User interest prediction method and apparatus, storage medium, and electronic device

Abstract

The application discloses a user interest prediction method and device, a storage medium and an electronic device. The method comprises: obtaining target task description data representing a target task, and target user description data representing a target user; performing multi-expert feature extraction on the target user description data to obtain multi-expert feature information; performing weight information extraction on the target task description data and the target user description data to obtain a first weight vector and a second weight vector; fusing the multi-expert feature information based on the first weight vector to obtain user fusion features; performing multi-reviewer interest prediction on the user fusion features to obtain a multi-reviewer interest prediction result; and fusing the multi-reviewer interest prediction result based on the second weight vector to obtain a target interest prediction result. The application can improve the interest prediction accuracy and reduce the training cost of related models by considering the multi-expert feature extraction result and the multi-reviewer interest prediction result.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and more particularly to user interest prediction methods, apparatus, storage media, and electronic devices. Background Technology

[0002] Currently, media content recommendation relying on candidate set expansion is becoming increasingly popular. Candidate set expansion can generate a prediction model based on the seeds corresponding to media content, and then predict the target users interested in that media content. By recommending the media content to these target users, the effectiveness of ad delivery can be improved. However, different media content can vary significantly, and correspondingly, their corresponding seed users can also vary considerably. This makes it difficult to port and reuse the prediction models for media content. As the amount of media content increases, the time and computing power costs of the prediction models also rise, and the prediction accuracy needs further improvement. Summary of the Invention

[0003] To improve the speed, cost, and accuracy of user interest prediction, embodiments of this application provide a user interest prediction method, apparatus, storage medium, and electronic device.

[0004] On one hand, embodiments of this application provide a user interest prediction method, the method comprising:

[0005] Obtain target task description data that represents the target task, and target user description data that represents the target user;

[0006] Multi-expert feature extraction is performed on the target user description data to obtain multi-expert feature information;

[0007] Weight information is extracted from the target task description data and the target user description data to obtain a first weight vector and a second weight vector.

[0008] The multi-expert feature information is fused based on the first weight vector to obtain the user fused feature;

[0009] Multi-commentator interest prediction is performed on the user fusion features to obtain multi-commentator interest prediction results;

[0010] The multi-commentator interest prediction results are fused based on the second weight vector to obtain the target interest prediction result, which represents the degree of interest of the target user in the target task.

[0011] On the other hand, embodiments of this application provide a user interest prediction device, the device comprising:

[0012] The target data acquisition module is used to acquire target task description data that represents the target task, and target user description data that represents the target user.

[0013] A multi-expert feature extraction module is used to extract multi-expert features from the target user description data to obtain multi-expert feature information.

[0014] The weight acquisition module is used to extract weight information from the target task description data and the target user description data to obtain a first weight vector and a second weight vector.

[0015] The feature fusion module is used to fuse the multi-expert feature information based on the first weight vector to obtain user fused features;

[0016] The multi-commentator interest prediction module is used to predict the multi-commentator interests of the user fusion features and obtain the multi-commentator interest prediction results.

[0017] The target interest prediction module is used to fuse the multi-commentator interest prediction results based on the second weight vector to obtain a target interest prediction result, which represents the degree of interest of the target user in the target task.

[0018] On the other hand, embodiments of this application provide a computer-readable storage medium, characterized in that the computer-readable storage medium stores at least one instruction or at least one program, wherein the at least one instruction or at least one program is loaded and executed by a processor to implement the above-described user interest prediction method.

[0019] On the other hand, embodiments of this application provide an electronic device, characterized in that it includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the at least one processor implements the above-described user interest prediction method by executing the instructions stored in the memory.

[0020] This application provides a user interest prediction method, apparatus, storage medium, and electronic device. This application can comprehensively consider the feature extraction results of multiple experts to obtain reasonable target fusion features. Based on these target fusion features, and comprehensively considering the interest prediction results of multiple commentators, a target interest prediction result is obtained. By considering the opinions of multiple experts and commentators during the interest prediction process, the final target interest prediction result is made objective and fair, improving the accuracy of interest prediction. Attached Figure Description

[0021] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a user interest prediction method provided in an embodiment of this application;

[0023] Figure 2 This is a schematic diagram of an interest prediction model for user interest prediction provided in an embodiment of this application;

[0024] Figure 3 This is a schematic flowchart of the method for extracting weight information provided in the embodiments of this application;

[0025] Figure 4 This is a flowchart of a method for fitting task features and average user features provided in an embodiment of this application;

[0026] Figure 5 This is a flowchart of a method for fusing prediction results of multiple commentators' interests, provided in an embodiment of this application.

[0027] Figure 6 This is a flowchart illustrating the interest prediction model training method provided in the embodiments of this application;

[0028] Figure 7 This is a block diagram of the user interest prediction device provided in the embodiments of this application;

[0029] Figure 8 This is a schematic diagram of the hardware structure of a device for implementing the method provided in the embodiments of this application. Detailed Implementation

[0030] The technical solutions of the embodiments 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 the embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the embodiments of this application.

[0031] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of the embodiments of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the present application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units 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 devices.

[0032] To make the objectives, technical solutions, and advantages disclosed in the embodiments of this application clearer, the embodiments of this application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of this application and are not intended to limit the embodiments of this application.

[0033] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this embodiment, unless otherwise stated, "multiple" means two or more. To facilitate understanding of the technical solutions described in the embodiments of this application and their resulting technical effects, the embodiments of this application first explain the relevant technical terms:

[0034] Artificial Intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0035] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0036] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or realize human learning interests to acquire new knowledge or skills, and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0037] Media content recommendation: This involves recommending media content to users based on their preferences. Media content can include product advertisements, marketing information, and event content, and can be presented in various formats such as images, GIFs, audio / video, and short videos. Media content recommendation can accelerate information flow, improve advertising conversion rates or event participation, and optimize marketing effectiveness.

[0038] Recommendation platform: A device that recommends media content. It can be a standalone device or it can run on a related application. It can recommend media content to users by integrating recommendation functions into the related application. For example, WeChat's "Discover", Toutiao's recommended news, and Sohu Browser's "You May Like" function are all media content recommendation functions.

[0039] Look-alike: A candidate set expansion technique. Given a candidate set, also known as a seed set (where users are identified by experts as potentially interested in a specific task), the candidate set is expanded by finding more potential users who might be interested in that specific task. Look-alike technology can be widely used in media content recommendation scenarios, supporting applications such as marketing and event strategies. Many companies have already adopted this technology in their products, such as Google, LinkedIn, and Ant Financial.

[0040] In related technologies, predictive models can be trained to identify target users interested in specific tasks, thereby delivering media content corresponding to those specific tasks to those users. However, different predictive models correspond to different specific tasks, and the number of specific tasks is increasing daily. For example, a company may conduct hundreds of marketing campaigns every day to promote different products. These hundreds of marketing campaigns correspond to hundreds of specific tasks, and correspondingly, hundreds of predictive models need to be trained, which obviously leads to excessively high model training costs.

[0041] To reduce the training cost of predictive models, related technologies have proposed a two-stage model training method. This method trains a generalization model on a large amount of data, and then trains a predictive model suitable for a specific task based on the generalization model. This can save training costs. However, the generalization model in these technologies does not consider the relationship between different tasks. For example, the data used to train the generalization model may come from different tasks, and the positive and negative samples of different tasks may be completely different, or even opposite. For instance, in a health product marketing task, positive samples are elderly people, and infants may be negative samples. However, in an infant formula marketing task, positive samples are infants, and elderly people may be negative samples. These technologies do not consider the sample differences caused by different tasks, resulting in the trained generalization model failing to converge well, exhibiting poor generalization ability, and leading to slow convergence speed and low accuracy in predictive models.

[0042] In view of this, embodiments of this application provide a user prediction method. This method can comprehensively consider the feature extraction results of multiple experts to obtain reasonable target fusion features. Based on these target fusion features, it comprehensively considers the interest prediction results of multiple commentators to obtain the target interest prediction result. By considering the opinions of multiple experts and multiple commentators during the interest prediction process, the final target interest prediction result is made objective and fair, improving prediction accuracy. The prediction model implementing this user prediction method in embodiments of this application can also be obtained through the above two-stage training method. However, because multiple experts and commentators are set in the model, considering task differences and sample differences, the generalization model converges faster and has stronger generalization ability. The prediction model obtained based on this generalization model has a fast convergence speed and high accuracy.

[0043] The methods provided in this application may relate to the field of cloud technology, such as the field of big data. Big data refers to data sets that cannot be captured, managed, and processed within a certain time frame using conventional software tools. It represents massive, rapidly growing, and diverse information assets that require new processing models to achieve stronger decision-making, insight discovery, and process optimization capabilities. With the advent of the cloud era, big data has attracted increasing attention. Big data requires specialized technologies to effectively process large amounts of data within a tolerable time frame. Technologies suitable for big data include massively parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the Internet, and scalable storage systems.

[0044] The methods provided in this application embodiment can also involve blockchain, meaning that the methods provided in this application embodiment can be implemented based on blockchain, or the data involved in the methods provided in this application embodiment can be stored based on blockchain, or the executing entity of the methods provided in this application embodiment can be located in blockchain. Blockchain is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include a blockchain underlying platform, a platform product service layer, and an application service layer.

[0045] The underlying blockchain platform can include processing modules such as user management, basic services, smart contracts, and operational monitoring. The user management module is responsible for managing the identity information of all blockchain participants, including maintaining public and private key generation (account management), key management, and maintaining the correspondence between user real identities and blockchain addresses (access management). Furthermore, under authorization, it monitors and audits transactions of certain real identities and provides risk control rule configuration (risk control audit). The basic services module is deployed on all blockchain node devices to verify the validity of business requests. After consensus is reached on valid requests, they are recorded in storage. For a new business request, the basic services first perform interface adaptation parsing and authentication (interface adaptation), and then encrypt the business information through a consensus algorithm (consensus management). After encryption, the data is transmitted completely and consistently to the shared ledger (network communication) and recorded and stored. The smart contract module is responsible for contract registration, issuance, triggering, and execution. Developers can define contract logic using a programming language and publish it to the blockchain (contract registration). According to the contract terms, the key or other events are invoked to trigger execution and complete the contract logic. It also provides functions for contract upgrades and cancellations. The operation monitoring module is mainly responsible for deployment, configuration modification, contract settings, cloud adaptation, and real-time status visualization output during product release, such as alarms, monitoring network conditions, and monitoring the health status of node devices.

[0046] The platform's product service layer provides the basic capabilities and implementation frameworks for typical applications. Developers can leverage these basic capabilities, along with the specific characteristics of their business needs, to implement blockchain-based business logic. The application service layer provides blockchain-based application services to business stakeholders.

[0047] This application can be applied to data processing devices, which can be terminal devices, such as smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, etc., but are not limited to these. The data processing device can also be a server, which can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Of course, the data processing device can be both a terminal device and a server, i.e., the two work together. The terminal device and the server can be directly or indirectly connected via wired or wireless communication, and this application does not impose any limitations on this.

[0048] The following describes a user interest prediction method according to an embodiment of this application. Figure 1 This document illustrates a flowchart of a user interest prediction method provided by an embodiment of this application. The embodiment provides the operational steps of the method as described in the embodiment or flowchart, but may include more or fewer operational steps based on conventional or non-inventive methods. The order of steps listed in the embodiment is merely one possible execution order among many and does not represent the only possible execution order. In actual system or server product execution, the method can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment) as shown in the embodiment or drawings. The above method may include:

[0049] S101. Obtain target task description data that represents the target task, and target user description data that represents the target user.

[0050] In this embodiment, the target task can be a task for recommending media content. For example, if we want to select a corresponding audience for marketing task 1 and recommend media content corresponding to marketing task 1 to that audience, then marketing task 1 is the target task. This embodiment does not limit the specific content of the target task description data; it can be some or all of the data needed to construct a target task profile. For example, in addition to the task identifier, the target task description data can also include the task domain, task content, and task audience profile.

[0051] In this embodiment, the target user is a user whose interests are to be predicted. For example, if user A's interest in marketing task 1 is predicted, then user A is the target user. This embodiment does not limit the specific content of the target user description data; it can be some or all of the data needed to construct a target user profile. For example, in addition to user identifiers, the target user description data may also include user age, user gender, user shopping history, user areas of interest, bloggers the user is interested in, users the user follows, and a list of the user's subscribed accounts.

[0052] S102. Perform multi-expert feature extraction on the above target user description data to obtain multi-expert feature information.

[0053] Please refer to Figure 2 The diagram illustrates an interest prediction model for user interest prediction provided in an embodiment of this application. Figure 2 The expert feature extraction layer in the model can be used to perform the aforementioned multi-expert feature extraction. This application embodiment does not limit the specific structure of the expert feature extraction layer; it can be a feedforward neural network. Each expert feature extraction layer can extract features from the target user description data to obtain the user feature information output by that expert feature extraction layer. In this application embodiment, feature vectors or feature matrices can be used to express the aforementioned user feature information, and the user feature information output by each expert feature extraction layer forms the aforementioned multi-expert feature information.

[0054] In one embodiment, feature extraction can be performed on the target user description data to obtain user feature extraction results. These results are then output to each expert feature extraction layer to obtain multi-expert feature information. In another embodiment, the target user description data can be directly output to each expert feature extraction layer to obtain the aforementioned multi-expert feature information.

[0055] S103. Extract weight information from the above target task description data and the above target user description data to obtain the first weight vector and the second weight vector.

[0056] In one embodiment, it can be based on Figure 2 The hybrid feature extraction layer, the first task-driven gate, and the second task-driven gate in the process extract weight information from the target task description data and the target user description data to obtain the first weight vector and the second weight vector.

[0057] Please refer to Figure 3 This document illustrates a flowchart of a method for extracting weight information according to an embodiment of this application. The method involves extracting weight information from the target task description data and the target user description data to obtain a first weight vector and a second weight vector, including:

[0058] S1031. Perform average pooling on the features obtained based on the target user description data to obtain the average user features.

[0059] In one embodiment, the hybrid feature extraction layer can extract features from the target user description data to obtain user feature extraction results, and then perform average pooling on the user feature extraction results to obtain the aforementioned average user features. In another embodiment, the hybrid feature extraction layer can also directly perform average pooling on the user feature extraction results obtained in step S101 to obtain the aforementioned average user features.

[0060] S1032. Perform feature extraction processing on the above target task description data to obtain task features.

[0061] In one embodiment, the hybrid feature extraction layer can also extract features from the target task description data. This application does not limit the specific feature extraction method. For example, the feature extraction can be performed based on a cascaded convolutional network, or the task features can be obtained based on a preset embedding network.

[0062] S1033. Fit the above task features and the above average user features to obtain the above first weight vector and the above second weight vector.

[0063] In one embodiment, please refer to Figure 4 The document illustrates a flowchart of a method for fitting the aforementioned task features and the aforementioned average user features according to an embodiment of this application. The method for fitting the aforementioned task features and the aforementioned average user features to obtain the aforementioned first weight vector and the aforementioned second weight vector includes:

[0064] S10331. Fit the above task features and the above user average features to obtain the first fitting result and the second fitting result.

[0065] S10332. Normalize the first fitting result and the second fitting result to obtain the first weight vector and the second weight vector.

[0066] Specifically, the task features and user average features output from the hybrid feature extraction layer can both be input into the first task-driven gate, which performs fitting processing to obtain a first fitting result. This first fitting result is then normalized to output the first weight vector. Similarly, the task features and user average features output from the hybrid feature extraction layer can both be input into the second task-driven gate, which performs fitting processing to obtain a second fitting result. This second fitting result is then normalized to obtain the second weight vector.

[0067] S104. Based on the first weight vector mentioned above, the multi-expert feature information is fused to obtain the user fused feature.

[0068] The embodiments of this application can be based on Figure 2The user fusion feature is obtained by a user feature fusion layer in the above-mentioned multi-expert feature information. The number of elements in the first weight vector is consistent with the number of user feature information elements in the multi-expert feature information, and the value of each element represents the weight of the corresponding user feature information. The above-mentioned fusion of the multi-expert feature information based on the first weight vector to obtain the user fusion feature includes: performing a weighted average of the multiple user feature information elements based on the first weight vector to obtain the user fusion feature. For example, the user fusion feature can be calculated using Formula 1, where Formula 1 is... Where r represents the user fusion feature, h represents the value of the i-th element in the aforementioned first weight vector. i (u) represents the user feature information output by the i-th expert feature extraction layer, u represents the common input of each expert user extraction layer, and n represents the number of expert feature extraction layers.

[0069] Furthermore, the aforementioned first weight vector can be calculated using Formula 2, which is: Among them, w expert Let represent the first weight vector, softmax represent the normalization process, g represent the fitting process in the first task-driven gate, which can be implemented based on a feedforward neural network, and c represent the task features. This represents the average user characteristic, and u has the same meaning as mentioned above.

[0070] This application embodiment believes that different expert feature extraction layers in the trained interest prediction model are good at extracting features from different user description data or different user features. By controlling the proportion of user feature information output by each expert feature extraction layer in the user fusion feature through the first weight vector, the user fusion feature can be made as objective and fair as possible, and the information effectiveness of the user fusion feature can be improved.

[0071] S105. Perform multi-commentator interest prediction on the above user fusion features to obtain the multi-commentator interest prediction results.

[0072] Figure 2 The commentator interest prediction layer in the document can be used to perform the aforementioned multi-commentator interest prediction. This embodiment does not limit the specific structure of the commentator interest prediction layer; it can be a feedforward neural network. Each commentator interest prediction layer can predict the user's interest based on the fused features, obtaining the commentator interest prediction result output by that layer. In this embodiment, feature vectors or feature matrices can be used to express the commentator interest prediction result, and the commentator interest prediction results output by each commentator interest prediction layer form the aforementioned multi-commentator interest prediction result.

[0073] S106. Based on the second weight vector mentioned above, the multi-commentator interest prediction results are fused to obtain the target interest prediction results, which represent the degree of interest of the target user in the target task.

[0074] The embodiments of this application can be based on Figure 2 The target interest prediction result is obtained by fusing the interest prediction results in the middle layer. The number of elements in the second weight vector is consistent with the number of commentator interest prediction results in the above multi-commentator interest prediction results, and the value of each element represents the weight of the corresponding commentator interest prediction result. Please refer to... Figure 5 The document illustrates a flowchart of a method for fusing multi-commentator interest prediction results provided in an embodiment of this application. The method involves fusing the multi-commentator interest prediction results based on the second weight vector to obtain a target user interest prediction result, including:

[0075] S1061. Activate the above multiple commentator interest prediction results to obtain multiple activation results.

[0076] S1062. Based on the second weight vector mentioned above, perform a weighted average of the multiple activation results to obtain the target interest prediction result.

[0077] For example, the above target interest prediction result can be calculated using Formula 3, which is: Where p represents the target interest prediction result mentioned above. t represents the value of the i-th element in the aforementioned second weight vector. i (r) represents the commentator interest prediction result output by the i-th commentator interest prediction layer, σ represents the activation process, and m represents the number of commentator interest prediction layers. m can be the same as or different from n mentioned above. Furthermore, the calculation method of the second weight vector can refer to the calculation method of the first weight vector, and will not be repeated here.

[0078] This application embodiment believes that different commentator interest prediction layers in a well-trained interest prediction model are good at predicting different targets. These different targets represent the degree of interest of different users in different tasks. By controlling the proportion of the commentator interest prediction results output by each commentator interest prediction layer in the target interest prediction results through the second weight vector, the target interest prediction results can be made objective and fair, thereby improving the prediction accuracy.

[0079] The target interest prediction results in this application embodiment can be represented by an interest score. For example, if the target task is task 1 and the target user is user A, the corresponding interest score can be understood as the degree of user A's interest in task 1. If the interest score is 1, it indicates that user A is very interested in task 1; if the interest score is 0, it indicates that the user is completely uninterested in task 1. Based on this prediction result, various applications can be performed, such as user profiling analysis, target task audience analysis, target task promotion, and target task prospect prediction, etc., which are not limited in this application. These applications can also be integrated into relevant software; for example, media content corresponding to the target task can be promoted to the target task audience in WeChat's "Discover" feature.

[0080] Based on the aforementioned user interest prediction method, interest prediction can be performed on a large number of target users to obtain an interest score for each target user. Based on this interest score, a target audience is determined. This target audience represents target users whose level of interest in the target task meets preset requirements. Media content corresponding to the target task is then recommended to this target audience. This application does not limit the method for determining target users. For example, all registered users of the recommendation platform can be considered as target users, or active users of the recommendation platform can be considered as target users.

[0081] This application does not limit the method for determining the target audience. In one embodiment, target users can be sorted in descending order of interest scores, and the top preset number of target users in the sorting results can be determined as the target audience. In another embodiment, a preset score threshold can be set, and target users with interest scores greater than the preset score threshold can be determined as the target audience. This application does not limit the specific content of the preset number and preset score threshold, and can be set according to actual conditions.

[0082] Based on the above-described target audience determination method, the embodiments of this application can quickly and accurately identify target audiences interested in the target task. By recommending media content corresponding to the target task to the target audience, a better recommendation effect can be achieved, thereby increasing the promotion of the target task.

[0083] The user interest prediction method described in this application can be implemented based on a trained interest prediction model. Please refer to [link / reference]. Figure 6 It illustrates a flowchart of the interest prediction model training method provided in an embodiment of this application. The interest prediction model training method includes:

[0084] S201. Obtain a first sample carrying a first label, wherein the first sample includes historical task description data representing historical tasks and historical user description data representing historical users, and the first label represents the degree of interest of the historical users in the historical tasks.

[0085] This application does not limit the type of historical task; it can be various types of completed tasks, such as health product marketing tasks, infant formula marketing tasks, e-commerce marketing tasks, and promotional tasks. For each historical task, the degree of interest of the historical user in that task can be clearly defined. For example, if the historical user participated in the historical task, the corresponding first tag can be 1; if the historical user did not participate in the historical task, the corresponding first tag can be 0.

[0086] S202. Input the first sample into a preset neural network to obtain the interest prediction result of the first sample.

[0087] The structure of the neural network in this embodiment can be referred to... Figure 2 The data processing logic of neural networks is as described above. Figure 2 Each neural network layer in the application (expert feature extraction layer, user feature fusion layer, commentator interest prediction layer, interest prediction result fusion layer, first task-driven gate, second task-driven gate, and hybrid feature extraction layer) can be constructed based on a feedforward neural network. In a feedforward neural network (FFN), parameters propagate unidirectionally from the input layer to the output layer and are adjusted by feedback. This application does not limit the specific structure of each neural network layer. For example, each neural network layer may include at least one convolutional layer, and the number of convolutional layers in different neural network layers may be the same or different.

[0088] Figure 2 Furthermore, the data output by each commentator interest prediction layer can be activated through an activation layer. This application does not limit the specific activation function used for activation processing. For example, it can be at least one of ReLU (Rectified Linear Unit), Sigmoid activation function, and Tanh activation function.

[0089] S203. Based on the above-mentioned first sample interest prediction results and the above-mentioned first label, adjust the parameters of the above-mentioned preset neural network to obtain a generalized network.

[0090] In one embodiment, the first loss generated by the preset neural network can be calculated based on the difference between the first sample interest prediction result and the first label. The parameters of the preset neural network can be adjusted based on the feedback of the first loss until a preset stopping condition is reached, thereby obtaining the generalization network.

[0091] This application embodiment does not limit the parameter adjustment method in the process of obtaining the generalization network. For example, gradient descent can be used for parameter tuning. When the first loss is less than the first loss threshold, it can be determined that the training of the preset neural network has met the requirements, and the preset neural network in this case is identified as the generalization network. When the first loss is greater than or equal to the first loss threshold, the parameters of each neural network layer in the preset neural network can be adjusted feedback until the obtained first loss is less than the first loss threshold. The first loss threshold can be a value set according to requirements, such as 0.1, but it is not a specific limitation of this embodiment.

[0092] The aforementioned generalization network is trained on a large number of initial samples. During the training process, the parameters of different expert feature extraction layers gradually differ, making each layer adept at extracting features from different users and obtaining corresponding user feature information. Similarly, the parameters of different commentator interest prediction layers also gradually differ, making each layer adept at predicting interests for different tasks. The parameters of the first and second task-driven gates also gradually change, allowing the first gate to output a weight vector for fusing multi-expert feature information, and the second gate to output a weight vector for fusing multi-commentator interest prediction results. By considering the differences between different tasks and the differences in user features under different tasks during the training process, this generalization network focuses more on the relationships between tasks compared to generalization networks in related technologies. It not only converges quickly but also significantly improves its generalization ability and stability. Therefore, based on this generalization network, interest prediction models for various target tasks can be trained quickly, and the convergence speed and prediction accuracy of the interest prediction models can be significantly improved.

[0093] S204. Obtain a second sample carrying a second label, wherein the second sample includes the target task description data and the sample user description data, and the second label represents the sample user's degree of interest in the target task.

[0094] Based on the generalized network, an interest prediction model can be trained for any target task. The target task description data in this step can be the same as the target task description data in step S101, which allows the trained interest prediction model to be used to execute steps S101-S106.

[0095] The second sample can include positive samples and negative samples. The users in the positive samples are all interested in the target task, and their corresponding second labels can be set to 1. The users in the negative samples are all not interested in the target task, and their corresponding second labels can be set to 0.

[0096] In this embodiment of the application, the source of the sample users in the positive sample is not limited. For example, they can come from a seed set. For each target task, a seed set provided by relevant experts can be obtained, and the users in the seed set are all interested in the target task.

[0097] This application does not limit the source of sample users in the negative samples. For example, users can be randomly selected from the users of the recommendation platform. If the selected user does not belong to the above seed set, the user can be set as a sample user in the negative samples.

[0098] S205. Input the second sample into the generalization network to obtain the interest prediction result of the second sample.

[0099] S206. Based on the above second sample interest prediction results and the above second label, adjust the parameters of the above generalization network to obtain the interest prediction model. The above interest prediction model is used to output the above target interest prediction results.

[0100] In one embodiment, a second loss generated by the generalization network can be calculated based on the difference between the second sample interest prediction result and the second label. The parameters of the generalization network are then adjusted based on this loss feedback until a preset stopping condition is met, thus obtaining the interest prediction model. This application does not limit the parameter adjustment method in obtaining the interest prediction model; the parameter tuning method in obtaining the generalization model can be referenced.

[0101] In this embodiment, a generalization model can be first trained based on a preset neural network, and then an interest prediction model can be trained based on the generalization model. By taking into account task differences and user differences between different tasks during the training process, the generalization ability of the obtained generalization model can be made stronger, thereby making it faster to obtain an interest prediction model based on the generalization model and with higher accuracy.

[0102] This application also discloses a user interest prediction device. Please refer to the embodiments therein. Figure 7 The diagram illustrates a block diagram of a user interest prediction device provided in an embodiment of this application. The device includes:

[0103] The target data acquisition module 101 is used to acquire target task description data that represents the target task and target user description data that represents the target user.

[0104] The multi-expert feature extraction module 102 is used to extract multi-expert features from the target user description data to obtain multi-expert feature information.

[0105] The weight acquisition module 103 is used to extract weight information from the target task description data and the target user description data to obtain a first weight vector and a second weight vector.

[0106] The feature fusion module 104 is used to fuse the multi-expert feature information based on the first weight vector to obtain the user fused feature.

[0107] The multi-commentator interest prediction module 105 is used to predict the multi-commentator interests of the above-mentioned user fusion features and obtain the multi-commentator interest prediction results.

[0108] The target interest prediction module 106 is used to fuse the multi-commentator interest prediction results based on the second weight vector to obtain the target interest prediction result, which represents the degree of interest of the target user in the target task.

[0109] Specifically, the user interest prediction device disclosed in this application and the corresponding method embodiment described above are both based on the same inventive concept. For details, please refer to the method embodiment; further elaboration will not be repeated here.

[0110] This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned user interest prediction method.

[0111] This application also provides a computer-readable storage medium that can store multiple instructions. These instructions are adaptable for a processor to load and execute the user interest prediction method described in this application.

[0112] Furthermore, Figure 8 A schematic diagram of a hardware structure for implementing the method provided in the embodiments of this application is shown. This device can participate in or include the apparatus or system provided in the embodiments of this application. Figure 8 As shown, device 10 may include one or more processors 102 (shown as 102a, 102b, ..., 102n in the figure) 102 (processor 102 may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 8 The structure shown is for illustrative purposes only and does not limit the structure of the electronic device described above. For example, device 10 may also include a... Figure 8 The more or fewer components shown, or having the same Figure 8 The different configurations shown.

[0113] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the device 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0114] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the method described in the embodiments of this application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the user interest prediction method described above. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the device 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0115] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of device 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a radio frequency (RF) module used for wireless communication with the Internet.

[0116] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows a user to interact with the user interface of device 10 (or a mobile device).

[0117] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0118] The embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and server embodiments are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0119] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0120] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present application should be included within the protection scope of the present application.

Claims

1. A method for predicting user interests, characterized in that, The method includes: Obtain a first sample carrying a first label. The first sample includes historical task description data representing historical tasks and historical user description data representing historical users. The first label represents the degree of interest of the historical user in the historical tasks. The first sample is input into a preset neural network to obtain the interest prediction result of the first sample; based on the interest prediction result of the first sample and the first label, the parameters of the preset neural network are adjusted to obtain a generalization network. Obtain a second sample carrying a second label. The second sample includes target task description data and sample user description data, which characterize the target task. The second label characterizes the sample user's level of interest in the target task. The second sample is input into the generalization network to obtain the interest prediction result of the second sample; based on the interest prediction result of the second sample and the second label, the parameters of the generalization network are adjusted to obtain the interest prediction model, which includes a first task-driven gate, a second task-driven gate, multiple expert feature extraction layers and multiple commentator interest prediction layers. Obtain target task description data that characterizes the target task, and target user description data that characterizes the target user; Based on the multiple expert feature extraction layers, multi-expert feature extraction is performed on the target user description data to obtain multi-expert feature information; The features obtained based on the target user description data are subjected to average pooling to obtain the average user features; the target task description data are subjected to feature extraction to obtain the task features. Based on the first task-driven gate, the task features and the average user features are input into a first preset formula for fitting and normalization, and a first weight vector is calculated. The first preset formula includes the task features. Based on the second task-driven gate, the task features and the average user features are input into a second preset formula for fitting and normalization, and a second weight vector is calculated. The second preset formula includes the task features. The multi-expert feature information is fused based on the first weight vector to obtain user fused features; based on the multiple commentator interest prediction layers, the user fused features are used to perform multi-commentator interest prediction for the target task to obtain multiple commentator interest prediction results that all represent the degree of interest of the target user in the target task. The multiple commentator interest prediction results are activated, and a weighted average of the multiple activation results is performed based on the second weight vector to obtain the target interest prediction result, which represents the degree of interest of the target user in the target task.

2. The method according to claim 1, characterized in that, The target interest prediction result is represented by an interest score. The method further includes: determining a target audience based on the interest score, wherein the target audience represents target users whose interest in the target task meets preset requirements; and recommending media content corresponding to the target task to the target audience.

3. The method according to claim 1, characterized in that, The first weight vector is calculated by inputting the task features and the average user features into a first preset formula based on the first task-driven gate for fitting and normalization, including: The task features and the average user features are both input into the first task driving gate, which performs fitting processing to obtain a first fitting result. The first fitting result is then normalized to output the first weight vector. Based on the second task-driven gate, the task features and the average user features are input into a second preset formula for fitting and normalization to calculate the second weight vector, including: The task features and the average user features are both input into the second task driving gate, which performs fitting processing to obtain a second fitting result. The second fitting result is then normalized to output the second weight vector.

4. The method according to claim 1 or 2, characterized in that, The multi-expert feature information includes multiple user feature information, and the elements in the first weight vector represent the weights of the corresponding user feature information. The step of fusing the multi-expert feature information based on the first weight vector to obtain user fusion features includes: performing a weighted average processing on the multiple user feature information based on the first weight vector to obtain the user fusion features.

5. A user interest prediction device, characterized in that, The device is used to perform the following operations: Obtain a first sample carrying a first label. The first sample includes historical task description data representing historical tasks and historical user description data representing historical users. The first label represents the degree of interest of the historical user in the historical tasks. The first sample is input into a preset neural network to obtain the interest prediction result of the first sample; based on the interest prediction result of the first sample and the first label, the parameters of the preset neural network are adjusted to obtain a generalization network. Obtain a second sample carrying a second label. The second sample includes target task description data and sample user description data, which characterize the target task. The second label characterizes the sample user's level of interest in the target task. The second sample is input into the generalization network to obtain the interest prediction result of the second sample; based on the interest prediction result of the second sample and the second label, the parameters of the generalization network are adjusted to obtain the interest prediction model, which includes a first task-driven gate, a second task-driven gate, multiple expert feature extraction layers and multiple commentator interest prediction layers. The device includes: The target data acquisition module is used to acquire target task description data that characterizes the target task, and target user description data that characterizes the target user; The multi-expert feature extraction module is used to perform multi-expert feature extraction on the target user description data based on the multiple expert feature extraction layers to obtain multi-expert feature information; The weight acquisition module is used to perform average pooling on the features obtained based on the target user description data to obtain user average features; to perform feature extraction on the target task description data to obtain task features; based on the first task-driven gate, to input the task features and the user average features into a first preset formula for fitting and normalization processing to calculate a first weight vector, wherein the first preset formula includes the task features; and based on the second task-driven gate, to input the task features and the user average features into a second preset formula for fitting and normalization processing to calculate a second weight vector, wherein the second preset formula includes the task features. The feature fusion module is used to fuse the multi-expert feature information based on the first weight vector to obtain user fused features; The multi-commentator interest prediction module is used to perform multi-commentator interest prediction on the user fusion features based on the multiple commentator interest prediction layers, and obtain multiple commentator interest prediction results that all represent the degree of interest of the target user in the target task. The target interest prediction module is used to activate the multiple commentators' interest prediction results and perform a weighted average of the multiple activation results based on the second weight vector to obtain the target interest prediction result, which represents the target user's degree of interest in the target task.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement a user interest prediction method as described in any one of claims 1 to 4.

7. An electronic device, characterized in that, The system includes at least one processor and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that are executed by the at least one processor, and the at least one processor implements a user interest prediction method as described in any one of claims 1 to 4 by executing the instructions stored in the memory.

8. A computer program product comprising computer instructions stored in a computer-readable storage medium, wherein a processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions to cause the computer device to perform a user interest prediction method as described in any one of claims 1 to 4.

Images