Object recommendation method, apparatus, device, storage medium, and computer program product

Abstract

The application relates to an object recommendation method, device, equipment, storage medium and computer program product. The method comprises the following steps: acquiring a first feature set comprising first features corresponding to each to-be-recommended object; extracting the relationship between the to-be-recommended objects based on the first feature set, obtaining second features corresponding to each to-be-recommended object; generating ranking estimation results of each to-be-recommended object under multiple ranking tasks according to the first features and the second features; for each to-be-recommended object, fusing the ranking estimation results of the to-be-recommended object under each ranking task to obtain a target ranking estimation result corresponding to the to-be-recommended object; ranking each to-be-recommended object according to the target ranking estimation result corresponding to each to-be-recommended object, and performing object recommendation based on the ranked result. The method can improve the recommendation accuracy of objects.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an object recommendation method, apparatus, device, storage medium, and computer program product. Background Art

[0002] With the development of computer technology, recommendation technology has emerged. Recommendation technology is a technology that can quickly and effectively filter out objects that meet user preferences from a large number of objects. Based on recommendation technology, recommendation systems can be built to recommend objects that meet user preferences.

[0003] Traditional recommendation systems typically consider only the single-point relationship between users and objects, using this as the sole criterion for ranking recommended objects. These objects are then recommended to users in the order of their ranking. Obviously, traditional methods of ranking recommended objects rely on a single criterion, resulting in inaccurate ranking results and low recommendation accuracy. Summary of the Invention

[0004] Based on this, it is necessary to provide an object recommendation method, device, equipment, storage medium and computer program product that can improve the recommendation accuracy in response to the above technical problems.

[0005] An object recommendation method, comprising:

[0006] Obtaining a set of first features; the set including first features corresponding to each object to be recommended; the first features including features representing the correlation between the corresponding object to be recommended and the user;

[0007] Extracting the relationship between the objects to be recommended based on the set of the first features to obtain second features corresponding to each object to be recommended; the second features are used to characterize the correlation between the objects to be recommended;

[0008] generating, based on the first feature and the second feature, estimated ranking results of each to-be-recommended object under multiple ranking tasks;

[0009] For each object to be recommended, the estimated ranking results of the object to be recommended under each of the ranking tasks are integrated to obtain the target ranking estimated result corresponding to the object to be recommended;

[0010] The objects to be recommended are sorted according to the target ranking estimation results corresponding to the objects to be recommended, and the objects are recommended based on the sorted results.

[0011] An object recommendation device, comprising:

[0012] An acquisition module is configured to acquire a set of first features, wherein the set includes first features corresponding to each object to be recommended, wherein the first features include features representing the correlation between the corresponding object to be recommended and the user;

[0013] an extraction module, configured to extract the relationship between the objects to be recommended based on the set of the first features, and obtain a second feature corresponding to each object to be recommended; the second feature is used to characterize the correlation between the objects to be recommended;

[0014] A generating module, configured to generate, based on the first feature and the second feature, an estimated ranking result of each to-be-recommended object under a plurality of ranking tasks;

[0015] A fusion module is used to fuse the ranking estimation results of each to-be-recommended object under each of the ranking tasks to obtain a target ranking estimation result corresponding to the to-be-recommended object;

[0016] The recommendation module is used to sort the objects to be recommended according to the target sorting estimation results corresponding to the objects to be recommended, and to recommend objects based on the sorted results.

[0017] In one embodiment, the acquisition module is also used to obtain the object attribute features and personalized semantic features corresponding to each object to be recommended; the personalized semantic features are used to characterize the correlation between the corresponding object to be recommended and the user; the object attribute features and the personalized semantic features corresponding to the same object to be recommended are combined to obtain the first features corresponding to each object to be recommended; and the set of first features is determined based on the first features corresponding to each object to be recommended.

[0018] In one embodiment, the first feature includes a first feature vector; the object attribute feature includes an object attribute feature vector; the personalized semantic feature includes a personalized semantic feature vector; the acquisition module is further used to perform vector splicing on the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended, to obtain the first feature vector corresponding to each object to be recommended.

[0019] In one embodiment, the objects to be recommended are objects to be recommended after initial sorting based on corresponding personalized semantic feature vectors; the acquisition module is also used to position encode each first feature vector according to the position of the object to be recommended after initial sorting to obtain a set of first feature vectors; in the set of first feature vectors, each first feature vector is arranged in the order of position between the corresponding objects to be recommended.

[0020] In one embodiment, the generation module is further used to fuse the first feature and the second feature corresponding to the same object to be recommended to obtain the fused feature corresponding to the object to be recommended; under each sorting task, the sorting of each object to be recommended is estimated according to the fused feature to obtain the sorting estimation results of each object to be recommended under multiple sorting tasks.

[0021] In one embodiment, the generation module is also used to adjust the first features corresponding to each object to be recommended according to the adjustment coefficient corresponding to the sorting task for each sorting task, so as to obtain the adjusted first features of each object to be recommended under different sorting tasks; under each sorting task, the adjusted first features and second features corresponding to the same object to be recommended are subjected to feature fusion, so as to obtain the fused features of each object to be recommended under multiple sorting tasks.

[0022] In one embodiment, the generation module is also used to predict the status information of each object to be recommended under the sorting task based on the fusion features for each sorting task; and to perform a sorting estimate on each object to be recommended according to the status information of each object to be recommended under the sorting task to obtain a sorting estimate result of each object to be recommended under the sorting task.

[0023] In one embodiment, the ranking estimation result includes a ranking estimation score; the target ranking estimation result includes a target ranking estimation score; the fusion module performs weighted fusion on the ranking estimation scores of each object to be recommended under each ranking task to obtain the target ranking estimation score corresponding to the object to be recommended.

[0024] In one embodiment, the extraction module is also used to input the set of first features into a trained object ranking model, so as to extract the relationship between the objects to be recommended based on the set of first features in the feature encoding layer of the trained object ranking model, and obtain the second features corresponding to each object to be recommended; the generation module is also used to obtain the ranking estimation results of each object to be recommended under the corresponding ranking task according to the first features and the second features corresponding to the same object to be recommended through the ranking estimation unit corresponding to each ranking task in the trained object ranking model.

[0025] In one embodiment, the apparatus further comprises:

[0026] A training module is used to obtain a sample set including multiple groups of training samples; each group of training samples includes a first sample feature of a corresponding sample object and a state label of the sample object under multiple sorting tasks; the first sample feature includes a feature that characterizes the correlation between the sample object and the user; based on the sample set, the object sorting model to be trained is trained to obtain an object sorting model for sorting objects according to the multiple sorting tasks.

[0027] In one embodiment, the training module is also used to input the sample set into the object sorting model to be trained in each round of iterative training, so as to extract the relationship between the sample objects based on the first sample features in each group of training samples in the feature coding layer of the object sorting model to be trained, and obtain the second sample features corresponding to each sample object; for each sorting task, through the sorting prediction unit corresponding to the sorting task in the object sorting model to be trained, based on the first sample features and the second sample features corresponding to the same sample object, predict the state information of each sample object under the corresponding sorting task; according to the difference between the predicted state information and the state label, update the model parameters of the object sorting model to be trained until the iteration stop condition is reached, and obtain the object sorting model for sorting objects according to the multiple sorting tasks.

[0028] In one embodiment, the training module is also used to determine, for each sorting task, a loss value between the predicted state information and the state label through an objective function corresponding to the task type to which the sorting task belongs; and update the model parameters of the object sorting model to be trained based on the loss value.

[0029] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0030] Obtaining a set of first features; the set including first features corresponding to each object to be recommended; the first features including features representing the correlation between the corresponding object to be recommended and the user;

[0031] Extracting the relationship between the objects to be recommended based on the set of the first features to obtain second features corresponding to each object to be recommended; the second features are used to characterize the correlation between the objects to be recommended;

[0032] generating, based on the first feature and the second feature, estimated ranking results of each to-be-recommended object under multiple ranking tasks;

[0033] For each to-be-recommended object, the ranking estimation results of the to-be-recommended object under each ranking task are fused to obtain a target ranking estimation result corresponding to the to-be-recommended object;

[0034] According to the target ranking estimation result corresponding to each to-be-recommended object, the to-be-recommended objects are ranked, and object recommendation is performed based on the ranked results.

[0035] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the following steps:

[0036] A set of first features is obtained; the set includes first features corresponding to each to-be-recommended object respectively; the first features include features representing the correlation between the corresponding to-be-recommended object and the user;

[0037] Based on the set of first features, relationships between the to-be-recommended objects are extracted to obtain second features corresponding to each to-be-recommended object respectively; the second features are used to represent the correlation between the to-be-recommended objects;

[0038] According to the first features and the second features, ranking estimation results of each to-be-recommended object under multiple ranking tasks are generated;

[0039] For each to-be-recommended object, the ranking estimation results of the to-be-recommended object under each ranking task are fused to obtain a target ranking estimation result corresponding to the to-be-recommended object;

[0040] According to the target ranking estimation result corresponding to each to-be-recommended object, the to-be-recommended objects are ranked, and object recommendation is performed based on the ranked results.

[0041] A computer program product comprising a computer program, the computer program being executed by a processor to implement the following steps:

[0042] A set of first features is obtained; the set includes first features corresponding to each to-be-recommended object respectively; the first features include features representing the correlation between the corresponding to-be-recommended object and the user;

[0043] Based on the set of first features, relationships between the to-be-recommended objects are extracted to obtain second features corresponding to each to-be-recommended object respectively; the second features are used to represent the correlation between the to-be-recommended objects;

[0044] According to the first features and the second features, ranking estimation results of each to-be-recommended object under multiple ranking tasks are generated;

[0045] For each object to be recommended, the estimated ranking results of the object to be recommended under each of the ranking tasks are integrated to obtain the target ranking estimated result corresponding to the object to be recommended;

[0046] The objects to be recommended are sorted according to the target ranking estimation results corresponding to the objects to be recommended, and object recommendations are performed based on the sorted results.

[0047] The object recommendation method, apparatus, device, storage medium and computer program product described above obtain a set of first features corresponding to each object to be recommended, wherein the first features include features that characterize the correlation between the corresponding object to be recommended and the user. Based on the set of first features, the relationship between the objects to be recommended can be extracted to obtain second features corresponding to each object to be recommended, and the second features can be used to characterize the correlation between the objects to be recommended. Based on the first features used to characterize the correlation between the corresponding object to be recommended and the user, and the second features used to characterize the correlation between the objects to be recommended, more accurate ranking estimation results can be generated for each object to be recommended under multiple ranking tasks. For each object to be recommended, the more accurate ranking estimation results of the object to be recommended under each ranking task can be fused to obtain a more accurate target ranking estimation result corresponding to the object to be recommended. By ranking each object to be recommended according to the more accurate target ranking estimation result corresponding to each object to be recommended, a more accurate ranking result can be obtained, and then object recommendation can be performed based on the more accurate ranking result, which can improve the recommendation accuracy of the object. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A diagram illustrating an application environment of an object recommendation method according to an embodiment;

[0049] Figure 2 A flowchart of an object recommendation method according to an embodiment;

[0050] Figure 3 A schematic diagram of a process for determining a set of first features in one embodiment;

[0051] Figure 4 A flowchart of an object recommendation method according to another embodiment;

[0052] Figure 5 is a structural block diagram of an object sorting model in one embodiment;

[0053] Figure 6 Schematic diagram of a flow chart of an object recommendation method in another embodiment;

[0054] Figure 7 A schematic diagram of the improvement rate of each sorting task after sorting objects using the object recommendation method of the present application in one embodiment;

[0055] Figure 8 is a structural block diagram of an object recommendation device in one embodiment;

[0056] Figure 9 is a structural block diagram of an object recommendation device in another embodiment;

[0057] Figure 10 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0059] The object recommendation method provided in this application can be applied to Figure 1 In the application environment shown, the terminal 102 communicates with the server 104 via a network. The terminal 102 may be, but is not limited to, various personal computers, laptops, smartphones, tablets, smart car networking, AR / VR, and portable wearable devices. The server 104 may be an independent physical server, or a server cluster or distributed system consisting of multiple physical servers. It may also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Each server may be a node in a blockchain. The terminal 102 and the server 104 may be directly or indirectly connected via wired or wireless communication, which is not limited in this application.

[0060] The server 104 can obtain a set of first features; the set includes first features corresponding to each object to be recommended; the first features include features that characterize the correlation between the corresponding object to be recommended and the user. The server 104 can extract the relationship between the objects to be recommended based on the set of first features, and obtain second features corresponding to each object to be recommended; the second features are used to characterize the correlation between the objects to be recommended. The server 104 can generate a ranking estimate result for each object to be recommended under multiple ranking tasks based on the first features and the second features, and for each object to be recommended, integrate the ranking estimate results of the object to be recommended under each ranking task to obtain a target ranking estimate result corresponding to the object to be recommended. The server 104 can rank each object to be recommended according to the target ranking estimate result corresponding to each object to be recommended, and recommend objects based on the ranked results. The terminal 102 can receive and display the objects recommended by the server 104.

[0061] It should be noted that the object recommendation methods in some embodiments of this application utilize artificial intelligence technology. For example, the second features corresponding to each recommended object are features encoded using artificial intelligence technology, and the estimated ranking results of each recommended object under multiple ranking tasks are also estimated ranking results predicted using artificial intelligence technology.

[0062] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to 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 seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also involves studying the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0063] Artificial intelligence technology is a comprehensive discipline covering a wide range of fields, encompassing both hardware and software technologies. Basic AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning. It is understood that the virtual characters in some embodiments of this application, which are not controlled by real users, are equivalent to intelligent robot models trained using machine learning techniques.

[0064] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0065] In one embodiment, Figure 2 As shown, an object recommendation method is provided, which can be applied to the server or to the interaction process between the terminal and the server. Figure 1 Taking the server 104 in the example as an example, the following steps are included:

[0066] At step 202, a set of first features is obtained, the set of first features comprising respective first features of respective to-be-recommended objects, and the first features comprising a feature representing a correlation between a corresponding to-be-recommended object and the user.

[0067] The to-be-recommended object is an object to be recommended.

[0068] Specifically, the server can obtain the to-be-recommended objects, and determine the set of first features according to the respective first features of the respective to-be-recommended objects.

[0069] In an embodiment, the server can obtain the feature representing the correlation between the to-be-recommended object and the user, and determine the respective first features of the respective to-be-recommended objects according to the feature representing the correlation between the to-be-recommended object and the user, to obtain the set of first features.

[0070] In an embodiment, the first feature can be the feature representing the correlation between the corresponding to-be-recommended object and the user. For each to-be-recommended object, the server can directly determine the feature representing the correlation between the to-be-recommended object and the user as the first feature corresponding to the to-be-recommended object.

[0071] In another embodiment, the first feature can comprise, in addition to the feature representing the correlation between the corresponding to-be-recommended object and the user, other features related to the to-be-recommended object (i.e., features other than the feature representing the correlation between the corresponding to-be-recommended object and the user and related to the to-be-recommended object). Furthermore, for each to-be-recommended object, the server can perform feature combination on the feature representing the correlation between the to-be-recommended object and the user and the other features related to the to-be-recommended object, to obtain the respective first features of the respective to-be-recommended objects, and determine the set of first features according to the respective first features of the respective to-be-recommended objects.

[0072] In an embodiment, the correlation between the to-be-recommended object and the user can comprise at least one of a historical preference behavior feature of the to-be-recommended object and the user, an attribute feature of the user, and the like.

[0073] In an embodiment, the to-be-recommended object can specifically comprise at least one of a video object, a picture object, a document object, an advertisement object, and a commodity object, and the like.

[0074] At step 204, a relationship between the to-be-recommended objects is extracted based on the set of first features, to obtain respective second features of the respective to-be-recommended objects, and the second features being used to represent a correlation between the to-be-recommended objects.

[0075] Specifically, the server can extract the relationship between the respective to-be-recommended objects based on the set of first features, to obtain the respective second features of the respective to-be-recommended objects.

[0076] In one embodiment, the server may input the set of first features into a feature extraction network that can extract the relationships between objects to be recommended. The feature extraction network may then extract the relationships between objects to be recommended based on the set of first features to obtain second features corresponding to each object to be recommended.

[0077] Step 206 : Generate estimated ranking results of each to-be-recommended object under multiple ranking tasks based on the first feature and the second feature.

[0078] The ranking task is to rank the recommended objects. Different ranking tasks rank the recommended objects along different dimensions. In other words, each ranking task corresponds to one dimension. The estimated ranking result is the result obtained after the ranking of each recommended object is estimated under each ranking task. In other words, each ranking task corresponds to one estimated ranking result.

[0079] Specifically, the server may perform a ranking estimate on each object to be recommended under multiple ranking tasks based on the first feature and the second feature, and generate a ranking estimate result for each object to be recommended under each ranking task.

[0080] In one embodiment, the server may perform feature processing on the first and second features corresponding to the same recommended object to obtain processed features corresponding to the recommended object. The server may then perform a ranking estimate for each recommended object based on the processed features for each ranking task, obtaining a ranking estimate for each recommended object under multiple ranking tasks.

[0081] In one embodiment, the server may input the first feature and the second feature corresponding to the same object to be recommended into a ranking estimation network that can perform ranking estimation on each object to be recommended. Then, the ranking estimation network may perform ranking estimation on each object to be recommended based on the first feature and the second feature corresponding to the same object to be recommended, and obtain the ranking estimation result of each object to be recommended under the corresponding ranking task.

[0082] In one embodiment, multiple sorting tasks can sort each object to be recommended based on at least two dimensions, such as likes, push, comments, attention, favorites, forwarding, and browsing time. For example, there are three sorting tasks A, B, and C. Sorting task A can correspond to the like dimension, that is, from the like dimension, sort the recommended objects according to the likes of each object to be recommended. Sorting task B corresponds to the push dimension, that is, from the comment dimension, sort the recommended objects according to the comments on each object to be recommended. Sorting task C corresponds to the forwarding dimension, that is, from the forwarding dimension, sort the recommended objects according to the forwarding status of each object to be recommended.

[0083] At step 208, for each to-be-recommended object, the ranking estimation results of the to-be-recommended object under the ranking tasks are fused to obtain a target ranking estimation result corresponding to the to-be-recommended object.

[0084] The target ranking estimation result is a ranking estimation result as a target, and is used for ranking estimation of each to-be-recommended object.

[0085] Specifically, the server can fuse, for each to-be-recommended object, the ranking estimation results of the to-be-recommended object under the ranking tasks to obtain a target ranking estimation result corresponding to the to-be-recommended object.

[0086] In one embodiment, the ranking estimation result includes a ranking estimation position, and the target ranking estimation result includes a target ranking estimation position. The server can fuse, for each to-be-recommended object, the ranking estimation positions of the to-be-recommended object under the ranking tasks to obtain a target ranking estimation position corresponding to the to-be-recommended object. The ranking estimation position is a position of the to-be-recommended object in a ranked list estimated. The target ranking estimation position is a position of the to-be-recommended object in a ranked list obtained after fusing the ranking estimation positions of the to-be-recommended object under the ranking tasks.

[0087] At step 210, each to-be-recommended object is ranked according to the target ranking estimation result corresponding to the to-be-recommended object, and object recommendation is performed based on the ranked result.

[0088] Specifically, the server can rank each to-be-recommended object according to the target ranking estimation result corresponding to the to-be-recommended object, and then the server can perform object recommendation based on the ranked result.

[0089] In one embodiment, the server can rank each to-be-recommended object according to the target ranking estimation result corresponding to the to-be-recommended object to obtain a ranked result, such as a ranked object list. Then, the server can send the ranked result to the terminal to implement object recommendation.

[0090] In the above-mentioned object recommendation method, a set of first features corresponding to each object to be recommended is obtained, and the first features include features that characterize the correlation between the corresponding object to be recommended and the user. Based on the set of first features, the relationship between the objects to be recommended can be extracted to obtain second features corresponding to each object to be recommended, and the second features can be used to characterize the correlation between the objects to be recommended. Based on the first features used to characterize the correlation between the corresponding object to be recommended and the user, and the second features used to characterize the correlation between the objects to be recommended, more accurate ranking estimation results can be generated for each object to be recommended under multiple sorting tasks. For each object to be recommended, the more accurate ranking estimation results of the object to be recommended under each sorting task can be fused to obtain a more accurate target ranking estimation result corresponding to the object to be recommended. By sorting the objects to be recommended according to the more accurate target ranking estimation results corresponding to each object to be recommended, a more accurate sorted result can be obtained, and then object recommendation can be performed based on the more accurate sorted result, which can improve the recommendation accuracy of the object.

[0091] In one embodiment, Figure 3 As shown, step 202, i.e., the step of obtaining a set of first features, specifically includes:

[0092] Step 302: Obtain object attribute features and personalized semantic features corresponding to each object to be recommended; the personalized semantic features are used to characterize the relevance between the corresponding object to be recommended and the user.

[0093] The object attribute features are attribute features corresponding to the object (item) to be recommended.

[0094] Specifically, the server may obtain object attribute features corresponding to each object to be recommended, and obtain personalized semantic features used to characterize the relevance between the corresponding object to be recommended and the user.

[0095] In one embodiment, the object attribute characteristics may include at least one of historical statistical characteristics of the object to be recommended and the object's own attribute characteristics.

[0096] Step 304 : combining the object attribute features and the personalized semantic features corresponding to the same object to be recommended to obtain first features corresponding to each object to be recommended.

[0097] Specifically, the server may combine the object attribute features and personalized semantic features corresponding to the same object to be recommended to obtain the first features corresponding to each object to be recommended.

[0098] In one embodiment, the first feature includes a first feature matrix, the object attribute feature includes an object attribute feature matrix, and the personalized semantic feature includes a personalized semantic feature matrix. The server may concatenate the object attribute feature matrix and the personalized semantic feature matrix corresponding to the same to-be-recommended object to obtain a first feature matrix corresponding to each to-be-recommended object.

[0099] Step 306: Determine a set of first features based on the first features corresponding to the objects to be recommended.

[0100] Specifically, the server may determine a set of first features according to the first features corresponding to the objects to be recommended.

[0101] In the above embodiment, by combining the object attribute features corresponding to the same object to be recommended and the personalized semantic features that can characterize the correlation between the corresponding object to be recommended and the user, a richer first feature can be obtained, which can further improve the accuracy of object recommendation.

[0102] In one embodiment, the first feature includes a first feature vector; the object attribute feature includes an object attribute feature vector; and the personalized semantic feature includes a personalized semantic feature vector. Step 304, which combines the object attribute feature and the personalized semantic feature corresponding to the same object to be recommended to obtain the first feature corresponding to each object to be recommended, specifically includes concatenating the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended to obtain the first feature vector corresponding to each object to be recommended.

[0103] Specifically, the server may perform vector concatenation of the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended, to obtain first feature vectors corresponding to each object to be recommended.

[0104] In the above embodiment, describing the first feature, object attribute features, and personalized semantic features in vector form makes their description more convenient and clear. Furthermore, by concatenating the object attribute feature vector and personalized semantic feature vector corresponding to the same recommended object, a richer first feature vector can be obtained, further improving the accuracy of object recommendations.

[0105] In one embodiment, each object to be recommended is an object to be recommended after initial sorting based on the corresponding personalized semantic feature vector; the step of determining a set of first features based on the first features corresponding to each object to be recommended specifically includes: position encoding each first feature vector based on the position of the object to be recommended after the initial sorting to obtain a set of first feature vectors; each first feature vector in the set of first feature vectors is arranged in order of position between the corresponding objects to be recommended.

[0106] Specifically, the server may perform an initial sorting of the objects to be recommended based on the corresponding personalized semantic feature vectors to obtain the sorted objects to be recommended. The server may obtain the object attribute feature vectors and personalized semantic feature vectors corresponding to each object to be recommended, and perform vector splicing of the object attribute feature vectors and personalized semantic feature vectors corresponding to the same object to be recommended to obtain the first feature vectors corresponding to each object to be recommended. Furthermore, the server may position-code each first feature vector according to the position of each object to be recommended after the initial sorting to obtain a set of first feature vectors. It can be understood that the server may specify the position order of each first feature vector in the set of first feature vectors according to the position of each object to be recommended after the initial sorting.

[0107] In the above embodiment, each first feature vector is position-encoded according to the position of the object to be recommended after the initial sorting to obtain a set of first feature vectors, so as to improve the accuracy of extracting the relationship between the objects to be recommended when the relationship between the objects to be recommended is subsequently extracted based on the set of first feature vectors.

[0108] In one embodiment, step 206, that is, the step of generating the estimated ranking results of each object to be recommended under multiple ranking tasks based on the first feature and the second feature, specifically includes: fusing the first feature and the second feature corresponding to the same object to be recommended to obtain the fused feature corresponding to the object to be recommended; under each ranking task, performing a ranking estimate on each object to be recommended based on the fused feature to obtain the estimated ranking results of each object to be recommended under multiple ranking tasks.

[0109] The fused feature is a feature obtained by fusing the first feature and the second feature.

[0110] Specifically, the server can fuse the first and second features corresponding to the same recommended object to obtain a fused feature corresponding to the recommended object. Furthermore, the server can estimate the ranking of each recommended object based on the fused feature in each ranking task, obtaining the estimated ranking results for each recommended object under multiple ranking tasks.

[0111] In one embodiment, the server may directly fuse the first and second features corresponding to the same object to be recommended to obtain a fused feature corresponding to the object to be recommended. It is understood that the server may not adjust the first feature, but instead directly fuse the first and second features corresponding to the same object to be recommended to obtain the same fused feature for each object to be recommended.

[0112] In the above embodiment, by fusing the first and second features corresponding to the same recommended object, a richer fused feature is obtained. In each ranking task, the ranking of each recommended object is estimated based on the richer fused feature, which can improve the accuracy of the ranking estimation results.

[0113] In one embodiment, the step of fusing the first feature and the second feature corresponding to the same object to be recommended to obtain the fused feature corresponding to the object to be recommended specifically includes: for each sorting task, adjusting the first feature corresponding to each object to be recommended according to the adjustment coefficient corresponding to the sorting task, to obtain the adjusted first feature of each object to be recommended under different sorting tasks; under each sorting task, fusing the adjusted first feature and the second feature corresponding to the same object to be recommended, to obtain the fused features of each object to be recommended under multiple sorting tasks.

[0114] The adjustment coefficient is a coefficient used to adjust the first feature.

[0115] Specifically, the server may predetermine the adjustment coefficient corresponding to each sorting task. For each sorting task, the server may adjust the first feature corresponding to each object to be recommended according to the adjustment coefficient corresponding to the sorting task, and obtain the adjusted first feature of each object to be recommended under different sorting tasks. Furthermore, the server may perform feature fusion on the adjusted first feature and second feature corresponding to the same object to be recommended under each sorting task, and obtain the fused features of each object to be recommended under multiple sorting tasks. It can be understood that the server first adjusts the first feature based on the adjustment coefficient, and then performs feature fusion on the adjusted first feature and second feature corresponding to the same object to be recommended, and the fused features of each object to be recommended obtained by fusion may be different.

[0116] In the above embodiment, for each sorting task, the first feature corresponding to each object to be recommended is adjusted according to the adjustment coefficient corresponding to the sorting task, so that the adjusted first feature is more suitable for the corresponding sorting task, thereby further improving the accuracy of object recommendation.

[0117] In an embodiment, the step of respectively estimating the ranking of each to-be-recommended object according to the fused features under each ranking task to obtain the ranking estimation result of each to-be-recommended object under each ranking task comprises: for each ranking task, predicting the state information of each to-be-recommended object under the ranking task according to the fused features; and estimating the ranking of each to-be-recommended object according to the state information of each to-be-recommended object under the ranking task to obtain the ranking estimation result of each to-be-recommended object under the ranking task.

[0118] The state information is information representing the state of the to-be-recommended object under the ranking task.

[0119] For example, if the to-be-recommended object includes a video A and the ranking task is a like, the state information of the video A under the like ranking task can be that the probability of the video A being liked is 90%.

[0120] Specifically, the server can predict the state information of each to-be-recommended object under each ranking task according to the fused features. Then, the server can estimate the ranking of each to-be-recommended object according to the state information of each to-be-recommended object under each ranking task to obtain the ranking estimation result of each to-be-recommended object under each ranking task.

[0121] In the above embodiment, the state information of each to-be-recommended object under each ranking task can be more accurately predicted according to the fused features, and the ranking of each to-be-recommended object can be more accurately estimated according to the more accurate state information of each to-be-recommended object under each ranking task to obtain a more accurate ranking estimation result of each to-be-recommended object under each ranking task, thereby further improving the accuracy of object recommendation.

[0122] In an embodiment, the ranking estimation result includes a ranking estimation score, and the target ranking estimation result includes a target ranking estimation score. The step 208, i.e., the step of fusing the ranking estimation results of each to-be-recommended object under each ranking task to obtain the target ranking estimation result corresponding to the to-be-recommended object, specifically comprises: for each to-be-recommended object, weighting and fusing the ranking estimation scores of the to-be-recommended object under each ranking task to obtain the target ranking estimation score corresponding to the to-be-recommended object.

[0123] The ranking estimation score is a score of the to-be-recommended object estimated and used for ranking. The target ranking estimation score is a score of the to-be-recommended object obtained by fusing the ranking estimation scores of the to-be-recommended object under each ranking task.

[0124] Specifically, the server can weighting and fuse the ranking estimation scores of each to-be-recommended object under each ranking task to obtain the target ranking estimation score corresponding to the to-be-recommended object.

[0125] In one embodiment, the target ranking estimated score of each recommended object in each ranking task can be expressed by the following formula:

[0126]

[0127] Among them, k represents the kth object to be recommended among all the objects to be recommended, n represents the number of sorting tasks, i represents the i-th sorting task among n sorting tasks, W represents the fusion weight vector corresponding to each sorting task, task ik Indicates the estimated ranking score of the kth recommended object under the i-th ranking task, item k It represents the estimated score of the target ranking of the kth recommended object in each ranking task.

[0128] In the above embodiment, by weighted fusion of the estimated ranking scores of the object to be recommended under each ranking task, a more accurate target ranking estimated score corresponding to the object to be recommended can be obtained, thereby further improving the accuracy of object recommendation.

[0129] In one embodiment, step 204, that is, the step of extracting the relationship between the objects to be recommended based on the set of first features to obtain the second features corresponding to each object to be recommended, specifically includes: inputting the set of first features into the trained object sorting model, so as to extract the relationship between the objects to be recommended based on the set of first features in the feature encoding layer of the trained object sorting model, and obtain the second features corresponding to each object to be recommended.

[0130] Specifically, the server can input the set of first features into the trained object sorting model, so as to extract the relationship between the objects to be recommended based on the set of first features in the feature encoding layer of the trained object sorting model, and obtain the second features corresponding to each object to be recommended.

[0131] In one embodiment, the feature encoding layer may be constructed using a deep network, which may include at least one of a multi-head attention mechanism and a pure multi-layer perceptron visual architecture (MLP-mixer).

[0132] In the above embodiment, by extracting the relationship between the objects to be recommended based on the set of first features in the feature encoding layer of the trained object ranking model, the second features corresponding to each object to be recommended can be quickly extracted, thereby improving the efficiency of object ranking estimation.

[0133] In one embodiment, step 206, that is, the step of generating the estimated ranking results of each object to be recommended under multiple ranking tasks based on the first feature and the second feature, specifically includes: respectively using the ranking estimation unit corresponding to each ranking task in the trained object ranking model, and obtaining the estimated ranking results of each object to be recommended under the corresponding ranking task based on the first feature and the second feature corresponding to the same object to be recommended.

[0134] Specifically, the server can sort each object to be recommended according to the first feature and the second feature corresponding to the same object to be recommended through the sorting estimation unit corresponding to each sorting task in the trained object sorting model, and obtain the sorting estimation result of each object to be recommended under the corresponding sorting task.

[0135] In the above embodiment, through the ranking estimation unit corresponding to each ranking task in the trained object ranking model, based on the first feature and the second feature corresponding to the same object to be recommended, a more accurate ranking estimation result of each object to be recommended under the corresponding ranking task can be obtained, thereby further improving the accuracy of object recommendation.

[0136] In one embodiment, the trained object ranking model includes a feature encoding layer and a ranking estimation unit corresponding to each ranking task. The server can input a set of first features into the trained object ranking model to extract the relationship between the objects to be recommended based on the set of first features in the feature encoding layer of the trained object ranking model, and obtain the second features corresponding to each object to be recommended. The server can sort each object to be recommended according to the first feature and the second feature corresponding to the same object to be recommended through the ranking estimation unit corresponding to each ranking task in the trained object ranking model, and obtain the ranking estimation result of each object to be recommended under the corresponding ranking task.

[0137] In one embodiment, the trained object ranking model includes, in addition to a feature encoding layer and a ranking estimation unit corresponding to each ranking task, at least one of a feature input layer, a feature adjustment layer, and a ranking result output layer. The feature input layer is used to obtain a set of first features. The feature enhancement layer is used to enhance the second features corresponding to each object to be recommended. The feature adjustment layer is used to adjust the first features. The ranking result output layer is used to output the ranking results.

[0138] In an embodiment, the trained object ranking model is obtained through a model training step; the model training step specifically includes: obtaining a sample set including multiple groups of training samples; each group of training samples includes first sample features of corresponding sample objects and state labels of the sample objects respectively under multiple ranking tasks; the first sample features include features representing correlations between the sample objects and users; based on the sample set, the object ranking model to be trained is trained to obtain the object ranking model for ranking objects according to the multiple ranking tasks.

[0139] In the formula, the training sample is sample data for training the object ranking model. The sample set is a set composed of multiple groups of training samples. The sample object is a to-be-ranked object used in the process of training the object ranking model. The state label is a label representing the state of the sample object under the ranking task.

[0140] Specifically, the server can obtain a sample set including multiple groups of training samples, and train the object ranking model to be trained based on the sample set to obtain the object ranking model for ranking objects according to multiple ranking tasks.

[0141] In an embodiment, the server can extract relationships between sample objects based on the first sample features in each group of training samples to obtain second sample features corresponding to each sample object. Further, the server can predict state information of each sample object under a corresponding ranking task based on the first sample features and the second sample features corresponding to the same sample object for each ranking task. The server can train the object ranking model to be trained based on differences between the predicted state information and the state labels to obtain the object ranking model for ranking objects according to multiple ranking tasks.

[0142] In an embodiment, each group of training samples can be represented as <first sample features of a sample object, state labels of the sample object respectively under multiple ranking tasks>.

[0143] In an embodiment, the first sample feature can include a sample object attribute feature that does not represent the relevance between the corresponding to-be-recommended object and the user, and a sample personalized semantic feature that represents the relevance between the corresponding to-be-recommended object and the user, and each set of training samples can be represented as <sample personalized semantic feature, sample object attribute feature, and sample object state label under each ranking task>. For example, if user_item_feature represents the sample personalized semantic feature, item_feature represents the sample object attribute feature, label1 represents the state information of the first ranking task, label2 represents the state information of the second ranking task, and label3 represents the state information of the third ranking task, each set of training samples can be represented as <user_item_feature, item_feature, label1, label2, label3>.

[0144] In an embodiment, if the ranking task is discrete, the state information of the corresponding state label can be set to 0 or 1. If the ranking task is continuous, the state information of the corresponding state label can be normalized to [0, 1]. For example, if the ranking task is like, the ranking task is discrete, and 1 can represent that the like has been given, and 0 can represent that the like has not been given. If the ranking task is the browsing time, the ranking task is continuous, and the browsing time is normalized to a value between [0, 1].

[0145] In the above embodiment, the object ranking model to be trained is trained by using the sample set including multiple sets of training samples, which can improve the accuracy of the object ranking model in object ranking prediction.

[0146] In an embodiment, the object ranking model to be trained is trained based on the sample set to obtain an object ranking model for ranking objects according to multiple ranking tasks, including: in each round of iterative training, inputting the sample set into the object ranking model to be trained, to extract the relationship between sample objects based on the first sample feature in each set of training samples in the feature encoding layer of the object ranking model to be trained, to obtain the second sample feature corresponding to each sample object; for each ranking task, predicting the state information of each sample object under the corresponding ranking task by using the ranking prediction unit corresponding to the ranking task in the object ranking model to be trained, based on the first sample feature and the second sample feature corresponding to the same sample object; updating the model parameters of the object ranking model to be trained according to the difference between the predicted state information and the state label, until the iterative stopping condition is reached, to obtain the object ranking model for ranking objects according to multiple ranking tasks.

[0147] Specifically, the object sorting model to be trained includes a feature encoding layer and a sorting prediction unit corresponding to the sorting task. In each round of iterative training, the server can input the sample set into the object sorting model to be trained, so as to extract the relationship between the sample objects based on the first sample features in each group of training samples in the feature encoding layer of the object sorting model to be trained, and obtain the second sample features corresponding to each sample object. For each sorting task, the server can predict the state information of each sample object under the corresponding sorting task based on the first sample features and the second sample features corresponding to the same sample object through the sorting prediction unit corresponding to the sorting task in the object sorting model to be trained. Furthermore, the server can update the model parameters of the object sorting model to be trained according to the difference between the predicted state information and the state label until the iteration stop condition is reached, thereby obtaining an object sorting model for sorting objects according to multiple sorting tasks.

[0148] In one embodiment, the server can determine the difference between the predicted state information and the state label by using the objective function corresponding to each sorting task. Then, the server can update the model parameters of the object sorting model to be trained based on the difference between the predicted state information and the state label.

[0149] In the above embodiment, by iteratively updating the model parameters of the object ranking model to be trained based on the difference between the predicted state information and the state labels included in the training samples, the accuracy of the object ranking model in estimating the object ranking can be improved.

[0150] In one embodiment, the model parameters of the object sorting model to be trained are updated based on the difference between the predicted state information and the state label, specifically including: for each sorting task, determining the loss value between the predicted state information and the state label through the objective function corresponding to the task type to which the sorting task belongs; and updating the model parameters of the object sorting model to be trained based on the loss value.

[0151] The task type is the category to which the sorting task belongs.

[0152] Specifically, for each sorting task, the server can calculate the loss between the predicted state information and the state label using the objective function corresponding to the task type to which the sorting task belongs. Furthermore, the server can update the model parameters of the object sorting model to be trained based on the calculated loss value.

[0153] In one embodiment, the task type to which the sorting task belongs may include at least one of a classification problem and a regression problem.

[0154] In the above embodiment, the sorting tasks of different task types have corresponding objective functions that are different. By using the objective function corresponding to the task type to which the sorting task belongs, the loss value between the predicted state information and the state label can be determined more accurately, so that the convergence speed of the object sorting model can be improved, thereby improving the training efficiency of the object sorting model.

[0155] In one embodiment, Figure 4 As shown, the server can obtain the object attribute features and personalized semantic features corresponding to each object to be recommended, and combine the object attribute features and personalized semantic features corresponding to the same object to be recommended to obtain the first features corresponding to each object to be recommended. The server can determine the set of first features based on the first features corresponding to each object to be recommended, and extract the relationship between the objects to be recommended based on the set of first features to obtain the second features corresponding to each object to be recommended. The server can score each object to be recommended under multiple sorting tasks based on the first features and the second features to generate a ranking estimate score for each object to be recommended under multiple sorting tasks. The server can fuse the ranking estimate scores of the object to be recommended under each sorting task for each object to be recommended to obtain the target ranking estimate score corresponding to the object to be recommended. Furthermore, the server can sort each object to be recommended according to the target ranking estimate score corresponding to each object to be recommended, and recommend objects based on the sorted results.

[0156] In one embodiment, Figure 5As shown, the object ranking model includes a feature input layer 501, a feature encoding layer 502 (including a multi-head attention mechanism network (Multi-Head Attention) and a feed-forward network (FFN, Feed-Forward Network)), a ranking estimation unit (Tower) corresponding to each ranking task, a feature adjustment layer (i.e., a gate mechanism (Gate)), and a ranking result output layer 503. The server can concatenate the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended through the feature input layer to obtain the first feature vector corresponding to each object to be recommended, and position encode each first feature vector to obtain a set of first feature vectors. The server can input the set of first feature vectors into the feature encoding layer. The server can extract the relationship between the objects to be recommended based on the set of first features through the multi-head attention mechanism network in the feature encoding layer, and obtain the second feature vector corresponding to each object to be recommended. The server can enhance the second feature vector through the feed-forward network in the feature encoding layer. The server can adjust the first feature vector through the feature adjustment layer. The server can generate the estimated ranking results of each object to be recommended under multiple ranking tasks based on the first eigenvector and the second eigenvector through the ranking estimation unit corresponding to each ranking task, and for each object to be recommended, fuse the estimated ranking results of the object to be recommended under each ranking task to obtain the target ranking estimation result corresponding to the object to be recommended. The server can sort each object to be recommended according to the target ranking estimation result corresponding to each object to be recommended to obtain the sorted result. The server can output the sorted result through the sorting result output layer. The server can recommend objects based on the sorted result output by the sorting result output layer.

[0157] For example, if the number of objects to be recommended is m=8 and the number of sorting tasks is n=3, then the set of first feature vectors includes 8 first feature vectors. The server can extract the relationship between the objects to be recommended based on the 8 first feature vectors in the set of first features, and obtain the second feature vectors corresponding to the 8 objects to be recommended. It can be understood that the number of second feature vectors is also 8. The server can generate the sorting estimation results of the 8 objects to be recommended under the 3 sorting tasks respectively according to the first feature vector and the second feature vector through the sorting estimation unit corresponding to each sorting task, and for each object to be recommended, fuse the sorting estimation results of the object to be recommended under the 3 sorting tasks to obtain the target sorting estimation result corresponding to the object to be recommended. The server can sort the 8 objects to be recommended based on the target sorting estimation results corresponding to the 8 objects to be recommended, and recommend objects to the 8 objects to be recommended based on the sorted results.

[0158] like Figure 6As shown, in one embodiment, a method for object recommendation is provided, which specifically includes the following steps:

[0159] Step 602: Obtain a sample set including multiple groups of training samples; each group of training samples includes a first sample feature of a corresponding sample object and a state label of the sample object under multiple sorting tasks; the first sample feature includes a feature that characterizes the correlation between the sample object and the user.

[0160] Step 604: In each round of iterative training, the sample set is input into the object sorting model to be trained, so that in the feature coding layer of the object sorting model to be trained, the relationship between the sample objects is extracted based on the first sample features in each group of training samples, and the second sample features corresponding to each sample object are obtained.

[0161] Step 606: For each sorting task, the sorting prediction unit corresponding to the sorting task in the object sorting model to be trained is used to predict the state information of each sample object under the corresponding sorting task based on the first sample feature and the second sample feature corresponding to the same sample object.

[0162] Step 608 : For each sorting task, determine the loss value between the predicted state information and the state label using the objective function corresponding to the task type to which the sorting task belongs.

[0163] Step 610 : Update the model parameters of the object sorting model to be trained according to the loss value until the iteration stop condition is reached, thereby obtaining an object sorting model for sorting objects according to multiple sorting tasks.

[0164] Step 612, obtain the object attribute feature vector and personalized semantic feature vector corresponding to each object to be recommended; the personalized semantic feature vector is used to characterize the correlation between the corresponding object to be recommended and the user; each object to be recommended is the object to be recommended after initial sorting based on the corresponding personalized semantic feature vector.

[0165] Step 614 : Concatenate the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended to obtain a first feature vector corresponding to each object to be recommended.

[0166] Step 616 , position encoding each first feature vector according to the position of the object to be recommended after the initial sorting, to obtain a set of first feature vectors; the first feature vectors in the set of first feature vectors are arranged in the order of their positions among the corresponding objects to be recommended.

[0167] Step 618: Input the set of first feature vectors into the trained object ranking model, so as to extract the relationship between the objects to be recommended based on the set of first feature vectors in the feature coding layer of the trained object ranking model, and obtain a second feature vector corresponding to each object to be recommended; the second feature vector is used to characterize the correlation between the objects to be recommended.

[0168] Step 620 , for each ranking task, adjust the first eigenvector corresponding to each object to be recommended according to the adjustment coefficient corresponding to the ranking task, and obtain the adjusted first eigenvector of each object to be recommended under different ranking tasks.

[0169] In step 622, under each sorting task, the adjusted first feature vector and second feature vector corresponding to the same object to be recommended are fused through the sorting estimation unit corresponding to each sorting task in the trained object sorting model to obtain the fused features of each object to be recommended under multiple sorting tasks.

[0170] Step 624 : For each ranking task, predict the status information of each object to be recommended under the ranking task based on the fusion features.

[0171] Step 626 , performing a ranking estimation on each object to be recommended according to the status information of each object to be recommended under the ranking task, and obtaining a ranking estimation score of each object to be recommended under the ranking task.

[0172] Step 628 : For each object to be recommended, weighted fusion is performed on the estimated ranking scores of the object to be recommended under each ranking task to obtain the target estimated ranking score corresponding to the object to be recommended.

[0173] Step 630 , sorting the objects to be recommended according to the target ranking estimated scores corresponding to the objects to be recommended, and performing object recommendations based on the sorted results.

[0174] In one embodiment, Figure 7 As shown in the figure, the sorting results after using the object recommendation method of the present application have improved in all sorting tasks compared to the sorting results before using the object recommendation method of the present application, namely, the recommendation rate, like rate, sharing rate, comment rate, attention rate and browsing time have all increased. For example, the recommendation rate increased by 16.03% and the like rate increased by 14.60%.

[0175] In one embodiment, for each sorting task, the improvement rate can be calculated by the change in MAP (Mean Average Precision, an object detection evaluation index) of the sorted results under the corresponding sorting task before and after adopting the object recommendation method of the present application.

[0176] The application further provides an application scenario of the object recommendation method. Specifically, the object recommendation method can be applied to a video object recommendation scenario. A server can obtain a sample set including multiple groups of training samples; each group of training samples includes a first sample feature of a corresponding sample video object and a state label of the sample video object under multiple ranking tasks respectively; the first sample feature includes a feature representing the correlation between the sample video object and a user. In each round of iterative training, the sample set is input into a video object ranking model to be trained, so as to extract the relationship between sample video objects based on the first sample features in each group of training samples in a feature encoding layer of the video object ranking model to be trained, and obtain second sample features corresponding to each sample video object respectively. For each ranking task, a ranking estimation unit corresponding to the ranking task in the video object ranking model to be trained is used to predict the state information of each sample video object under the corresponding ranking task based on the first sample feature and the second sample feature corresponding to the same sample video object. For each ranking task, a target function corresponding to the task type to which the ranking task belongs is used to determine the loss value between the predicted state information and the state label. According to the loss value, the model parameters of the video object ranking model to be trained are updated until an iterative stop condition is reached, and a video object ranking model for object ranking according to multiple ranking tasks is obtained.

[0177] The server can obtain an object attribute feature vector and a personalized semantic feature vector corresponding to each to-be-recommended video object respectively; the personalized semantic feature vector is used to represent the correlation between the corresponding to-be-recommended video object and the user; each to-be-recommended video object is a to-be-recommended video object after initial ranking based on the corresponding personalized semantic feature vector. The object attribute feature vector and the personalized semantic feature vector corresponding to the same to-be-recommended video object are concatenated to obtain a first feature vector corresponding to each to-be-recommended video object. According to the position of the to-be-recommended video object after the initial ranking, the first feature vectors are positionally encoded to obtain a set of first feature vectors; the first feature vectors in the set of first feature vectors are arranged in the order of positions of the corresponding to-be-recommended video objects.

[0178] The server can input the set of first feature vectors into the trained video object ranking model to extract the relationship between the to-be-recommended video objects based on the set of first feature vectors in a feature encoding layer of the trained video object ranking model, and obtain second feature vectors corresponding to each to-be-recommended video object respectively; the second feature vectors are used to represent the correlation between the to-be-recommended video objects. For each ranking task, the first feature vectors corresponding to each to-be-recommended video object are adjusted according to an adjustment coefficient corresponding to the ranking task, to obtain adjusted first feature vectors of each to-be-recommended video object under different ranking tasks.

[0179] Under each sorting task, the server can perform feature fusion on the adjusted first feature vector and second feature vector corresponding to the same video object to be recommended, and obtain the fusion features of each video object to be recommended under multiple sorting tasks. For each sorting task, the status information of each video object to be recommended under the sorting task is predicted based on the fusion features. According to the status information of each video object to be recommended under the sorting task, the ranking of each video object to be recommended is estimated to obtain the ranking estimation result of each video object to be recommended under the sorting task. For each video object to be recommended, the ranking estimation scores of the video object to be recommended under each sorting task are weighted and fused to obtain the target ranking estimation score corresponding to the video object to be recommended. According to the target ranking estimation score corresponding to each video object to be recommended, each video object to be recommended is sorted, and video objects are recommended based on the sorted results.

[0180] The present application also provides an application scenario, which applies the above-mentioned object recommendation method. Specifically, the object recommendation method can be applied to the scenario of document object recommendation. The server can obtain a set of first features; the set includes first features corresponding to each document object to be recommended; the first features include features that characterize the correlation between the corresponding document object to be recommended and the user. The server can extract the relationship between the document objects to be recommended based on the set of first features, and obtain second features corresponding to each document object to be recommended; the second feature is used to characterize the correlation between the document objects to be recommended. The server can generate the ranking estimation results of each document object to be recommended under multiple sorting tasks based on the first feature and the second feature, and for each document object to be recommended, fuse the ranking estimation results of the document object to be recommended under each sorting task to obtain the target ranking estimation result corresponding to the document object to be recommended. According to the target ranking estimation result corresponding to each document object to be recommended, each document object to be recommended is sorted, and document objects are recommended based on the sorted results.

[0181] It should be understood that, although the steps in the above flowchart are shown in sequence as indicated by the arrows, these steps are not necessarily performed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be performed in other orders. Moreover, at least a portion of the steps in the above flowchart may include multiple steps or multiple stages, and these steps or stages are not necessarily performed at the same time, but can be performed at different times. The execution order of these steps or stages is not necessarily to be performed in sequence, but can be performed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0182] In one embodiment, Figure 8As shown, an object recommendation apparatus 800 is provided, which can be a part of a computer device in the form of a software module or a hardware module, or a combination of both. The apparatus specifically comprises:

[0183] An acquisition module 801 is configured to acquire a set of first features, wherein the set of first features comprises first features corresponding to each of the to-be-recommended objects, and each of the first features comprises a feature representing the relevance between the corresponding to-be-recommended object and the user.

[0184] An extraction module 802 is configured to extract the relationship between the to-be-recommended objects based on the set of first features, to obtain second features corresponding to each of the to-be-recommended objects, wherein each of the second features represents the relevance between the to-be-recommended objects.

[0185] A generation module 803 is configured to generate ranking estimation results of each of the to-be-recommended objects in a plurality of ranking tasks according to the first features and the second features.

[0186] A fusion module 804 is configured to fuse the ranking estimation results of each of the to-be-recommended objects in the plurality of ranking tasks to obtain a target ranking estimation result corresponding to each of the to-be-recommended objects.

[0187] A recommendation module 805 is configured to rank each of the to-be-recommended objects according to the target ranking estimation result corresponding to each of the to-be-recommended objects, and to recommend the objects based on the ranked results.

[0188] In an embodiment, the acquisition module 801 is further configured to acquire object attribute features and personalized semantic features corresponding to each of the to-be-recommended objects, wherein the personalized semantic features represent the relevance between the corresponding to-be-recommended object and the user, and the object attribute features and the personalized semantic features corresponding to the same to-be-recommended object are combined to obtain the first features corresponding to each of the to-be-recommended objects, and the set of first features is determined according to the first features corresponding to each of the to-be-recommended objects.

[0189] In an embodiment, the first features comprise first feature vectors, the object attribute features comprise object attribute feature vectors, and the personalized semantic features comprise personalized semantic feature vectors, and the acquisition module 801 is further configured to concatenate the object attribute feature vectors and the personalized semantic feature vectors corresponding to the same to-be-recommended object to obtain the first feature vectors corresponding to each of the to-be-recommended objects.

[0190] In an embodiment, each of the to-be-recommended objects is a to-be-recommended object ranked initially based on the corresponding personalized semantic feature vector, and the acquisition module 801 is further configured to encode each of the first feature vectors according to the position of the to-be-recommended object after the initial ranking to obtain a set of first feature vectors, and each of the first feature vectors in the set of first feature vectors is arranged in the order of positions between the corresponding to-be-recommended objects.

[0191] In an embodiment, the generating module 803 is further configured to perform feature fusion on the first feature and the second feature corresponding to the same to-be-recommended object to obtain a fusion feature corresponding to the to-be-recommended object; and perform ranking estimation on each to-be-recommended object according to the fusion feature under each ranking task to obtain ranking estimation results of each to-be-recommended object under the multiple ranking tasks.

[0192] In an embodiment, the generating module 803 is further configured to, for each ranking task, adjust the first feature corresponding to each to-be-recommended object according to an adjustment coefficient corresponding to the ranking task to obtain an adjusted first feature of each to-be-recommended object under the different ranking tasks; and perform feature fusion on the adjusted first feature and the second feature corresponding to the same to-be-recommended object under each ranking task to obtain a fusion feature of each to-be-recommended object under the multiple ranking tasks.

[0193] In an embodiment, the generating module 803 is further configured to, for each ranking task, predict state information of each to-be-recommended object under the ranking task according to the fusion feature; and perform ranking estimation on each to-be-recommended object according to the state information of each to-be-recommended object under the ranking task to obtain a ranking estimation result of each to-be-recommended object under the ranking task.

[0194] In an embodiment, the ranking estimation result includes a ranking estimation score, the target ranking estimation result includes a target ranking estimation score, and the fusion module 804 is configured to, for each to-be-recommended object, perform weighted fusion on the ranking estimation scores of the to-be-recommended object under the ranking tasks to obtain a target ranking estimation score corresponding to the to-be-recommended object.

[0195] In an embodiment, the extracting module 802 is further configured to input the set of first features into the trained object ranking model to extract relationships between the to-be-recommended objects based on the set of first features in a feature encoding layer of the trained object ranking model to obtain the second features corresponding to each to-be-recommended object; and the generating module 803 is further configured to obtain, by each ranking estimation unit corresponding to each ranking task in the trained object ranking model, the ranking estimation result of each to-be-recommended object under the corresponding ranking task according to the first feature and the second feature corresponding to the same to-be-recommended object.

[0196] Reference Figure 9 In an embodiment, the object recommendation apparatus 800 further includes:

[0197] The training module 806 is used to obtain a sample set including multiple groups of training samples; each group of training samples includes the first sample features of the corresponding sample objects and the state labels of the sample objects under multiple sorting tasks; the first sample features include features that characterize the correlation between the sample objects and the users; based on the sample set, the object sorting model to be trained is trained to obtain an object sorting model for sorting objects according to multiple sorting tasks.

[0198] In one embodiment, the training module 806 is also used to input the sample set into the object sorting model to be trained in each round of iterative training, so as to extract the relationship between sample objects based on the first sample features in each group of training samples in the feature coding layer of the object sorting model to be trained, and obtain the second sample features corresponding to each sample object; for each sorting task, through the sorting prediction unit corresponding to the sorting task in the object sorting model to be trained, based on the first sample features and the second sample features corresponding to the same sample object, predict the state information of each sample object under the corresponding sorting task; according to the difference between the predicted state information and the state label, update the model parameters of the object sorting model to be trained until the iteration stop condition is reached, and obtain the object sorting model for sorting objects according to multiple sorting tasks.

[0199] In one embodiment, the training module 806 is also used to determine the loss value between the predicted state information and the state label for each sorting task through the objective function corresponding to the task type to which the sorting task belongs; and update the model parameters of the object sorting model to be trained based on the loss value.

[0200] The above-mentioned object recommendation device obtains a set of first features corresponding to each object to be recommended, and the first features include features that characterize the correlation between the corresponding object to be recommended and the user. Based on the set of first features, the relationship between the objects to be recommended can be extracted to obtain second features corresponding to each object to be recommended, and the second features can be used to characterize the correlation between the objects to be recommended. According to the first features used to characterize the correlation between the corresponding object to be recommended and the user, and the second features used to characterize the correlation between the objects to be recommended, more accurate ranking estimation results of each object to be recommended under multiple sorting tasks can be generated. For each object to be recommended, the more accurate ranking estimation results of the object to be recommended under each sorting task can be fused to obtain a more accurate target ranking estimation result corresponding to the object to be recommended. By sorting each object to be recommended according to the more accurate target ranking estimation result corresponding to each object to be recommended, a more accurate sorted result can be obtained, and then object recommendation can be performed based on the more accurate sorted result, which can improve the recommendation accuracy of the object.

[0201] The specific limitations of the object recommendation apparatus can refer to the limitations of the object recommendation method described above, and will not be repeated here. Each module in the object recommendation apparatus described above can be implemented by software, hardware and a combination thereof in whole or in part. Each module described above can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor invokes and executes the operations corresponding to each module.

[0202] In one embodiment, a computer device, which can be a server, is provided, and an internal structure diagram of the computer device can be as shown in Figure 10 The computer device includes a processor, a memory and a network interface connected through a system bus. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is configured to communicate with an external terminal through a network connection. The computer program is executed by the processor to implement an object recommendation method.

[0203] Those skilled in the art can understand that Figure 10 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0204] In one embodiment, a computer device is also provided, including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps in each method embodiment described above.

[0205] In one embodiment, a computer readable storage medium is provided, storing a computer program, and the computer program is executed by the processor to implement the steps in each method embodiment described above.

[0206] In one embodiment, a computer program product is provided, including a computer program, and the computer program is executed by the processor to implement the steps in each method embodiment described above.

[0207] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory or optical memory, etc. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0208] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0209] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. An object recommendation method, characterized in that: The method comprises: Obtain a set of first feature vectors; the set includes a first feature vector corresponding to each object to be recommended; the first feature vectors include a feature vector representing the correlation between the corresponding object to be recommended and the user; the first feature vector is obtained by combining the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended; the first feature vectors in the set of first feature vectors are arranged in the order of their positions after the initial sorting of the corresponding object to be recommended; Extracting the relationship between the objects to be recommended based on the set of the first feature vectors to obtain a second feature vector corresponding to each object to be recommended; the second feature vector is used to represent the correlation between the objects to be recommended; For each ranking task, the first eigenvector corresponding to each object to be recommended is adjusted according to the adjustment coefficient corresponding to the ranking task, to obtain the adjusted first eigenvector of each object to be recommended under different ranking tasks. Under each ranking task, the adjusted first eigenvector and the second eigenvector corresponding to the same object to be recommended are fused to obtain the fused eigenvector of each object to be recommended under multiple ranking tasks. The ranking task is a task of ranking each object to be recommended, and different ranking tasks rank the objects to be recommended from different dimensions. Under each ranking task, the ranking of each object to be recommended is estimated according to the fused feature vector, and the ranking estimation results of each object to be recommended under multiple ranking tasks are obtained; For each object to be recommended, the estimated ranking results of the object to be recommended under each of the ranking tasks are integrated to obtain the target ranking estimated result corresponding to the object to be recommended; The objects to be recommended are sorted according to the target ranking estimation results corresponding to the objects to be recommended, and the objects are recommended based on the sorted results.

2. The method according to claim 1, characterized in that The obtaining of a set of first feature vectors includes: Obtaining object attribute feature vectors and personalized semantic feature vectors corresponding to each object to be recommended; the personalized semantic feature vectors are used to characterize the correlation between the corresponding object to be recommended and the user; Performing feature vector combination on the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended, to obtain first feature vectors corresponding to each object to be recommended; A set of first feature vectors is determined according to the first feature vectors corresponding to the objects to be recommended.

3. The method according to claim 2, characterized in that The step of combining the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended to obtain first feature vectors corresponding to each object to be recommended includes: The object attribute feature vector corresponding to the same object to be recommended and the personalized semantic feature vector are concatenated to obtain first feature vectors corresponding to each object to be recommended.

4. The method according to claim 3, characterized in that The objects to be recommended are objects to be recommended after initial sorting based on the corresponding personalized semantic feature vectors; The determining of a set of first feature vectors according to the first feature vectors corresponding to the respective objects to be recommended includes: According to the position of the objects to be recommended after the initial sorting, position encoding is performed on each of the first feature vectors to obtain a set of first feature vectors.

5. The method according to claim 1, wherein In each ranking task, the ranking of each object to be recommended is estimated based on the fused feature vector, and the ranking estimation results of each object to be recommended under multiple ranking tasks are obtained, including: For each sorting task, predicting the status information of each to-be-recommended object under the sorting task according to the fused feature vector; According to the status information of each object to be recommended under the sorting task, a sorting estimate is performed on each object to be recommended to obtain a sorting estimate result of each object to be recommended under the sorting task.

6. The method according to claim 1, characterized in that The ranking estimation result includes a ranking estimation score; the target ranking estimation result includes a target ranking estimation score; For each object to be recommended, fusing the estimated ranking results of the object to be recommended under each of the ranking tasks to obtain an estimated target ranking result corresponding to the object to be recommended, including: For each object to be recommended, the estimated ranking scores of the object to be recommended under each of the ranking tasks are weighted and fused to obtain the target ranking estimated score corresponding to the object to be recommended.

7. The method according to any one of claims 1 to 6, characterized in that The extracting the relationship between the objects to be recommended based on the set of the first feature vectors to obtain the second feature vectors corresponding to each object to be recommended includes: The set of the first feature vectors is input into a trained object ranking model, so that in the feature encoding layer of the trained object ranking model, the relationship between the objects to be recommended is extracted based on the set of the first feature vectors to obtain the second feature vectors corresponding to each object to be recommended.

8. The method according to any one of claims 1 to 6, characterized in that For each ranking task, adjusting the first feature vectors corresponding to each to-be-recommended object according to the adjustment coefficient corresponding to the ranking task to obtain the adjusted first feature vectors of each to-be-recommended object under different ranking tasks; performing feature vector fusion on the adjusted first feature vector and second feature vector corresponding to the same to-be-recommended object under each ranking task to obtain fused feature vectors of each to-be-recommended object under multiple ranking tasks, including: Through the ranking estimation unit corresponding to each ranking task in the trained object ranking model, for each ranking task, the first feature vector corresponding to each object to be recommended is adjusted according to the adjustment coefficient corresponding to the ranking task to obtain the adjusted first feature vector of each object to be recommended under different ranking tasks. Under each ranking task, the adjusted first feature vector and the second feature vector corresponding to the same object to be recommended are fused to obtain the fused feature vectors of each object to be recommended under multiple ranking tasks.

9. The method according to claim 8, characterized in that The trained object sorting model is obtained through a model training step; the model training step includes: Acquire a sample set including multiple groups of training samples; each group of training samples includes a first sample feature vector of a corresponding sample object and a state label of the sample object under multiple sorting tasks; the first sample feature vector includes a feature vector representing the correlation between the sample object and the user; Based on the sample set, the object sorting model to be trained is trained to obtain an object sorting model for sorting objects according to the multiple sorting tasks.

10. The method according to claim 9, characterized in that The training of the object sorting model to be trained based on the sample set to obtain the object sorting model for sorting objects according to the plurality of sorting tasks includes: In each round of iterative training, the sample set is input into the object ranking model to be trained, so that in the feature encoding layer of the object ranking model to be trained, the relationship between the sample objects is extracted based on the first sample feature vector in each group of the training samples, and the second sample feature vector corresponding to each sample object is obtained; For each sorting task, using a sorting prediction unit corresponding to the sorting task in the object sorting model to be trained, based on the first sample feature vector and the second sample feature vector corresponding to the same sample object, predict the state information of each sample object under the corresponding sorting task; According to the difference between the predicted state information and the state label, the model parameters of the object sorting model to be trained are updated until an iteration stop condition is reached, thereby obtaining an object sorting model for sorting objects according to the multiple sorting tasks.

11. The method according to claim 10, characterized in that The updating of the model parameters of the object sorting model to be trained according to the difference between the predicted state information and the state label includes: For each sorting task, determine the loss value between the predicted state information and the state label through the objective function corresponding to the task type to which the sorting task belongs; According to the loss value, the model parameters of the object sorting model to be trained are updated.

12. An object recommendation device, characterized in that: The device comprises: An acquisition module is configured to acquire a set of first feature vectors; the set includes first feature vectors corresponding to respective objects to be recommended; the first feature vectors include feature vectors representing the correlation between the respective objects to be recommended and the user; the first feature vectors are obtained by combining the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended; the first feature vectors in the set of first feature vectors are arranged in the order of their positions after the initial sorting of the respective objects to be recommended; an extraction module, configured to extract the relationship between the objects to be recommended based on the set of the first feature vectors, and obtain a second feature vector corresponding to each object to be recommended; the second feature vector is used to represent the correlation between the objects to be recommended; A generation module is used to adjust the first feature vectors corresponding to each object to be recommended according to the adjustment coefficient corresponding to the sorting task for each sorting task, to obtain the adjusted first feature vectors of each object to be recommended under different sorting tasks, and to fuse the adjusted first feature vectors and second feature vectors corresponding to the same object to be recommended under each sorting task, to obtain the fused feature vectors of each object to be recommended under multiple sorting tasks; under each sorting task, the sorting of each object to be recommended is estimated according to the fused feature vector, to obtain the estimated sorting results of each object to be recommended under multiple sorting tasks; wherein the sorting task is a task of sorting each object to be recommended, and different sorting tasks sort the objects to be recommended from different dimensions; A fusion module is used to fuse the ranking estimation results of each to-be-recommended object under each of the ranking tasks to obtain a target ranking estimation result corresponding to the to-be-recommended object; The recommendation module is used to sort the objects to be recommended according to the target sorting estimation results corresponding to the objects to be recommended, and to recommend objects based on the sorted results.

13. The object recommendation device according to claim 12, wherein: The acquisition module is also used to obtain the object attribute feature vector and personalized semantic feature vector corresponding to each object to be recommended; the personalized semantic feature vector is used to characterize the correlation between the corresponding object to be recommended and the user; the object attribute feature vector and the personalized semantic feature vector corresponding to the same object to be recommended are combined to obtain the first feature vector corresponding to each object to be recommended; and a set of first feature vectors is determined based on the first feature vectors corresponding to each object to be recommended.

14. The object recommendation device according to claim 13, wherein: The acquisition module is further configured to perform vector concatenation on the object attribute feature vector corresponding to the same object to be recommended and the personalized semantic feature vector to obtain first feature vectors corresponding to each object to be recommended.

15. The object recommendation device according to claim 14, wherein: The objects to be recommended are objects to be recommended after initial sorting based on the corresponding personalized semantic feature vectors; the acquisition module is also used to position encode each first feature vector according to the position of the objects to be recommended after initial sorting to obtain a set of first feature vectors.

16. The object recommendation device according to claim 12, wherein: The generation module is also used to predict the status information of each object to be recommended under the sorting task based on the fused feature vector for each sorting task; and to perform a sorting estimate on each object to be recommended according to the status information of each object to be recommended under the sorting task to obtain a sorting estimate result of each object to be recommended under the sorting task.

17. The object recommendation device according to claim 12, wherein: The ranking estimation result includes a ranking estimation score; the target ranking estimation result includes a target ranking estimation score; the fusion module is also used to perform weighted fusion of the ranking estimation scores of each object to be recommended under each ranking task for each object to be recommended, to obtain the target ranking estimation score corresponding to the object to be recommended.

18. The object recommendation device according to any one of claims 12 to 17, characterized in that: The extraction module is also used to input the set of the first feature vectors into a trained object ranking model, so as to extract the relationship between the objects to be recommended based on the set of the first feature vectors in the feature encoding layer of the trained object ranking model, and obtain the second feature vectors corresponding to each object to be recommended.

19. The object recommendation device according to any one of claims 12 to 17, characterized in that: The generation module is also used to adjust the first feature vectors corresponding to each object to be recommended according to the adjustment coefficient corresponding to each sorting task in the trained object sorting model, respectively, to obtain the adjusted first feature vectors of each object to be recommended under different sorting tasks. Under each sorting task, the adjusted first feature vector and the second feature vector corresponding to the same object to be recommended are subjected to feature vector fusion to obtain the fused feature vectors of each object to be recommended under multiple sorting tasks.

20. The object recommendation device according to claim 19, wherein: The device also includes a training module, which is used to obtain a sample set including multiple groups of training samples; each group of training samples includes a first sample feature vector of a corresponding sample object and a state label of the sample object under multiple sorting tasks; the first sample feature vector contains a feature vector representing the correlation between the sample object and the user; based on the sample set, the object sorting model to be trained is trained to obtain an object sorting model for sorting objects according to the multiple sorting tasks.

21. The object recommendation device according to claim 20, wherein: The training module is also used to input the sample set into the object sorting model to be trained in each round of iterative training, so as to extract the relationship between the sample objects in the feature coding layer of the object sorting model to be trained based on the first sample feature vector in each group of training samples, and obtain the second sample feature vector corresponding to each sample object; for each sorting task, through the sorting prediction unit corresponding to the sorting task in the object sorting model to be trained, based on the first sample feature vector and the second sample feature vector corresponding to the same sample object, predict the state information of each sample object under the corresponding sorting task; according to the difference between the predicted state information and the state label, update the model parameters of the object sorting model to be trained until the iteration stop condition is reached, and obtain the object sorting model for sorting objects according to the multiple sorting tasks.

22. The object recommendation device according to claim 21, wherein: The training module is also used to determine, for each sorting task, a loss value between the predicted state information and the state label through an objective function corresponding to the task type to which the sorting task belongs; and update the model parameters of the object sorting model to be trained based on the loss value.

23. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 11 are implemented.

24. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

25. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

Images