A data recommendation method, a computer device, and a storage medium

Abstract

Embodiments of the present application disclose a data recommendation method, a computer device and a storage medium. The method comprises: obtaining attribute information of a plurality of candidate data, and obtaining fatigue information of a target object for each candidate data; determining target data embedding features of the each candidate data based on the attribute information and the fatigue information of the each candidate data; obtaining historical behavior embedding features corresponding to historical behavior data of the target object, and predicting click rates of the each candidate data based on the historical behavior embedding features and the target data embedding features of the each candidate data; and performing data recommendation for the target object by using the click rates of the each candidate data. The data recommendation effect can be improved. The present application relates to the field of artificial intelligence. The prediction of the click rate can be implemented by using a neural network model, and is used in the scene of data recommendation.

Description

Technical Field

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

[0002] With the rapid development of electronic and internet technologies, multimedia data is increasing dramatically, leading to the rapid development of recommender systems. These systems can extract user-relevant data from massive amounts of multimedia data and recommend it to users. Therefore, improving the effectiveness of data-driven recommendations has become a current research hotspot. Summary of the Invention

[0003] This application provides a data recommendation method, computer device, and storage medium, which can improve the data recommendation effect.

[0004] In a first aspect, embodiments of this application provide an image super-resolution processing method, including:

[0005] Obtain attribute information from multiple candidate data sets, and obtain fatigue information of the target object for each candidate data set;

[0006] Based on the corresponding attribute information and fatigue information of each candidate data, the target data embedding features of each candidate data are determined;

[0007] Obtain the historical behavior embedding features corresponding to the historical behavior data of the target object, and predict the click-through rate of each candidate data based on the historical behavior embedding features and the target data embedding features of each candidate data respectively;

[0008] Data recommendations are made for the target object based on the click-through rate of each candidate data.

[0009] Secondly, embodiments of this application provide a data recommendation device, including:

[0010] The acquisition unit is used to acquire attribute information of multiple candidate data and acquire fatigue information of the target object for each candidate data.

[0011] The determining unit is used to determine the target data embedding features of each candidate data based on the corresponding attribute information and fatigue information of each candidate data.

[0012] The prediction unit is used to obtain the historical behavior embedding features corresponding to the historical behavior data of the target object, and predict the click rate of each candidate data based on the historical behavior embedding features and the target data embedding features of each candidate data.

[0013] The recommendation unit is used to recommend data to the target object based on the click-through rate of each candidate data.

[0014] Thirdly, embodiments of this application provide a computer device, the computer device including: a processor and a memory, the processor being configured to execute the method described in the first aspect above.

[0015] Fourthly, embodiments of this application also provide a computer-readable storage medium storing program instructions that, when executed, implement the method described in the first aspect above.

[0016] Fifthly, embodiments of this application also provide a computer program product or computer program, which includes program instructions that, when executed by a processor, implement the method described in the first aspect above.

[0017] This application embodiment can acquire attribute information of multiple candidate data and fatigue information of the target object for each candidate data. Then, based on the corresponding attribute information and fatigue information of each candidate data, the target data embedding features of each candidate data can be determined. It can also acquire historical behavior embedding features corresponding to the historical behavior data of the target object, and predict the click-through rate of each candidate data based on the historical behavior embedding features and the target data embedding features of each candidate data. Furthermore, the click-through rate of each candidate data can be used to recommend data to the target object. In this way, multiple dimensions of features (such as the historical behavior data of the target object and fatigue information related to fatigue) can be combined to predict the click-through rate of each candidate data, thereby improving prediction accuracy. Simultaneously, data recommendation based on the predicted click-through rate of each candidate data ensures that the data ultimately recommended to the target object has a higher click-through rate, thus increasing the click-through rate of the recommended data and consequently increasing the target object's data consumption time, thereby improving the data recommendation effect. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the architecture of a data recommendation system provided in an embodiment of this application;

[0020] Figure 2This is a flowchart illustrating a data recommendation method provided in an embodiment of this application;

[0021] Figure 3a This is a schematic diagram illustrating the performance changes in global fatigue analysis provided in an embodiment of this application;

[0022] Figure 3b This is a schematic diagram illustrating the performance changes in coarse-grained fatigue analysis provided in an embodiment of this application;

[0023] Figure 4 This is a flowchart illustrating another data recommendation method provided in an embodiment of this application;

[0024] Figure 5a This is a schematic diagram of the structure of a target model provided in an embodiment of this application;

[0025] Figure 5b This is a schematic diagram of the structure of a feature extraction module provided in an embodiment of this application;

[0026] Figure 5c This is a schematic diagram of the structure of a feature weighting module provided in an embodiment of this application;

[0027] Figure 5d This is a schematic diagram of the structure of a correlation calculation module provided in an embodiment of this application;

[0028] Figure 5e This is a schematic diagram of the structure of a feature extraction submodule provided in an embodiment of this application;

[0029] Figure 5f This is a schematic diagram of another feature extraction submodule provided in an embodiment of this application;

[0030] Figure 5g This is a schematic diagram of the structure of another feature extraction submodule provided in the embodiments of this application;

[0031] Figure 5h This is a schematic diagram of another target model provided in an embodiment of this application;

[0032] Figure 6 This is a schematic diagram of a performance analysis curve provided in an embodiment of this application;

[0033] Figure 7 This is a schematic diagram of the structure of a data recommendation device provided in an embodiment of this application;

[0034] Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0035] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0036] With the continuous development of internet technology, artificial intelligence (AI) technology has also seen significant advancements. AI technology refers to the theories, methods, techniques, and application systems that utilize digital computers or computers-controlled machines to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science; it primarily aims to understand the essence of intelligence and produce new intelligent machines that can react in a manner similar to human intelligence, enabling these machines to possess multiple functions such as perception, reasoning, and decision-making. Accordingly, AI technology is a multidisciplinary field, mainly encompassing computer vision (CV), speech processing, natural language processing, and machine learning (ML) / deep learning.

[0037] Computer vision (CV) is a science that studies how to enable machines to "see." More specifically, it refers to machine vision, which uses cameras and computers to replace human eyes in recognizing and measuring targets, and then performs image processing to create images more suitable for human observation or transmission to instruments. As a scientific discipline, computer vision studies related theories and technologies, attempting to build artificial intelligence systems capable of extracting information from images or multidimensional data. Computer vision technologies typically include image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, 3D object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping (SLAM), autonomous driving, intelligent transportation, and other technologies, as well as common biometric recognition technologies such as facial recognition and fingerprint recognition.

[0038] Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of AI and the fundamental approach to enabling computer devices to possess intelligence. Deep learning, on the other hand, is a technique that utilizes deep neural network systems for machine learning. Machine learning / deep learning typically includes artificial neural networks, reinforcement learning (RL), supervised learning, and unsupervised learning. Supervised learning refers to training models using training samples with known categories (labeled categories), while unsupervised learning refers to training models using training samples with unknown categories (unlabeled categories).

[0039] Based on the aforementioned AI technologies, including computer vision and machine learning / deep learning, this application proposes a data recommendation scheme. The principle of this scheme is as follows: Fatigue information of the target object for each candidate data can be obtained, and data recommendation can be performed based on this fatigue information. For example, attribute information of multiple candidate data can be obtained, and target data embedding features of each candidate data can be determined based on the corresponding attribute information and fatigue information. Furthermore, historical behavior embedding features corresponding to the target object's historical behavior data can be obtained, and the click-through rate of each candidate data can be predicted based on the historical behavior embedding features and the target data embedding features of each candidate data. Therefore, data recommendation can be performed using the predicted click-through rates of each candidate data.

[0040] By implementing the above scheme, the click-through rate of each candidate data can be predicted by combining features from multiple dimensions (such as users' historical behavior data and fatigue information related to user fatigue) to improve prediction accuracy. At the same time, data recommendation based on the predicted click-through rates of each candidate data can also ensure that the data ultimately recommended to the target audience is data with a high click-through rate, thereby increasing the click-through rate of the recommended data and thus increasing the target audience's data consumption time, thereby improving the data recommendation effect.

[0041] In practical implementation, the data recommendation scheme mentioned above can be executed by a computer device, which can be a terminal or a server. The terminal mentioned here can include, but is not limited to, smartphones, tablets, laptops, desktop computers, smart TVs, etc. Various clients (applications, apps) can run on the terminal, such as multimedia playback clients, social media clients, browser clients, news feed clients, educational clients, and so on. The server mentioned here can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms, etc.

[0042] For example, when the computer device is a server, in a data recommendation scenario (such as a feed stream recommendation scenario, which is a content recommendation of aggregated information, and the feed stream can dynamically and in real time spread to subscribers, which is an effective way for objects to obtain information streams), when any object has a data recommendation request, it can be uploaded to the server through any terminal, so that the server can respond to the data recommendation request, use the data recommendation scheme to determine the click-through rate of each candidate data, and sort the multiple candidate data in descending order of click-through rate to obtain the corresponding ranking result. After obtaining the ranking result, the server can display the ranking result on the data recommendation interface of the terminal, such as... Figure 1 As shown.

[0043] Based on the above description of the data recommendation scheme, this application proposes a data recommendation method, which is mainly described using a computer device as the execution subject; please refer to... Figure 2 The data recommendation method may include the following steps S201-S204:

[0044] S201, obtain attribute information of multiple candidate data, and obtain fatigue information of the target object for each candidate data.

[0045] Candidate data can include various types of data such as articles, videos, products, and advertisements.

[0046] In this context, the attribute information of a candidate data point refers to the attributes corresponding to that candidate data. For example, attribute information may include data name, data category, data subcategory, etc. For instance, if the candidate data is a product, the product's attribute information may include the product name, product category, product subcategory, the store where the product is located, etc.

[0047] Fatigue (or user fatigue) refers to the fatigue experienced by users after entering a recommendation system, including browsing, clicking, and reading. These actions consume user energy, leading to fatigue. Fatigue can be categorized into fine-grained, coarse-grained, and global fatigue. Fine-grained fatigue refers to fatigue caused by repeatedly displayed data (items); coarse-grained fatigue refers to fatigue caused by the same type of items; and global fatigue refers to fatigue experienced after using recommended products for a period of time. Fatigue information refers to behavioral information generated by users within the recommendation system, such as reading volume and reading time. In this embodiment, fatigue information may include one or more of the following: fatigue information corresponding to global fatigue (e.g., simply referred to as global fatigue information) and fatigue information corresponding to coarse-grained fatigue (e.g., simply referred to as attribute fatigue information or coarse-grained fatigue information). In summary, global fatigue information can refer to the behavioral information generated by a target object (which can refer to any user) within a target time period; such behavioral information can be referred to as first behavioral information. Attribute fatigue information can refer to the behavioral information generated by a target object within a target time period that is associated with any candidate data; such behavioral information can be referred to as second behavioral information. The target time period can refer to the target object's current session. Therefore, the fatigue information in this embodiment can include one or more of the first and second behavioral information.

[0048] The first set of data can include one or more of the following: target data views, target data viewing duration, target data fetch counts, and target data exposure. For example, in a scenario where the data is articles, target data views can refer to the number of articles read by the target object in the current session, such as 17 views; target data viewing duration can refer to the time spent by the target object reading articles in the current session, such as 100 minutes; target data fetch counts can refer to the number of times the target object fetches data (or refreshes the data recommendation page), such as 8 fetch counts; and target data exposure can refer to the number of times the target object sees the articles in the current session, such as 25 exposures.

[0049] For any candidate data, the second line of information can include one or more of the following: account data views, category data views, and subcategory data views. For example, in a recommendation scenario where the data is articles, account data views can refer to the number of articles published by the account that publishes the candidate data in the current session (or the number of times the target audience reads the candidate data's account in the current session), where "account" can refer to the author account that publishes the candidate data. Category data views can refer to the number of articles under the category that publishes the candidate data in the current session (or the number of times the target audience reads the candidate data's category in the current session); where "category" can refer to the category to which the candidate data belongs, such as sports, music, or film. Subcategory data views can refer to the number of articles under the subcategory that publishes the candidate data in the current session (or the number of times the target audience reads the candidate data's subcategory in the current session); where "subcategory" can refer to a sub-category under a category, such as basketball or football under the category of sports.

[0050] For example, if a candidate data point is an article, and the account that published the article is AAA, then the number of articles published by account AAA can be counted from the articles read. This count represents the account's view count; for example, the account view count could be 16. If the article's category is sports, then the number of articles published in the sports category can be counted from the articles read. This count represents the category view count; for example, the category view count could be 26. If the article's subcategory is basketball, then the number of articles published in the sports subcategory can be counted from the articles read. This count represents the subcategory view count; for example, the category view count could be 19.

[0051] In one implementation, to verify the reliability of using the above information to characterize fatigue information, both global fatigue analysis and coarse-grained fatigue analysis can be performed. For example, taking an article as an example, such as... Figure 3a This is the analysis result for overall fatigue, such as... Figure 3b This is the analysis result for coarse-grained fatigue. Among them, Figure 3a In the graph, (a) represents the curve of click-through rate as browsing depth, and (b) represents the curve of average reading time per article as browsing depth. Figure 3b (a), (b), and (c) in the figure represent the comparison curves of the click-through rate of articles in the same category compared to the click-through rate of articles in other categories when reading the same category repeatedly.

[0052] from Figure 3aThe two graphs show that as browsing depth increases, both click-through rate and average reading time per article decline rapidly. From... Figure 3b The two graphs show that as users continue to read articles in the same category, the click-through rate of articles in that category declines faster than that of articles in other categories. Based on the above analysis, it can be concluded that using data related to reading and clicking behaviors to characterize fatigue information is reliable; at the same time, it can also be seen that users experience both coarse-grained fatigue and global fatigue during the consumption process of the recommendation system.

[0053] In one implementation, step S201 can be triggered upon receiving a data recommendation request. For example, when a data recommendation page on the terminal is detected to be displayed, it can be determined that a data recommendation request exists; exemplarily, when an opening operation of the data recommendation page on the terminal by a target object is detected, it can be determined that the data recommendation page on the terminal has been triggered to be displayed. Similarly, when a data recommendation page on the terminal is detected to be refreshed, it can be determined that a data recommendation request exists; exemplarily, when a refresh operation of the data recommendation page by a target object is detected, it can be determined that the data recommendation page on the terminal has been refreshed.

[0054] S202, based on the corresponding attribute information and fatigue information of each candidate data, determine the target data embedding features of each candidate data.

[0055] In one implementation, considering that the method for determining the target data embedding feature corresponding to any candidate data is consistent, this embodiment of the application uses a candidate data as an example to illustrate the method for determining the target data embedding feature. Specifically, for any candidate data, step S202 may include the following steps s1-s3.

[0056] s1, determine the data embedding features of any candidate data based on its attribute information. For example, vector embedding can be performed on the attribute information to obtain the corresponding data embedding features.

[0057] s2, determine the fatigue embedding feature of any candidate data based on the fatigue information of any candidate data.

[0058] As previously known, fatigue information may include one or more of the first behavioral information and the second behavioral information. Therefore, the determination method of the corresponding fatigue embedding feature differs under different circumstances. The following describes the specific implementation of determining the fatigue embedding feature when the fatigue information contains different information.

[0059] Case (1): Fatigue information only includes the first behavior information.

[0060] In one implementation, for any candidate data, the global fatigue feature of any candidate data can be determined based on the information in the first row; then, the global fatigue feature can be vector-embedded to obtain the fatigue embedding feature of any candidate data.

[0061] As previously known, the first line of information may include one or more of the following: target data views, target data viewing duration, target data fetch count, target data exposure. For a candidate data, the global fatigue feature corresponding to each type of first line of information can be determined first; then, the global fatigue feature corresponding to each type of first line of information can be vector-embedded to obtain the fatigue embedding feature corresponding to each type of first line of information; finally, the fatigue embedding feature of any candidate data can be determined based on the fatigue embedding feature corresponding to each type of first line of information. For example, the fatigue embedding features corresponding to each type of first line of information can be concatenated, and the result of the concatenation process is the fatigue embedding feature of any candidate data. It should also be noted that, considering that the first line of information is fatigue information for global fatigue, the fatigue embedding feature of each candidate data is the same in case (1).

[0062] Taking the first line of information as the target data view count as an example, this section explains how to determine the global fatigue level feature. Specifically, multiple bucket categories can be constructed for the target data view count, and the bucket features corresponding to each bucket category can be determined. Each bucket category corresponds to a view count range for the data view count. For example, taking articles as candidate data, the target data view count can be bucketed according to the numerical division of "-1, 0, 1, 2, 3, 4, 5, 8, 10, 15, 20, 30, 50, 100" to create features. The numerical range between two adjacent values ​​can constitute a bucket category, such as -1-0, 0-1, and 10-15, which can represent different bucket categories. Considering that the probability of data view counts corresponding to smaller values ​​is greater than that corresponding to larger values, the difference between adjacent values ​​increases progressively when constructing the bucket categories for the target data view count. In the numerical division "-1,0,1,2,3,4,5,8,10,15,20,30,50,100" mentioned above, the numerical differences between the first part of the values ​​are relatively small, while the numerical differences between the second part of the values ​​are relatively large.

[0063] In one embodiment, the binning feature corresponding to each bin category can be a series of positive integers starting from 0, i.e., the binning feature can be a value such as 0, 1, 2, etc.; for example, the binning feature corresponding to -1-0 is 0, the binning feature corresponding to 0-1 is 1, the binning feature corresponding to 1-2 is 2, and so on, thus determining the binning feature for each bin category. In another embodiment, the binning feature corresponding to each bin category can be a value within the corresponding numerical range of the bin category, and this value can be any value within the corresponding numerical range; for example, for bin category 10-15, the binning feature corresponding to this bin category can be a value such as 10, 13, or 15.

[0064] After constructing multiple bucket categories and their respective bucket features, the target bucket category corresponding to the target data view count can be found from these categories. This means identifying the numerical range within which the target data view count falls; the bucket category representing this range is the target bucket category. Once the target bucket category is determined, its corresponding bucket features can be used as the global fatigue feature for the target data view count. For example, if the target data view count is 13, then the numerical range of 13 is 10-15, meaning the target bucket category for the target data view count is 10-15. The bucket feature corresponding to 10-15 can be 9, therefore the global fatigue feature for the target data view count is 9.

[0065] Similarly, the determination method for the global fatigue features corresponding to the target data viewing time, the number of target data fetches, and the target data exposure can refer to the method for determining the global fatigue features of the target data viewing volume mentioned above. For example, taking articles as candidate data, the target data viewing time can be divided into buckets according to the values ​​"-1, 0, 10, 30, 50, 80, 120, 180, 250, 400, 600, 900, 1200, 1500, 1800, 3000" to create the corresponding global fatigue features; the target data exposure can be divided into buckets according to the values ​​"-1, 0, 1, 3, 8, 15, 30, 50, 80, 120, 150, 200, 300, 500" to create the corresponding global fatigue features. Considering that the number of times various target data are retrieved is relatively dense, the number of times the target data is retrieved can be directly used as the discrete feature input without the need for bucketing. That is, the number of times the target data is retrieved can be used as the corresponding global fatigue feature.

[0066] Case (2): Fatigue information only includes second behavior information.

[0067] In one implementation, for any candidate data, the attribute fatigue feature of any candidate data can be determined based on the second row information; then, the attribute fatigue feature can be vector-embedded to obtain the fatigue embedding feature of any candidate data.

[0068] As previously known, the second behavior information may include one or more of the following: account data views, category data views, and subcategory data views. For a candidate data, the attribute fatigue feature corresponding to each type of second behavior information can be determined first; then, the attribute fatigue feature corresponding to each type of second behavior information can be vector-embedded to obtain the fatigue embedding feature corresponding to each type of second behavior information; finally, the fatigue embedding feature of any candidate data can be determined based on the fatigue embedding feature corresponding to each type of second behavior information. For example, the fatigue embedding features corresponding to each type of second behavior information can be concatenated, and the result of the concatenation process is the fatigue embedding feature of any candidate data. It should also be noted that, considering that the second behavior information is fatigue information for attribute fatigue, and attribute fatigue is fatigue generated for the same type of data in each candidate data, the fatigue generated for different candidate data is also different, so the fatigue embedding feature of each candidate data in case (2) is different.

[0069] The specific implementation method for determining the attribute fatigue feature of any candidate data based on the second-behavior information is similar to that for determining the global fatigue feature of any candidate data based on the first-behavior information. For example, let's take account view counts as an example of the second-behavior information. Multiple bucket categories can be constructed for account view counts, and bucket features corresponding to each bucket category can be determined. Each bucket category corresponds to a view count range for account view counts. For example, taking articles as candidate data, account view counts can be bucketed according to the numerical values ​​"-1, 0, 1, 2, 3, 4, 5, 8, 10, 15, 30, 40, 70, 100" to create bucket features. The specific implementation of constructing bucket categories and corresponding bucket features can be found in the above description. After constructing multiple bucket categories and corresponding bucket features for account view counts, the target bucket category corresponding to the account view count can be found from the multiple bucket categories, and the bucket feature corresponding to the target bucket category is used as the attribute fatigue feature of the account view count.

[0070] Similarly, the method for determining the attribute fatigue features corresponding to category data views and subcategory data views can refer to the method for determining the attribute fatigue features of account data views mentioned above. For example, taking articles as candidate data, the category data views can be divided into buckets according to the values ​​"-1,0,1,2,3,4,5,8,10,15,20,30,50,100" to create the corresponding attribute fatigue features; the subcategory data views can be divided into buckets according to the values ​​"-1,0,1,2,3,4,5,8,10,15,20,30,60,100" to create the corresponding attribute fatigue features.

[0071] Case (3): Fatigue information includes first behavior information and second behavior information.

[0072] In one implementation, for any candidate data, the global fatigue feature of the candidate data can be determined based on the first row of information, and the attribute fatigue feature of the candidate data can be determined based on the second row of information. After obtaining the global fatigue feature and the attribute fatigue feature, vector embedding can be performed on the global fatigue feature and the attribute fatigue feature to obtain the fatigue embedding feature of any candidate data. That is, vector embedding can be performed on the global fatigue feature and the attribute fatigue feature separately, and the result of vector embedding of the global fatigue feature and the result of vector embedding of the attribute fatigue feature can be concatenated to obtain the fatigue embedding feature of any candidate data.

[0073] In this embodiment, bucketing is performed when determining the global fatigue feature corresponding to the first row of information and the attribute fatigue feature corresponding to the second row of information. Considering that vector embedding is required later, if vector embedding is directly performed on the original data (such as the target data view count, target data view duration, etc.), a large number of different values ​​need to be embedded, resulting in a large amount of computation and affecting the data processing speed. Therefore, the original data can be bucketed (as described above, the target data view count is converted into the corresponding bucket feature based on the bucket category). The bucketing operation can be understood as quantization, that is, the values ​​within a certain range can be grouped into one value to reduce the computation of subsequent vector embedding and thus improve the data processing speed. It should also be noted that the values ​​corresponding to the bucket categories described above are only illustrative examples, and bucket categories can also be constructed according to other values. No specific limitation is made here.

[0074] s3, the data embedding feature and the fatigue embedding feature are fused to obtain the target data embedding feature for any candidate data. This fusion process can refer to concatenating the two embedding features; the concatenation order is not limited. For example, the data embedding feature and the fatigue embedding feature can be concatenated in the same order to obtain the corresponding target data embedding feature; similarly, the fatigue embedding feature and the data embedding feature can be concatenated in the same order to obtain the corresponding target data embedding feature.

[0075] S203, obtain the historical behavior embedding features corresponding to the historical behavior data of the target object, and predict the click-through rate of each candidate data based on the historical behavior embedding features and the target data embedding features of each candidate data.

[0076] Historical behavior data refers to behavioral data within a specific historical time period. This historical time period is defined as the period preceding the current time and spanning a preset time interval (e.g., 1 day, 2 days). The current time can refer to the moment when a data recommendation request was detected. Historical behavior data may include one or more historical recommendation data sets, which can refer to data viewed by the target user within the historical time period. Each historical recommendation data set corresponds to a historical behavior embedding vector.

[0077] In one implementation, determining the historical behavior embedding features corresponding to the historical behavior data of the target object is equivalent to determining the historical behavior embedding features of each historical recommendation data point. Specifically, determining the historical behavior embedding feature of a historical recommendation data point can be achieved by: acquiring the attribute information of the historical recommendation data and using this attribute information to determine the corresponding historical behavior embedding feature; for example, vector embedding can be performed on the attribute information to obtain the corresponding historical behavior embedding feature. Alternatively, the attribute information and fatigue information of the historical recommendation data can be acquired and used to determine the corresponding historical behavior embedding feature; for example, vector embedding can be performed on the attribute information and fatigue information to obtain the corresponding historical behavior embedding feature.

[0078] In one implementation, after determining the historical behavior embedding features of the target object and the target data embedding features of each candidate data, the target interest features of the target object for each candidate data can be determined based on these features. Furthermore, the click-through rate of each candidate data can be predicted based on these target interest features.

[0079] S204, use the click-through rate of each candidate data to recommend data to the target audience.

[0080] In one implementation, candidate data can be sorted sequentially from highest to lowest click-through rate to obtain a ranking result. After obtaining this ranking result, data recommendations can be made for the target object based on the ranking. For example, when making data recommendations, candidate data with higher click-through rates can be displayed first to increase user click-through rates.

[0081] For example, the click-through rate of each candidate data can be compared with a click-through rate threshold, and candidate data with a click-through rate lower than the threshold can be filtered out. This ensures that the data recommended to the target audience is candidate data with a higher click-through rate. This can reduce the data recommendation process while minimizing the impact on the target audience's viewing experience, thereby improving the data recommendation speed.

[0082] In this embodiment, features can be used to model both coarse-grained and global user fatigue levels. This involves incorporating fatigue-related information and corresponding features into historical user behavior data to predict the click-through rate (CTR) of each candidate data point. This multi-dimensional feature-based comprehensive prediction improves accuracy. Furthermore, the introduction of fatigue information and corresponding features helps to minimize user fatigue during consumption, slowing the decline in CTR and consumption duration, thus enhancing data recommendation effectiveness. Simultaneously, recommending data based on the predicted CTR of each candidate data point ensures that the data recommended to the target audience has a higher CTR, further increasing the CTR of the recommended data and consequently extending the target audience's consumption duration, thus improving overall data recommendation effectiveness.

[0083] Please see Figure 4 This is a flowchart illustrating another data recommendation method provided in this application. This application primarily describes a computer device as the execution subject; please refer to... Figure 4 The data recommendation method may include the following steps S401-S406:

[0084] S401, obtain attribute information of multiple candidate data, and obtain fatigue information of the target object for each candidate data.

[0085] S402, Based on the corresponding attribute information and fatigue information of each candidate data, determine the target data embedding features of each candidate data.

[0086] In one implementation, the target data embedding features can be obtained by calling the encoding module (or fatigue encoding module) in the target model. For example, the target model can be as follows: Figure 5aAs shown in the diagram, the encoding module can fuse fatigue features and candidate data (such as target items) to obtain the target data embedding features of the candidate data. For example, the encoding module can use a DNN network structure, taking the concatenated data of the candidate data's feature embeddings and the fatigue feature's embeddings as input. The output of the encoding module can be the encoded feature embedding of the candidate data (such as target item embedding). Based on this, the specific implementation of step S402 can be as follows: For any candidate data, the corresponding data embedding features and fatigue embedding features can be determined based on the candidate data's attribute information and fatigue information, respectively. The determination methods for these two embedding features are described in step S202 above and will not be repeated here. After obtaining the data embedding features and fatigue embedding features, they can be concatenated, and then the encoding module can encode the concatenated result to obtain the target data embedding features of any candidate data.

[0087] S403, Obtain the historical behavior embedding features corresponding to the historical behavior data of the target object.

[0088] In one implementation, the historical behavior embeddings (or sequence embeddings) can be obtained by calling the embedding module in the target model. In this case, step S403 can be implemented as follows: obtain the historical behavior data of the target object; after obtaining the historical behavior data of the target object, construct the corresponding sparse features based on the historical behavior data; then, input the sparse features into the embedding module to perform vector embedding on the sparse features to obtain the corresponding historical behavior embeddings.

[0089] The specific implementation of steps S401-S403 can be referred to the description in steps S201-S203 above, and will not be repeated here.

[0090] S404, based on historical behavior embedding features and target data embedding features of each candidate data, determine the target interest features of the target object for each candidate data.

[0091] In one implementation, step S404 can be obtained by calling the feature extraction module in the target model. For example, for any candidate data, the historical behavior embedding feature and the target data embedding feature of the candidate data can be input into the feature extraction module to extract the target interest feature, thereby obtaining the target interest feature of the target object for the candidate data.

[0092] In one implementation, for any candidate data, firstly, the initial interest embedding features of the target object for any candidate data can be determined based on the target data embedding features and the historical behavior embedding features of the candidate data. For example, the correlation between the target data embedding features and the historical behavior embedding features of the candidate data can be calculated separately to obtain the weights of each historical behavior embedding feature; then, a weighted summation can be performed based on each historical behavior embedding feature and its corresponding weight to obtain the initial interest embedding features of the target object for any candidate data. After obtaining the initial interest embedding features, the target interest features of the target object for any candidate data can be determined based on the initial interest embedding features. Therefore, the feature extraction module described above can specifically include a feature weighting module and a feature extraction submodule, such as... Figure 5b As shown in the diagram, the feature weighting module can determine the initial interest embedding features of the target object for any candidate data by embedding the target data embedding features and the historical behavior embedding features of any candidate data; the feature extraction submodule can determine the target interest features of the target object for any candidate data based on the initial interest embedding features.

[0093] The network structure of the feature weighting module can be as follows: Figure 5c As shown, the processing mechanism of this relevance calculation module is an attention mechanism. This attention mechanism, in its modeling, not only considers the historical behavior data of the target object but also incorporates factors related to user fatigue, thereby enhancing the modeling effect. As previously known, historical behavior data can include one or more historical recommendation data sets. The attention mechanism here can calculate the relevance weights (weights) between candidate data and each historical recommendation data set. These weights can then be used as weighting coefficients to perform a weighted summation of the historical behavior embedding features of each historical recommendation data set, and the target object's interest can be reflected by the result of this weighted summation. In this application, when calculating the relevance weights (weights) between candidate data and each historical recommendation data set, the representation of the candidate data not only utilizes the data's own attribute features but also incorporates features related to fatigue, thereby strengthening the data representation.

[0094] The correlation calculation module in the feature weighting module can be as follows: Figure 5d As shown, Figure 5d As shown, the historical behavior embedding feature and the target data embedding feature can be multiplied by the outer product calculation model to obtain an outer product calculation result. Then, the historical behavior embedding feature, the target data embedding feature, and the outer product calculation result can be input into the connection module to generate a feature. This feature is then processed by the activation module and the linear module to obtain the required weights.

[0095] Optionally, when determining the target interest features of the target object for any candidate data based on the initial interest embedding features, feature extraction can be performed directly on the initial interest embedding features, and the result of feature extraction can be used as the target interest features of the target object for any candidate data.

[0096] Optionally, sparse features can be constructed based on the historical behavior data of the target object. After obtaining these sparse features, vector embedding can be performed to obtain one or more of wide embeddings and dense embeddings. Different vector embedding processes can be set according to requirements to obtain different types of embedding features; wide and dense embeddings are embedding features of different dimensions, with the dimension of dense embeddings being greater than that of wide embeddings. Then, the target interest features of the target object for any candidate data can be determined based on the embedding features obtained from the vector embedding process (such as one or more of wide and dense embeddings) and the initial interest embedding features. The following details the determination of the target interest features under different conditions using the embedding features obtained from the vector embedding process. It should be noted that the historical behavior embedding vectors mentioned above are also obtained by performing vector embedding on the sparse features here.

[0097] By combining the above implementation with the target model, it can be seen that sparse features can be input into the embedding module of the target model. The output of the embedding module can be one or more of broad embedding features and dense embedding features. Then, the output of the embedding module and the output of the feature weighting module (i.e. the initial interest embedding features) can be input into the feature extraction submodule. The output of the feature extraction submodule is the target interest feature.

[0098] Case (1): When the embedding features obtained from vector embedding processing only include broad embedding features, the target interest features of the target object for any candidate data can be determined based on the broad embedding features and the initial interest embedding features. In one embodiment, a first feature extraction can be performed on the broad embedding features to obtain first-order features; and a third feature extraction can be performed on the initial interest embedding features to obtain higher-order features; further, the first-order features and higher-order features can be concatenated to obtain the target interest features of the target object for any candidate data.

[0099] In this case, the feature extraction submodule in the target model can include a first feature module, a third feature module, and a concatenation module, such as... Figure 5eAs shown in the diagram, the first feature module can extract first features from the broad embedding features to obtain first-order features; the third feature module can extract third features from the initial interest embedding features to obtain higher-order features; and the concatenation module can concatenate the first-order features and higher-order features to obtain the target interest features of the target object for any candidate data.

[0100] Case (2): When the embedding features obtained from vector embedding processing only include dense embedding features, the target interest features of the target object for any candidate data can be determined based on the dense embedding features and the initial interest embedding features. In one embodiment, a second feature extraction can be performed on the dense embedding features to obtain second-order features; and a third feature extraction can be performed on the initial interest embedding features to obtain higher-order features; further, the second-order features and higher-order features can be concatenated to obtain the target interest features of the target object for any candidate data. In another embodiment, a second feature extraction can be performed on the dense embedding features and the initial interest embedding features to obtain second-order features; and a third feature extraction can be performed on the dense embedding features and the initial interest embedding features to obtain higher-order features; further, the second-order features and higher-order features can be concatenated to obtain the target interest features of the target object for any candidate data.

[0101] In this scenario, the feature extraction submodule in the target model can include a second feature module, a third feature module, and a concatenation module, such as... Figure 5f As shown in the diagram. The second feature module can extract second features from the dense embedding features to obtain second-order features; or, it can extract second features from both the dense embedding features and the initial interest embedding features to obtain second-order features. The third feature module can extract third features from the initial interest embedding features to obtain higher-order features; or, it can extract third features from both the dense embedding features and the initial interest embedding features to obtain higher-order features. The concatenation module can concatenate the second-order features and the higher-order features to obtain the target interest features of the target object for any candidate data.

[0102] Case (3): When the embedding features obtained from vector embedding processing include broad embedding features and dense embedding features, the target interest features of the target object for any candidate data can be determined based on the broad embedding features, dense embedding features, and initial interest embedding features. In one embodiment, a first feature extraction can be performed on the broad embedding features to obtain first-order features; a second feature extraction can be performed on the dense embedding features to obtain second-order features; a third feature extraction can be performed on the initial interest embedding features to obtain higher-order features; further, the first-order features, second-order features, and higher-order features can be concatenated to obtain the target interest features of the target object for any candidate data. In another embodiment, a first feature extraction can be performed on the broad embedding features to obtain first-order features; a second feature extraction can be performed on the dense embedding features and initial interest embedding features to obtain second-order features; a third feature extraction can be performed on the dense embedding features and initial interest embedding features to obtain higher-order features; further, the first-order features, second-order features, and higher-order features can be concatenated to obtain the target interest features of the target object for any candidate data.

[0103] In this context, the feature extraction submodule in the target model can include a first feature module, a second feature module, a third feature module, and a concatenation module, such as... Figure 5g As shown in the diagram. The first feature module extracts first features from the broad embedding features to obtain first-order features. The second feature module extracts second features from the dense embedding features to obtain second-order features; or, it extracts second features from both the dense embedding features and the initial interest embedding features to obtain second-order features. The third feature module extracts third features from the initial interest embedding features to obtain higher-order features; or, it extracts third features from both the dense embedding features and the initial interest embedding features to obtain higher-order features. The concatenation module concatenates the second-order features and higher-order features to obtain the target interest features of the target object for any candidate data.

[0104] The first feature module can be a linear network, such as a Wide network. The second and third feature modules can be deep neural networks, such as an FM network for the second feature module and an MLP network for the third feature module. By combining linear networks and deep neural networks, the memory capacity of shallow networks and the generalization capacity of deep neural networks can be comprehensively utilized. This allows the target model to simultaneously acquire both memory and generalization capabilities, thereby improving the model's performance. In other words, it can improve the accuracy of click-through rate prediction, and consequently, improve the accuracy and effectiveness of data recommendation. The first-order, second-order, and higher-order features mentioned above are generated based on different networks. Generally, if the network processing involves a linear combination, the network outputs a first-order feature; if the network processing involves n linear combinations, the network outputs an nth-order feature.

[0105] S405, based on the target object's target interest features for each candidate data, predicts the click-through rate of each candidate data separately.

[0106] In one implementation, the click-through rate (CTR) for any candidate data can be obtained by calling the prediction module of the target model. For example, for any candidate data, the target interest features of the candidate data can be input into the prediction module to predict the CTR of that candidate data. That is, the output of the prediction module is the CTR of that candidate data. Optionally, the prediction module can be a Sigmoid network structure.

[0107] S406 recommends data based on the click-through rate of each candidate data.

[0108] The specific implementation of this step can be found in the description of step S204 above, and will not be repeated here.

[0109] Based on the above description of the target model, the target model can also be as follows: Figure 5h As shown.

[0110] In one implementation, to verify the effectiveness of data recommendation based on fatigue information proposed in this application, the recommendation effects in two scenarios are compared below. The first scenario involves obtaining the click-through rate of candidate data without utilizing fatigue information and then recommending data based on that click-through rate. The second scenario involves obtaining the click-through rate of candidate data using fatigue information and then recommending data based on that click-through rate. The model used in the data recommendation process can be a target model. For example... Figure 6(a) in the graph represents the change in average clicks per user with refresh count, and (b) represents the change in average reading time per user with refresh count. It can be seen that compared to the first scenario, the second scenario shows an increase in both average clicks per user and average reading time per user. Furthermore, comparing the recommendation performance in the first and second scenarios yields the following data: compared to the first scenario, the second scenario increases the average click-through rate by 1.23%, the average number of clicks per user by 1.73%, the average reading time per user by 1.34%, the average number of articles exposed per user by 0.59%, and the offline AUC by 0.44%. Based on the above data, it can be seen that modeling user fatigue in article-based data recommendation scenarios can reduce user fatigue, slow down the decline in click-through rate and reading time, and also enhance the dynamic adjustment capability of data in fine-tuning, thereby improving exposure click-through rate and average consumption (reading) time per user.

[0111] In one implementation, the fatigue feature proposed in this application embodiment can be applied to any recommendation model. Offline experiments can be used to evaluate the effectiveness of other recommendation models utilizing the fatigue feature. Specifically, the computer device can also acquire one or more reference recommendation models, such as DeepFM, AutoInt, and DFN. Then, each reference recommendation model can be used to process traditional features in traditional recommendation methods, and each reference recommendation model can be used to process the target features in the data recommendation method proposed in this application embodiment, to obtain performance data of each reference recommendation model under different features. Traditional features can refer to object attribute features, object historical behavior features, candidate data attribute features, and context features, etc. The target feature is the fatigue feature proposed in this application embodiment added to the traditional features. After obtaining the performance data of each reference recommendation model under different feature data, performance analysis can be performed based on this performance data to verify the effectiveness of adding fatigue features to the data recommendation scenario proposed in this application embodiment.

[0112] For example, Table 1 shows the performance data for different feature processing using the three models DeepFM, AutoInt, and DFN.

[0113] Table 1

[0114]

[0115] Here, AUC represents the model's prediction accuracy, and RelaImpr measures the model's relative improvement. Table 1 shows that, under different model conditions, the AUC corresponding to processing the target features using the model is higher than the AUC corresponding to processing traditional features using the model. Furthermore, the RelaImpr values ​​obtained from processing the target features using the model are all positive. Therefore, processing the target features using the model is more effective than processing traditional features, and this also verifies that fusing fatigue features in data recommendation processing, as proposed in this application, can improve data recommendation performance.

[0116] In this embodiment, features can be used to model both coarse-grained and global user fatigue levels. This involves incorporating fatigue-related information and corresponding features into historical user behavior data to predict the click-through rate (CTR) of each candidate data point. This multi-dimensional feature-based comprehensive prediction improves accuracy. Furthermore, the introduction of fatigue information and corresponding features helps to minimize user fatigue during consumption, slowing the decline in CTR and consumption duration, thus enhancing data recommendation effectiveness. Simultaneously, recommending data based on the predicted CTR of each candidate data point ensures that the data recommended to the target audience has a higher CTR, further increasing the CTR of the recommended data and consequently extending the target audience's consumption duration, thus improving overall data recommendation effectiveness.

[0117] Please see Figure 7 This is a schematic diagram of a data recommendation device provided in an embodiment of this application. The data recommendation device described in this embodiment includes:

[0118] The acquisition unit 701 is used to acquire attribute information of multiple candidate data and acquire fatigue information of the target object for each candidate data.

[0119] The determining unit 702 is used to determine the target data embedding features of each candidate data based on the corresponding attribute information and fatigue information of each candidate data;

[0120] The prediction unit 703 is used to obtain the historical behavior embedding features corresponding to the historical behavior data of the target object, and predict the click rate of each candidate data based on the historical behavior embedding features and the target data embedding features of each candidate data.

[0121] The recommendation unit 704 is used to recommend data to the target object based on the click rate of each candidate data.

[0122] In one implementation, the determining unit 702 is specifically used for:

[0123] For any candidate data, the data embedding feature of the candidate data is determined based on the attribute information of the candidate data, and the fatigue embedding feature of the candidate data is determined based on the fatigue information of the candidate data.

[0124] The data embedding features and the fatigue embedding features are fused together to obtain the target data embedding features for any candidate data.

[0125] In one implementation, the fatigue information of any candidate data includes: first behavioral information generated by the target object within a target time period, and second behavioral information generated by the target object within the target time period that is associated with any candidate data; the determining unit 702 is specifically used for:

[0126] The global fatigue feature of any candidate data is determined based on the first behavioral information, and the attribute fatigue feature of any candidate data is determined based on the second behavioral information.

[0127] Vector embedding is performed on the global fatigue feature and the attribute fatigue feature to obtain the fatigue embedding feature of any candidate data.

[0128] In one implementation, the first behavioral information includes the target data viewing volume; the determining unit 702 is specifically used for:

[0129] Construct multiple bucket categories for the target data viewing volume, and determine the bucket characteristics corresponding to each bucket category; each bucket category corresponds to a viewing volume range for the data viewing volume.

[0130] Find the target bucket category corresponding to the target data view count from the multiple bucket categories, and use the bucket feature corresponding to the target bucket category as the global fatigue feature of the target data view count.

[0131] In one implementation, the prediction unit 703 is specifically used for:

[0132] Based on the historical behavior embedding features and the target data embedding features of each candidate data, the target interest features of the target object for each candidate data are determined.

[0133] Based on the target object's interest features for each candidate data, the click-through rate of each candidate data is predicted.

[0134] In one implementation, the historical behavior data includes one or more historical recommendation data, with each historical recommendation data corresponding to a historical behavior embedding vector; the prediction unit 703 is specifically used for:

[0135] For any candidate data, the correlation between the target data embedding feature and each historical behavior embedding feature of the candidate data is calculated to obtain the weight of each historical behavior embedding feature.

[0136] Based on the historical behavior embedding features and their corresponding weights, a weighted summation process is performed to obtain the initial interest embedding features of the target object for any candidate data.

[0137] Based on the initial interest embedding features, the target interest features of the target object for any candidate data are determined.

[0138] In one implementation, the prediction unit 703 is specifically used for:

[0139] Sparse features are constructed based on the historical behavior data of the target object;

[0140] Vector embedding is performed on the sparse features to obtain broad embedding features and dense embedding features;

[0141] Based on the broad embedding features, the dense embedding features, and the initial interest embedding features, the target interest features of the target object for any candidate data are determined.

[0142] In one implementation, the prediction unit 703 is specifically used for:

[0143] The broad embedding features are subjected to first feature extraction to obtain first-order features;

[0144] A second feature extraction is performed on the dense embedding features and the initial interest embedding features to obtain second-order features;

[0145] A third feature extraction is performed on the dense embedding features and the initial interest embedding features to obtain higher-order features;

[0146] The first-order features, the second-order features, and the higher-order features are concatenated to obtain the target interest features of the target object for any candidate data.

[0147] It is understood that the division of units in this embodiment is illustrative and merely a logical functional division; in actual implementation, there may be other division methods. The functional units in this embodiment 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. The integrated units can be implemented in hardware or as software functional units.

[0148] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor 801 and a memory 802. Optionally, the computer device may also include a network interface 803. The processor 801, memory 802, and network interface 803 can exchange data.

[0149] The processor 801 described above can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0150] The aforementioned memory 802 may include read-only memory and random access memory, and provides program instructions and data to the processor 801. A portion of the memory 802 may also include non-volatile random access memory. The processor 801 executes the program instructions by calling them.

[0151] Obtain attribute information from multiple candidate data sets, and obtain fatigue information of the target object for each candidate data set;

[0152] Based on the corresponding attribute information and fatigue information of each candidate data, the target data embedding features of each candidate data are determined;

[0153] Obtain the historical behavior embedding features corresponding to the historical behavior data of the target object, and predict the click-through rate of each candidate data based on the historical behavior embedding features and the target data embedding features of each candidate data respectively;

[0154] Data recommendations are made for the target object based on the click-through rate of each candidate data.

[0155] In one implementation, the processor 801 is specifically used for:

[0156] For any candidate data, the data embedding feature of the candidate data is determined based on the attribute information of the candidate data, and the fatigue embedding feature of the candidate data is determined based on the fatigue information of the candidate data.

[0157] The data embedding features and the fatigue embedding features are fused together to obtain the target data embedding features for any candidate data.

[0158] In one implementation, the fatigue information of any candidate data includes: first behavioral information generated by the target object within a target time period, and second behavioral information generated by the target object within the target time period and associated with any candidate data; the processor 801 is specifically used for:

[0159] The global fatigue feature of any candidate data is determined based on the first behavioral information, and the attribute fatigue feature of any candidate data is determined based on the second behavioral information.

[0160] Vector embedding is performed on the global fatigue feature and the attribute fatigue feature to obtain the fatigue embedding feature of any candidate data.

[0161] In one implementation, the first behavioral information includes the target data viewing volume; the processor 801 is specifically used for:

[0162] Construct multiple bucket categories for the target data viewing volume, and determine the bucket characteristics corresponding to each bucket category; each bucket category corresponds to a viewing volume range for the data viewing volume.

[0163] Find the target bucket category corresponding to the target data view count from the multiple bucket categories, and use the bucket feature corresponding to the target bucket category as the global fatigue feature of the target data view count.

[0164] In one implementation, the processor 801 is specifically used for:

[0165] Based on the historical behavior embedding features and the target data embedding features of each candidate data, the target interest features of the target object for each candidate data are determined.

[0166] Based on the target object's interest features for each candidate data, the click-through rate of each candidate data is predicted.

[0167] In one implementation, the historical behavior data includes one or more historical recommendation data, with each historical recommendation data corresponding to a historical behavior embedding vector; the processor 801 is specifically used for:

[0168] For any candidate data, the correlation between the target data embedding feature and each historical behavior embedding feature of the candidate data is calculated to obtain the weight of each historical behavior embedding feature.

[0169] Based on the historical behavior embedding features and their corresponding weights, a weighted summation process is performed to obtain the initial interest embedding features of the target object for any candidate data.

[0170] Based on the initial interest embedding features, the target interest features of the target object for any candidate data are determined.

[0171] In one implementation, the processor 801 is specifically used for:

[0172] Sparse features are constructed based on the historical behavior data of the target object;

[0173] Vector embedding is performed on the sparse features to obtain broad embedding features and dense embedding features;

[0174] Based on the broad embedding features, the dense embedding features, and the initial interest embedding features, the target interest features of the target object for any candidate data are determined.

[0175] In one implementation, the processor 801 is specifically used for:

[0176] The broad embedding features are subjected to first feature extraction to obtain first-order features;

[0177] A second feature extraction is performed on the dense embedding features and the initial interest embedding features to obtain second-order features;

[0178] A third feature extraction is performed on the dense embedding features and the initial interest embedding features to obtain higher-order features;

[0179] The first-order features, the second-order features, and the higher-order features are concatenated to obtain the target interest features of the target object for any candidate data.

[0180] This application also provides a computer storage medium storing program instructions, which, when executed, may include, for example... Figure 2 or Figure 4 Some or all of the steps of the data recommendation method in the corresponding embodiments.

[0181] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0182] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0183] This application also provides a computer program product or computer program that includes program instructions that, when executed by a processor, can implement some or all of the steps in the methods described above. For example, the program instructions are stored in a computer-readable storage medium. A processor of a computer device reads the program instructions from the computer-readable storage medium and executes the program instructions, causing the computer device to perform the steps executed in the embodiments of the methods described above.

[0184] The data recommendation method, computer device, and storage medium provided in the embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A data recommendation method, characterized in that, The method includes: Obtain attribute information from multiple candidate data sets, and obtain fatigue information of the target object for each candidate data set; Based on the corresponding attribute information and fatigue information of each candidate data, the target data embedding feature of each candidate data is determined; wherein, the fatigue information of any candidate data includes: first behavioral information generated by the target object within a target time period, and second behavioral information generated by the target object within the target time period that is associated with any candidate data; the method for determining the target data embedding feature of any candidate data includes: for any candidate data, determining the data embedding feature of any candidate data based on the attribute information of any candidate data; determining the global fatigue feature of any candidate data based on the first behavioral information, and determining the attribute fatigue feature of any candidate data based on the second behavioral information; performing vector embedding on the global fatigue feature and the attribute fatigue feature to obtain the fatigue embedding feature of any candidate data; and fusing the data embedding feature and the fatigue embedding feature to obtain the target data embedding feature of any candidate data; Obtain the historical behavior embedding features corresponding to the historical behavior data of the target object, and predict the click-through rate of each candidate data based on the historical behavior embedding features and the target data embedding features of each candidate data respectively; Data recommendations are made for the target object based on the click-through rate of each candidate data.

2. The method according to claim 1, characterized in that, The first behavioral information includes the target data viewing volume; determining the global fatigue feature of any candidate data based on the first behavioral information includes: Construct multiple bucket categories for the target data viewing volume, and determine the bucket characteristics corresponding to each bucket category; each bucket category corresponds to a viewing volume range for the data viewing volume. Find the target bucket category corresponding to the target data view count from the multiple bucket categories, and use the bucket feature corresponding to the target bucket category as the global fatigue feature of the target data view count.

3. The method according to any one of claims 1-2, characterized in that, The step of predicting the click-through rate of each candidate data based on the historical behavior embedding features and the target data embedding features of each candidate data includes: Based on the historical behavior embedding features and the target data embedding features of each candidate data, the target interest features of the target object for each candidate data are determined. Based on the target object's interest features for each candidate data, the click-through rate of each candidate data is predicted.

4. The method according to claim 3, characterized in that, The historical behavior data includes one or more historical recommendation data sets, with each historical recommendation data set corresponding to a historical behavior embedding vector; determining the target interest features of the target object for each candidate data set based on the historical behavior embedding features and the target data embedding features of each candidate data set includes: For any candidate data, the correlation between the target data embedding feature and each historical behavior embedding feature of the candidate data is calculated to obtain the weight of each historical behavior embedding feature. Based on the historical behavior embedding features and their corresponding weights, a weighted summation process is performed to obtain the initial interest embedding features of the target object for any candidate data. Based on the initial interest embedding features, the target interest features of the target object for any candidate data are determined.

5. The method according to claim 4, characterized in that, Determining the target interest features of the target object for any candidate data based on the initial interest embedding features includes: Sparse features are constructed based on the historical behavior data of the target object; Vector embedding is performed on the sparse features to obtain broad embedding features and dense embedding features; Based on the broad embedding features, the dense embedding features, and the initial interest embedding features, the target interest features of the target object for any candidate data are determined.

6. The method according to claim 5, characterized in that, The step of determining the target interest features of the target object for any candidate data based on the broad embedding features, the dense embedding features, and the initial interest embedding features includes: The broad embedding features are subjected to first feature extraction to obtain first-order features; A second feature extraction is performed on the dense embedding features and the initial interest embedding features to obtain second-order features; A third feature extraction is performed on the dense embedding features and the initial interest embedding features to obtain higher-order features; The first-order features, the second-order features, and the higher-order features are concatenated to obtain the target interest features of the target object for any candidate data.

7. A computer device, characterized in that, The device includes a processor and a memory, wherein the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to perform the method as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed, are used to implement the method as described in any one of claims 1-6.

9. A data recommendation device, characterized in that, Includes a unit for performing the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, The computer program product includes program instructions that, when executed by a processor, implement the method as described in any one of claims 1-6.

Images