaussian distribution among the N initial Gaussian distributions, thus obtaining the adjusted k-th initial Gaussian distribution; i is a positive integer less than N; when i ranges from 1 to N, the adjusted N initial Gaussian distributions are obtained, and the evaluation coefficients are determined based on the adjusted N initial Gaussian distributions; when the adjusted N initial Gaussian distributions satisfy the convergence condition based on the evaluation coefficients, the adjusted N initial Gaussian distributions are determined as N target Gaussian distributions.

[0064] This can be achieved by determining that the adjusted N initial Gaussian distributions meet the convergence condition when the change in the evaluation coefficients is less than or equal to a preset change value. The change in the evaluation coefficients can be the difference between the evaluation coefficients obtained after the previous adjustment of the initial Gaussian distributions and the evaluation coefficients obtained after the current adjustment. Alternatively, it can be determined that the adjusted N initial Gaussian distributions meet the convergence condition when the number of adjustments to the initial Gaussian distributions reaches a preset number.

[0065] Taking the i-th first object as an example, the posterior probability of determining that the i-th first object belongs to the k-th initial Gaussian distribution can be:

[0066]

[0067] Where, x i Let N(X|μ) be the first object vector of the i-th first object; k ,Σ k ) represents the probability density function of the Gaussian distribution corresponding to the k-th initial Gaussian distribution; π k μ k Σ k This represents the correlation distribution parameters of the k-th initial Gaussian distribution.

[0068] Specifically, the distribution parameters of the k-th initial Gaussian distribution can be adjusted based on the posterior probability that each first object (e.g., the number of first objects is P, where P is a positive integer) belongs to the k-th initial Gaussian distribution, to obtain the adjusted k-th initial Gaussian distribution.

[0069] For example, adjusting μ in the distribution parameter k The way (u k ′ is the adjusted μ k It could be:

[0070]

[0071] Among them, adjusting the Σ in the distribution parameter k The way (Σ) k ′ is the adjusted Σ k It could be:

[0072]

[0073] Among them, adjusting the π in the distribution parameter k The way (π) k ′ is the adjusted π k It could be:

[0074]

[0075] The evaluation coefficients can be determined based on the adjusted distribution parameters, and these evaluation coefficients can be ELBO (Evidence Lower Bound). Specifically, the current evaluation coefficients can be determined based on the first object vectors of multiple first objects and the distribution parameters of N adjusted initial Gaussian distributions.

[0076] This involves obtaining the change in evaluation coefficients to determine whether the adjusted N initial Gaussian distributions have reached the convergence condition. For example, historical evaluation coefficients (i.e., the evaluation coefficients obtained after the previous round of adjusting the initial Gaussian distributions) can be obtained based on the first object vectors of multiple first objects and the distribution parameters of the N initial Gaussian distributions before adjustment. The difference between the historical evaluation coefficients and the current evaluation coefficients is then used as the change value to determine whether the convergence condition is met (e.g., ΔELBO=|ELBO). t -ELBO t-1 |≤Preset change value).

[0077] The evaluation coefficients can be determined in the following ways:

[0078] ELBO = E q(z|x) [logp(x|z)]-KL(q(z|x)||p(z))

[0079] The variational distribution is composed of a Gaussian distribution, where x represents the user vector and z is from 1 to N. E q(z|x) [logp(x|z)] represents the reconstruction loss, which measures the difference between the object vector and the center of the Gaussian distribution (i.e., μ). kThe degree of matching between the prototypes is denoted by KL(q(z|x)||p(z)). KL(q(z|x)||p(z)) represents the KL divergence, which measures the difference between the variational distribution and the prior distribution. E represents the expectation. It can be understood that this can maximize the likelihood probability (variational distribution) of the interaction data behavior of old users, while adding the diversity constraint between prototypes (KL divergence) to ensure that the learned interest prototypes can effectively reconstruct user behavior and have sufficient discriminative power.

[0080] Here, the reconstruction loss refers to calculating the distance between each object vector x and the distribution center, and taking the logarithmic probability, such as:

[0081]

[0082] Where, μ k Σ k Let represent the distribution parameters of the k-th Gaussian distribution, and D be the dimension of the object vector. KL divergence refers to the ratio of the variational distribution q(z|x) to the prior distribution p(z). KL Divergence.

[0083] The N interest prototypes can also be obtained through k-means clustering. For example, by clustering the first object vectors of multiple first objects using the k-means algorithm, the multiple first objects can be divided into N object categories. Each object category includes at least one first object. In this case, an object category can be identified as an interest prototype, and the mean vector of the first object vectors of the first objects in an object category can be used as the prototype vector of that interest prototype. The method for determining the interest prototypes is not limited here.

[0084] S103. When the second interaction data sequence of the second object is obtained, the second interaction data sequence is mapped to the interest prototype space to obtain the weight distribution used to represent the second object under N interest prototypes.

[0085] The second interaction data sequence of the second object consists of the interaction data of the second object with respect to the resource. One piece of interaction data is obtained by an interaction operation performed by the second object on the resource. For example, interaction data can be used to indicate clicked published content, purchased goods, etc. If the interaction data can be an identifier of a resource for which an interaction operation has been performed, then the second interaction data sequence can constitute the behavior sequence of the second object.

[0086] The second object refers to an object whose interaction data volume is less than or equal to a preset threshold (i.e., it can be considered a new user, a user who performs few interactive operations and whose interests in resources are difficult to predict). Therefore, the second interaction data sequence can be a sequence consisting of a specified number of interaction data of the second object. For example, it can be a sequence consisting of the content identifiers of 3-5 recently clicked published content, or it can be a sequence of all clicked published content.

[0087] For example, when a target account needs to be recommended resources, the amount of interaction data of the target account can be obtained to determine whether the target account belongs to the second object or the first object, and the object vector can be determined in different ways to make resource recommendations.

[0088] Specifically, the second interaction data sequence is input into the first business model, and the first business model maps the second interaction data sequence to the interest prototype space to obtain the weight distribution used to represent the second object under N interest prototypes in the interest prototype space.

[0089] For example, the second interaction data sequence can be input into the first business model to map the second interaction data sequence to the interest prototype space containing N interest prototypes, thereby obtaining the weight distribution of the second object under the N interest prototypes. The training method of the first business model can be found in the relevant description of the following embodiments.

[0090] Specifically, this can involve inputting the second interaction data sequence into the first business model, where the first business model processes the second interaction data sequence to obtain the latent vector of the second object, and then determining the weight distribution of the second object among the N interest prototypes based on the latent vector of the second object and the prototype vectors of the N interest prototypes.

[0091] This can be achieved by extracting features from the second interaction data sequence in the first business model to obtain the latent vector of the second object. Optionally, in addition to inputting the second interaction data sequence of the second object into the first business model, object information of the second object can also be input simultaneously. In this way, the latent vector can be determined not only by combining user behavior but also by combining user-related attribute and other feature information.

[0092] For example, the first business model may include a transformer encoder, or it may also include a variational autoencoder, etc. No restrictions are placed on the processing method for the first business model here.

[0093] Taking a variational autoencoder as an example, the specific steps can be as follows: inputting the second interaction data sequence into the first business model, mapping the second interaction data sequence to the interest prototype space by the first business model to obtain the latent Gaussian distribution of the first interaction data; resampling the latent Gaussian distribution to obtain the latent vector of the second object.

[0094] The variational autoencoder (VAE) is characterized by its ability to map the sparse behavior of a second object to a probability distribution in the interest prototype space, thus explicitly modeling interest uncertainty. For example, by outputting a latent Gaussian distribution (also known as probability distribution parameters) through a VAE, the VAE framework can be used to encode user interaction data sequences into a latent Gaussian distribution (e.g., generating the mean and variance).

[0095] Here, variance represents the most likely position of the second object in the interest prototype space, and variance also represents the degree of uncertainty. Users with higher uncertainty correspond to larger variances, indicating that their interests and preferences are not yet clear.

[0096] One approach is to resample the potential Gaussian distribution to obtain the potential vector of the second object.

[0097] The determination of the weight distribution using the latent vector of the second object and the prototype vectors of N interest prototypes can be achieved by obtaining the similarity between the latent vector and the prototype vector of each interest prototype, and then normalizing the similarity to obtain the normalized similarity of the latent vector to each interest prototype. This normalized similarity can be used as the weight distribution. In other words, the weight distribution includes the interest weights of the second object under each interest prototype, and the interest weight under an interest prototype is the normalized similarity of the second object under that interest prototype.

[0098] Alternatively, if the interest prototype is obtained from a Gaussian distribution, it can also be obtained by substituting the latent vector x of the second object into the probability density function N(x|μ) of the interest prototype (such as the k-th interest prototype). k ,Σ k In the above, the probability density value of the second object under the k-th interest prototype is obtained. This can be achieved by normalizing the probability density values ​​of the K interest prototypes after obtaining them, and then using the normalized probability density values ​​as the weight distribution.

[0099] Optionally, the weight distribution can be further determined by combining the mixing coefficient π, for example:

[0100]

[0101] Among them, w k This represents the interest weight of the second object under the k-th interest prototype.

[0102] Alternatively, weight distribution can be predicted using neural networks. For example, the first business model can include not only vector extraction networks (such as variational autoencoders) but also weight prediction networks (such as networks composed of multiple layers of neurons, such as feedforward neural networks or multilayer perceptrons).

[0103] For example, the weight prediction network could consist of an input layer, hidden layers, and an output layer. The input layer is used to input the second object vector and the prototype parameters of N interest prototypes; the hidden layers are used for feature fusion and transformation through fully connected layers, and there can be one or more layers; the output layer is used to output the weight distribution (at this time, the Softmax function is used to ensure that the output values ​​are between (0,1) and sum to 1).

[0104] Here, the method of determining the weight distribution is not limited. The weight distribution can represent the probability that the second object's interest in the resource belongs to a certain interest prototype. The larger the interest weight, the more likely the second object is to be interested in resources related to this type of interest prototype.

[0105] S104. Using the weight distribution and prototype vectors of N interest prototypes, obtain the second object vector used to characterize the second object.

[0106] The second object vector is used to recommend resources for the second object.

[0107] The weight distribution includes the interest weights for each interest prototype under the second object. This can be achieved by weighted summation of the prototype vectors of the N interest prototypes according to their interest weights, resulting in the second object vector; or by using the prototype vector of the interest prototype with the highest interest weight as the second object vector. No specific limitation is imposed here.

[0108] The resource recommendation process can involve: obtaining resource vectors of multiple resources from a resource library; determining the similarity between the second object vector and the resource vectors of the multiple resources; and determining recommended resources from the multiple resources based on the similarity to recommend to the second object.

[0109] Resource vectors can be determined based on multimodal data of resources (such as text, images, categories, etc. The multimodal data can differ for different resources). For example, multimodal data can be input into a second business model (such as a Transformer-based multimodal model, a vision-language pre-trained model, etc.), and the second business model can output the corresponding resource vectors. The specific training method for the second business model can be found in the relevant descriptions of the following embodiments.

[0110] For example, the L resources with the highest similarity could be recommended to the second object.

[0111] Optionally, for the first object, the resource recommendation method is similar. It can be to obtain the first object vector of the first object through the third business model, determine the similarity between the first object vector and the resource vectors of multiple resources, and determine the recommended resources to recommend to the first object from multiple resources based on the similarity.

[0112] Optionally, since the variance in the potential Gaussian distribution determined by the first business model is used to characterize the interest uncertainty of the second object when determining the second object vector, the variance can be combined to appropriately increase the diversity of recommendations when selecting recommended resources.

[0113] For example, for a second object with a large variance (e.g., variance greater than a preset value), the selection of recommended resources can not be strictly based on similarity. For instance, a subset of resources can be randomly selected from the resource library. Alternatively, a subset of recommended resources can be selected from those with lower similarity (e.g., in addition to selecting the L most similar resources, a subset can be randomly selected from the L-2L most similar resources). The method for determining recommended resources is not limited here.

[0114] For example, such as Figure 4 As shown, Figure 4 This is a schematic diagram of a resource recommendation scenario provided in an embodiment of this application. The scenario involves obtaining first interaction data sequences for multiple first objects (e.g., interaction data sequence 42a for object 41a, interaction data sequence 42b for object 41b, ..., interaction data sequence 42k for object 41k), inputting these first interaction data sequences into a third business model to obtain first object vectors for the first objects (e.g., object vector 43a for object 41a, object vector 43b for object 41b, ..., object vector 43k for object 41k). Each first interaction data sequence includes multiple interaction data, such as [item1, item2, ..., itemr1].

[0115] In this process, the first object vectors of multiple first objects are clustered to obtain the distribution parameters (including prototype vectors) of N interest prototypes.

[0116] Specifically, when the second interaction data sequence of the second object is obtained, the second interaction data sequence and the distribution parameters of N interest prototypes can be input into the first business model to obtain the weight distribution used to represent the second object under the N interest prototypes in the interest prototype space.

[0117] A second interactive data sequence includes multiple interactive data, such as [item1, item2, ..., itemr2]. r1 is greater than r2.

[0118] In the first business model, the second interaction data sequence can be mapped to the interest prototype space to obtain the potential Gaussian distribution (mean, variance) of the first interaction data, and then resampled to obtain the potential vector of the second object.

[0119] At this point, the weight distribution of the second object among the N interest prototypes can be obtained based on the latent vector of the second object and the distribution parameters of the N interest prototypes.

[0120] Alternatively, the second interactive data sequence can be input into the first business model, and the first business model outputs a latent Gaussian distribution. Then, the weight distribution can be determined based on the latent Gaussian distribution (i.e., this process can be performed in the first business model or not).

[0121] The second object vector of the second object can be obtained by the weight distribution of the second object among N interest prototypes and the prototype vectors of the N interest prototypes.

[0122] This involves retrieving resource data (multimodal data) from the resource repository, such as resource data 45a for resource 44a, resource data 45b for resource 44b, ..., resource data 45v for resource 44v. The resource data from multiple resources can be input into the second business model to obtain resource vectors for multiple resources (such as resource vector 46a for resource 44a, resource vector 46b for resource 44b, ..., resource vector 46v for resource 44v).

[0123] Specifically, recommended resources for recommending to the second object can be determined from multiple resources based on the similarity between the second object vector and the resource vectors of multiple resources.

[0124] In this embodiment, multiple first objects (i.e., old users, such as users with a large amount of interaction data) can be pre-acquired with their first interaction data sequences. N interest prototypes representing the object interests of the first objects are then obtained by clustering the first object vectors of these first objects. At this point, a second interaction data sequence of a second object (i.e., a new user, such as a user with a small amount of interaction data, where the amount of interaction data in the first interaction data sequence is greater than the amount of interaction data in the second interaction data sequence) can be obtained to determine the weight distribution of the second object under the N interest prototypes in the interest prototype space. Then, based on the weight distribution and the prototype vectors of the N interest prototypes, the second object vector of the second object is determined. This allows for the prediction of the weight distribution of new users under the interest prototypes based on the interest prototypes representing user interests abstracted from old users. The new user is represented by the prototype vector combining the weight distribution and the interest prototypes. This effectively utilizes the general interest patterns contained in the behavioral data of old users, enabling new users to quickly obtain accurate second object vectors with minimal information. Subsequent downstream tasks can be performed using the second object vectors to improve application performance, such as resource recommendations for the first object, achieving high-quality personalized recommendations.

[0125] Based on the foregoing description, this application proposes a data processing method that can be executed by the aforementioned electronic device. Please refer to... Figure 5 , Figure 5 A flowchart illustrating a data processing method provided in this application embodiment. Figure 2 Specifically, it can be Figure 1 The server shown. (As shown) Figure 5 As shown, the data processing method of this application embodiment may include the following flow:

[0126] S201. Obtain the first interaction data sequence of multiple sample first objects.

[0127] A sample first interaction data sequence consists of M interaction data points of a sample first object, where M is a positive integer greater than 1. Each interaction data point indicates a sample resource for the interaction operation performed by a sample first object. These multiple sample first objects can be used to train the first business model, the third business model, and the second business model.

[0128] S202. Train a third initial business model using the sample first interaction data sequence of multiple sample first objects to obtain a trained third business model, and use the third business model to determine the reference object vector of multiple sample first objects based on the sample first interaction data sequence of multiple sample first objects.

[0129] Specifically, a sequence of first interaction data for multiple sample first objects (optionally, it may also include sample object information of the sample first objects) can be input into a third initial business model. The third initial business model outputs predicted object vectors for multiple sample first objects. The third initial business model can be trained using these predicted object vectors, and the trained third business model also outputs reference object vectors for multiple sample first objects. That is, the trained third business model is used to output the first object vector of the first object.

[0130] The third business model can be an encoder in the target model (such as a transformer model) used to encode object vectors. The target model may also include an encoder used to output prediction results based on the object vectors. Since the purpose of training the third business model is to encode the corresponding object vectors from the sample first interaction data sequence, the encoder's prediction result can be the interaction data obtained by the next interaction operation of the sample first object (e.g., the next interaction data obtained after the sample first interaction data sequence is used as a label, and the prediction goal is to make the encoder predict and output the next interaction data based on the object vector obtained from the sample first interaction data sequence; or it can use a specified interaction data in the sample first interaction data sequence as a label, and the prediction goal is to make the encoder predict and output the next interaction data based on the object vector obtained from the sample first interaction data sequence other than the specified interaction data).

[0131] For example, the first sample interaction data sequence is constructed from M interaction data (identifiers of the resources performing the interaction operations) obtained from M interaction operations of the first sample object. The label can be the interaction data obtained from the (M+1)th interaction operation. The third initial business model encodes the first sample interaction data sequence to obtain a prediction object vector, and decodes the prediction object vector to obtain the prediction result, which is the prediction result for the (M+1)th interaction data.

[0132] For example, if the Mth interaction data in the first interaction data sequence of the sample is used as the label, the first M-1 interaction data can be used as the new first interaction data sequence of the sample. The third initial business model encodes the new first interaction data sequence of the sample to obtain the prediction object vector, and the prediction result is obtained by decoding the prediction object vector. The prediction result is the prediction result for the Mth interaction data.

[0133] At this point, the target model can be trained based on the deviation between the prediction results and the labels. The encoder in the trained target model is the third business model that has been trained.

[0134] Specifically, the complete first interaction data sequence of the sample (i.e., the data sequence containing M interaction data) can be input into the trained third business model to obtain the reference object vector of the first sample object.

[0135] S203. Cluster the reference object vectors of the first object of multiple samples to obtain N prototype vectors of interest prototypes.

[0136] Specifically, the reference object vectors of the multiple sample first objects can be clustered to obtain N interest prototypes for training the first and second business models. These N interest prototypes can be interest prototypes used in the application phase, or, in the application phase, the latest second interaction data sequences of multiple first objects can be obtained by clustering to obtain N interest prototypes. No limitation is made here. The clustering method can be found in the relevant description of the above embodiments.

[0137] S204. Divide the first interaction data sequence of each sample first object into a support interaction data sequence and a query interaction data sequence.

[0138] The sample first interaction data sequence can be divided into two interaction data sequences. A support interaction data sequence is obtained from the first D interaction data in the sample first interaction data sequence (e.g., the first 5 interaction data), and a query interaction data sequence is obtained from the interaction data in the sample first interaction data sequence excluding the first D interaction data. N is a positive integer less than M, and the support interaction data sequence refers to the sequence used to simulate the second interaction data sequence.

[0139] It is understandable that since the second interactive data sequence is the interactive data sequence of the second object (such as the sequence composed of 5 interactive data of the second object), the first D interactive data in the first interactive data sequence of the sample are divided into supporting interactive data sequences. In this way, when the first object of the sample is regarded as the second object, the supporting interactive data sequence can be equivalent to the corresponding second interactive data sequence (i.e., the sequence composed of a small number of interactive data).

[0140] Therefore, the initial business model can be trained using the supporting interactive data sequence and the query interactive data sequence.

[0141] S205. Train the first initial business model based on the supporting interaction data sequence and query interaction data sequence corresponding to the first object of each sample, and obtain the trained first business model.

[0142] The specific method for determining the first initial business model can be through meta-learning. This meta-learning method includes an inner loop and an outer loop. The supporting interaction data sequence is applied in the inner loop phase, while the query interaction data sequence is applied in the outer training phase.

[0143] Taking any first sample object as the target sample object as an example, specifically, the supporting interaction data sequence of the target sample object is input into a first initial business model. The first initial business model maps the supporting interaction data sequence to an interest prototype space, obtaining a sample weight distribution representing the target sample object under N interest prototypes in the interest prototype space. Through the sample weight distribution and the prototype vectors of the N interest prototypes, a sample object vector representing the target sample object is obtained. This sample object vector is equivalent to the second object vector of the second object.

[0144] Among them, the first initial business model can be trained based on the similarity between the sample object vector and the sample resource vectors of the D sample resources indicated by the D interactive data in the supporting interactive data sequence, so as to obtain the first transitional business model.

[0145] Since the D interaction data points are used to indicate the resources clicked by the first sample object, the training objective is to make the sample object vector as similar as possible to the sample resource vectors of the D sample resources. Therefore, the cross-entropy loss value used to train the first initial business model can be determined by the similarity between the sample object vector and the sample resource vectors of the D sample resources (for example, the sum or average of the similarity values ​​between each sample object vector and each sample resource vector can be used as the cross-entropy loss value). The first initial business model is then trained using this cross-entropy loss value until the model converges (or after a specified number of training iterations), resulting in the first transitional business model. In other words, the model parameters in the first initial business model are adjusted to obtain the adjusted model parameters determined through the inner loop, which is the first transitional business model.

[0146] The sample resource vector can be obtained by processing the resource data (multimodal data) of the sample resources using the second initial business model.

[0147] Within this process, after the inner loop completes, the model can be trained again using the outer loop. Specifically, a first transitional business model or a first initial business model is trained based on the supporting interaction data sequence and query interaction data sequence corresponding to the first object for each sample, resulting in a trained first business model. At this point, the first transitional business model can be trained based on the inner loop, or the first initial business model can be trained directly; no specific limitation is imposed here.

[0148] The training method for the outer loop is based on the same principle as that for the inner loop, namely, the first transitional business model or the first initial business model can be trained based on the similarity between the object vector obtained from the first transitional business model and the resource vector of the resource indicated by the query interaction data sequence.

[0149] For example, the supporting interaction data sequence of the target sample object can be input into the first transitional business model. The first transitional business model maps the supporting interaction data sequence to the interest prototype space to obtain the transition weight distribution used to represent the target sample object under N interest prototypes in the interest prototype space. Through the transition weight distribution and the prototype vectors of the N interest prototypes, the transition object vector used to represent the target sample object is obtained. The first transitional business model or the first initial business model is trained based on the similarity between the transition object vector and the sample resource vectors of the MD sample resources indicated by the MD interaction data in the query interaction data sequence until the model converges, thus obtaining the trained first business model.

[0150] The specific method for obtaining the corresponding object vector through the first business model (first initial business model, first transitional business model) can be found in the relevant description of the above embodiments.

[0151] For example, the sum or average of the similarities between all transition object vectors and the sample resource vectors of each of the MD sample resources can be used as the cross-entropy loss value, so as to train the first transition business model or the first initial business model through the cross-entropy loss function.

[0152] Optionally, the first initial business model and the second initial business model can be trained simultaneously. Therefore, the first transitional business model and the second initial business model can be trained based on the similarity between the transitional object vector and the sample resource vectors of MD sample resources to obtain the trained first business model and the second business model. The second business model is used to obtain the corresponding resource vector based on the resource data of the resource.

[0153] Optionally, the first initial business model can be trained multiple times by alternating between inner and outer loops (while the second initial business model can be trained simultaneously), or the final first and second business models can be obtained after reaching a specified number of iterations.

[0154] For example, such as Figures 6a-6b As shown, Figures 6a-6b This is a schematic diagram of a model training scenario provided in an embodiment of this application; wherein, in Figure 6a In this process, multiple sample first object first interaction data sequences can be obtained (such as the interaction data sequence 62a of object 61a, the interaction data sequence 62b of object 61b, ..., the interaction data sequence 62k of object 61k), and labels for sample first interaction data sequences can be constructed (such as the label 63a of the interaction data sequence 62a of object 61a, the label 63b of the interaction data sequence 62b of object 61b, ..., the label 63k of the interaction data sequence 62k of object 61k).

[0155] In this process, the sample first interaction data sequence of multiple sample first objects is input into the third initial business model to obtain the predicted object vectors of multiple sample first objects (such as object vector 64a of object 61a, object vector 64b of object 61b, ..., object vector 64k of object 61k).

[0156] Specifically, the third initial business model is trained by using the predicted object vectors of multiple sample first objects and the labels of the sample first interaction data sequences of multiple sample first objects, resulting in the trained third business model.

[0157] In this process, the first interaction data sequence of multiple sample first objects is input into the trained third business model to obtain reference object vectors of multiple sample first objects. These reference object vectors can then be clustered to obtain prototype vectors of N interest prototypes.

[0158] Specifically, the sample first interaction data sequence of multiple sample first objects can be input into the first initial business model to obtain the relevant object vector of multiple sample first objects, and the resource data of the sample resource indicated by the sample first interaction data sequence can be obtained. The resource data of the sample resource can be input into the second initial business model to obtain the sample resource vector of the sample resource.

[0159] Specifically, the first and second initial business models can be trained using relevant object vectors and sample resource vectors to obtain the trained first business model and the trained third business model.

[0160] For example, in Figure 6bIn this context, the first initial business model and the second initial business model can be, for example, taking the target sample object as an example, the first interaction data sequence ([item1,item2,...,itemN]) of the target sample object can be divided into the support interaction data sequence ([item1,item2,...,item5]) and the query interaction data sequence ([item6,item7,...,itemN]).

[0161] At this point, the supporting interactive data sequence (which may also include object information of the target sample object) can be input into the first initial business model to obtain the sample object vector; the sample resource data of the sample resources (resource g11, resource g12, ..., resource g15) corresponding to the supporting interactive data sequence ([item1, item2, ..., item5]) can be input into the second initial business model to obtain the sample resource vector (resource vector g21 of resource g11, resource vector g22 of resource g12, ..., resource vector g25 of resource g15).

[0162] Specifically, the first loss value (cross-entropy loss value) can be determined by the similarity between the sample object vector and the sample resource vectors (resource vector g21 of resource g11, resource vector g22 of resource g12, ..., resource vector g25 of resource g15), and the first initial business model can be trained by the first loss value to obtain the first transitional business model.

[0163] Specifically, the supporting interactive data sequence (which may also include object information of the target sample object) is input into the first transitional business model to obtain the transitional object vector; the sample resource data of the sample resources (resource g16, resource g17, ..., resource g1N) corresponding to the query interactive data sequence ([item6, item7, ..., itemN]) is input into the second initial business model to obtain the sample resource vector (resource vector g26 of resource g16, resource vector g27 of resource g17, ..., resource vector g2N of resource g1N).

[0164] Specifically, the second loss value (cross-entropy loss value) can be determined by the similarity between the transition object vector and the sample resource vectors (resource vector g26 of resource g16, resource vector g27 of resource g17, ..., resource vector g2N of resource g1N). The first initial business model and the second initial business model are then trained using the second loss value to obtain the trained first business model and the trained second business model.

[0165] S206. After obtaining the second interaction data sequence of the second object, the second interaction data sequence is input into the first business model. The first business model maps the second interaction data sequence to the interest prototype space to obtain the weight distribution used to represent the second object under N interest prototypes in the interest prototype space. The specific implementation of step S206 can be found in the relevant description of the above embodiments, and will not be repeated here.

[0166] S207. Using the weight distribution and prototype vectors of N interest prototypes, obtain the second object vector to represent the second object, and use the second object vector to make resource recommendations for the second object.

[0167] The second object vector is used for resource recommendation for the second object. The specific method for determining the second object vector can be found in the relevant description of the above embodiments.

[0168] The technical solution of this application lies in the ability to construct a recommendation model that can quickly adapt to new users. By designing a brand-new user interest modeling paradigm, new users can quickly obtain accurate interest representations (object vectors) with very little behavioral data, thus achieving high-quality personalized recommendations. Specifically, it can achieve: (1) rapid adaptability, that is, new users only need 3-5 behavioral records to obtain reliable interest representations; (2) knowledge transfer capability, effectively utilizing the general interest patterns contained in old user data; (3) uncertainty handling, explicitly modeling and handling the uncertainty of new user interests, that is, the potential Gaussian distribution of new users can be obtained through the third business model; (4) scalability, supporting efficient recommendations for large-scale users and resources.

[0169] The technical solution of this application can extract reusable interest patterns from massive behavioral data of existing users. Through clustering, it abstracts the diverse interest behaviors of existing users into multiple representative interest prototypes. These interest prototypes can cover the user group's interest space (i.e., the direction of interest for resources, such as travel, food, etc.) and have good generalization ability to adapt to new users. Furthermore, by determining the weight distribution, new users can quickly match suitable interest prototype combinations. This allows new users to quickly determine their weight distribution in the interest prototype space based on minimal behavioral data, forming a personalized interest representation. Moreover, by introducing probabilistic modeling methods when representing the interests of new users, the uncertainty of new user interests is explicitly represented (i.e., the third business model can output variance). This allows for reasonable recommendations in recommendation decisions based on the high uncertainty of new user interests. Furthermore, during training, a meta-learning framework can achieve end-to-end rapid learning capabilities, enabling the model to "learn how to quickly learn new users," achieving a rapid transition from few samples to accurate recommendations.

[0170] It is understood that the new user recommendation method based on interest prototype meta-learning proposed in this application has an overall architecture comprising four core modules: a global interest prototype library construction module, a variational user encoding module, a meta-learning rapid adaptation module, and a user resource matching module. These modules work collaboratively to achieve end-to-end learning from sparse new user behavior to accurate recommendation results.

[0171] The global interest prototype library construction module is used to encode the interaction data sequences of old users and abstract the complex interest behaviors of all old users into multiple sets of common interest prototypes to form a shared interest representation basis (such as the user interest space can be represented by a mixture of K Gaussian distributions, with each Gaussian distribution corresponding to an interest prototype).

[0172] Therefore, the global interest prototype library construction module proposes an interest prototype discovery method based on a variational Gaussian mixture model, which can automatically learn a set of representative and discriminative interest prototypes from massive amounts of old user data. Unlike traditional simple clustering methods, the prototype learning process of this invention considers the probability distribution characteristics of user interests and ensures the coverage of prototypes through diversity constraints. That is, user interests are modeled as probabilistic combinations of interest prototypes, rather than single-category classification. In this way, a set of interest prototypes can serve all new users, avoiding repeated learning, and can represent a complex and diverse user interest space. Each prototype can represent an interest semantic, which is easy to understand and optimize. In this way, the rich behavioral data and interest patterns accumulated by a large number of old users can be effectively reused by new users, reducing the waste of data resources and improving the efficiency and accuracy of interest representation for new users.

[0173] The variational user coding module is used to map the sparse behavior of new users to a weight distribution on the interest prototype space, explicitly modeling interest uncertainty. That is, for the behavior sequence of new users, instead of directly outputting a definite interest representation, it can output probability distribution parameters (i.e., using the variational autoencoder framework to encode user behavior into a latent distribution). In this way, latent vectors can be obtained through resampling, and the weight distribution of users on each interest prototype can be calculated based on the latent variables obtained by sampling.

[0174] Therefore, the variational user encoding module can model new users and explicitly quantify the uncertainty of their interests. By outputting probability distribution parameters rather than deterministic vectors through the variational encoder, it can accurately reflect the fuzzy state of new user interests. Specifically, a variational encoder can be designed to output the mean and variance, and latent vectors can be obtained through reparameterization sampling, while ensuring end-to-end differentiable training. Furthermore, uncertainty information can be incorporated into recommendation decisions to achieve adaptive recommendation strategies. This improves robustness (able to handle noise and inconsistencies in new user behavior), ensures reasonable decision-making (more conservative recommendations under high uncertainty, avoiding misguidance), and ensures efficient learning (uncertainty naturally decreases with continuous optimization of the variational autoencoder, and the learning process is adaptive).

[0175] The meta-learning fast adaptation module is used to train a model that can "learn how to quickly learn new users", enabling it to quickly infer user interests from a small number of samples. The training steps include: (1) training an encoder (first business model) that performs sequence encoding on the historical behavior sequences of old users to obtain interest prototypes through clustering of the historical behavior sequences of old users; (2) constructing a "simulating new user" task from the old user data, that is, dividing the historical behavior sequences of old users into a support set (such as the first 5 behaviors, simulating sparse data of new users) and a query set (such as the remaining behaviors, as the evaluation target of learning effect). In this case, 2.1 the variational autoencoder is trained through the inner loop, that is, the prototype weight distribution of users is quickly learned, with the goal of making the user interest vector (object vector) match the resource vector of the resources in its support set as much as possible; 2.2 the variational autoencoder (first business model) and the model for extracting item vectors (second business model) are trained through the outer loop, thereby realizing end-to-end meta-learning training. The inner loop can be used to adapt user representations and evaluate the recommendation effect on the query set. The outer loop aims to optimize the ability to adapt quickly, so that the user interest vector should be similar to the resource vector of the resources in the query set, thereby maximizing the accuracy of user resource matching.

[0176] The meta-learning rapid adaptation module addresses the cold start problem in recommender systems by employing a dual-loop optimization framework. Through meta-training on the "simulating new users" task, the model learns the general ability to quickly infer user interests from a small number of samples. An innovative task construction strategy is proposed: transforming old user data into a new user simulation task; dual-loop optimization: the inner loop learns the support set, and the outer loop learns the query set, achieving general processing capabilities; parameter space constraints: rapid adaptation is performed in the prototype weight space. This improves adaptation speed (accurate interest representation can be obtained from only 3-5 actions for new users), generalization ability (adaptation strategies learned in meta-training can be applied to various types of new users), and computational efficiency (rapid adaptation and good real-time performance).

[0177] The user resource matching module is used for resource recommendation, such as mapping the multimodal features of resources (text, images, categories, etc.) to the same space as the object vector (i.e., the same dimension). Candidate resources can be sorted by matching score using similarity calculation to generate a recommendation list.

[0178] The user resource matching module uses a unified representation space, where user interest representations and resource vectors are mapped to the same dimension. This design greatly simplifies matching calculations and supports efficient vector retrieval. Specifically, it implements multimodal resource feature fusion: unified encoding of features such as text, images, and categories; and a dynamic matching mechanism: supporting real-time updates of user interest representations and resource vector matching. This results in high matching accuracy: more accurate similarity calculations in the unified space; high retrieval efficiency: supporting rapid filtering of large-scale candidate sets; and good scalability: new resources can be directly encoded into the existing space.

[0179] Therefore, the technical solution of this application systematically solves the cold start problem for new users from multiple levels, including data representation, model architecture, training methods, and recommendation strategies. It integrates techniques such as Gaussian mixture modeling and meta-learning, possessing a solid mathematical foundation and theoretical guarantee, and ensuring the feasibility of the solution: all models are differentiable neural networks, supporting end-to-end training, facilitating rapid engineering deployment, meeting the needs of real-time personalized recommendations, and ensuring broad application scenarios: not limited to specific domains, it can be applied to various recommendation scenarios such as e-commerce, content recommendation, and social networks. Therefore, this application can significantly improve the service capability of recommendation systems for new users, solve the cold start problem, and improve recommendation performance.

[0180] In this embodiment, multiple sample first interaction data sequences of first objects can be obtained to train a third business model. The reference object vectors of multiple sample first objects can be determined through the third business model, thereby constructing N interest prototypes. At this time, the sample first interaction data sequence of each sample first object can be divided into a support interaction data sequence and a query interaction data sequence. The inner loop training for the first business model is achieved through the support interaction data sequence, and the outer loop training for the first and second business models is achieved through the query interaction data sequence. This results in multiple models that can determine the second object vector corresponding to the second interaction data sequence of a new user based on the first interaction data sequence of an old user. Furthermore, the weight distribution of a new user under the interest prototype can be predicted based on the interest prototype representing user interests abstracted from the old user. The prototype vector combining the weight distribution and the interest prototype represents the new user. This effectively utilizes the general interest patterns contained in the behavioral data of old users, allowing new users to quickly obtain accurate second object vectors with minimal information data. Subsequently, downstream tasks can be performed using the second object vectors to improve the application effect, such as resource recommendation for the first object to achieve high-quality personalized recommendations.

[0181] Please see Figure 7 , Figure 7 This is a schematic diagram of a data processing apparatus provided in an embodiment of this application. It should be noted that... Figure 7 The data processing apparatus shown is used to execute this application. Figure 3 , Figure 5 The methods in the illustrated embodiments are shown only in the parts relevant to the embodiments of this application for ease of explanation; specific technical details are not disclosed. Reference to this application is required. Figure 3 , Figure 5 The illustrated embodiment. The data processing device 900 may include: a sequence acquisition module 901 and a sequence mapping module 902. Wherein:

[0182] The sequence acquisition module 901 is used to acquire the first interaction data sequence of multiple first objects, and to obtain the first object vector of multiple first objects based on the first interaction data sequence of multiple first objects;

[0183] The sequence acquisition module 901 is also used to construct prototype vectors of N interest prototypes in the interest prototype space through the first object vectors of multiple first objects; N is a positive integer greater than 1.

[0184] The sequence mapping module 902 is used to map the second interaction data sequence of the second object to the interest prototype space when the second interaction data sequence of the second object is obtained, so as to obtain the weight distribution used to represent the second object under N interest prototypes; the interaction data volume of the first interaction data sequences of multiple first objects is greater than the interaction data volume of the second interaction data sequence of the second object;

[0185] The sequence mapping module 902 is also used to obtain a second object vector to represent the second object through the weight distribution and the prototype vectors of N interest prototypes; the second object vector is used to make resource recommendations for the second object.

[0186] Specifically, when the sequence mapping module 902 is used to construct prototype vectors of N interest prototypes in the interest prototype space using the first object vectors of multiple first objects, it is used for:

[0187] Construct N initial Gaussian distributions, and determine the posterior probability of each first object belonging to each initial Gaussian distribution based on the first object vector of each first object;

[0188] Based on the posterior probability, N initial Gaussian distributions are adjusted to obtain N target Gaussian distributions;

[0189] Based on the Gaussian distribution of N targets, we obtain prototype vectors for N interest prototypes;

[0190] Wherein, an initial Gaussian distribution includes an initial center vector with the same dimension as the first object vector of the first object, and a target Gaussian distribution includes a target center vector obtained by adjusting the initial center vector. A target center vector can be used as a prototype vector of an interest prototype.

[0191] Specifically, when the sequence mapping module 902 is used to adjust N initial Gaussian distributions based on posterior probabilities to obtain N target Gaussian distributions, it is used for:

[0192] Based on the posterior probability of each first object belonging to the kth initial Gaussian distribution among N initial Gaussian distributions, the distribution parameters of the kth initial Gaussian distribution are adjusted to obtain the adjusted kth initial Gaussian distribution; k is a positive integer less than N.

[0193] As k ranges from 1 to N, N adjusted initial Gaussian distributions are obtained, and the evaluation coefficients are determined based on these N adjusted initial Gaussian distributions.

[0194] When the N initial Gaussian distributions after adjustment satisfy the convergence condition based on the evaluation coefficients, the N initial Gaussian distributions after adjustment are determined as N target Gaussian distributions.

[0195] Specifically, the sequence mapping module 902, when mapping the second interaction data sequence to the interest prototype space to obtain the weight distribution for representing the second object under N interest prototypes, is used for:

[0196] The second interactive data sequence is input into the first business model, and the first business model maps the second interactive data sequence to the interest prototype space to obtain the potential Gaussian distribution corresponding to the second interactive data sequence.

[0197] Resampling the latent Gaussian distribution yields the latent vector of the second object;

[0198] The weight distribution of the second object among the N interest prototypes is determined based on the latent vector of the second object and the prototype vectors of the N interest prototypes.

[0199] The weight distribution includes the interest weights for each interest prototype under the second object;

[0200] The sequence mapping module 902, when used to obtain a second object vector representing the second object through a weight distribution and N interest prototype vectors, specifically performs the following:

[0201] The second object vector is obtained by weighting and summing the prototype vectors of the N interest prototypes according to their interest weights; or,

[0202] Use the prototype vector of the interest prototype corresponding to the maximum interest weight as the second object vector.

[0203] The sequence mapping module 902 is also used for:

[0204] Obtain the resource vectors of multiple resources;

[0205] Determine the similarity between the second object vector and the resource vectors of multiple resources;

[0206] Based on similarity, recommend resources are determined from multiple resources to be recommended to the second object.

[0207] The weight distribution is obtained through the first business model based on the second interactive data sequence;

[0208] The sequence mapping module 902 is also used for:

[0209] Obtain the first interaction data sequence of multiple sample first objects; a sample first interaction data sequence consists of M interaction data of a sample first object, where M is a positive integer greater than 1, and an interaction data is used to indicate a sample resource for the interaction operation performed by a sample first object;

[0210] The sample first interaction data sequence of each sample first object is divided into a support interaction data sequence and a query interaction data sequence. A support interaction data sequence is obtained from the first D interaction data in a sample first interaction data sequence, and a query interaction data sequence is obtained from the interaction data in a sample first interaction data sequence excluding the first D interaction data. N is a positive integer less than M. The support interaction data sequence refers to the sequence used to simulate the second interaction data sequence.

[0211] The first initial business model is trained based on the supporting interaction data sequence and query interaction data sequence corresponding to the first object of each sample, and the trained first business model is obtained.

[0212] In this context, the first object of any sample is the target sample object;

[0213] When the sequence mapping module 902 is used to train the first initial business model based on the supporting interaction data sequence and query interaction data sequence corresponding to each sample first object, and to obtain the trained first business model, it is specifically used for:

[0214] The supporting interaction data sequence of the target sample object is input into the first initial business model, and the first initial business model maps the supporting interaction data sequence to the interest prototype space to obtain the sample weight distribution used to represent the target sample object under N interest prototypes in the interest prototype space.

[0215] By using the sample weight distribution and the prototype vectors of N interest prototypes, a sample object vector is obtained to characterize the target sample object.

[0216] The first initial business model is trained based on the similarity between the sample object vector and the sample resource vectors of the D sample resources indicated by the D interaction data in the supporting interaction data sequence, and the first transition business model is obtained.

[0217] The first transitional business model or the first initial business model is trained based on the supporting interaction data sequence and query interaction data sequence corresponding to the first object of each sample, so as to obtain the trained first business model.

[0218] Specifically, when the sequence mapping module 902 is used to train the first transitional business model or the first initial business model based on the supporting interaction data sequence and the query interaction data sequence corresponding to each sample first object, and to obtain the trained first business model, it is used for:

[0219] The supporting interaction data sequence of the target sample object is input into the first transition business model. The first transition business model maps the supporting interaction data sequence to the interest prototype space to obtain the transition weight distribution used to characterize the target sample object under N interest prototypes in the interest prototype space.

[0220] The transition object vector, used to characterize the target sample object, is obtained through the transition weight distribution and the prototype vectors of N interest prototypes.

[0221] The first transition business model or the first initial business model is trained based on the similarity between the sample resource vectors of the MD sample resources indicated by the MD interaction data in the transition object vector and the query interaction data sequence, so as to obtain the trained first business model.

[0222] Among them, the sample resource vector is obtained by processing the resource data of the sample resources by the second initial business model;

[0223] When the sequence mapping module 902 trains the first transition business model based on the similarity between the sample resource vectors of the MD sample resources indicated by the MD interaction data in the transition object vector and the query interaction data sequence, and obtains the trained first business model, it is specifically used for:

[0224] The first transitional business model and the second initial business model are trained based on the similarity between the transitional object vector and the sample resource vectors of MD sample resources, resulting in the trained first business model and the second business model; the second business model is used to obtain the corresponding sample resource vector based on the resource data of the sample resources.

[0225] The specific implementation methods of the sequence acquisition module and the sequence mapping module can be found in the description of the above embodiments, and will not be repeated here. It should be understood that the beneficial effects obtained by using the same method will also not be repeated here.

[0226] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0227] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8As shown, the electronic device 2600 includes at least one processor 2601 and a memory 2602. Optionally, the electronic device may also include a network interface. The processor 2601, memory 2602, and network interface can exchange data. The network interface, controlled by the processor 2601, is used to send and receive messages. The memory 2602 stores computer programs, including program instructions. The processor 2601 executes the program instructions stored in the memory 2602. The processor 2601 is configured to invoke the program instructions to execute the aforementioned method. The memory 2602 may include volatile memory, such as random-access memory (RAM); the memory 2602 may also include non-volatile memory, such as flash memory, solid-state drive (SSD), etc.; the memory 2602 may also include combinations of the above types of memory.

[0228] Processor 2601 may be a central processing unit (CPU). In one embodiment, processor 2601 may also be a graphics processing unit (GPU). Processor 2601 may also be a combination of a CPU and a GPU. Processor 2601 may be used to invoke device control applications stored in memory 2602 to perform the above-described tasks. Figure 3 , Figure 5 The description of the data processing method in the corresponding embodiments can also be executed as described above. Figure 7 The description of the data processing apparatus in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0229] In specific implementations, the devices, processors, memory, etc., described in the embodiments of this application can execute the implementation methods described in the above method embodiments, or they can execute the implementation methods described in the embodiments of this application, which will not be repeated here.

[0230] This application also provides a computer-readable storage medium storing a computer program. The computer program includes program instructions, which, when executed by a processor, enable the processor to perform some or all of the steps described in the above method embodiments. Optionally, the computer storage medium can be volatile or non-volatile. The computer-readable storage medium may primarily include a program storage area and a data storage area. The program storage area may store an operating system, at least one application program required for a given function, etc.; the data storage area may store data created based on the use of blockchain nodes, etc.

[0231] This application provides a computer program product, which may include a computer program. When the computer program is executed by a processor, it can implement some or all of the steps in the above method, which will not be elaborated here.

[0232] In this article, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0233] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a computer storage medium, which can be a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0234] The above-disclosed embodiments are merely some of the embodiments of this application, and should not be construed as limiting the scope of this application. Those skilled in the art can understand that all or part of the processes for implementing the above embodiments, and equivalent changes made in accordance with the claims of this application, still fall within the scope of this application.

Claims

1. A data processing method, characterized in that, The method includes: Obtain a first interaction data sequence of multiple first objects, and based on the first interaction data sequence of the multiple first objects, obtain a first object vector of the multiple first objects; The prototype vectors of N interest prototypes are constructed in the interest prototype space using the first object vectors of the multiple first objects; N is a positive integer greater than 1. When the second interaction data sequence of the second object is obtained, the second interaction data sequence is mapped to the interest prototype space to obtain the weight distribution used to characterize the second object under the N interest prototypes; the interaction data volume of the first interaction data sequences of the plurality of first objects is greater than the interaction data volume of the second interaction data sequence of the second object; A second object vector is obtained by using the weight distribution and the prototype vectors of the N interest prototypes to characterize the second object; the second object vector is used to recommend resources for the second object.

2. The method according to claim 1, characterized in that, The prototype vectors of the N interest prototypes constructed in the interest prototype space through the first object vectors of the plurality of first objects include: Construct the N initial Gaussian distributions, and determine the posterior probability of each first object belonging to each initial Gaussian distribution based on the first object vector of each first object; The N initial Gaussian distributions are adjusted based on the posterior probability to obtain the N target Gaussian distributions; Based on the Gaussian distribution of the N targets, the prototype vectors of the N interest prototypes are obtained; Wherein, an initial Gaussian distribution includes an initial center vector with the same dimension as the first object vector of the first object, and a target Gaussian distribution includes a target center vector obtained by adjusting the initial center vector. A target center vector can serve as a prototype vector of an interest prototype.

3. The method according to claim 2, characterized in that, The step of adjusting the N initial Gaussian distributions based on the posterior probability to obtain the N target Gaussian distributions includes: Based on the posterior probability that each first object belongs to the kth initial Gaussian distribution among the N initial Gaussian distributions, the distribution parameters of the kth initial Gaussian distribution are adjusted to obtain the adjusted kth initial Gaussian distribution; k is a positive integer less than N; As k ranges from 1 to N, N adjusted initial Gaussian distributions are obtained, and evaluation coefficients are determined based on these N adjusted initial Gaussian distributions. When the adjusted N initial Gaussian distributions satisfy the convergence condition based on the evaluation coefficients, the adjusted N initial Gaussian distributions are determined as the N target Gaussian distributions.

4. The method according to claim 1, characterized in that, The step of mapping the second interaction data sequence to the interest prototype space to obtain the weight distribution used to characterize the second object under the N interest prototypes includes: The second interaction data sequence is input into the first business model, and the first business model maps the second interaction data sequence to the interest prototype space to obtain the latent Gaussian distribution corresponding to the second interaction data sequence. The potential Gaussian distribution is resampled to obtain the potential vector of the second object; The weight distribution of the second object among the N interest prototypes is determined based on the latent vector of the second object and the prototype vectors of the N interest prototypes.

5. The method according to claim 1, characterized in that, The weight distribution includes the interest weights of the second object under each interest prototype; The process of obtaining a second object vector to characterize the second object through the weight distribution and the prototype vectors of the N interest prototypes includes: The prototype vectors of the N interest prototypes are weighted and summed according to the interest weights to obtain the second object vector; or, The prototype vector of the interest prototype corresponding to the maximum interest weight is used as the second object vector.

6. The method according to claim 1, characterized in that, The method further includes: Obtain the resource vectors of multiple resources; Determine the similarity between the second object vector and the resource vectors of the plurality of resources; Based on the similarity, recommended resources are determined from the plurality of resources for recommendation to the second object.

7. The method according to claim 1, characterized in that, The weight distribution is obtained through the first business model based on the second interactive data sequence; The method further includes: Obtain the first interaction data sequence of multiple sample first objects; a sample first interaction data sequence consists of M interaction data of a sample first object, where M is a positive integer greater than 1, and an interaction data is used to indicate a sample resource for the interaction operation performed by a sample first object; The sample first interaction data sequence of each sample first object is divided into a support interaction data sequence and a query interaction data sequence. A support interaction data sequence is obtained from the first D interaction data in a sample first interaction data sequence, and a query interaction data sequence is obtained from the interaction data in a sample first interaction data sequence excluding the first D interaction data. D is a positive integer less than M. The support interaction data sequence refers to the sequence used to simulate the second interaction data sequence. A first initial business model is trained based on the supporting interaction data sequence and query interaction data sequence corresponding to each sample first object, and the trained first business model is obtained.

8. The method according to claim 7, characterized in that, The first object in any sample is the target sample object; The step of training a first initial business model based on the supporting interaction data sequence and query interaction data sequence corresponding to each sample first object to obtain the trained first business model includes: The supporting interaction data sequence of the target sample object is input into the first initial business model, and the first initial business model maps the supporting interaction data sequence to the interest prototype space to obtain the sample weight distribution used to characterize the target sample object under N interest prototypes in the interest prototype space. The sample object vector representing the target sample object is obtained by using the sample weight distribution and the prototype vectors of the N interest prototypes. The first initial business model is trained based on the similarity between the sample object vector and the sample resource vectors of the D sample resources indicated by the D interaction data in the supporting interaction data sequence, to obtain the first transitional business model. The first transitional business model or the first initial business model is trained based on the supporting interaction data sequence and query interaction data sequence corresponding to the first object of each sample, so as to obtain the trained first business model.

9. The method according to claim 8, characterized in that, The step of training the first transitional business model or the first initial business model based on the supporting interaction data sequence and query interaction data sequence corresponding to each sample first object to obtain the trained first business model includes: The supporting interaction data sequence of the target sample object is input into the first transition service model, and the first transition service model maps the supporting interaction data sequence to the interest prototype space to obtain the transition weight distribution used to characterize the target sample object under N interest prototypes in the interest prototype space. The transition object vector, used to characterize the target sample object, is obtained through the transition weight distribution and the prototype vectors of the N interest prototypes. The first transition business model or the first initial business model is trained based on the similarity between the transition object vector and the sample resource vectors of the MD sample resources indicated by the MD interaction data in the query interaction data sequence, so as to obtain the trained first business model.

10. The method according to claim 9, characterized in that, The sample resource vector is obtained by processing the resource data of the sample resources using the second initial business model; The step of training the first transition service model based on the similarity between the transition object vector and the sample resource vectors of the MD sample resources indicated by the MD interaction data in the query interaction data sequence, to obtain the trained first service model, includes: The first transition business model and the second initial business model are trained based on the similarity between the transition object vector and the sample resource vectors of the MD sample resources, resulting in the trained first business model and second business model; the second business model is used to obtain the corresponding sample resource vector based on the resource data of the sample resources.

11. A data processing apparatus, characterized in that, The device includes: The sequence acquisition module is used to acquire a first interaction data sequence of multiple first objects, and to obtain a first object vector of the multiple first objects based on the first interaction data sequence of the multiple first objects; The sequence acquisition module is also used to construct prototype vectors of N interest prototypes in the interest prototype space through the first object vectors of the plurality of first objects; N is a positive integer greater than 1. The sequence mapping module is used to map the second interaction data sequence of the second object to the interest prototype space when the second interaction data sequence of the second object is obtained, so as to obtain the weight distribution used to represent the second object under the N interest prototypes; the interaction data volume of the first interaction data sequences of the plurality of first objects is greater than the interaction data volume of the second interaction data sequence of the second object; The sequence mapping module is further configured to obtain a second object vector representing the second object through the weight distribution and the prototype vectors of the N interest prototypes; the second object vector is used to make resource recommendations for the second object.

12. An electronic device, characterized in that, The system includes a processor and a memory, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to perform the method as described in any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1-10.

14. A computer program product, characterized in that, The method includes a computer program comprising program instructions that, when executed by a processor, implement the method according to any one of claims 1-10.

Images