Data set construction method, model training method, object recommendation method and recommendation system

Abstract

The embodiment of the invention provides a data set construction method, a model training method, an object recommendation method and a recommendation system. The data set construction method comprises the steps of obtaining a support set of a target user; extracting a target user representation of the target user from the historical interaction data; performing conditional mapping on the prior interaction distribution based on target user representation through a conditional generation model to obtain potential interaction distribution; performing interaction data reconstruction on the potential interaction distribution based on the target user representation through a condition generation model, and generating potential interaction data between the target user and the potential object; and constructing a query set of the target user based on the potential interaction data. According to the method, a high-quality query set conforming to potential object interests of a target user is constructed by sparse historical interaction data, so that a model adaptive to the target user can be effectively trained by a meta-learning normal form, and the cold start problem of the target user in a personalized recommendation scene is effectively solved.

Description

Technical Field

[0001] This specification relates to the technical field of machine learning, and in particular to a dataset construction, model training, object recommendation method, and recommendation system. Background Technology

[0002] With the rapid development of machine learning technology, personalized recommendations based on machine learning models have become a key technical means to improve user experience and platform stickiness, such as content recommendation, product recommendation, and service recommendation.

[0003] Currently, through the meta-learning paradigm, a support set containing historical interaction data and a query set containing future interaction data are constructed. For example, historical interaction data of users watching movies and future interaction data of users purchasing peripheral products after watching movies are used to train the object ranking model. This allows the object ranking model to effectively learn users' object interaction interests on the support set and to verify and optimize the model performance on the query set, thereby obtaining a personalized object ranking model adapted to the target user.

[0004] However, meta-learning paradigms heavily rely on the richness of support sets and query sets. For target users lacking historical behavior, such as new users or inactive users, the sparse nature of their future interaction data makes it impossible to directly construct rich query sets from this sparse future interaction data, hindering the direct application of traditional meta-learning paradigms. Therefore, there is an urgent need for a method to construct datasets capable of generating high-quality query sets. Summary of the Invention

[0005] In view of this, embodiments of this specification provide a method for constructing a dataset. One or more embodiments of this specification also relate to a model training method, an object recommendation method, a content recommendation method, a recommendation system, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0006] According to a first aspect of the embodiments of this specification, a method for constructing a dataset is provided, comprising:

[0007] Obtain the support set of the target user, wherein the support set includes historical interaction data between the target user and historical objects;

[0008] Extract target user representations from historical interaction data;

[0009] By using a conditional generation model, prior interaction distributions are conditionally mapped based on target user representations to obtain potential interaction distributions, where the potential interaction distributions represent the target user's object interaction interests.

[0010] By using a conditional generation model, based on the target user representation, the potential interaction distribution is reconstructed to generate potential interaction data between the target user and potential objects.

[0011] Based on potential interaction data, construct a query set for the target user.

[0012] According to a second aspect of the embodiments of this specification, a model training method is provided, comprising:

[0013] Obtain the support set and query set for the target user, wherein the query set is constructed according to the above dataset construction method;

[0014] Based on the support set and query set, meta-learning training is performed on the object ranking model to obtain a target object ranking model adapted to the target user.

[0015] According to a third aspect of the embodiments of this specification, an object recommendation method is provided, comprising:

[0016] Obtain the target user representation of the target user and the candidate object representation of multiple candidate objects;

[0017] The target object ranking model predicts the candidate scores corresponding to the candidate interaction data between the target user and multiple candidate objects based on the target user representation and the candidate object representation of multiple candidate objects. The target object ranking model is trained according to the above model training method.

[0018] Based on candidate scores, at least one target object is determined from multiple candidate objects;

[0019] Recommend at least one target object to the target user.

[0020] According to a fourth aspect of the embodiments of this specification, a content recommendation method is provided, comprising:

[0021] Obtain the target user representation of the target user and the candidate content representation of multiple candidate content;

[0022] The target content ranking model predicts the candidate scores corresponding to the candidate interaction data between the target user and multiple candidate contents based on the target user's target user representation and the candidate content representation of multiple candidate contents. The target content ranking model is trained according to the above model training method.

[0023] Based on candidate scores, at least one target content is determined from multiple candidate contents;

[0024] Recommend at least one piece of targeted content to the target user.

[0025] According to a fifth aspect of the embodiments of this specification, a recommendation system is provided, including a condition generation module;

[0026] The condition generation module is used to obtain the support set of the target user, which includes historical interaction data between the target user and historical objects; extract the target user representation from the historical interaction data; use the condition generation model to perform conditional mapping on the prior interaction distribution based on the target user representation to obtain the potential interaction distribution, where the potential interaction distribution represents the target user's object interaction interests; use the condition generation model to reconstruct the interaction data of the potential interaction distribution based on the target user representation to generate potential interaction data between the target user and potential objects; and construct the query set of the target user based on the potential interaction data.

[0027] According to a sixth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0028] Memory and processor;

[0029] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the above method.

[0030] According to a seventh aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0031] According to an eighth aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0032] One embodiment of this specification provides a method for constructing a dataset, comprising: obtaining a support set for a target user, wherein the support set includes historical interaction data between the target user and historical objects; extracting a target user representation of the target user from the historical interaction data; using a conditional generation model, performing conditional mapping on the prior interaction distribution based on the target user representation to obtain a potential interaction distribution, wherein the potential interaction distribution represents the target user's object interaction interests; using a conditional generation model, reconstructing the interaction data on the potential interaction distribution based on the target user representation to generate potential interaction data between the target user and potential objects; and constructing a query set for the target user based on the potential interaction data.

[0033] By extracting target user representations from the target user's support set and sparse historical interaction data, the user attributes and object interaction interests of the target user are abstracted, laying the data foundation for subsequent conditional generation model to perform conditional mapping and reconstruction. The conditional generation model performs conditional mapping based on the target user representation, which can accurately capture the potential interaction distribution of the target user from sparse data, thereby representing its deep and rich object interaction interests. By reconstructing the potential interaction distribution, high-quality and diverse potential interaction data is generated, effectively simulating the potential interaction behaviors that the target user may have with potential objects but which have not actually occurred. Based on the generated potential interaction data, a query set for the target user is constructed, breaking through the dependence of meta-learning on real and rich query sets. Only sparse historical interaction data is used to construct a high-quality query set for cold-start target users for model validation and optimization.

[0034] By using the target user representation in sparse historical interaction data as the generation condition, a high-quality query set that matches the potential object interests of the target user is constructed through a conditional generation model. This enables the meta-learning paradigm to effectively train a model adapted to the target user, effectively solving the cold start problem of the target user in personalized recommendation scenarios. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating a method for constructing a dataset according to one embodiment of this specification;

[0036] Figure 2 This is a schematic diagram of the structure of a conditional generation model in a dataset construction method provided in one embodiment of this specification;

[0037] Figure 3 This is a flowchart of a model training method provided in one embodiment of this specification;

[0038] Figure 4 This is a flowchart illustrating an object recommendation method provided in one embodiment of this specification;

[0039] Figure 5 This is a flowchart illustrating a content recommendation method provided in one embodiment of this specification;

[0040] Figure 6 This is a front-end schematic diagram of a content recommendation method provided in one embodiment of this specification;

[0041] Figure 7 This is a schematic diagram of the structure of a recommendation system provided in one embodiment of this specification;

[0042] Figure 8 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0043] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0044] The terminology used in one or more embodiments of the present invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention. The singular forms “a,” “the,” and “the” used in one or more embodiments of the invention are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” used in one or more embodiments of the invention refers to and includes any or all possible combinations of one or more associated listed items.

[0045] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of the present invention, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of the present invention, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0046] Furthermore, it should be noted that the data involved in one or more embodiments of the present invention are all information and data authorized by the user or fully authorized by all parties, and the statistics, use and processing of the relevant data need to comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0047] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0048] Meta-learning is a machine learning paradigm whose core goal is to teach models how to learn. Unlike traditional machine learning models that focus on learning a single task from a specific dataset, meta-learning models acquire a high-level, transferable learning strategy or prior knowledge by training on a large number of different but related tasks. This allows the model to quickly adapt and demonstrate good performance when faced with a completely new task, even with only a very small number of training samples, through very few iterations.

[0049] Support set: A small set of samples that provides the model with the contextual information needed to learn a new task. It is equivalent to the training set in traditional machine learning, but much smaller. The model uses samples from the support set to understand the target and boundaries of the current task, and builds a classifier or prediction model for that specific task based on this understanding.

[0050] Query set: In the meta-learning paradigm, this is the set of samples used to evaluate and optimize the model. It's equivalent to the test set in traditional machine learning, but its role is more crucial. After learning task information from the support set, the model needs to predict the samples in the query set. The loss values ​​generated by these predictions not only evaluate the model's performance on the new task, but more importantly, this loss serves as a feedback signal, directly used through gradient backpropagation to optimize the parameters of the meta-learning model itself, enabling it to learn better from the support set.

[0051] Model-Agnostic Meta-Learning (MAML) is a general meta-learning algorithm framework. Its core objective is to find model initialization parameters θ that are sensitive to task variations. The "model-independent" nature of this algorithm lies in its ability to be combined with any model architecture (such as neural networks) optimized using gradient descent. The training process of MAML consists of an inner loop and an outer loop: In the inner loop, the model uses the support set for each task to perform one or more gradient updates on the initialization parameters θ, obtaining the target parameters θ' adapted to that specific task; in the outer loop, the model uses the query set corresponding to each task to calculate the performance loss based on the target parameters θ', and optimizes the initialization parameters θ using gradient descent. The ultimate goal is to ensure that the target parameters θ' obtained after fine-tuning in the inner loop, starting from the initialization parameters θ, achieve excellent performance across various tasks.

[0052] Variational Autoencoder (VAE): A generative neural network architecture based on a probabilistic graphical model, consisting of an encoder layer and a decoder layer. Its core idea is to map input data to a probability distribution in a latent space (usually assumed to be Gaussian), and then reconstruct the original input or generate new data samples using a decoder by sampling from this distribution. Unlike traditional autoencoders, VAEs apply regularization (such as KL divergence constraints) to the latent space, giving it good generalization ability and continuity, thus enabling high-quality data generation, interpolation, and denoising tasks. VAEs are widely used in image generation, text modeling, and recommender systems.

[0053] Conditional Variational Autoencoder (CVAE) is an extension of VAE that introduces generation conditions (such as category labels, user features, and contextual data) into the VAE process. CVAE uses these generation conditions as input during both encoding and decoding, thus achieving precise control over the generation process. Its goal is to learn a conditional probability distribution so that, given specific conditions, the model can generate data samples that conform to the given context. CVAE is widely used in tasks that rely on additional information for generation, such as conditional image synthesis, sequence generation, and interactive data simulation in personalized recommendations.

[0054] Generative Adversarial Networks (GANs) are game-theory-based generative deep learning models consisting of two neural networks: a generator and a discriminator. The core idea is to enable the generator to learn the distribution of real data through adversarial training between the generator and discriminator, thereby generating high-quality new samples. The generator's goal is to generate data realistic enough to fool the discriminator, while the discriminator's goal is to accurately distinguish between real and generated data. This adversarial training mechanism makes GANs excellent in areas such as image generation, data augmentation, and style transfer. In recommender systems, GANs can be used to generate user interaction data and solve the cold start problem.

[0055] Diffusion Model: A generative model based on Markov chains, its core idea is to learn the data distribution by progressively adding noise (forward process) and progressively denoising (backward process). The forward process gradually transforms the original data into Gaussian noise, while the backward process reconstructs the data from pure noise by learning noise predictions. Through this progressive generative approach, the diffusion model can generate high-quality and diverse samples, achieving breakthrough results in tasks such as image generation, audio synthesis, and text generation. In recommender system applications, the diffusion model can be used to generate user interaction data that conforms to the distribution of user interests, building high-quality training samples for cold-start users.

[0056] Kullback-Leibler Divergence (KL Divergence) is an asymmetric measure of the difference between two probability distributions. In information theory, KL divergence represents the expected amount of information lost when using a probability distribution Q to approximate the true distribution P. Its value is non-negative and is zero if and only if the two distributions are identical. A larger value indicates a greater difference between the two distributions. In variational autoencoders (VAEs) and their derivative models (such as CVAEs), KL divergence acts as a regularization term.

[0057] Object ranking model: The object ranking model is the core prediction model in recommender systems, used to personalize the ranking of candidate objects. Based on user and object identifiers, this model uses machine learning algorithms to learn user interests and preferences, thereby predicting the user's preference for each object and generating a recommended ranking list.

[0058] Dual-Tower Model: A classic recommender system architecture that uses a dual-tower structure to learn representation vectors for users and objects respectively. The user tower learns the representation of user identifiers, and the object tower learns the representation of object identifiers. Finally, the prediction score is obtained by calculating the similarity between the two representation vectors.

[0059] Deep Interest Network (DIN) model: An attention-based recommendation model that adaptively learns how user interests change with different objects. This model dynamically calculates the relevance between users and candidate objects through an attention network, thereby better capturing users' diverse interests and preferences.

[0060] Transformer: A deep learning model architecture based on self-attention mechanism, commonly used in recommender systems to process user behavior sequence data. Through self-attention, Transformer can capture long-term dependencies in user behavior sequences and effectively learn the dynamic evolution of user interests.

[0061] Contrastive training: Contrastive training is a learning paradigm that trains a model to distinguish between positive and negative samples, aiming to narrow the gap between positive samples and widen the gap between negative samples. In recommender systems, by constructing contrasting pairs of positive and negative samples representing user preferences, the model learns better representations.

[0062] Currently, in the field of personalized recommendation systems, the cold start problem for new users is a core technical bottleneck restricting system performance. Traditional recommendation algorithms heavily rely on users' historical interaction data for collaborative filtering or content matching. When new users first enter the system, due to a lack of sufficient behavioral records, the system cannot accurately capture their personalized preferences, leading to a significant decline in recommendation quality. Specifically, there are two main problems:

[0063] Limitations of meta-learning methods: While existing recommendation methods based on model-agnostic meta-learning can achieve rapid adaptation, they rely on a clear division between the support set and the query set. In real-world cold-start scenarios, new users only have a limited number of current interaction records, making it impossible to obtain "future query samples" for model validation and optimization. This makes it difficult to directly apply traditional meta-learning paradigms.

[0064] Sample scarcity and quality issues: In cold start scenarios, the available training samples are extremely limited. Existing data augmentation methods are mostly based on simple random sampling or rule generation, which cannot effectively simulate the true distribution of user preferences.

[0065] To address the aforementioned issues, this specification proposes a query set construction method based on a conditional generation model, the process of which is as follows: Figure 1 As shown, this method first obtains the support set of the target user and extracts user representations. Then, it uses a conditional generative model to conditionally map the prior interaction distribution, obtaining a potential interaction distribution reflecting the user's potential interaction interests. This data is then reconstructed to generate high-quality, diverse potential interaction data, constructing a query set suitable for meta-learning optimization. This method effectively alleviates the problems of scarce and low-quality query sets in cold-start scenarios, providing a feasible data foundation for adapting meta-learning models to new users.

[0066] This specification provides a method for constructing a dataset, and also relates to a model training method, an object recommendation method, a content recommendation method, a recommendation system, a computing device, a computer-readable storage medium, and a computer program product, which are described in detail in the following embodiments.

[0067] See Figure 1 , Figure 1 A flowchart illustrating a method for constructing a dataset according to an embodiment of this specification is shown, including the following specific steps:

[0068] Step 102: Obtain the support set of the target user, wherein the support set includes historical interaction data between the target user and historical objects.

[0069] The embodiments in this specification are applied to a sample data processing platform with query data construction function. It can be a general sample data processing platform or a dedicated sample data processing module in recommendation systems such as e-commerce platforms, community content platforms, or social media platforms.

[0070] The target users are low-activity users with sparse historical behavioral data on the platform, which can be new users or low-activity users. For example, new users who have registered within the last 7 days and have fewer than 10 behavioral records, or low-activity users who have logged in less than 3 times in the last 30 days.

[0071] The support set for the target user is a sample dataset used for inner-layer updates during meta-learning training. It is a small-scale sample dataset constructed using sparse, real-world historical interaction data of the target user, including historical interaction data between the target user and historical objects.

[0072] Historical interaction data between target users and historical objects consists of data tuples that record historical interaction behaviors between target users and historical objects. This data may include user identifiers of target users, object identifiers of historical objects, interaction types of historical interaction behaviors, interaction timestamps of historical interaction behaviors, and label scores of historical interaction behaviors (interaction intensity, used to determine the degree of user preference and as a supervision signal for model training. This label score quantifies the target user's object interest intensity towards historical objects; the higher the value, the higher the interest, which directly affects the optimization direction of object recommendation ranking).

[0073] Historical objects are resource objects representing the target user's past interactions on the platform. These can include goods, services, content, social media, etc. For example, in an e-commerce platform, historical objects could be specific goods (such as mobile phones) that the target user has purchased, viewed, or saved; in an online education platform, historical objects could be course services (such as introductory programming courses) that the target user has purchased, viewed, or saved; in a content platform, objects could be articles or videos (such as articles like "Ten Tips for Efficient Learning") that the target user has purchased, viewed, or saved; and in a social media platform, objects could be user profiles or topic pages that the target user has purchased, viewed, or saved.

[0074] For example, the support set for newly registered users includes n historical interaction data: S u ={(i1,r1),(i2,r2),……,(i n ,r n )}, where (i i ,r i Let be the i-th historical interaction data, i i For the i-th historical object (item), r i Let be the label score of the i-th historical object.

[0075] One possible way to obtain the support set of the target user is to collect historical interaction data between the target user and historical objects from the platform's user behavior logs based on the target user's user identifier, and then construct the support set of the target user. Another possible way is to query and import historical interaction data between the target user and historical objects from a third-party authorized data platform based on the target user's user identifier, and then construct the support set of the target user. No particular method is specified here.

[0076] For example, taking a community content platform as an example, the target users are multiple newly registered users. The platform exposes popular community content to these users, including content on travel, food, technology, fitness, and fashion. Newly registered user "User A" interacted with some of this content (clicking, reading, sharing, saving, etc.), generating 5 historical interaction records, which are kept in the platform's user behavior log.

[0077] (User ID: User A, Object ID: Tourism 001, Interaction Type: Click, Timestamp t1, Tag Score 0.6);

[0078] (User ID: User A, Object ID: Tourism 002, Interaction Type: Reading, Timestamp t2, Tag Score 0.8);

[0079] (User ID: User A, Object ID: Travel 003, Interaction Type: Sharing, Timestamp t3, Tag Score 0.9);

[0080] (User ID: User A, Object ID: Food001, Interaction Type: Favorite, Timestamp t4, Tag Score 0.85);

[0081] (User ID: User A, Object ID: Technology_001, Interaction Type: Click, Timestamp t5, Tag Score 0.3).

[0082] Among them, sharing and saving behaviors received high tag scores (0.9 and 0.85 respectively), indicating that users have a strong interest in travel and food content; while short-lived clicks on technology content only received 0.3 points, indicating weaker interest.

[0083] A support set was constructed by collecting historical interaction data between multiple newly registered users and historical objects.

[0084] In step 102, by obtaining the support set of the target user, a data foundation of historical interaction data is provided for extracting the target user's representation.

[0085] Step 104: Extract the target user representation from historical interaction data.

[0086] The target user representation is the user interaction interest of the target user as represented in historical interaction data. It is usually represented by a vector and can be used to characterize user profiles such as gender, location, preferences, activity level, and spending power. The target user representation is typically used as the user-side input to an object ranking model obtained through meta-learning training based on the support set and query set. Because the generation conditions of the target user representation highly abstract the attribute features of the target user, it can be used as a generation condition in the process of reconstructing the generation conditions from the potential interaction data of the conditional generation model, thereby controlling the generated potential interaction data. For example, the target user representation e of a newly registered user... u ∈R d The criteria for new user generation are young women located in a certain area with a certain level of spending power, who show a significant preference for "makeup tutorials" and "fashion outfits" content.

[0087] For example, target user representations for multiple newly registered users were extracted from their historical interaction data. Specifically, the target user representation for newly registered user "User A" is that of a young user with high spending power, showing a strong preference for travel and food-related content, and a weaker interest in technology-related content. This representation will serve as a key generation condition, guiding subsequent conditional generation models to generate potential interaction data that conforms to this user's interest distribution.

[0088] In step 104, by extracting the target user representation from the target user's support set and sparse historical interaction data, the user attributes and object interaction interests of the target user are abstracted, laying the data foundation for the subsequent conditional generation model to perform conditional mapping and reconstruction.

[0089] Step 106: Using a conditional generation model, based on the target user representation, conditionally map the prior interaction distribution to obtain the potential interaction distribution, where the potential interaction distribution represents the target user's object interaction interests.

[0090] Conditional generative models are generative neural network models based on conditional probability. Their architecture consists of three parts: an encoding layer, a latent space, and a decoding layer. Core functions include, but are not limited to, conditional mapping and reconstruction generation. They achieve precise control over the data generation process by learning from conditional inputs. Conditional generative models include, but are not limited to, conditional variational autoencoders, generative adversarial networks, and diffusion models.

[0091] The prior interaction distribution is the underlying probability distribution of all users' historical interaction behaviors on the platform when no user conditions are introduced. It can be represented by a standard Gaussian distribution z ~ N(0,I) and serves as the baseline distribution for conditional mapping.

[0092] The latent interaction distribution is the spatial distribution of potential interactive behaviors in the latent space obtained by transforming the prior interaction distribution based on the generation conditions of the target user's representation in the conditional generative model. The latent interaction distribution represents the target user's object interaction interests. If the prior interaction distribution is a standard Gaussian distribution, the latent interaction distribution is also Gaussian. For example, the latent interaction distribution... e u Let z be the target user representation, and z be the latent space of the prior interaction distribution. The mean of the distribution is... The variance of the distribution, This refers to the conditional mapping operation for the conditional generation model.

[0093] By using a conditional generation model, the prior interaction distribution is conditionally mapped based on the target user representation to obtain the potential interaction distribution. One possible approach is to use the encoding layer of the conditional generation model to conditionally map the prior interaction distribution based on the target user representation to obtain the potential interaction distribution.

[0094] Based on the target user representation, the prior interaction distribution is conditionally mapped to obtain the latent interaction distribution. Specifically, through the conditional generation model, the generation conditions based on the target user representation are mapped to a prior interaction distribution that is unrelated to the target user to obtain the target user's specific posterior interaction distribution, i.e., the latent interaction distribution, which can represent the target user's object interaction interests in the latent space.

[0095] For example, for a newly registered user "User A", the encoding layer of the conditional variational autoencoder uses its target user representation as the generation condition, maps it to the latent space of a standard Gaussian prior distribution, and obtains the latent interaction distribution N(μ,σ). 2 ), where the mean μ and variance σ 2 The distribution, calculated by the encoding layer based on the generation conditions, is concentrated in the potential areas of tourism and food-related content, while staying away from technology content areas, accurately reflecting the user's interactive interests.

[0096] In step 106, by performing conditional mapping based on the target user representation through the conditional generation model, the potential interaction distribution of the target user can be accurately captured from sparse data, thereby representing its deep and rich object interaction interests.

[0097] Step 108: Using a conditional generation model, based on the target user representation, reconstruct the interaction data of the potential interaction distribution to generate potential interaction data between the target user and potential objects.

[0098] Potential objects are resource objects that the target user is likely to interact with on the platform but has not yet done so. Their type is usually consistent with historical objects and can be goods, services, content, social media, etc. For example, in a community content platform, potential objects could be other travel guides or food videos that the target user has not yet encountered; in an e-commerce platform, potential objects could be similar products that the user may be interested in but has not yet browsed.

[0099] The potential interaction data between the target user and potential objects consists of data tuples of potential interaction behaviors simulated through a conditional generation model. This data can include the target user's user identifier, the potential object's object identifier, the interaction type of the potential interaction behavior, the interaction timestamp of the potential interaction behavior, and the label score of the potential interaction behavior (interaction strength, used to determine the degree of user preference and as a supervision signal for model training; this label score quantifies the target user's interest in the potential object; a higher value indicates a higher interest and directly affects the optimization direction of object recommendation ranking). For example, potential interaction data p... θ (x|z,e u ), where x is the generated conditional query sample, i.e., the potential interaction data x. gen =G(z|e u ), R d →R m p θ (·|z,e_u) represents the interactive data reconstruction operation of the conditional generation model.

[0100] By using a conditional generation model, based on the target user representation, the potential interaction distribution is reconstructed to generate potential interaction data between the target user and potential objects. One possible approach is to use the decoding layer of the conditional generation model to reconstruct the interaction data between the potential interaction distribution based on the target user representation to generate potential interaction data between the target user and potential objects.

[0101] Based on the target user representation, the potential interaction distribution is reconstructed to generate potential interaction data between the target user and potential objects. Specifically, through a conditional generation model, based on the generation conditions of the target user representation, the target user representation and potential interaction distribution are used as joint inputs to perform a nonlinear transformation and output the potential interaction data between the target user and potential objects.

[0102] For example, for target user "User A" (whose target user representation e) u This indicates a strong preference for travel and food-related content, and a weaker interest in technology-related content. The decoding layer of the conditional variational autoencoder, based on its target user representation e... u As a condition for generation, from the potential interaction distribution Sample the latent variable z as a joint input, perform a nonlinear transformation, and generate 5 latent interaction data points:

[0103] (User ID: User A, Object ID: Tourism 004, Interaction Type: Deep Reading, Timestamp t6, Tag Score 0.87);

[0104] (User ID: User A, Object ID: Travel 005, Interaction Type: Favorite, Timestamp t7, Tag Score 0.91);

[0105] (User ID: User A, Object ID: Food 002, Interaction Type: Sharing, Timestamp t8, Tag Score 0.89);

[0106] (User ID: User A, Object ID: Food003, Interaction Type: Deep Reading, Timestamp t9, Tag Score 0.84);

[0107] (User ID: User A, Object ID: Technology_002, Interaction Type: Click, Timestamp t10, Tag Score 0.35).

[0108] In step 108, by reconstructing the potential interaction distribution, high-quality and diverse potential interaction data is generated, effectively simulating the potential interaction behaviors that the target user may have with potential objects but which have not actually occurred.

[0109] Step 110: Construct a query set for the target user based on potential interaction data.

[0110] The query set for the target user is a sample dataset used for outer layer updates during meta-learning training. It uses a large-scale sample dataset of generated, rich data that simulates potential interactions of the target user that have not actually occurred, including potential interaction data between the target user and potential objects.

[0111] Based on potential interaction data, a query set for the target user can be constructed. One possible approach is to directly organize the potential interaction data into a query set for the target user. Another possible approach is to filter the potential interaction data to obtain the filtering results and construct the query set for the target user based on the filtering results. Yet another possible approach is to mix the potential interaction data with interaction data from other sources to construct the query set for the target user. No specific approach is specified here.

[0112] For example, based on 5 potential interaction data points generated for the target user "User A", a query set is constructed by filtering with a tag score threshold of 0.8 and retaining 4 high-quality interaction data points (Tourism 004, Tourism 005, Food 002, Food 003):

[0113] (User ID: User A, Object ID: Tourism 004, Interaction Type: Deep Reading, Timestamp t6, Tag Score 0.87);

[0114] (User ID: User A, Object ID: Travel 005, Interaction Type: Favorite, Timestamp t7, Tag Score 0.91);

[0115] (User ID: User A, Object ID: Food 002, Interaction Type: Sharing, Timestamp t8, Tag Score 0.89);

[0116] (User ID: User A, Object ID: Food003, Interaction Type: Deep Reading, Timestamp t9, Tag Score 0.84).

[0117] The query set and support set are consistent with the target users' interactive interests (mainly travel and food-related content), while avoiding interference from low-quality samples (technology_002, label score 0.35), providing high-quality and diverse optimization targets for the meta-learning model.

[0118] It should be noted that the network architecture design of the Conditional Variational Autoencoder is original in its encoder-decoder structure, latent space design, and conditional input method. This architecture maps the user support set to a Gaussian-distributed latent space through the encoder layer, and the decoder layer generates diverse interaction samples based on user representation conditions. This effectively solves the core problem of query set scarcity in cold start scenarios, ensuring a core technological advantage in sample generation capabilities, which directly determines the overall performance of the recommendation system.

[0119] In step 110, a query set for the target user is constructed based on the generated potential interaction data. This breaks through the dependence of meta-learning on real and rich query sets. It can construct a high-quality query set for cold-start target users using only sparse historical interaction data for model validation and optimization.

[0120] In the embodiments of this specification, the target user representation in sparse historical interaction data is used as the generation condition. Through the conditional generation model, a high-quality query set that conforms to the potential object interests of the target user is constructed, so that the meta-learning paradigm can effectively train a model adapted to the target user and effectively solve the cold start problem of the target user in the personalized recommendation scenario.

[0121] In one optional embodiment of this specification, step 104 includes the following specific steps:

[0122] The target user representation is extracted from historical interaction data through the representation extraction layer.

[0123] The representation extraction layer is a network layer structure used to extract multi-dimensional feature representations from user historical behavior. It can include representation extraction layers with multiple dimensions: preference representation extraction layer, activity representation extraction layer, spending power representation extraction layer, spatiotemporal feature representation extraction layer, social relationship representation extraction layer, and device environment representation extraction layer. The number of dimensions is typically set according to the user's representation needs. The representation extraction layer can be a structure combining embedding layers and aggregation networks, with each embedding layer corresponding to one dimension of the representation extraction layer.

[0124] It should be noted that the representation extraction layer usually maintains the same or similar parameters and structure as the representation extraction layer of the model that needs to be trained by meta-learning. This ensures that the generated target user representation is consistent with the representation in the subsequent meta-learning process, thereby ensuring that the generated query set and support set are consistent in the distribution of the feature space and improving the training effect of the meta-learning model.

[0125] For example, for target user "User A", the target user representation is extracted from its 5 historical interaction data through the representation extraction layer:

[0126] The preference representation extraction layer generates a preference representation vector e based on the interaction object type (tourism, food, technology) and tag scores. pref This indicates a strong preference for tourism and food; the activity representation extraction layer generates an activity representation vector e based on timestamp intervals and interaction frequency. act The first layer represents moderate activity; the second layer, based on historical payment behavior, generates a consumption capacity representation vector, representing high consumption capacity; other representation layers also generate corresponding feature vectors, which are finally concatenated to form a complete target user representation e. u .

[0127] In the embodiments described in this specification, the characterization extraction layer can comprehensively and accurately extract user features from sparse historical interaction data, providing rich generation conditions for the subsequent conditional generation model generation process.

[0128] Considering the user representations used in existing recommender systems, there is a common problem of overfitting to highly active users. The model easily misinterprets user interaction frequency as preference strength, leading to poor recommendation performance for inactive users with clear preferences, thus reducing the overall fairness and personalization level of the recommender system. In one optional embodiment of this specification, the target user representation includes a target user preference representation and a target user activity representation.

[0129] The target user representation is extracted from historical interaction data through the representation extraction layer, including the following specific steps:

[0130] Through the preference representation extraction layer, target user preference representations are extracted from historical interaction data.

[0131] The activity representation extraction layer extracts the target user's activity representation from historical interaction data. The weight matrices of the preference representation extraction layer and the activity representation extraction layer are orthogonally constrained.

[0132] The preference representation extraction layer is a neural network layer used to extract interest preference features from users' historical interaction behavior. Its input is features such as object type and label score of historical interaction data, and its output is a vector of user preference intensity for different types of content.

[0133] The activity representation extraction layer is a neural network layer used to extract behavioral activity features from users' historical interaction behavior. Its inputs are features such as timestamp intervals and interaction frequencies of historical interaction data, and its output is a vector of user activity on the platform.

[0134] The weight matrices of the preference representation extraction layer and the activity representation extraction layer are orthogonally constrained, specifically by adding an orthogonal regularization loss term, so that the weight matrices of the two extraction layers satisfy... The orthogonality condition ensures that preference features and activity features are independent of each other in the representation space, avoiding feature coupling.

[0135] The target user preference is represented as a vector representation that quantifies the target user's interest in different types of content. The magnitude of the value reflects the intensity of the preference, and the higher the value, the stronger the interest.

[0136] Target user activity is represented as a vector representation that quantifies the level of activity of target users on the platform. Its value reflects the intensity of activity, and the higher the value, the more active the user.

[0137] It should be noted that the subspace decomposition constraint mechanism, through techniques such as orthogonal constraint implementation, activity-supervised learning, and subspace dimension allocation, innovatively solves the activity bias problem commonly found in the recommendation industry. This mechanism ensures that user preference features and activity features are independent of each other in the representation space, avoiding misjudging high activity as strong preference, and significantly improving the fairness of object recommendation for users with different activity levels.

[0138] For example, for target user "User A", the preference representation is extracted from its 5 historical interaction data through the preference representation extraction layer:

[0139] Input features: object type (tourism, food, technology) and corresponding tag scores (0.6, 0.8, 0.9, 0.85, 0.3);

[0140] The output vector [0.85, 0.75, 0.1] represents a strong preference for travel and food, and a weaker interest in technology.

[0141] Extract target user activity characteristics through the activity characterization extraction layer:

[0142] Input features: timestamp interval (t2-t1, t3-t2, t4-t3, t5-t4) and interaction frequency (5 times / time period);

[0143] Output vector: [0.7, 0.3], representing moderate activity.

[0144] Orthogonal constraints are used to ensure that the weight matrices of the preference representation extraction layer and the activity representation extraction layer satisfy the following: Ensure the independence of the two feature spaces.

[0145] In the embodiments of this specification, by extracting dual representations of preference and activity through orthogonal constraints, the two relatively independent feature dimensions of user interest preferences and behavioral patterns can be decoupled from sparse data. This explicitly separates user preference signals from activity signals, providing more accurate and richer condition information for subsequent condition generation. In turn, a high-quality query set that reflects both the user's true interest preferences and is adapted to their activity level is constructed for meta-learning, thereby improving the fairness and personalized recommendation of the recommendation system for users with different activity levels.

[0146] In one optional embodiment of this specification, the training steps of the preference representation extraction layer include the following specific steps:

[0147] Obtain sample interaction data between reference users and sample objects, and the corresponding label preference values ​​for the sample interaction data. The label preference values ​​are determined and labeled based on whether the sample interaction data is a factual interaction.

[0148] The preference representation extraction layer extracts the user preference representation of the sample users from the sample interaction data.

[0149] The object representation extraction layer extracts sample object representations from sample interaction data.

[0150] Based on the sample user preference representation and sample object representation, calculate the predicted preference degree corresponding to the sample interaction data;

[0151] The contrast loss is determined based on the predicted preference and the label preference corresponding to the sample interaction data.

[0152] Based on the contrastive loss, the parameter weights of the preference representation extraction layer are adjusted until the trained preference representation extraction layer is obtained.

[0153] The reference users are highly active users with rich historical behavioral data on the platform, which can be either long-term users or highly active users. For example, highly active users who have registered more than 3 months ago and have more than 5,000 behavioral records, or highly active users who have logged in more than 20 times in the past 30 days and have rich click and return visit records.

[0154] The sample objects are resource objects that reference users have interacted with on the platform. Their types are consistent with historical objects and can include goods, services, content, social media, etc. For example, in a community content platform, sample objects could be travel guides or food videos that reference users have viewed; in an e-commerce platform, sample objects could be specific products that reference users have purchased.

[0155] The sample interaction data between the reference user and the sample object is a data set that records the interaction behavior between the reference user and the sample object, including positive samples (objects that the reference user has actually interacted with) and negative samples (objects that the reference user has not interacted with but have been exposed by the platform), which are used to train the preference representation extraction layer.

[0156] The label preference degree corresponding to the sample interaction data is the preference intensity value labeled according to the user's actual interaction behavior. Positive samples are usually labeled as 1 or assigned a score between 0 and 1 according to the interaction depth, while negative samples are labeled as 0, which is used to supervise the learning process of preference representation.

[0157] Tag preference is determined and labeled based on whether the sample interaction data is actual interaction. Specifically, for positive interaction behaviors such as actual user clicks, browsing, favorites, and purchases, a preference score between 0.1 and 1.0 is assigned according to the interaction type and duration; for objects that are exposed but do not generate interaction, a preference score of 0 is assigned.

[0158] The sample user preference characterization is a vector representation that quantifies the sample users' interest preferences for different types of content. The magnitude of the value reflects the intensity of the preference, and the higher the value, the stronger the interest.

[0159] The object representation extraction layer is a network layer structure used to extract feature representations from the content of sample objects. It typically employs an embedding layer or a convolutional neural network to map object identifiers, content features, etc., into low-dimensional dense vectors.

[0160] The sample object representation is a vector representation output by the object representation extraction layer, which encodes the semantic content and attribute features of the object.

[0161] The predicted preference score corresponding to the sample interaction data is a prediction score obtained by calculating the similarity between the sample user preference representation and the sample object representation, reflecting the degree of user preference for the object predicted by the model.

[0162] Contrastive loss is a loss function used to optimize preference representation. By bringing positive sample pairs closer together and pushing negative sample pairs further apart, the model learns to distinguish user preferences.

[0163] One possible way to obtain sample interaction data between reference users and sample objects is to collect complete interaction records of highly active users from the platform's user behavior logs, including positive samples and randomly sampled negative samples.

[0164] One possible way to obtain the tag preference scores corresponding to sample interaction data is to set preference mapping rules based on the interaction type, such as mapping clicks to 0.6, reading to 0.8, favorites to 0.9, and purchases to 1.0.

[0165] Based on the user preference representation and object representation of the samples, the predicted preference score corresponding to the sample interaction data is calculated. One possible approach is to calculate the inner product similarity of the two representation vectors and output the predicted preference score based on the inner product similarity score = f. pref (e pref e item ), where e pref e represents the preferences of the sample users item To represent the sample object.

[0166] Based on the predicted preference and the label preference corresponding to the sample interaction data, the contrast loss is determined. One possible approach is to use the binary cross-entropy loss function to calculate the difference between the predicted value and the true label.

[0167] For example, for the reference user "User B" (a highly active user), 1000 sample interaction data points are collected, including 500 positive samples (containing actions such as clicks, reading, and favorites) and 500 negative samples (randomly sampled non-interacting objects). A preference representation extraction layer extracts sample user preference representations from these data, while an object representation extraction layer extracts corresponding sample object representations. After calculating the predicted preference degree for each sample pair, a binary cross-entropy loss is used for optimization.

[0168] Positive sample: (User B, Tourism 123) True label 0.9, predicted value 0.85;

[0169] Negative sample: (User B, Technology 456) True label 0.0, predicted value 0.2.

[0170] By minimizing the loss function and continuously adjusting the parameters of the preference representation extraction layer, a trained preference representation extraction layer that can accurately capture user preferences is finally obtained.

[0171] In the embodiments of this specification, by optimizing the preference representation extraction layer through a contrastive learning framework, it is possible to effectively distinguish between the user's true preferences and the interference signals brought about by activity levels, thereby improving the purity and accuracy of preference representation and providing more reliable preference signals for subsequent condition generation.

[0172] In one optional embodiment of this specification, the training steps of the activity representation extraction layer include the following specific steps:

[0173] Obtain sample interaction data between reference users and sample objects and the corresponding tag activity of the sample interaction data. The tag activity is determined and labeled based on the interaction frequency between reference users and sample objects.

[0174] The activity representation extraction layer extracts the activity representation of sample users from the sample interaction data.

[0175] The activity prediction layer predicts the activity level corresponding to the sample interaction data based on the sample user activity representation.

[0176] The autoregressive supervision loss is determined based on predicted activity and labeled activity.

[0177] Based on autoregressive supervised loss, the parameter weights of the activity representation extraction layer are adjusted until the trained activity representation extraction layer is obtained.

[0178] The tag activity level corresponding to the sample interaction data is a quantified value of activity calculated based on the user's historical interaction frequency. It is typically the logarithm of the total number of user interactions, used to supervise the learning process of the activity representation. For example, if a user has 100 interactions, the tag activity level is log(|S u |), where |S u | is the logarithm of the number of user interactions: log(100) = 2.0.

[0179] The activity prediction layer is a neural network layer used to predict user activity from activity representations. It typically employs a linear transformation plus activation function structure to map the activity representations to predicted activity values, which can be represented as follows:

[0180] The predicted activity level corresponding to the sample interaction data is the value output by the activity prediction layer, which represents the user activity level predicted by the model.

[0181] Autoregressive supervised loss is a loss function used to optimize the activity representation. By minimizing the difference between predicted activity and labeled activity, the model accurately captures user activity patterns. For example, the mean squared error loss value L... act =|h act (e act )-log(|S u|)||2.

[0182] One possible way to obtain the tag activity corresponding to the sample interaction data is to calculate the total number of interactions of the user within a specific time window, and determine the tag activity corresponding to the sample interaction data based on the total number of interactions. The natural logarithm can be used as the tag activity.

[0183] For example, for the reference user "User B" (a highly active user), their interaction records from the past 30 days, totaling 2000 times, are collected, and the tag activity is calculated as log(2000)≈3.3. The activity representation extraction layer extracts the sample user activity representation [0.8, 0.6] from the sample interaction data, and the predicted activity is 3.1 obtained through the activity prediction layer (a single-layer fully connected network). The mean squared error loss is calculated as (3.1-3.3). 2 =0.04, and the parameters of the activity representation extraction layer are adjusted through backpropagation to finally obtain a trained activity representation extraction layer that can accurately capture user activity patterns.

[0184] In the embodiments of this specification, the activity representation extraction layer is optimized by using an autoregressive supervised learning framework, which can accurately capture the user's activity pattern characteristics, avoid mistaking activity signals for preference signals, and improve the adaptability and fairness of the recommendation system to users with different activity levels.

[0185] In one optional embodiment of this specification, the training steps of the conditional generation model include the following specific steps:

[0186] Obtain sample interaction data between the reference user and the sample object;

[0187] Extract reference user representations from the sample interaction data;

[0188] By using a conditional generation model, based on reference user representations, conditional mapping is performed on the prior interaction distribution to obtain the potential interaction distribution, where the potential interaction distribution represents the object interaction interests of the reference user.

[0189] By using a conditional generation model, based on the reference user representation, the potential interaction distribution is reconstructed to generate potential interaction data between the reference user and potential objects.

[0190] The reconstruction loss term is determined based on the similarity between potential interaction data and sample interaction data;

[0191] The distribution loss term is determined based on the similarity between the potential interaction distribution and the prior interaction distribution;

[0192] The loss value is determined based on the reconstruction loss term and the distribution loss term;

[0193] Based on the loss value, adjust the parameter weights of the conditional generative model until a fully trained conditional generative model is obtained.

[0194] The reference user representation is the user interaction interest of the reference user represented in the sample interaction data. It is usually represented by a vector and can be used to represent the user profile of the reference user, such as gender, location, preferences, activity level, and spending power. Its structure is consistent with the target user representation and can include dimensions such as preference representation and activity level representation.

[0195] The reconstruction loss term is a loss function that measures the difference between the generated data and the real data. It is used to ensure the quality of the generated data, and its calculation formula is: L recon =-E[logp θ (x|z,e u ]], where x is the sample interaction data, z is the latent variable, and e u For reference user representation.

[0196] The distribution loss term is a regularization term that measures the difference between the latent distribution and the prior distribution. It is used to constrain the structure of the latent space, and its calculation formula is as follows: in Let p(z) be the potential interaction distribution and p(z) be the prior interaction distribution.

[0197] The loss value is a weighted sum of the reconstruction loss term and the distribution loss term, used in the global optimization conditional generation model, and its calculation formula is as follows:

[0198] Based on the similarity between potential interaction data and sample interaction data, the reconstruction loss term is determined. One possible approach is to use the negative log-likelihood function to calculate the difference between the generated data and the real data.

[0199] Based on the similarity between the latent interaction distribution and the prior interaction distribution, the distribution loss term is determined. One possible approach is to use KL divergence to calculate the difference between the two distributions.

[0200] Based on the reconstruction loss term and the distribution loss term, the loss value is determined. One possible approach is to perform a weighted summation of the two loss terms to obtain the final optimization objective.

[0201] The embodiments in this specification and steps 102 to 110 are based on the same inventive concept. Steps not discussed in detail can be found in the embodiments in the above specification, and will not be repeated here.

[0202] For example, for the reference user "User B" (a highly active user), 1000 sample interaction data points are collected. A latent interaction distribution is generated based on the user's representation using a conditional variational autoencoder, from which 500 latent interaction data points are sampled. The reconstruction loss is calculated as: L recon=2.3; Calculate distributed loss: L KL =0.8, so the total loss is L VAE =2.3 + 0.5 * 0.8 = 2.7. By optimizing the parameters through backpropagation, a trained conditional variational autoencoder capable of generating high-quality query sets is finally obtained.

[0203] In the embodiments described in this specification, by jointly optimizing the reconstruction loss and the distribution loss, the conditional generation model can generate high-quality interactive data that matches user interests while maintaining a good latent space structure. This allows for the construction of realistic and diverse query sets for cold-start users, effectively supporting the training and optimization of the meta-learning model.

[0204] In one optional embodiment of this specification, the following specific steps are included before step 110:

[0205] Acquire subsequent interaction data between the target user and subsequent objects, and similar interaction data between similar users and historical objects. The subsequent interaction data occurs after the historical interaction data, and the similarity between the similar user and the target user exceeds a preset threshold.

[0206] Step 110 includes the following specific steps:

[0207] Based on preset distribution weights, subsequent interaction data, similar interaction data, and potential interaction data are mixed to construct a query set for the target user.

[0208] Subsequent interaction data consists of records of the actual interactive behaviors of the target user after the historical interaction data time point, including interaction information between the target user and subsequent objects, used to provide real signals of future behavior. For example, user A is newly generated after timestamp t11 (user A, object identifier: travel006, interaction type: collection, timestamp t11, tag score 0.92).

[0209] Similar users are other users who are highly similar to the target user in terms of user attributes, behavioral patterns, or interests. They are usually calculated using a user similarity algorithm, with a similarity exceeding a preset threshold (such as 0.8). For example, user C, who, like user A, has a strong preference for travel and food content.

[0210] Similar interaction data consists of historical interaction records between similar users and historical objects, used to provide supplementary information on the real behavioral patterns of similar users. For example, user C's historical interaction data: (User C, Object ID: Travel 007, Interaction Type: Sharing, Timestamp t12, Tag Score 0.88).

[0211] The preset distribution weights are weight parameters used to balance the importance of different data sources in the query set, and are typically dynamically adjusted based on data quality and quantity. For example, setting the actual subsequent interaction weight λ... real =1.0, generate sample weights λ gen =0.3, similar user replay weight λ replay =0.5.

[0212] One possible way to obtain subsequent interaction data between the target user and subsequent objects is to collect such data from the platform's user behavior logs based on the target user's user identifier. Another possible way is to query and import the subsequent interaction data between the target user and subsequent objects from a third-party authorized data platform based on the target user's user identifier.

[0213] One possible way to obtain similar interaction data between similar users and historical objects is to collect similar interaction data between similar users and historical objects from the platform's user behavior logs based on the user identifiers of similar users. Another possible way is to query and import similar interaction data between similar users and historical objects from a third-party authorized data platform based on the user identifiers of similar users. No specific method is specified here.

[0214] It should be noted that the weight allocation algorithm for multi-source queries, based on a dynamic weight adjustment strategy for sample quality assessment, achieves intelligent fusion of generated samples, real subsequent interactions, and similar user replay data through innovative designs such as quality assessment indicators, weight update rules, and distribution consistency measures. It can dynamically adjust the weight ratio of each data source according to the quality of the generated samples, effectively preventing distribution drift and ensuring the stability and reliability of the system in actual deployment.

[0215] For example, for the target user "User A", two subsequent interaction data were collected:

[0216] (User A, Travel 006, Collection, t11, 0.92);

[0217] (User A, Food004, Share, t12, 0.89);

[0218] By calculating user similarity, we identified a similar user, "User C," and collected three historical interaction data points from him / her.

[0219] (User C, Travel007, Share, t12, 0.88);

[0220] (User C, Food005, Collection, t13, 0.91);

[0221] (User C, Travel 008, In-depth Reading, t14, 0.87);

[0222] Five potential interaction data points were generated. A query set was constructed by sampling from various data sources and combining them according to the weights λ_real = 1.0, λ_gen = 0.3, and λ_replay = 0.5.

[0223] Two data points are sampled from subsequent interaction data, two from potential interaction data (5*0.3≈2), and two from similar interaction data (3*0.5≈2). This results in a hybrid query set containing six interaction data points, encompassing both real future behavior and generated samples and similar user behavior, providing a more comprehensive and diverse set of optimization objectives.

[0224] In the embodiments described in this specification, a query set is constructed through a multi-source data mixing strategy. This effectively utilizes the accuracy of real subsequent interactions, the richness of generated samples, and the diversity of similar user behaviors, providing more comprehensive and balanced sample data for the meta-learning model, thereby further improving the performance of the recommendation system in cold-start scenarios. Simultaneously, through a dynamic weight adjustment mechanism, the mixing ratio can be optimized in real time based on the quality of the generated samples, avoiding distribution drift caused by poor quality from a single data source.

[0225] In one optional embodiment of this specification, the following specific steps are included before step 108:

[0226] The potential interaction distribution is sampled to obtain the sampled potential interaction distribution.

[0227] The potential interaction distribution is sampled to obtain the sampled potential interaction distribution. One possible way is to randomly sample the potential interaction distribution to obtain the sampled potential interaction distribution. Another possible way is to sample the potential interaction distribution by importance to obtain the sampled potential interaction distribution. Yet another possible way is to sample the potential interaction distribution using a Markov chain to obtain the sampled potential interaction distribution. No particular method is specified here.

[0228] For example, the potential interaction distribution N(μ,σ) for target user "User A" 2 Random sampling is used.

[0229] 100 potential vectors z are randomly selected from the distribution, covering the main region of the distribution;

[0230] In the embodiments of this specification, by sampling the potential interaction distribution, a sampled potential interaction distribution is obtained, which can extract representative discrete sample points from the continuous probability distribution, providing diverse potential variable inputs for subsequent interaction data reconstruction, and effectively enhancing the coverage and diversity of the generated samples.

[0231] In one optional embodiment of this specification, the historical interaction data has a corresponding label score, which is a pre-labeled historical interaction data;

[0232] Step 110 includes the following specific steps:

[0233] Label the potential interaction data and obtain the label scores corresponding to the potential interaction data;

[0234] Based on potential interaction data and the corresponding tag scores, construct a query set for the target user.

[0235] The label score corresponding to the historical interaction data is a quantitative value of interest intensity determined based on the user's actual interaction behavior. It is used to characterize the target user's preference for historical objects. Its value range is usually set between 0 and 1. The higher the value, the stronger the interest. As a supervision signal for model training, it directly affects the optimization direction of recommendation ranking.

[0236] The label score corresponding to the potential interaction data is the score of the predicted target user's interest intensity towards the potential object. Its numerical range is consistent with the label score of the historical interaction data. It is used to quantify the quality of the generated interaction data and serve as a supervision signal for meta-learning training.

[0237] To label potential interaction data and obtain corresponding label scores, one option is to directly predict the label scores through the output layer of the conditional generation model. Another option is to calculate the label scores based on the similarity between user representations and object representations. Yet another option is to use a pre-trained preference prediction model to label the generated data with label scores. No specific method is specified here.

[0238] For example, for the generated potential interaction data (user A, travel 004, deep reading, t6): using a pre-trained deep interest network to predict, a label score of 0.86 is obtained.

[0239] In the embodiments described in this specification, by accurately quantifying the label scores of historical interaction data, a reliable supervision signal is provided for the training of the conditional generation model, ensuring that the generated potential interaction data is consistent with the distribution of real data in the dimension of interest intensity, thereby constructing a high-quality query set to optimize the meta-learning model.

[0240] In one optional embodiment of this specification, a query set for the target user is constructed based on potential interaction data and the corresponding tag scores, including the following specific steps:

[0241] Based on the tag scores corresponding to the potential interaction data, the potential interaction data is filtered to obtain the filtering results;

[0242] Based on the filtering results, construct a query set for the target users.

[0243] The filtering results are a subset of high-quality potential interaction data retained after filtering based on label scores. These results have high confidence and representativeness, effectively supporting the optimization process of the meta-learning model. The filtering results typically include complete information such as user identifier, object identifier, interaction type, timestamp, and corresponding label scores.

[0244] The potential interaction data is filtered based on the tag scores corresponding to the potential interaction data to obtain the filtering results. Optionally, at least one of the following implementation methods can be used: filtering the potential interaction data based on the tag scores and score thresholds corresponding to the potential interaction data to obtain the filtering results; sorting the potential interaction data based on the tag scores corresponding to the potential interaction data; and using a Top-K selection strategy to filter the potential interaction data to obtain the filtering results. No limitation is imposed here.

[0245] For example, here are 5 potential interaction data points generated for target user "User A":

[0246] (User A, Travel 004, In-depth Reading, t6, 0.87);

[0247] (User A, Travel 005, Collection, t7, 0.91);

[0248] (User A, Food002, Share, t8, 0.89);

[0249] (User A, Foodie 003, In-depth Reading, t9, 0.84);

[0250] (User A, Technology_002, Clicks, t10, 0.35);

[0251] The three data points with the highest scores are retained (Tourism 005, Food 002, Tourism 004).

[0252] In the embodiments described in this specification, a smart filtering strategy based on label scores can be used to extract high-quality, high-confidence samples from the generated potential interaction data to construct a query set. This ensures the reliability of the generated data and maintains the diversity of interest distribution, providing optimized sample data for the meta-learning model and significantly improving the recommendation effect in cold start scenarios.

[0253] Corresponding to the above-mentioned embodiments in the specification, Figure 2 The diagram illustrates the structure of a conditional generation model in a dataset construction method provided in one embodiment of this specification, as shown below. Figure 2 As shown:

[0254] Input historical interaction data from the target user's support set into the representation extraction layer to extract target user representations in multiple dimensions;

[0255] The target user representation and prior interaction distribution are input into the encoding layer of the conditional generation model. Based on the target user representation, the prior interaction distribution is conditionally mapped to obtain the potential interaction distribution.

[0256] The target user representation and prior interaction distribution are input into the conditional generation model. Based on the target user representation, the interaction data of the potential interaction distribution is reconstructed to generate potential interaction data between the target user and potential objects.

[0257] Based on potential interaction data, construct a query set for the target user.

[0258] Through the above Figure 1 and Figure 2 The dataset construction method shown obtains a query set. Since the query set reflects the true interest distribution of the target users and contains a rich variety of high-quality interaction samples, the object ranking model can be specifically trained based on this query set. This allows the trained target object ranking model to quickly adapt to new user interests and preferences, accurately capture potential interest patterns, and stably output personalized recommendation results. Figure 3 A flowchart of a model training method provided in one embodiment of this specification is shown, including the following specific steps:

[0259] Step 302: Obtain the support set and query set for the target user, wherein the query set is constructed according to the above dataset construction method.

[0260] The embodiments in this specification are applied to training platforms with model training functions. They can be general training platforms or dedicated training platforms for e-commerce platforms, community content platforms, or social media platforms.

[0261] For example, pre-built support set and query set files are read from the distributed storage system of the training platform. The support set file contains complete historical user interaction data, and may also include information such as historical object representations and label scores of historical interaction data. The query set file contains complete potential user interaction data, and may also include information such as potential object representations and label scores of potential interaction data.

[0262] In step 302, by acquiring a high-quality support set and a generated query set, sufficient and balanced sample data is provided for subsequent meta-learning training.

[0263] Step 304: Based on the support set and query set, perform meta-learning training on the object ranking model to obtain a target object ranking model adapted to the target user.

[0264] The object ranking model is a predictive model used for personalized recommendation ranking of candidate objects. The trained target object ranking model is adapted to the target user. The target object recommendation model calculates a prediction score based on the target user's target user representation and the candidate object representations of multiple candidate objects, thereby determining at least one target object. The object ranking model can be a dual-tower model, a deep interest network model, or a Transformer-based sequence recommendation model; no specific limitation is made here.

[0265] The target object ranking model is a personalized recommendation model obtained through meta-learning training. Its model parameters have been specifically optimized, enabling it to quickly adapt to the new user interest characteristics. This model can accurately capture users' potential interest patterns and generate high-quality personalized recommendation lists even with only a small amount of user behavior data.

[0266] Based on the support set and query set, meta-learning training is performed on the object ranking model to obtain a target object ranking model adapted to the target user. One possible approach is to use a model-independent meta-learning framework, perform parameter adaptation in the inner loop on the support set, and perform meta-optimization in the outer loop on the query set. Meta-learning training is performed on the object ranking model through double loop training to obtain a target object ranking model adapted to the target user.

[0267] For example, the training process of the object ranking model for the target user "User A":

[0268] Inner loop adaptation: The initial model weights θ of the dual-tower model are updated in 3 steps using the support set (5 historical interaction data) to obtain the adapted model weights θ'.

[0269] Outer loop optimization: Using a dual-tower model with adapted model weights θ', the meta-loss is calculated on the query set (4 generated high-quality interaction data), and the initial parameters θ are updated through backpropagation.

[0270] Multi-task iteration: Repeat the above process on multiple similar tasks to enable the dual-tower model to learn to quickly adapt to newly registered users. After 5 epochs of training, the target object ranking model is obtained.

[0271] In step 304, through the dual-loop optimization mechanism of the meta-learning framework, the model not only learns general recommendation patterns, but more importantly, it gains the ability to quickly capture users' personalized interests from a small number of samples, thus realizing personalized recommendations.

[0272] In the embodiments described in this specification, by combining the generated high-quality query set and the advanced meta-learning training paradigm, the technical goal of training a target object ranking model with powerful personalized recommendation capabilities using only the sparse support set of the target user is achieved, effectively solving the core problem of recommendation systems in the cold start scenario for new users.

[0273] In one optional embodiment of this specification, historical interaction data has corresponding label scores, and potential interaction data has corresponding label scores;

[0274] Step 304 includes the following specific steps:

[0275] Based on the support set and query set, a meta-learning training is performed on the object ranking model to obtain a target object ranking model adapted to the target user, including:

[0276] Extract the primary target user representation of the target user and the historical object representation of the historical object from the historical interaction data;

[0277] Using an object ranking model, based on the first target user representation and historical object representation, predict the predicted score corresponding to historical interaction data;

[0278] Based on the predicted scores and label scores corresponding to historical interaction data, the support loss is determined.

[0279] Based on the support loss, the weights of the inner model of the object ranking model are adjusted to obtain the first object ranking model.

[0280] Extract the second target user representation of the target user and the potential object representation of the potential object from the potential interaction data;

[0281] Using the first object ranking model, based on the second target user representation and potential object representation, predict the predicted score corresponding to the potential interaction data;

[0282] The meta-loss is determined based on the predicted score and label score corresponding to the potential interaction data.

[0283] Based on the meta-loss, the weights of the outer model of the first object ranking model are adjusted to obtain a target object ranking model adapted to the target user.

[0284] From historical interaction data, the primary target user representation of the target user and the historical object representation of historical objects are extracted. One possible approach is to use a representation extraction layer to extract these representations from the historical interaction data. This representation extraction layer can be an independent feature extraction network, or it can be a user-side representation extraction layer and an object-side representation extraction layer embedded in the object ranking model. This is suitable for... Figure 1According to the embodiments in the specification, the representation extraction layer may include a multi-dimensional representation extraction layer: a preference representation extraction layer and an activity representation extraction layer, which respectively extract the target user's preference representation and the target user's activity representation.

[0285] From potential interaction data, the second target user representation of the target user and the potential object representation of potential objects are extracted. One possible approach is to extract the first target user representation of the target user and the historical object representation of historical objects from historical interaction data through a representation extraction layer. This representation extraction layer can be an independent feature extraction network, or it can be a user-side representation extraction layer and an object-side representation extraction layer embedded in the object ranking model. This is suitable for... Figure 1 According to the embodiments in the specification, the representation extraction layer may include a multi-dimensional representation extraction layer: a preference representation extraction layer and an activity representation extraction layer, which respectively extract the target user's preference representation and the target user's activity representation.

[0286] Support loss is a loss function that measures the accuracy of a model's predictions on the support set. For example, Among them, L support () represents the loss function on the support set, which can be either mean squared error or cross-entropy loss, where α is the inner loop learning rate, and f θ For the model with the current parameters, S u For the support set of the target users, For gradient.

[0287] Meta-loss is a loss function that measures the generalization performance of a model on a query set. For example, meta-loss L... meta =Σ u [λ real ·L(f θu Q real )+λ gen L(f θu Q gen )+λ replay ·L(f θu Q replay )]; Where, λ real , λ gen , λ replay Here are the weight coefficients for each data source, β is the outer loop learning rate, and Q... real Q gen Q replay These represent subsequent interaction data, potential interaction data, and similar interaction data, respectively.

[0288] It should be noted that the end-to-end meta-learning training process, through innovative solutions such as an alternating inner-outer layer optimization training pipeline, gradient calculation order, loss function design, and hyperparameter setting strategies, forms a complete technical solution that can be directly industrialized. This process achieves a complete transformation from sparse support sets to high-quality personalized recommendation models, solving key technical challenges in the practical application of meta-learning in recommendation systems.

[0289] For example, the meta-learning training process for the target user "User A" is as follows:

[0290] Representations are extracted from historical interaction data in the support set. Through a pre-trained embedding layer, the first target user representation and historical object representation are extracted from five historical interaction data.

[0291] The predicted score is calculated using a dual-tower model, and the support loss L is calculated based on the predicted score and the label score. support =MSE(predicted, labeled)=0.0084.

[0292] Inner loop parameter updates, using gradient descent with a learning rate α = 0.01: After the update, the adaptation parameter θ is obtained. u .

[0293] Using the same embedding layer, a second target user representation is extracted from four subsequent interaction data, potential interaction data, and similar interaction data of the query set.

[0294] Using the adaptation parameter θ u The model calculates the predicted score, and based on the predicted score and label score, it calculates the meta-loss:

[0295] Preset weighting coefficient: λ real =0.5, λ gen =0.3, λ replay =0.2.

[0296] L meta =0.5MSE real +0.3MSE gen +0.2*MSE replay =0.5*0.0002+0.3*0.0003+0.2*0.0004=0.00027

[0297] Outer loop parameters are updated, using a learning rate β = 0.001 to update the initial parameters:

[0298] After repeating this process for 5 epochs, the target object ranking model meets the requirements on the test set.

[0299] In the embodiments described in this specification, a two-level optimization approach using support loss and meta-loss enables the model to achieve excellent generalization performance while maintaining personalized adaptability. Support loss ensures that the model can quickly adapt to the specific interest patterns of the target user, while meta-loss guarantees stable performance of the model across different data sources. This training method not only significantly improves recommendation accuracy in cold-start scenarios but also enhances the model's adaptability to changes in user interests.

[0300] Through the above Figure 3 The target object ranking model trained by the model training method shown can quickly adapt to the interest characteristics of new users and accurately capture the potential interest patterns of users because it has been optimized by meta-learning. Therefore, using the target object ranking model for object recommendation can achieve accurate personalized recommendations and effectively improve user experience and recommendation results. Figure 4 A flowchart illustrating an object recommendation method according to an embodiment of this specification is shown, including the following specific steps:

[0301] Step 402: Obtain the target user representation of the target user and the candidate object representation of multiple candidate objects.

[0302] The embodiments in this specification are applied to a recommendation system platform with object recommendation function, which can be an e-commerce platform, a community content platform, or a social media platform.

[0303] Candidate objects are a set of potential options to be recommended to the target user, including all products, services, content, or social media resources that could potentially be recommended on the platform. For example, in an e-commerce platform, candidate objects could be all products sold on the platform; in a community content platform, candidate objects could be all recommendable articles and videos.

[0304] The candidate object representation is a vector representation output by the object representation extraction layer, which encodes the semantic content and attribute features of the candidate object.

[0305] One method for obtaining target user representations is to directly read pre-computed target user representations from a user profile database. Another option is to obtain target user interaction data and extract target user representations from it. Furthermore, another method involves obtaining target user interaction data and extracting target user representations from it using a representation extraction layer. This approach is suitable for... Figure 1 According to the embodiments in the specification, the representation extraction layer may include a multi-dimensional representation extraction layer: a preference representation extraction layer and an activity representation extraction layer, which respectively extract the target user's preference representation and the target user's activity representation.

[0306] To obtain candidate object representations for multiple candidate objects, one option is to read pre-computed candidate object representations in batches from an object feature library, and another option is to compute deep feature representations of candidate objects in real time.

[0307] For example, for the target user "User A", the target user representation is extracted from its 5 most recent interaction records, and the representation vectors of 1000 candidate products are obtained from the product database.

[0308] In step 402, the representations of the target user and candidate objects are obtained, providing the necessary input data for subsequent recommendation calculations.

[0309] Step 404: Using the target object ranking model, based on the target user representation of the target user and the candidate object representations of multiple candidate objects, predict the candidate scores corresponding to the candidate interaction data between the target user and multiple candidate objects. The target object ranking model is trained according to the model training method described above.

[0310] The candidate scores corresponding to the candidate interaction data between the target user and multiple candidate objects are the quantitative values ​​of the user's predicted interest in each candidate object, ranging from 0 to 1. The higher the value, the stronger the interest.

[0311] Based on the target user's target user representation and the candidate object representations of multiple candidate objects, the candidate scores corresponding to the candidate interaction data between the target user and the multiple candidate objects are predicted. One possible approach is to use a dual-tower model to calculate the inner product similarity between the user representation and the object representation. Another possible approach is to use a deep neural network to calculate complex nonlinear interaction features. No specific approach is specified here.

[0312] For example, the representation of user A and the representations of 1,000 candidate products are input into a trained target object ranking model to calculate the predicted score for each product. Among them, the scores of tourism products are generally above 0.8, and the scores of technology products are around 0.3.

[0313] In step 404, batch predictions are performed using the trained model, which can quickly and accurately assess the user's interest in a large number of candidate objects.

[0314] Step 406: Based on the candidate scores, determine at least one target object from multiple candidate objects.

[0315] Based on candidate scores, at least one target object is determined from multiple candidate objects. One possible approach is to use a Top-K strategy to determine at least one target object from multiple candidate objects based on candidate scores. Another possible approach is to determine at least one target object from multiple candidate objects based on candidate scores and a preset score threshold. No specific approach is required here.

[0316] For example, using the Top-10 strategy, the 10 products with the highest scores are selected as the recommendation targets, including 6 travel products and 4 food products, all with scores above 0.85.

[0317] In step 406, an intelligent filtering strategy is used to accurately locate the target object that matches the user's interests from a massive number of candidate objects.

[0318] Step 408: Recommend at least one target object to the target user.

[0319] Recommending at least one target object to a target user can be done in two ways: one is to sort the at least one target object and then recommend it to the target user; the other is to directly recommend at least one target object to the target user. No particular choice is made here.

[0320] For example, 10 target objects are sorted from highest to lowest according to their predicted scores to generate a "You May Like" recommendation list, which is then displayed to user A via an app push notification.

[0321] In step 408, personalized object recommendation services are provided to target users to improve user experience.

[0322] In the embodiments described in this specification, a trained target object ranking model is used to automate the entire process of personalized recommendation. From representation extraction to candidate scoring, target selection, and final recommendation, the whole process is efficient and accurate. It is particularly suitable for solving the object recommendation problem in the cold start scenario for new users, and significantly improves the performance and user experience of the recommendation system.

[0323] Taking community content recommendation as an example, Figure 5 A flowchart illustrating a content recommendation method according to an embodiment of this specification is shown, including the following specific steps:

[0324] Step 502: Obtain the target user representation of the target user and the candidate content representation of multiple candidate content.

[0325] The embodiments in this specification are applied to community content platforms with content recommendation functions.

[0326] Candidate content is a set of potential content to be recommended to the target user, including all content that might be recommended on the platform. For example, in a community content platform, candidate content could be all recommendable articles and videos.

[0327] In step 502, the representations of the target user and candidate content are obtained, providing the necessary input data for subsequent recommendation calculations.

[0328] Step 504: Using the target content ranking model, based on the target user representation of the target user and the candidate content representation of multiple candidate content, predict the candidate scores corresponding to the candidate interaction data between the target user and multiple candidate content. The target content ranking model is trained according to the model training method described above.

[0329] In step 504, batch predictions are performed using the trained model, which can quickly and accurately assess the user's interest in a large number of candidate contents.

[0330] Step 506: Based on the candidate scores, determine at least one target content from multiple candidate contents.

[0331] The target content is one or more recommended content items selected from the candidate content that are most likely to arouse the interest and interaction of the target users.

[0332] In step 506, an intelligent filtering strategy is used to accurately locate target content that matches the user's interests from a massive amount of candidate content.

[0333] Step 508: Recommend at least one piece of target content to the target user.

[0334] In step 508, personalized content recommendation services are provided to target users to improve user experience.

[0335] Figure 6 This specification illustrates a front-end diagram of a content recommendation method according to an embodiment of this specification, such as... Figure 6 As shown:

[0336] The front-end interface is the target user's homepage. The top navigation bar includes buttons to switch between three channels: "Following," "Discover," and "Nearby," used to filter community content from different sources. The central content display area presents four image and text recommendation cards in a vertical flow format, namely:

[0337] The "Create Your Versatile Autumn OOTD" card displays autumn outfit images, titles, and summaries, corresponding to fashion-related community content. The "Explore Hidden Gems in the City" card displays food images, titles, and summaries, corresponding to food-related community content. The "How to Take Cinematic Photos with Your Phone" card displays photography images, titles, and summaries, corresponding to photography-related community content. The "Unmissable Off-the-Beaten-Path Domestic Travel Destinations" card displays travel images, titles, and summaries, corresponding to travel-related community content.

[0338] Each card contains a content image, title text, summary text, and interactive buttons (such as a "Details" button). The bottom navigation bar has five icon buttons: "Home," "Videos," "Publish Notes," "Messages," and "Me," used for in-app navigation. The background is white, the layout is clean and visually clear, facilitating user browsing and interaction. The target user's homepage displays the targeted content generated through the recall model, and the target user can obtain personalized information through channel switching and click actions.

[0339] In the embodiments described in this specification, a trained target content ranking model is used to automate the entire process of personalized recommendation. From representation extraction to candidate scoring, target selection, and final recommendation, the whole process is efficient and accurate. It is particularly suitable for solving the content recommendation problem in the cold start scenario for new users, and significantly improves the performance and user experience of the recommendation system.

[0340] Corresponding to the above method embodiments, this specification also provides recommended system embodiments. Figure 7 A schematic diagram of the structure of a recommendation system provided in one embodiment of this specification is shown. Figure 7 As shown, the recommendation system 700 includes a condition generation module 710;

[0341] The condition generation module 710 is used to obtain the support set of the target user, wherein the support set includes historical interaction data between the target user and historical objects; extract the target user representation of the target user from the historical interaction data; through the condition generation model, based on the target user representation, perform conditional mapping on the prior interaction distribution to obtain the potential interaction distribution, wherein the potential interaction distribution represents the target user's object interaction interests; through the condition generation model, based on the target user representation, reconstruct the interaction data of the potential interaction distribution to generate potential interaction data between the target user and potential objects; and construct the query set of the target user based on the potential interaction data.

[0342] In one optional embodiment of this specification, the recommendation system 700 further includes a subspace decomposition module 720;

[0343] The subspace decomposition module 720 is used to extract the target user's preference representation from historical interaction data through the preference representation extraction layer; and to extract the target user's activity representation from historical interaction data through the activity representation extraction layer, wherein the weight matrices of the preference representation extraction layer and the activity representation extraction layer are orthogonally constrained.

[0344] In one optional embodiment of this specification, the recommendation system 700 further includes a multi-source query fusion module 730;

[0345] The multi-source query fusion module 730 is used to obtain subsequent interaction data between the target user and subsequent objects and similar interaction data between similar users and historical objects. The subsequent interaction data occurs after the historical interaction data, and the similarity between similar users and the target user exceeds a preset threshold. According to the preset distribution weight, the subsequent interaction data, similar interaction data and potential interaction data are mixed to construct the query set of the target user.

[0346] In one optional embodiment of this specification, the recommendation system 700 further includes a meta-learning adaptation module 740;

[0347] The meta-learning adaptation module 740 is used to perform meta-learning training on the object ranking model based on the support set and query set to obtain a target object ranking model adapted to the target user.

[0348] Recommender System 700 is a personalized object recommendation system based on conditional generative models and meta-learning, designed to address the cold start problem for new users. By generating high-quality query sets and undergoing meta-learning training, Recommender System 700 achieves the technical goal of providing accurate personalized object recommendations even with sparse user data. For example, in a community content platform, this system can quickly generate a list of articles and videos that match the interests of newly registered users.

[0349] The condition generation module 710 is the core generation module in the recommender system 700, responsible for generating the query set for the target user through the condition generation model. The condition generation module 710 first obtains the user support set and extracts user representations, then performs conditional mapping on the prior distribution based on the user representations to obtain the potential interaction distribution, and finally reconstructs and generates potential interaction data to construct the query set. For example, this module can generate potential interaction data for community content that the target user is likely to be interested in.

[0350] The subspace decomposition module 720 is the feature processing module in the recommender system 700, responsible for decomposing the user representation into two orthogonal subspaces: preference and activity. The subspace decomposition module 720 extracts user content preference features and activity features through a preference representation extraction layer and an activity representation extraction layer, respectively, and ensures the independence of the two feature spaces through orthogonal constraints. For example, this module can effectively separate the user's preference features for "travel guides" ([0.85, 0.75, 0.1]) from the content consumption activity features ([0.7, 0.3]).

[0351] The multi-source query fusion module 730 is the data fusion module in the recommendation system 700, responsible for intelligently fusing interaction data from multiple sources. The multi-source query fusion module 730 collects real subsequent interaction data, generated potential interaction data, and similar interaction data from similar users, and performs mixed sampling according to preset weights to construct diverse query sets. For example, this module can perform λ...real =1.0, λ gen =0.3, λ replay The ratio of 0.5 integrates users' actual reading records, generated content interaction data, and browsing history of similar users.

[0352] The meta-learning adaptation module 740 is the model training module in the recommender system 700, responsible for training a personalized content recommendation model through a meta-learning framework. The meta-learning adaptation module 740 uses the support set for inner-loop parameter adaptation and the query set for outer-loop meta-optimization, obtaining a target ranking model that quickly adapts to new users through double-loop training. For example, this module can significantly improve the model's performance metrics for community content recommendations within 5 epochs.

[0353] In the embodiments described in this specification, a complete personalized recommendation solution is constructed through the collaborative work of the condition generation module 710, the subspace decomposition module 720, the multi-source query fusion module 730, and the meta-learning adaptation module 740. This system innovatively combines a condition generation model with meta-learning, effectively solving the challenges of recommendation systems in the cold start scenario for new users.

[0354] The above is an illustrative scheme of a recommendation system according to this embodiment. It should be noted that the technical solution of this recommendation system belongs to the same concept as the technical solutions of the dataset construction method and the model training method described above. For details not described in detail in the technical solution of the recommendation system, please refer to the description of the technical solutions of the dataset construction method or the model training method described above.

[0355] Figure 8 This specification illustrates a structural block diagram of a computing device according to one embodiment. The components of the computing device 800 include, but are not limited to, a memory 810 and a processor 820. The processor 820 is connected to the memory 810 via a bus 830, and a database 850 is used to store data.

[0356] The computing device 800 also includes an access device 840, which enables the computing device 800 to communicate via one or more networks 860. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 840 may include one or more of any type of wired or wireless network interface (e.g., Network Interface Controller (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or Near Field Communication (NFC).

[0357] In one embodiment of this specification, the above-described components of the computing device 800 and Figure 8 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 8 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0358] The computing device 800 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 800 can also be a mobile or stationary server.

[0359] The processor 820 is used to execute the following computer program / instruction, which, when executed by the processor, implements the steps of the above-mentioned dataset construction method, model training method, object recommendation method, or content recommendation method.

[0360] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device belongs to the same concept as the technical solutions of the dataset construction method, model training method, object recommendation method, and content recommendation method described above. For details not described in detail in the technical solution of the computing device, please refer to the descriptions of the technical solutions of the dataset construction method, model training method, object recommendation method, or content recommendation method described above.

[0361] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the above-described dataset construction method, model training method, object recommendation method, or content recommendation method.

[0362] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solutions of the dataset construction method, model training method, object recommendation method, and content recommendation method described above. Details not described in detail in the technical solution of the storage medium can be found in the descriptions of the technical solutions of the dataset construction method, model training method, object recommendation method, or content recommendation method described above.

[0363] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described dataset construction method, model training method, object recommendation method, or content recommendation method.

[0364] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product belongs to the same concept as the technical solutions of the dataset construction method, model training method, object recommendation method, and content recommendation method described above. For details not described in detail in the technical solution of the computer program product, please refer to the descriptions of the technical solutions of the dataset construction method, model training method, object recommendation method, or content recommendation method described above.

[0365] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0366] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content contained in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0367] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0368] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0369] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method of constructing a data set, characterized by, The method comprises: obtaining a support set of a target user, wherein the support set comprises historical interaction data between the target user and a historical object; extracting a target user representation of the target user from the historical interaction data; mapping a prior interaction distribution to a potential interaction distribution based on the target user representation by a conditional generation model, wherein the potential interaction distribution represents an object interaction interest of the target user; reconstructing the potential interaction distribution based on the target user representation by the conditional generation model to generate potential interaction data between the target user and a potential object; constructing a query set of the target user based on the potential interaction data.

2. The method of claim 1, wherein, The extracting a target user representation of the target user from the historical interaction data comprises: extracting a target user representation of the target user from the historical interaction data by a representation extraction layer.

3. The method of claim 2, wherein, The target user representation of the target user comprises a target user preference representation and a target user activity representation. The extracting a target user representation of the target user from the historical interaction data by the representation extraction layer comprises: extracting a target user preference representation of the target user from the historical interaction data by a preference representation extraction layer; extracting a target user activity representation of the target user from the historical interaction data by an activity representation extraction layer, wherein weight matrices of the preference representation extraction layer and the activity representation extraction layer are orthogonally constrained.

4. The method of claim 3, wherein, The training step of the preference representation extraction layer comprises: obtaining sample interaction data between a reference user and a sample object and a label preference degree corresponding to the sample interaction data, wherein the label preference degree is determined and labeled based on whether the sample interaction data is factual interaction; extracting a sample user preference representation of the sample user from the sample interaction data by the preference representation extraction layer; extracting a sample object representation from the sample interaction data by an object representation extraction layer; calculating a predicted preference degree corresponding to the sample interaction data based on the sample user preference representation and the sample object representation; determining a contrast loss based on the predicted preference degree and the label preference degree corresponding to the sample interaction data; adjusting parameter weights of the preference representation extraction layer based on the contrast loss until a trained preference representation extraction layer is obtained.

5. The method of claim 3, wherein, The training step of the activity representation extraction layer comprises: obtaining sample interaction data between a reference user and a sample object and a label activity degree corresponding to the sample interaction data, wherein the label activity degree is determined and labeled based on an interaction frequency between the reference user and the sample object; extracting a sample user activity representation of the sample user from the sample interaction data by the activity representation extraction layer; predicting a predicted activity degree corresponding to the sample interaction data based on the sample user activity representation by an activity prediction layer; determining a self-recurrence supervision loss based on the predicted activity degree and the label activity degree; Based on the autoregressive supervision loss, the parameter weight of the activity representation extraction layer is adjusted until the activity representation extraction layer is trained.

6. The method of claim 1, wherein, The training step of the conditional generation model comprises: obtaining sample interaction data between the reference user and a sample object; extracting a reference user representation of the reference user from the sample interaction data; performing conditional mapping on the prior interaction distribution based on the reference user representation through the conditional generation model to obtain a potential interaction distribution, wherein the potential interaction distribution represents the object interaction interest of the reference user; reconstructing the potential interaction distribution based on the reference user representation through the conditional generation model to generate potential interaction data between the reference user and a potential object; determining a reconstruction loss term based on the similarity between the potential interaction data and the sample interaction data; determining a distribution loss term based on the similarity between the potential interaction distribution and the prior interaction distribution; determining a loss value based on the reconstruction loss term and the distribution loss term; adjusting the parameter weight of the conditional generation model based on the loss value until the conditional generation model is trained.

7. The method of claim 1, wherein, Before the query set of the target user is constructed based on the potential interaction data, the method further comprises: obtaining subsequent interaction data between the target user and a subsequent object and similar interaction data between a similar user and a historical object, wherein the subsequent interaction data occurs after the historical interaction data, and the similarity between the similar user and the target user exceeds a preset threshold; constructing the query set of the target user based on the potential interaction data comprises: sampling the subsequent interaction data, the similar interaction data, and the potential interaction data according to a preset distribution weight to construct the query set of the target user.

8. The method according to any one of claims 1 to 7, characterized in that, Before the potential interaction data between the target user and a potential object is generated by reconstructing the potential interaction distribution based on the target user representation through the conditional generation model, the method further comprises: sampling the potential interaction distribution to obtain a sampled potential interaction distribution.

9. The method according to any one of claims 1 to 7, characterized in that, The historical interaction data has a corresponding label score, which is annotated in advance; constructing the query set of the target user based on the potential interaction data comprises: annotating the potential interaction data to obtain a label score corresponding to the potential interaction data; constructing the query set of the target user based on the potential interaction data and the label score corresponding to the potential interaction data.

10. The method of claim 9, wherein, constructing the query set of the target user based on the potential interaction data and the label score corresponding to the potential interaction data comprises: filtering the potential interaction data based on the label score corresponding to the potential interaction data to obtain a filtering result; constructing the query set of the target user based on the filtering result.

11. A model training method, comprising: comprises: obtaining a support set and a query set of a target user, wherein the query set is constructed according to the method of any one of claims 1-10; Based on the support set and the query set, the object ranking model is meta-learning trained to obtain a target object ranking model adapted to the target user.

12. The method of claim 11, wherein, The historical interaction data has a corresponding label score, and the potential interaction data has a corresponding label score; The meta-learning training of the object ranking model based on the support set and the query set to obtain the target object ranking model adapted to the target user comprises: From the historical interaction data, a first target user representation of the target user and a historical object representation of a historical object are extracted; Through the object ranking model, a predicted score corresponding to the historical interaction data is predicted based on the first target user representation and the historical object representation; Based on the predicted score and the label score corresponding to the historical interaction data, a support loss is determined; Based on the support loss, the inner model weight of the object ranking model is adjusted to obtain a first object ranking model; From the potential interaction data, a second target user representation of the target user and a potential object representation of a potential object are extracted; Through the first object ranking model, a predicted score corresponding to the potential interaction data is predicted based on the second target user representation and the potential object representation; Based on the predicted score and the label score corresponding to the potential interaction data, a meta-loss is determined; Based on the meta-loss, the outer model weight of the first object ranking model is adjusted to obtain a target object ranking model adapted to the target user.

13. An object recommendation method characterized by comprising: Comprising: Obtaining a target user representation of a target user and a candidate object representation of a plurality of candidate objects; Through the target object ranking model, a candidate score corresponding to candidate interaction data between the target user and the plurality of candidate objects is predicted based on the target user representation of the target user and the candidate object representation of the plurality of candidate objects, wherein the target object ranking model is trained according to the method of claim 11 or 12; Based on the candidate score, at least one target object is determined from the plurality of candidate objects; The at least one target object is recommended to the target user.

14. A content recommendation method characterized by, Comprising: Obtaining a target user representation of a target user and a candidate content representation of a plurality of candidate contents; Through the target content ranking model, a candidate score corresponding to candidate interaction data between the target user and the plurality of candidate contents is predicted based on the target user representation of the target user and the candidate content representation of the plurality of candidate contents, wherein the target content ranking model is trained according to the method of claim 11 or 12; Based on the candidate score, at least one target content is determined from the plurality of candidate contents; The at least one target content is recommended to the target user.

15. A recommendation system characterized in that, Comprising a condition generation module; The condition generation module is configured to: obtain a support set of a target user, wherein the support set comprises historical interaction data between the target user and a historical object; extract a target user representation of the target user from the historical interaction data; perform conditional mapping on a prior interaction distribution based on the target user representation by using a condition generation model, to obtain a potential interaction distribution, wherein the potential interaction distribution represents an object interaction interest of the target user; perform interaction data reconstruction on the potential interaction distribution based on the target user representation by using the condition generation model, to generate potential interaction data between the target user and a potential object; and construct a query set of the target user based on the potential interaction data.

16. The system of claim 15, wherein, The subspace decomposition module is further configured to: The subspace decomposition module is configured to: extract a target user preference representation of the target user from the historical interaction data by using a preference representation extraction layer; and extract a target user activity representation of the target user from the historical interaction data by using an activity representation extraction layer, wherein weight matrices of the preference representation extraction layer and the activity representation extraction layer are subject to orthogonal constraints.

17. The system of claim 15, wherein, The multi-source query fusion module is further configured to: The multi-source query fusion module is configured to: obtain subsequent interaction data between the target user and a subsequent object, and similar interaction data between a similar user and the historical object, wherein the subsequent interaction data occurs after the historical interaction data, and a similarity between the similar user and the target user exceeds a preset threshold; and sample the subsequent interaction data, the similar interaction data, and the potential interaction data according to a preset distribution weight, to construct the query set of the target user.

18. The system of any of claims 15-17, wherein, The meta-learning adaptation module is further configured to: The meta-learning adaptation module is configured to: perform meta-learning training on an object ranking model based on the support set and the query set, to obtain a target object ranking model adapted to the target user.

19. A computing device, comprising: The apparatus comprises: The memory is configured to store computer programs / instructions, and the processor is configured to execute the computer programs / instructions, so as to implement the steps of the method according to any one of claims 1 to 14. The apparatus has computer programs / instructions stored therein, and the computer programs / instructions are configured to be executed by the processor, so as to implement the steps of the method according to any one of claims 1 to 14.

20. A computer-readable storage medium, characterized in that, The apparatus comprises computer programs / instructions, and the computer programs / instructions are configured to be executed by the processor, so as to implement the steps of the method according to any one of claims 1 to 14.

21. A computer program product, characterised in that, ​

Images