Information migration method and apparatus
By identifying and utilizing dependencies between multiple tasks, especially similarity, complementarity, and causal relationships, and using methods such as Bayesian networks to calculate conditional probabilities and mutual information, the information transfer problem in multi-task models is solved, achieving more efficient and flexible information transfer and improving the model's learning efficiency and applicability.
Patent Information
- Application Number
- PCT/CN2025/144321
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-07
- Filing Date
- 2025-12-22
- Publication Date
- 2026-11-12
AI Technical Summary
In existing technologies, the inter-task information transfer methods in multi-task models limit the flexibility and wide applicability of the models, affecting learning efficiency and performance.
By obtaining the dependencies between multiple tasks, using graph networks such as Bayesian networks to represent the similarity, complementarity, or causal relationships between tasks, calculating conditional probabilities and mutual information, and obtaining task weights, efficient information transfer can be achieved.
It improves the model's learning efficiency and performance, increases the model's flexibility and broad applicability, and ensures the adaptability and accuracy of information transfer.
Smart Images

Figure CN2025144321_12112026_PF_FP_ABST
Abstract
Description
Information migration method and apparatus Technical Field
[0001] This application relates to artificial intelligence (AI) technology, and more particularly to an information transfer method and apparatus. Background Technology
[0002] Neural network transfer learning involves transferring a neural network model trained on one dataset to another with some adjustments. Transfer learning (TL) is a machine learning (ML) technique that transfers the parameters of a trained model to a new model to aid in its training. Given that most tasks are related, transfer learning can share learned model parameters (or the knowledge the model has acquired) with a new model in some way. For example, if a model can recognize images of dogs, a smaller image set can be used to train a new model that recognizes cats, highlighting only the feature differences between dogs and cats. This can accelerate and optimize the learning efficiency of the new model.
[0003] Integrating the tasks of multiple models results in a multi-task model. As the name suggests, a multi-task model can handle multiple tasks simultaneously, enabling information sharing between tasks and establishing sequential dependencies between tasks in a pre-defined manner, thereby facilitating information transfer between tasks. However, this method of information transfer between tasks limits the model's flexibility and broad applicability. Summary of the Invention
[0004] This application provides an information transfer method and apparatus to increase the flexibility and applicability of a model, and improve the model's learning efficiency and performance.
[0005] In a first aspect, this application provides an information migration method, comprising: obtaining a dataset, the dataset being used to describe user interaction behaviors corresponding to multiple tasks; obtaining the dependencies between the multiple tasks based on the dataset; and completing the information migration between the multiple tasks based on the dependencies between the multiple tasks.
[0006] This application obtains the dependencies between multiple tasks by describing a dataset of multiple tasks, and realizes efficient information transfer between tasks based on these dependencies, thereby improving the learning efficiency and performance of the model. In addition, the dependencies based on the task dataset can also increase the flexibility and wide applicability of the model.
[0007] In this application, a dataset X,Y = [(x1,y],y] can be formed from samples of all user interaction behaviors in a specific application scenario (e.g., music streaming service). 1,1,y 1,2 ,y 1,3 ,y 1,4 ),(x2,y 2,1 ,y 2,2 ,y 2,3 ,y 2,4 ),(x3,y 3,1 ,y 3,2 ,y 3,3 ,y 3,4 ),…,(x n ,y n,1 ,y n,2 ,y n,3 ,y n,4 )], where (x i ,y i,1 ,y i,2 ,y i,3 ,y i,4 ) represents a sample, i∈[1,n], and there are n samples in this dataset; the label can represent different tasks, such as download, conversion, payment, activation, etc. i,j Let represent the task data corresponding to the j-th label in the i-th sample, where j∈[1,4]. Therefore, the i-th sample in this dataset contains four task data items, namely y i,1 y i,2 y i,3 y i,4 It should be noted that the dataset obtained in this application adopts corresponding expression methods and data content based on the application scenario. Therefore, there can be multiple ways to represent the dataset, and this application does not make any specific limitations on this.
[0008] In one possible implementation, the potential relationships between multiple tasks are first obtained based on the dataset, which include at least one of similarity, complementarity, or causality; then the dependencies between multiple tasks are obtained based on the potential relationships.
[0009] This application can construct a graph network based on data labels. By analyzing the label information in the aforementioned dataset, it can identify potential relationships between different tasks. These potential relationships can include at least one of similarity, complementarity, or causal relationship between tasks. For example, using the aforementioned dataset as an example, it identifies potential relationships between two tasks, such as identifying the potential relationship between tasks corresponding to labels j=1 and j=2, where the task data corresponding to label j=1 includes {y... 1,1 ,y 2,1 ,y 3,1 ,…,y n,1 The task data corresponding to the label j=2 includes {y}. 1,2 ,y 2,2 ,y 3,2 ,…,yn,2 By analyzing the similarity, complementarity, or causal relationship between the two task datasets, the potential relationship between the two tasks can be obtained. For example, the similarity between two tasks can be calculated by examining the variance of the datasets to obtain their errors. The smaller the error, the higher the similarity between the two tasks. In this case, information from one task (knowledge of the source domain) can help the learning of the other task (learning the target domain), thus achieving high-quality information transfer. Similarly, the complementarity between two tasks can be obtained through the data distribution or modality in the datasets. In this case, information from one task can supplement the other, improving model robustness, thus achieving high-quality information transfer. Furthermore, the causal relationship between two tasks can be obtained through function call relationships or the order of function calls in the datasets. In this case, information from the preceding task is very useful for the execution of the following task, thus achieving high-quality information transfer.
[0010] Optionally, the dependencies between multiple tasks can be represented using a Bayesian relational network. Based on the aforementioned latent relationships, the dependencies of task keys can be obtained. For example, based on the causal relationship between task A and task B, it is known that the completion of task A contributes to the success of task B. Therefore, a dependency relationship between task B and task A can be established, thus creating an arrow from task A to task B in the graph network. For instance, if the registration task occurs 10 times and the payment task occurs 5 times, and payment depends on registration, then the weights between the two tasks can be calculated based on their occurrence probabilities.
[0011] A Bayesian network is a graph model based on probabilistic reasoning used to represent dependencies between variables. A Bayesian network consists of a directed acyclic graph (DAG) and a conditional probability table. The DAG consists of nodes (variables) and arrowed edges (dependencies), with no loops. The conditional probability table stores the probability distribution for each node given its parent node.
[0012] It should be noted that, in addition to Bayesian networks, this application can also use other graph networks to represent the dependencies between multiple tasks, and no specific limitations are imposed on this.
[0013] As described above, this application establishes dependencies between multiple tasks based on a dataset. These dependencies can reflect at least one of the similarity, complementarity, or causal relationship between tasks, thereby achieving efficient information transfer between tasks, improving the learning efficiency and performance of the model, and eliminating the need for manually setting dependencies between tasks. Dependencies based on task datasets can increase the flexibility and wide applicability of the model.
[0014] In this application, the weights of multiple tasks in information migration can be obtained first based on the dependencies between multiple tasks; then, the information migration between multiple tasks can be completed based on the weights of multiple tasks.
[0015] In one possible implementation, the conditional probabilities and / or mutual information between multiple tasks can be obtained based on the dependencies between them. The conditional probabilities represent the probability that one task is in a specific state and the other task is in a specific state among any two tasks with dependencies. The mutual information represents the degree of dependency between any two tasks with dependencies. The weights of the multiple tasks are obtained based on the conditional probabilities and / or mutual information between them.
[0016] Conditional probability is defined as follows: Given two tasks A and B, the probability that task A will occur given that task B will occur is called conditional probability, denoted as P(A|B). Its formula is: P(A|B) = P(A∩B) / P(B)
[0017] Where P(A∩B) represents the probability that task A and task B occur simultaneously, and P(B) represents the probability that task B occurs.
[0018] The dependency relationship between Task A and Task B means that the two tasks have at least one potential relationship, such as high similarity, mutual complementarity, or causal relationship. In terms of conditional probability table, it means that the conditional probability of the two tasks is not 0. In terms of graph network, it means that there is an edge with an arrow between the nodes corresponding to the two tasks to represent their dependency relationship.
[0019] A specific state can refer to the occurrence of a task. For example, if task B occurs and task A occurs, then the conditional probability is used to represent the probability that task A occurs when task B occurs, given that tasks A and B have a dependency relationship. Alternatively, a specific state can refer to the completion of a task. For example, if task A is completed and task B is completed, then the conditional probability is used to represent the probability that task B is completed when task A is completed, given that tasks A and task B have a dependency relationship. It should be noted that a specific state can also include other task states, but this application does not specifically limit this.
[0020] Weights reflect the dependencies between tasks, enabling more adaptive and accurate information transfer. For example, when processing multiple parallel tasks, these tasks can be intelligently scheduled based on weights to ensure that all tasks are executed sequentially according to the correct process, thereby improving overall efficiency and the accuracy of task completion.
[0021] In one possible implementation, the input vector of the second task can be obtained based on the weights and output vector of the first task, and the input vector of the second task can be processed by a multilayer perceptron (MLP) to obtain the output vector of the second task. The first task is the preceding task of the second task.
[0022] Optionally, information transfer includes feature vector transfer between multiple tasks. In feature vector transfer, an improved method using vector function units (VFUs) can be employed. That is, a multi-layer perceptron (MLP) is used as the backbone network, taking the output vector of the preceding task (e.g., task A above) as the input vector of the subsequent task (e.g., task B above). After processing by the MLP, the output vector of the subsequent task is obtained. The output vector of the subsequent task captures the deep features of multiple tasks, which helps to improve information transfer. For example, in natural language processing tasks, the output vector of the preceding task can contain lexical, syntactic, and semantic information, which is crucial for the execution of the subsequent task. In this way, the effectiveness of information transfer can be improved, making information transfer between tasks more efficient and accurate.
[0023] In one possible implementation, the predicted value of the second task can be obtained based on the weight of the first task and the predicted value of the first task, where the first task is the preceding task of the second task.
[0024] Optionally, information transfer includes the transfer of predicted values between multiple tasks. This transfer can be performed using predicted values (Logits information). That is, based on the constructed Bayesian network, the conditional probabilities and mutual information between tasks are calculated. Then, the task weights are used as the basis for weighting. Logits are a centralized representation of task information, containing deep features and predicted information. As an intermediate result of model prediction, they are often used for subsequent probability transformations or knowledge transfer. Using Logits for transfer enables more stable and efficient transfer. For example, in image classification tasks, Logits can represent the deep features and classification probabilities of images. This information can be transferred to subsequent image subclass classification tasks, utilizing a shared classification probability space to achieve image subclass classification. Similarly, in natural language recognition tasks, Logits can represent the deep features and semantic features of images. This information can be transferred to subsequent speech recognition tasks, utilizing a shared semantic feature space to achieve speech recognition. Weighting measures the strength of dependencies between tasks, making the transfer effect more significant and improving the collaborative efficiency between tasks.
[0025] The key to information transfer in machine learning is to reduce the difference in conditional probability distributions between the source and target domains, thereby achieving effective knowledge transfer. This application utilizes the dependencies between tasks to achieve this goal. For example, Bayesian networks used to represent dependencies between tasks can efficiently compute the conditional probabilities of complex tasks.
[0026] Secondly, this application provides an information migration apparatus, comprising: an acquisition module for acquiring a dataset, the dataset being used to describe user interaction behaviors corresponding to multiple tasks; acquiring the dependency relationships between the multiple tasks based on the dataset; and a processing module for completing the information migration between the multiple tasks based on the dependency relationships between the multiple tasks.
[0027] In one possible implementation, the processing module is specifically configured to obtain the weights of the multiple tasks in the information migration based on the dependencies between the multiple tasks; and to complete the information migration between the multiple tasks based on the weights of the multiple tasks.
[0028] In one possible implementation, the processing module is specifically configured to obtain conditional probabilities and / or mutual information between the multiple tasks based on the dependencies between the multiple tasks. The conditional probability represents the probability that one task is in a specific state and the other task is in the same specific state among any two tasks with dependencies. The mutual information represents the degree of dependency between any two tasks with dependencies. The module also obtains the weights of the multiple tasks based on the conditional probabilities and / or mutual information between the multiple tasks.
[0029] In one possible implementation, the information transfer includes feature vector transfer between the plurality of tasks; or, the information transfer includes prediction value transfer between the plurality of tasks.
[0030] In one possible implementation, the processing module is specifically used to obtain the input vector of the second task based on the weight of the first task and the output vector of the first task, and to obtain the output vector of the second task after processing the input vector of the second task through a multilayer perceptron (MLP), wherein the first task is the preceding task of the second task.
[0031] In one possible implementation, the processing module is specifically used to obtain the predicted value of the second task based on the weight of the first task and the predicted value of the first task, wherein the first task is a preceding task of the second task.
[0032] In one possible implementation, the acquisition module is specifically used to acquire potential relationships between the multiple tasks based on the dataset, the potential relationships including at least one of similarity, complementarity, or causality; and to acquire dependencies between the multiple tasks based on the potential relationships.
[0033] In one possible implementation, the acquisition module is specifically used to identify potential relationships between the multiple tasks by analyzing the label information in the dataset.
[0034] In one possible implementation, the dependencies between the multiple tasks are represented using a Bayesian relational network.
[0035] Thirdly, this application provides a computer device, comprising: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to perform the method as described in any one of the first aspects above.
[0036] Fourthly, this application provides a computer-readable storage medium including a computer program that, when executed on a computer, causes the computer to perform the method described in any one of the first aspects above.
[0037] Fifthly, this application provides a computer program product comprising computer program code, which, when run on a computer, causes the computer to perform the method described in any one of the first aspects above. Attached Figure Description
[0038] Figure 1 is a schematic block diagram of a device 100 according to this application;
[0039] Figure 2 is a flowchart of process 200 of the information migration method provided in an embodiment of this application;
[0040] Figures 3a and 3b are schematic diagrams of Bayesian networks;
[0041] Figure 4 is a schematic diagram of information migration;
[0042] Figure 5a is a schematic diagram of Huawei App Market;
[0043] Figure 5b is a schematic diagram of the Huawei Mall;
[0044] Figure 5c is a schematic diagram of Huawei Music;
[0045] Figure 6 is a structural schematic diagram of the information migration device 600 of this application. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0047] The terms "first," "second," etc., used in the specification, embodiments, claims, and drawings of this application are for distinguishing purposes only and should not be construed as indicating or implying relative importance or order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as including a series of steps or units. A method, system, product, or apparatus is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.
[0048] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.
[0049] Figure 1 is a schematic block diagram of a device 100 according to the present application. As shown in Figure 1, the device 100 may include a processor 101 and a transceiver / transceiver pin 102, and optionally, a memory 103.
[0050] The various components of device 100 are coupled together via bus 104, which includes a data bus, a power bus, a control bus, and a status signal bus. For clarity, all buses are referred to as bus 104 in the figure.
[0051] Optionally, the memory 103 can be used to store instructions in the embodiments of this application. The processor 101 can be used to execute the instructions in the memory 103, control the receive pin to receive signals, and control the transmit pin to transmit signals.
[0052] Device 100 may be a computer device or a chip of a computer device as described in the embodiments of this application.
[0053] All relevant content of each step involved in the embodiments of this application can be referenced from the functional description of the corresponding functional module, and can be referred to the embodiments below.
[0054] Since this application involves artificial intelligence (AI) technology, for ease of understanding, some AI-related terms or concepts used in this application are explained below, and these terms or concepts are also part of the content of the invention.
[0055] 1. Transfer learning in machine learning
[0056] Transfer learning in machine learning is a learning method that utilizes knowledge from a source domain to aid in learning tasks in a target domain. Transfer learning plays a crucial role in machine learning, especially in addressing domain adaptation problems.
[0057] The goal of transfer learning is to learn a classifier from labeled data in the source domain to predict labels in the target domain, given a certain degree of similarity between the source and target domains (despite potential differences in marginal probability distributions). This typically involves increasing the similarity between the two domains and quantitatively measuring this similarity. Transfer learning enables machine learning models to leverage previously learned knowledge when faced with new tasks, thereby accelerating the learning process and improving the model's generalization ability. It has significant value in many practical applications, such as natural language processing, computer vision, and recommender systems.
[0058] 2. Conditional probability in machine learning
[0059] Conditional probability describes the likelihood of a task occurring under a specific condition. It is widely used in various machine learning tasks, such as classification, regression, and clustering. The definition of conditional probability is: given two tasks A and B, the probability that task A will occur given that task B will occur is called the conditional probability, denoted as P(A|B). Its formula is: P(A|B) = P(A∩B) / P(B)
[0060] Where P(A∩B) represents the probability that task A and task B occur simultaneously, and P(B) represents the probability that task B occurs.
[0061] Applications of conditional probability in machine learning:
[0062] (1) Classification task:
[0063] In classification tasks, conditional probability is often used to calculate the probability that a sample belongs to a certain class. For example, in the Naive Bayes classifier, the conditional probability of a sample belonging to each class is calculated, and the class with the highest probability is selected as the prediction result.
[0064] (2) Bayesian networks:
[0065] A Bayesian network is a graphical model based on conditional probability used to represent dependencies between variables. Bayesian networks allow for the efficient computation of conditional probabilities for complex tasks.
[0066] (3) Decision tree:
[0067] Each node in a decision tree represents a conditional decision, and the next node or leaf node is selected based on the result of the conditional decision. Conditional probability plays an important role in the construction and pruning of decision trees.
[0068] (4) Hidden Markov Model:
[0069] Hidden Markov Models (HMMs) are statistical models used to describe time series data, where conditional probabilities are used to represent the probability of observing a particular observation given the current state.
[0070] (5) Reinforcement learning:
[0071] In reinforcement learning, conditional probability is often used to calculate the probability of transitioning to the next state after taking a certain action in a certain state, as well as the probability of obtaining a certain reward.
[0072] Methods for calculating conditional probability:
[0073] (1) Direct calculation: If the probabilities of tasks A and B and their joint probability are known, the definition formula of conditional probability can be used directly for calculation.
[0074] (2) Bayes' theorem: When you need to calculate P(A|B) but only know P(B|A), you can use Bayes' theorem to transform it: P(A|B)=P(B|A)*P(A) / P(B).
[0075] (3) Maximum Likelihood Estimation: In some machine learning scenarios, it may be necessary to estimate conditional probabilities based on observed data. Maximum likelihood estimation is a commonly used method that estimates parameters by maximizing the likelihood function of the observed data.
[0076] (4) Monte Carlo method: For complex problems, the Monte Carlo method can be used to approximate the conditional probability through simulation.
[0077] In summary, conditional probability can be used to understand the dependencies between tasks and make predictions and decisions accordingly.
[0078] 3. Conditional Probability Transfer in Machine Learning
[0079] In machine learning, conditional probability transfer is an important aspect of transfer learning. Conditional probability transfer specifically refers to reducing the distance between the conditional probability distributions of the source and target domains, thereby achieving transfer learning. Conditional distribution adaptation methods approximate the difference between the two domains by calculating the distance between their conditional probability distributions (e.g., P(ys|Xs) and P(yt|Xt)) and attempt to reduce this difference to achieve transfer. This method is particularly important in machine learning because machine learning models typically rely on large amounts of labeled data for training, while in many practical applications, labeled data for the target domain is often very limited. Conditional distribution adaptation can leverage the abundant labeled data in the source domain to improve the performance of tasks in the target domain. Therefore, the key to combining conditional probability transfer in machine learning is to utilize conditional distribution adaptation methods to reduce the difference in conditional probability distributions between the source and target domains, thereby achieving effective knowledge transfer.
[0080] In machine learning, a classic example of conditional probability transfer is image classification in computer vision. Suppose we have a source domain dataset containing a large number of labeled animal images, primarily taken in zoo environments. The goal is to classify images on a target domain dataset containing images of animals taken in the wild, which also need to be classified, but with relatively fewer labeled images. In this example, transfer learning is applied as follows:
[0081] (1) Conditional probability of the source domain:
[0082] In the source domain, an image classification model has been trained that can calculate the conditional probability, P(animal species|zoo image), that a given animal image (feature) in a zoo setting belongs to a certain animal species (e.g., “lion”, “tiger”, etc.).
[0083] (2) Conditional probability of the target domain:
[0084] In the target domain, we want to predict the probability that a given image of an animal in the wild belongs to a certain animal species, i.e., P(animal species|wild image).
[0085] (3) Transfer learning:
[0086] Although animal images from zoo and natural environments may differ significantly in terms of background and lighting conditions, the visual characteristics of the animals themselves (e.g., shape, texture, color, etc.) are generally similar between the source and target domains. Therefore, transfer learning can be used to leverage animal feature representations learned in the source domain to help better estimate the conditional probability P(animal species | wild image) in the target domain.
[0087] (4) Specific procedures:
[0088] Feature-based transfer learning methods can be used to transform the image features of the source and target domains into a common feature space. In this feature space, the animal feature representations of the source and target domain images should be similar, thus allowing the use of a classifier learned in the source domain to classify the target domain image.
[0089] In addition, model-based transfer learning methods, such as fine-tuning, can be considered. In this method, a deep learning model is pre-trained using source domain data, and then fine-tuned on the target domain data. The fine-tuning process preserves the animal feature representations learned in the source domain and appropriately adjusts the model according to the characteristics of the target domain data to improve classification performance in the target domain.
[0090] Therefore, conditional probability transfer can more effectively utilize labeled data in the source domain to improve classification tasks in the target domain.
[0091] 4. Bayesian Network
[0092] Bayesian networks are graph models based on probabilistic reasoning used to represent dependencies between variables and handle uncertainty. The network consists of a directed acyclic graph (DAG) and a conditional probability table. The DAG consists of nodes (variables) and arrowed edges (dependencies), with no loops. The conditional probability table stores the probability distribution for each node given its parent node.
[0093] In transfer learning, structural transfer is achieved by sharing the Bayesian network topology of the source and target domains, parameter transfer is achieved by adjusting the conditional probability table to adapt to the data distribution of the target domain, and adaptive learning is achieved by dynamically adjusting the parameter distribution of the target domain.
[0094] A typical example of applying Bayesian networks to transfer learning is a cross-domain recommender system. The source domain is e-commerce user behavior data, and the target domain is short video content preferences. By transferring the probabilistic dependencies of the "user-item" interaction relationship, the content feature nodes are retrained.
[0095] The information transfer method provided in this application can be applied to recommendation systems, and the application scenarios for these recommendation systems include, but are not limited to:
[0096] 1. E-commerce platforms: In e-commerce websites, recommendation systems can be used for personalized product recommendations to improve user satisfaction and purchase rates.
[0097] 2. Video streaming services: Video platforms can use recommendation systems to recommend content that users are interested in, thereby enhancing the user experience.
[0098] 3. Social Media: Social media platforms can use recommendation systems to recommend news, articles, or users that users may be interested in.
[0099] 4. Music streaming services: Music platforms can use recommendation systems to recommend new songs or music of similar styles to users, improving the user's listening experience.
[0100] 5. Educational Platforms: Educational platforms can use recommendation systems to recommend relevant learning resources to users, helping them improve their learning efficiency.
[0101] In addition, recommendation systems can be used in other scenarios to provide users with accurate and personalized recommendation services, and this application does not make any specific limitations on this.
[0102] The aforementioned recommendation system can be implemented in several forms. One form is as Software as a Service (SaaS), allowing users to subscribe and use the system via the internet without having to build and maintain a complex recommendation system infrastructure themselves. Another form is as a customized solution, developed to meet specific user needs and providing a recommendation system more suited to their business requirements. Furthermore, key components or modules of the recommendation system can be provided to users as libraries for integration into their applications. It should be understood that the recommendation system can also be provided to users in other forms, which this application does not specifically limit.
[0103] Based on the above content, the technical solution of the information migration method provided in this application will be described below.
[0104] Figure 2 is a flowchart of process 200 of the information migration method provided in an embodiment of this application. Process 200 can be executed by the apparatus 100 described above. Process 200 is described as a series of steps or operations. It should be understood that process 200 can be executed in various orders and / or occur simultaneously, and is not limited to the execution order shown in Figure 2. Process 200 may include:
[0105] Step 201: Obtain the dataset, which describes the user interaction behavior corresponding to multiple tasks.
[0106] In this application, a dataset X,Y = [(x1,y],y] can be formed from samples of all user interaction behaviors in a specific application scenario (e.g., music streaming service). 1,1 ,y 1,2 ,y 1,3 ,y 1,4 ),(x2,y 2,1 ,y 2,2 ,y 2,3 ,y 2,4 ),(x3,y 3,1 ,y 3,2 ,y 3,3 ,y 3,4 ),…,(x n ,y n,1 ,y n,2 ,y n,3 ,y n,4 )], where (x i ,y i,1 ,y i,2 ,y i,3 ,y i,4 ) represents a sample, i∈[1,n], and there are n samples in this dataset; the label can represent different tasks, such as download, conversion, payment, activation, etc. i,j Let represent the task data corresponding to the j-th label in the i-th sample, where j∈[1,4]. Therefore, the i-th sample in this dataset contains four task data items, namely y i,1 y i,2 y i,3 y i,4 .
[0107] It should be noted that the dataset obtained in this application adopts corresponding expression methods and data content based on the application scenario. Therefore, the above dataset does not constitute a limitation, and this application does not make any specific limitations in this regard.
[0108] Step 202: Obtain the dependencies between multiple tasks based on the dataset.
[0109] In one possible implementation, the potential relationships between multiple tasks are first obtained based on the dataset, which include at least one of similarity, complementarity, or causality; then the dependencies between multiple tasks are obtained based on the potential relationships.
[0110] This application can construct a graph network based on data labels. By analyzing the label information in the aforementioned dataset, it can identify potential relationships between different tasks. These potential relationships can include at least one of similarity, complementarity, or causal relationship between tasks. For example, using the aforementioned dataset as an example, it identifies potential relationships between two tasks, such as identifying the potential relationship between tasks corresponding to labels j=1 and j=2, where the task data corresponding to label j=1 includes {y... 1,1 ,y 2,1 ,y 3,1 ,…,y n,1 The task data corresponding to the label j=2 includes {y}. 1,2 ,y 2,2 ,y 3,2 ,…,y n,2 By analyzing the similarity, complementarity, or causal relationship between the two task datasets, the potential relationship between the two tasks can be obtained. For example, the similarity between two tasks can be calculated by examining the variance of the datasets to obtain their errors. The smaller the error, the higher the similarity between the two tasks. In this case, information from one task (knowledge of the source domain) can help the learning of the other task (learning the target domain), thus achieving high-quality information transfer. Similarly, the complementarity between two tasks can be obtained through the data distribution or modality in the datasets. In this case, information from one task can supplement the other, improving model robustness, thus achieving high-quality information transfer. Furthermore, the causal relationship between two tasks can be obtained through function call relationships or the order of function calls in the datasets. In this case, information from the preceding task is very useful for the execution of the following task, thus achieving high-quality information transfer.
[0111] Optionally, the dependencies between multiple tasks can be represented using a Bayesian relational network. Based on the aforementioned latent relationships, the dependencies of task keys can be obtained. For example, based on the causal relationship between task A and task B, it is known that the completion of task A contributes to the success of task B. Therefore, a dependency relationship between task B and task A can be established, thus creating an arrow from task A to task B in the Bayesian network. For instance, if the registration task occurs 10 times and the payment task occurs 5 times, and payment depends on registration, then the weights between the two tasks can be calculated based on their occurrence probabilities.
[0112] Referring to the introduction of AI technology above, a Bayesian network is a graph model based on probabilistic reasoning used to represent dependencies between variables. A Bayesian network consists of a directed acyclic graph (DAG) and a conditional probability table. The DAG consists of nodes (variables) and arrowed edges (dependencies), with no loops. The conditional probability table stores the probability distribution for each node given its parent node. For example, Figure 3a is a schematic diagram of a Bayesian network. As shown in Figure 3a, the Bayesian network describes six tasks (task-1—task-6). The conditional probability of tasks-1, task-4, and task-6 is 1, while the conditional probability of tasks-2, task-3, and task-5 is 0. This indicates that there are dependencies between tasks-1, task-4, and task-6; that is, task-4 depends on task-1, and task-6 depends on both tasks-1 and task-4.
[0113] It should be noted that, in addition to Bayesian networks, this application can also use other graph networks to represent the dependencies between multiple tasks, and no specific limitations are imposed on this.
[0114] As described above, this application establishes dependencies between multiple tasks based on a dataset. These dependencies can reflect at least one of the similarity, complementarity, or causal relationship between tasks, thereby achieving efficient information transfer between tasks, improving the learning efficiency and performance of the model, and eliminating the need for manually setting dependencies between tasks. Dependencies based on task datasets can increase the flexibility and wide applicability of the model.
[0115] Step 203: Complete the information migration between multiple tasks based on the dependencies between them.
[0116] In this application, the weights of multiple tasks in information migration can be obtained first based on the dependencies between multiple tasks; then, the information migration between multiple tasks can be completed based on the weights of multiple tasks.
[0117] In one possible implementation, the conditional probabilities and / or mutual information between multiple tasks can be obtained based on the dependencies between them. The conditional probabilities represent the probability that one task is in a specific state and the other task is in a specific state among any two tasks with dependencies. The mutual information represents the degree of dependency between any two tasks with dependencies. The weights of the multiple tasks are obtained based on the conditional probabilities and / or mutual information between them.
[0118] Conditional probability is defined as follows: Given two tasks A and B, the probability that task A will occur given that task B will occur is called conditional probability, denoted as P(A|B). Its formula is: P(A|B) = P(A∩B) / P(B)
[0119] Where P(A∩B) represents the probability that task A and task B occur simultaneously, and P(B) represents the probability that task B occurs.
[0120] The dependency relationship between tasks A and B means that the two tasks have at least one potential relationship, such as high similarity, complementarity, or causality. Represented by a conditional probability table, this means the conditional probability of the two tasks is not zero. In a graph network, it means there is an edge with an arrow between the nodes corresponding to the two tasks, indicating their dependency. For example, in Figure 3a, tasks-1,-4, and-6 have dependencies, where task-4 depends on task-1, and task-6 depends on both task-1 and task-4. As another example, Figure 3b is a schematic diagram of a Bayesian network. As shown in Figure 3b, a sample contains five tasks: Activation (ACT), Registration (Regist), Engagement (Engage), Retention (ReTent), and Payment (Pay). Dependencies between tasks can be identified based on their co-occurrence relationships. Activation (ACT), retention (ReTent), and payment (Pay) occur simultaneously, therefore there is an edge from activation to retention, an edge from activation to payment, and an edge from retention to payment. Each node in ACT, Retent, and Payment has a self-loop. Statistical analysis of all samples constitutes a Bayesian graph. Based on this Bayesian graph, mutual information, conditional probabilities, and weights can be calculated.
[0121] A specific state can refer to the occurrence of a task. For example, if task B occurs and task A occurs, then the conditional probability is used to represent the probability that task A occurs when task B occurs, given that tasks A and B have a dependency relationship. Alternatively, a specific state can refer to the completion of a task. For example, if task A is completed and task B is completed, then the conditional probability is used to represent the probability that task B is completed when task A is completed, given that tasks A and task B have a dependency relationship. It should be noted that a specific state can also include other task states, but this application does not specifically limit this.
[0122] Weights reflect the dependencies between tasks, enabling more adaptive and accurate information transfer. For example, when processing multiple parallel tasks, these tasks can be intelligently scheduled based on weights to ensure that all tasks are executed sequentially according to the correct process, thereby improving overall efficiency and task completion accuracy. For instance, in this application, mutual information can be calculated based on the occurrence probabilities and conditional probabilities of upstream task S and downstream task K. The weights for transfer between two tasks are the mutual information between them; that is, once the mutual information is determined, the weights are also determined. To measure the strength of the dependency between two tasks, their mutual information can be calculated as: MI(K|S) = H(K) - H(K|S)
[0123] Where H(K) is the information entropy of task K; H(K|S) is the conditional entropy, which is the weighted average of the entropies of the target task based on task S. The formulas for calculating the aforementioned two can be expressed as:
[0124] Where k represents different values of K; s represents different values of S. For a binary classification task, k∈{0,1}, s∈{0,1}. Accordingly, the mutual information set of task K is represented as: MI S→K ={MI(K|S)}
[0125] Furthermore, the weight of information transfer from upstream task S to downstream task K is: W S→K =MI S→K
[0126] In one possible implementation, the input vector of the second task can be obtained based on the weights and output vector of the first task, and the input vector of the second task can be processed by a multilayer perceptron (MLP) to obtain the output vector of the second task. The first task is the preceding task of the second task.
[0127] Optionally, information transfer includes feature vector transfer between multiple tasks. In feature vector transfer, a vector function unit (VFU) improvement method can be employed. That is, a multi-layer perceptron (MLP) is used as the backbone network, and the output vector of the preceding task (e.g., task A above) is used as the input vector of the subsequent task (e.g., task B above). After passing through the MLP, the output vector of the subsequent task is obtained. The output vector of the subsequent task captures the deep features of multiple tasks, which helps to better transfer information. When using the VFU improvement method in feature vector transfer, the core idea is to achieve optimized transfer of cross-task feature representation through MLP. The MLP in VFU can adjust the feature distribution through nonlinear mapping, reducing the domain difference between the preceding task (task A) and the subsequent task (task B). For example, in cross-sensor or cross-device scenarios, MLP can alleviate the covariate shift problem between the source and target domains, reducing the difference in their marginal distributions. MLP is a feedforward artificial neural network composed of multiple neuron layers, capable of solving complex pattern recognition problems through nonlinear mapping. MLPs overcome the limitation of single-layer perceptrons, which can only handle linearly separable tasks, by introducing hidden layers and activation functions to model nonlinear relationships. For example, in sequence recommendation, a user's historical behavior (such as browsing and purchase records) is encoded into a low-dimensional dense vector by an MLP, and the influence weight of key behaviors is strengthened by an attention mechanism. The multi-layer nonlinear activation structure of MLPs can enhance the discriminative power of features. For example, in image stitching tasks, by optimizing the feature generation process, the accuracy of feature matching is improved. Therefore, MLPs can denoise or compress the feature dimension of the output vector of task A, making it more suitable for the input requirements of task B. For example, in natural language processing tasks, the output vector of a preceding task can contain lexical, syntactic, and semantic information, which is important for the execution of subsequent tasks. In this way, the effect of information transfer can be improved, making information transfer between tasks more efficient and accurate.
[0128] In one possible implementation, the predicted value of the second task can be obtained based on the weight of the first task and the predicted value of the first task, where the first task is the preceding task of the second task.
[0129] Optionally, information transfer includes the transfer of predicted values between multiple tasks. This transfer can be performed using predicted values (Logits information). That is, based on the constructed Bayesian network, the conditional probabilities and mutual information between tasks are calculated. Then, the task weights are used as the basis for weighting. Logits are a centralized representation of task information, containing deep features and predicted information. As an intermediate result of model prediction, they are often used for subsequent probability transformations or knowledge transfer. Using Logits for transfer enables more stable and efficient transfer. For example, in image classification tasks, Logits can represent the deep features and classification probabilities of images. This information can be transferred to subsequent image subclass classification tasks, utilizing a shared classification probability space to achieve image subclass classification. Similarly, in natural language recognition tasks, Logits can represent the deep features and semantic features of images. This information can be transferred to subsequent speech recognition tasks, utilizing a shared semantic feature space to achieve speech recognition. Weighting measures the strength of dependencies between tasks, making the transfer effect more significant and improving the collaborative efficiency between tasks. The core idea of prediction transfer is to transfer predicted values from the first task to the second task, achieving knowledge reuse by adjusting model parameters or output distribution. An adaptive mechanism is employed to dynamically adjust the weights of the predicted values from the first task based on the data distribution of the second task. For example, an attention mechanism can be used to assign differentiated fusion weights to the predicted values for different tasks. By reusing the prediction knowledge from the first task, prediction transfer significantly reduces the labeled data requirements and training costs for the second task.
[0130] For example, Figure 4 is a schematic diagram of information transfer. As shown in Figure 4, feature vector transfer between multiple tasks can transfer the vectors of the preceding task to the subsequent task. After dimensional transformation and information fusion by MLP, the predicted value is output.
[0131] As described above regarding AI technology, the key to information transfer in machine learning is to reduce the difference in conditional probability distributions between the source and target domains, thereby achieving effective knowledge transfer. This application utilizes the dependencies between tasks to achieve this goal. For example, Bayesian networks used to represent dependencies between tasks can efficiently calculate the conditional probabilities of complex tasks. Therefore, this application obtains the dependencies between multiple tasks and achieves information transfer based on these dependencies. Flexible dependencies can increase the flexibility and applicability of the model, more effectively transfer knowledge, improve the model's learning efficiency and performance, and enhance the model's prediction accuracy.
[0132] This application obtains the dependencies between multiple tasks by describing a dataset of multiple tasks, and realizes efficient information transfer between tasks based on these dependencies, thereby improving the learning efficiency and performance of the model. In addition, the dependencies based on the task dataset can also increase the flexibility and wide applicability of the model.
[0133] Based on this, the information migration method provided in this application embodiment can sort the results by combining multiple tasks in a large number of personalized service scenarios.
[0134] For example, as shown in Figure 5a (Figure 5a is a schematic diagram of Huawei App Market), the various lists (such as Featured Apps and Featured New Games) on the "Recommendations" page of Huawei App Market need to recommend the apps that users are most likely to be interested in from tens of thousands of candidate applications (APPs). Based on the technical solution of this application, the user's interactive behaviors such as downloading and clicking on different apps on different lists can be migrated to the apps, which can realize the prediction of the apps that users are most likely to be interested in.
[0135] For example, as shown in Figure 5b (Figure 5b is a schematic diagram of Huawei Mall), different lists in Huawei Mall (such as Recommended for You, Mobile Phones) need to recommend products that users are most likely to be interested in. Based on the technical solution of this application, the user's interactive behaviors such as purchasing, collecting, and adding to cart can be transferred to products, and the modeling of products that users are most likely to be interested in can be realized.
[0136] For example, as shown in Figure 5c (Figure 5c is a schematic diagram of Huawei Music), different charts in Huawei Music (such as Private FM and Daily Playlists) need to recommend music that users are most likely to be interested in. Based on the technical solution of this application, the user's interactive behaviors such as collecting, liking, playing, and switching songs on different charts can be transferred to music, and the modeling of the music that users are most likely to be interested in can be realized.
[0137] Figure 6 is a structural schematic diagram of the information migration device 600 of this application. As shown in Figure 6, the information migration device 600 of this embodiment can be applied to the aforementioned smart terminal. The information migration device 600 may include: an acquisition module 601 and a processing module 602.
[0138] The acquisition module 601 is used to acquire a dataset, which describes user interaction behaviors corresponding to multiple tasks; and to acquire the dependency relationships between the multiple tasks based on the dataset. The processing module 602 is used to complete the information migration between the multiple tasks based on the dependency relationships between the multiple tasks.
[0139] In one possible implementation, the processing module 602 is specifically used to obtain the weights of the multiple tasks in the information migration based on the dependencies between the multiple tasks; and to complete the information migration between the multiple tasks based on the weights of the multiple tasks.
[0140] In one possible implementation, the processing module 602 is specifically used to obtain the conditional probabilities and / or mutual information between the multiple tasks based on the dependencies between the multiple tasks. The conditional probability is used to represent the probability that one task is in a specific state and the other task is in the specific state among any two tasks with dependencies. The mutual information is used to represent the degree of dependency between any two tasks with dependencies. The weights of the multiple tasks are obtained based on the conditional probabilities and / or mutual information between the multiple tasks.
[0141] In one possible implementation, the information transfer includes feature vector transfer between the plurality of tasks; or, the information transfer includes prediction value transfer between the plurality of tasks.
[0142] In one possible implementation, the processing module 602 is specifically used to obtain the input vector of the second task based on the weight of the first task and the output vector of the first task, and to obtain the output vector of the second task after processing the input vector of the second task through a multilayer perceptron (MLP), wherein the first task is the preceding task of the second task.
[0143] In one possible implementation, the processing module 602 is specifically used to obtain the predicted value of the second task based on the weight of the first task and the predicted value of the first task, wherein the first task is a preceding task of the second task.
[0144] In one possible implementation, the acquisition module 601 is specifically used to acquire potential relationships between the multiple tasks based on the dataset, the potential relationships including at least one of similarity, complementarity, or causality; and to acquire dependencies between the multiple tasks based on the potential relationships.
[0145] In one possible implementation, the acquisition module 601 is specifically used to identify potential relationships between the multiple tasks by analyzing the label information in the dataset.
[0146] In one possible implementation, the dependencies between the multiple tasks are represented using a Bayesian relational network.
[0147] The apparatus in this embodiment can be used to execute the technical solution of the method embodiment shown in FIG2. Its implementation principle and technical effect are similar, and will not be described again here.
[0148] In implementation, each step of the above method embodiments can be completed by integrated logic circuits in the processor hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this application can be directly implemented by a hardware encoding processor, or by a combination of hardware and software modules in the encoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.
[0149] The memory mentioned in the above embodiments can be volatile memory or non-volatile memory, or may include both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.
[0150] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0151] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0152] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.
[0153] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0154] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0155] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0156] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. An information migration method, characterized in that, include: Obtain a dataset that describes user interaction behaviors corresponding to multiple tasks; Based on the dataset, obtain the dependencies between the multiple tasks; The information migration between the multiple tasks is completed based on the dependencies between them.
2. The method according to claim 1, characterized in that, The step of performing information migration between the multiple tasks based on their dependencies includes: The weights of the multiple tasks in the information migration are obtained based on the dependencies between the multiple tasks; Information transfer between the multiple tasks is completed based on their weights.
3. The method according to claim 2, characterized in that, The step of obtaining the weights of the multiple tasks in information migration based on the dependencies between the multiple tasks includes: Based on the dependencies between the multiple tasks, the conditional probabilities and / or mutual information between the multiple tasks are obtained. The conditional probabilities are used to represent the probability that one task is in a specific state and the other task is in the specific state among any two tasks with dependencies. The mutual information is used to represent the degree of dependency between any two tasks with dependencies. The weights of the multiple tasks are obtained based on the conditional probabilities and / or mutual information among the multiple tasks.
4. The method according to claim 2 or 3, characterized in that, The information transfer includes feature vector transfer between the multiple tasks; or, the information transfer includes prediction value transfer between the multiple tasks.
5. The method according to claim 4, characterized in that, The step of performing information transfer between the multiple tasks based on their weights includes: The input vector of the second task is obtained based on the weight of the first task and the output vector of the first task. The input vector of the second task is then processed by a multilayer perceptron (MLP) to obtain the output vector of the second task. The first task is the preceding task of the second task.
6. The method according to claim 4, characterized in that, The step of performing information transfer between the multiple tasks based on their weights includes: The predicted value of the second task is obtained based on the weight of the first task and the predicted value of the first task, wherein the first task is the preceding task of the second task.
7. The method according to any one of claims 1-6, characterized in that, The step of obtaining the dependencies between the multiple tasks based on the dataset includes: Based on the dataset, potential relationships between the multiple tasks are obtained, and the potential relationships include at least one of similarity, complementarity, or causality. The dependencies between the multiple tasks are obtained based on the potential relationships.
8. The method according to claim 7, characterized in that, The step of obtaining the potential relationships between the multiple tasks based on the dataset includes: By analyzing the label information in the dataset, the potential relationships between the multiple tasks can be identified.
9. The method according to any one of claims 1-8, characterized in that, The dependencies between the multiple tasks are represented using a Bayesian relational network.
10. An information transfer device, characterized in that, include: The acquisition module is used to acquire a dataset, which describes user interaction behaviors corresponding to multiple tasks. Based on the dataset, obtain the dependencies between the multiple tasks; The processing module is used to complete the information migration between the multiple tasks based on the dependencies between the multiple tasks.
11. The apparatus according to claim 10, characterized in that, The processing module is specifically used to obtain the weights of the multiple tasks in the information migration based on the dependencies between the multiple tasks; and to complete the information migration between the multiple tasks based on the weights of the multiple tasks.
12. The apparatus according to claim 11, characterized in that, The processing module is specifically used to obtain the conditional probabilities and / or mutual information between the multiple tasks based on the dependencies between the multiple tasks. The conditional probability is used to represent the probability that one task is in a specific state and the other task is in the specific state among any two tasks with dependencies. The mutual information is used to represent the degree of dependency between any two tasks with dependencies. The module also obtains the weights of the multiple tasks based on the conditional probabilities and / or mutual information between the multiple tasks.
13. The apparatus according to claim 11 or 12, characterized in that, The information transfer includes feature vector transfer between the multiple tasks; or, the information transfer includes prediction value transfer between the multiple tasks.
14. The apparatus according to claim 13, characterized in that, The processing module is specifically used to obtain the input vector of the second task based on the weight of the first task and the output vector of the first task, and to obtain the output vector of the second task after processing the input vector of the second task through a multilayer perceptron (MLP). The first task is the preceding task of the second task.
15. The apparatus according to claim 13, characterized in that, The processing module is specifically used to obtain the predicted value of the second task based on the weight of the first task and the predicted value of the first task, wherein the first task is the preceding task of the second task.
16. The apparatus according to any one of claims 10-15, characterized in that, The acquisition module is specifically used to acquire potential relationships between the multiple tasks based on the dataset, wherein the potential relationships include at least one of similarity, complementarity, or causal relationship; and to acquire dependency relationships between the multiple tasks based on the potential relationships.
17. The apparatus according to claim 16, characterized in that, The acquisition module is specifically used to identify the potential relationships between the multiple tasks by analyzing the label information in the dataset.
18. The apparatus according to any one of claims 1-17, characterized in that, The dependencies between the multiple tasks are represented using a Bayesian relational network.
19. A computer device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-9.
20. A computer-readable storage medium, characterized in that, Includes a computer program, which, when executed on a computer, causes the computer to perform the method of any one of claims 1-9.
21. A computer program product, characterized in that, The computer program product includes computer program code that, when run on a computer, causes the computer to perform the method of any one of claims 1-9.
