Data pushing method and data sorting model training method and device
By adopting multi-objective feature fusion processing in data push, the problem of low data push accuracy in traditional technology is solved, and more efficient data push accuracy and sorting reference information accuracy are achieved.
Patent Information
- Application Number
- CN202310665175.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-06
- Publication Date
- 2024-12-06
AI Technical Summary
In traditional technology, the accuracy of data push is low, mainly because the basis is single, and it cannot effectively meet the multi-objective prediction needs.
By obtaining the object information of the target object and the data information of the data to be processed in the data pool, the preliminary feature data of multiple interactive prediction targets is used for forward fusion processing, and more accurate prediction results are generated, thereby improving the accuracy of data push.
Through multi-objective feature fusion processing, the accuracy of data push and the accuracy of sorting reference information are significantly improved, making data push more accurate and meeting the interest needs of target objects.
Smart Images

Figure CN119094600A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of Internet technology, and in particular to a data push method, a data sorting model training method, an apparatus, a computer device, a storage medium and a computer program product. Background Art
[0002] With the development of Internet technology, Internet platforms have become diversified, such as social platforms, video platforms, news platforms, shopping platforms, etc. Internet platforms can push data of interest to objects.
[0003] In traditional technology, a multi-target prediction method is used to determine the push priority score of the data to be processed. Each prediction target determines the prediction result of the prediction target based on the object information of the current object and the data information of the data to be processed. This method is based on a relatively single dimension and the push accuracy is not high. Summary of the invention
[0004] Based on this, it is necessary to provide a data push method, a data sorting model training method, device, computer equipment, computer-readable storage medium and computer program product that can improve the push accuracy in response to the above technical problems.
[0005] In a first aspect, the present application provides a data push method. The method comprises:
[0006] Obtain object information of the target object and data information corresponding to each piece of data to be processed in the data pool;
[0007] Based on the object information and the data information corresponding to each of the to-be-processed data, preliminary feature data of each of the to-be-processed data on at least two interactive prediction targets are obtained, and at least two of the interactive prediction targets have a sequential dependency relationship;
[0008] For each data to be processed, the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion feature corresponding to each interactive prediction target, and the forward fusion feature is the fusion feature corresponding to the interactive prediction target on which the corresponding interactive prediction target depends; based on the fusion feature corresponding to each interactive prediction target, the prediction result corresponding to each interactive prediction target is determined; based on the prediction result corresponding to each interactive prediction target, the sorting reference information of the current data to be processed is determined;
[0009] Based on the sorting reference information corresponding to each piece of data to be processed in the data pool, the push data corresponding to the target object is determined.
[0010] On the other hand, the present application also provides a data push device. The device includes:
[0011] A first acquisition module is used to acquire object information of a target object and data information corresponding to each piece of data to be processed in a data pool;
[0012] A second acquisition module is used to acquire preliminary feature data of each to-be-processed data on at least two interactive prediction targets based on the object information and the data information corresponding to each to-be-processed data, wherein the at least two interactive prediction targets have a sequential dependency relationship;
[0013] A fusion module is used to fuse the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature for each data to be processed, so as to obtain the fusion feature corresponding to each interactive prediction target, wherein the forward fusion feature is the fusion feature corresponding to the interactive prediction target on which the corresponding interactive prediction target depends; based on the fusion feature corresponding to each interactive prediction target, determine the prediction result corresponding to each interactive prediction target; based on the prediction result corresponding to each interactive prediction target, determine the sorting reference information of the current data to be processed;
[0014] The determination module is used to determine the push data corresponding to the target object based on the sorting reference information corresponding to each piece of data to be processed in the data pool.
[0015] In some embodiments, the second acquisition module is specifically used to perform feature mapping processing on the object information, the data information corresponding to the current data to be processed, and the interaction description information for each data to be processed to obtain each mapping feature; perform splicing processing on each mapping feature to obtain a spliced feature; perform feature extraction on the spliced feature through each feature extraction network to obtain the feature extraction results corresponding to each feature extraction network; obtain the weight information corresponding to each feature extraction network; determine the weighted features corresponding to each interaction prediction target according to the target feature extraction results corresponding to each target feature extraction network corresponding to each interaction prediction target and the target weight information corresponding to each target feature extraction network; perform nonlinear transformation on the weighted features corresponding to each interaction prediction target to obtain preliminary feature data of the current data to be processed on each interaction prediction target.
[0016] In some embodiments, the second acquisition module is specifically used to perform dimension transformation processing on the weight parameter matrix based on the splicing features to obtain a target vector whose dimension matches the number of feature extraction networks; and activate the target vector to obtain the weight information corresponding to each feature extraction network.
[0017] In some embodiments, the second acquisition module is specifically used to perform linear processing on the target feature extraction results and corresponding target weight information of each target feature extraction network corresponding to the current interaction prediction target for each interaction prediction target, to obtain the linear processing results of each target feature extraction network corresponding to the current interaction prediction target; perform statistical processing on the linear processing results of each target feature extraction network corresponding to the current interaction prediction target, to obtain the weighted features corresponding to the current interaction prediction target.
[0018] In some embodiments, the fusion module is specifically used to calculate, for each interactive prediction target, information filtering parameters based on preliminary feature data of the current data to be processed on the current interactive prediction target and the corresponding forward fusion features; calculate candidate output features based on preliminary feature data of the current data to be processed on the current interactive prediction target, the corresponding forward fusion features and information filtering parameters; calculate the fusion features corresponding to the current interactive prediction target based on the information filtering parameters, preliminary feature data of the current data to be processed on the current interactive prediction target and the candidate output features.
[0019] In some embodiments, the fusion module is specifically used to aggregate the prediction results corresponding to each interactive prediction target to obtain sorting reference information of the current data to be processed.
[0020] In some embodiments, the second acquisition module is specifically used to, for each data to be processed, based on the object information and the data information corresponding to the current data to be processed, obtain the preliminary feature data of the current data to be processed on at least two interaction prediction targets respectively through the preliminary feature extraction network in the data sorting model; the fusion module is specifically used to, for each interaction prediction target, fuse the preliminary feature data of the current data to be processed on the current interaction prediction target and the corresponding forward fusion feature through the state transition network corresponding to the current interaction prediction target in the data sorting model to obtain the fusion feature corresponding to the current interaction prediction target.
[0021] In some embodiments, the data push device also includes a training module, which is used to obtain training samples, and the training samples include sample object information of the sample object and sample data information of the sample data; based on the sample object information and the sample data information, the sample preliminary feature data of the sample data on at least two interactive prediction targets are obtained through the preliminary feature extraction network in the data sorting model to be trained, and there is a chronological dependency relationship between the at least two interactive prediction targets; for each interactive prediction target, the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion features are fused through the state transition network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion features corresponding to the current interactive prediction target; based on the sample fusion features corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are determined; based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target, the model loss is determined; based on the model loss, the parameters of the data sorting model to be trained are adjusted until the stopping condition of the model training is met, and the trained data sorting model is obtained.
[0022] In some embodiments, the first acquisition module is also used to respond to a login request from the target application, extract a login object identifier from the login request, and use the login object identifier as the target object; the determination module is also used to respond to a trigger operation on the data display entrance in the target application, and display the push data corresponding to the target object in sequence.
[0023] In some embodiments, the at least two interaction prediction targets include: executing a trigger operation on a data display portal and data interaction satisfying a preset condition, wherein the satisfying of the preset condition by the data interaction depends on the occurrence of executing a trigger operation on the data display portal.
[0024] On the other hand, the present application further provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the method provided in the first aspect when executing the computer program.
[0025] On the other hand, the present application further provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the method provided in the first aspect is implemented.
[0026] On the other hand, the present application further provides a computer program product, including a computer program, which implements the method provided in the first aspect when executed by a processor.
[0027] The above-mentioned data push method, device, computer equipment, storage medium and computer program product, after obtaining the object information of the target object and the data information corresponding to each data to be processed in the data pool, first obtain the preliminary feature data of each data to be processed on at least two interactive prediction targets based on the object information and the data information corresponding to each data to be processed; then, for each data to be processed, the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion feature corresponding to each interactive prediction target; then, based on the fusion feature corresponding to each interactive prediction target, the prediction result corresponding to each interactive prediction target is determined; This method of fusing the forward fusion features with the preliminary feature data corresponding to the current interactive prediction target makes the prediction result of the current interactive prediction target based on more information, greatly improving the accuracy of the prediction result corresponding to the current interactive prediction target; based on the prediction results corresponding to each interactive prediction target, the sorting reference information of the current data to be processed is determined, and the improvement in the accuracy of the prediction results can further promote the improvement in the accuracy of the sorting reference information; based on the sorting reference information corresponding to each data to be processed in the data pool, the push data corresponding to the target object is determined, and the improvement in the accuracy of the sorting reference information makes it more likely that the target object is interested in the pushed data, thereby improving the accuracy of data push.
[0028] In a second aspect, the present application provides a method for training a data sorting model. The method comprises:
[0029] Acquire a training sample, where the training sample includes sample object information of the sample object and sample data information of the sample data;
[0030] Based on the sample object information and the sample data information, obtaining sample preliminary feature data of the sample data on at least two interactive prediction targets respectively through a preliminary feature extraction network in the data sorting model to be trained, wherein at least the two interactive prediction targets have a sequential dependency relationship;
[0031] For each interactive prediction target, the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion features are fused through the state transfer network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion features corresponding to the current interactive prediction target;
[0032] Based on the sample fusion features corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are determined;
[0033] Based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target, the model loss is determined; based on the model loss, the parameters of the data sorting model to be trained are adjusted until the stopping condition of the model training is met, and the trained data sorting model is obtained. The data sorting model is used to determine the push data corresponding to the target object.
[0034] On the other hand, the present application also provides a training device for a data sorting model. The device comprises:
[0035] A third acquisition module is used to acquire a training sample, where the training sample includes sample object information of the sample object and sample data information of the sample data;
[0036] A fourth acquisition module is used to acquire sample preliminary feature data of the sample data on at least two interactive prediction targets respectively through a preliminary feature extraction network in a data sorting model to be trained based on the sample object information and the sample data information, wherein the at least two interactive prediction targets have a sequential dependency relationship;
[0037] The second fusion module is used to fuse the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion feature through the state transfer network corresponding to the current interactive prediction target in the data sorting model to be trained for each interactive prediction target, so as to obtain the sample fusion feature corresponding to the current interactive prediction target;
[0038] A second determination module is used to determine the prediction results corresponding to each interactive prediction target based on the sample fusion features corresponding to each interactive prediction target;
[0039] The training module is used to determine the model loss based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target; adjust the parameters of the data sorting model to be trained based on the model loss until the stopping condition of the model training is met, and obtain the trained data sorting model, which is used to determine the push data corresponding to the target object.
[0040] On the other hand, the present application further provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the method provided in the second aspect when executing the computer program.
[0041] On the other hand, the present application further provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the method provided in the second aspect is implemented.
[0042] On the other hand, the present application further provides a computer program product, including a computer program, which implements the method provided in the second aspect when executed by a processor.
[0043] The training method, apparatus, computer equipment, storage medium and computer program product of the above-mentioned data sorting model first obtain a training sample, which includes sample object information of the sample object and sample data information of the sample data. Then, based on the sample object information and the sample data information, the preliminary feature data of the sample data on at least two interactive prediction targets are obtained through the preliminary feature extraction network in the data sorting model to be trained, and there is a chronological dependency relationship between the at least two interactive prediction targets. Then, for each interactive prediction target, the preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion features are fused through the state transition network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion features corresponding to the current interactive prediction target. Based on the sample fusion features corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are determined. This method of performing feature fusion processing between targets with a dependency relationship makes the amount of information based on which the interactive prediction target is predicted increase, greatly improving the accuracy of the prediction results corresponding to the interactive prediction target. Finally, the model loss is determined based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target. The parameters of the data sorting model to be trained are adjusted based on the model loss until the stopping condition of the model training is met, and the trained data sorting model is obtained. During the above model training process, the fused features corresponding to the interactive prediction targets that the current interactive prediction targets depend on will participate in the prediction process of the current interactive prediction targets. This modeling method expands the feature richness of the interactive prediction targets and improves the prediction accuracy of the data sorting model. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 An application environment diagram of a data push method in an embodiment;
[0045] Figure 2 A schematic diagram of a flow chart of a data push method in an embodiment;
[0046] Figure 3 A schematic diagram of a model structure of a data sorting model in one embodiment;
[0047] Figure 4 A schematic diagram of a model structure of a data sorting model in another embodiment;
[0048] Figure 5 A schematic diagram of a data display entry in an embodiment;
[0049] Figure 6 A schematic diagram of a video display interface in one embodiment;
[0050] Figure 7 A schematic diagram of a flow chart of a method for training a data sorting model in one embodiment;
[0051] Figure 8 A schematic diagram of a process for determining push data in one embodiment;
[0052] Fig. 9 is a structural block diagram of a data push device in an embodiment;
[0053] Fig.10 is a structural block diagram of a training device for a data sorting model in one embodiment;
[0054] Fig.11 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.
[0056] The data push method provided in the embodiment of the present application can be applied to Figure 1In the application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or it can be placed on the cloud or other servers. The data push method provided in the embodiment of the present application can be executed by the terminal 102 or the server 104 alone, or it can be executed by the terminal 102 and the server 104 in collaboration, and the embodiment of the present application is not limited to this. During the execution process, the object information of the target object and the data information corresponding to each data to be processed in the data pool can be obtained first; then, based on the object information and the data information corresponding to each data to be processed, the preliminary feature data of each data to be processed on at least two interactive prediction targets are obtained, and there is a sequential dependency relationship between at least two interactive prediction targets; for each data to be processed, the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion feature corresponding to each interactive prediction target, and the forward fusion feature is the fusion feature corresponding to the interactive prediction target on which the corresponding interactive prediction target depends; based on the fusion feature corresponding to each interactive prediction target, the prediction result corresponding to each interactive prediction target is determined; based on the prediction result corresponding to each interactive prediction target, the sorting reference information of the current data to be processed is determined; finally, based on the sorting reference information corresponding to each data to be processed in the data pool, the push data corresponding to the target object is determined.
[0057] The terminal 102 may be, but is not limited to, various personal computers, laptops, smart phones, tablet computers, IoT devices, and portable wearable devices. The IoT devices may be smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. The portable wearable devices may be smart watches, smart bracelets, head-mounted devices, etc. The server 104 may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides cloud computing services.
[0058] In some embodiments, Figure 2 As shown, a data push method is provided, which can be Figure 1 The method may be executed by a terminal or server alone, or by a system consisting of a terminal and a server. Taking the method executed by a system consisting of a terminal and a server as an example, the data push method provided in the embodiment of the present application includes the following steps:
[0059] Step 202: Obtain object information of the target object and data information corresponding to each piece of data to be processed in the data pool.
[0060] Among them, the data in the data push method provided in the embodiment of the present application can be any type of data. Exemplarily, the data in the data push method provided in the embodiment of the present application can be video, news, message, etc., which is not limited in the embodiment of the present application.
[0061] Among them, different applications (Application, referred to as APP) can be installed on the terminal, and the APP corresponding to the data push method provided in the embodiment of the present application can be called the target application. The target application can be a social APP, a video APP, or a news APP, etc., which is not limited in the embodiment of the present application.
[0062] The target object may be an object identifier that currently needs to push data, and the object identifier that currently needs to push data may be an object identifier that is in an online state.
[0063] In a specific implementation, after detecting that the object has performed a login operation, the terminal can generate a login request and send the login request to the server. The server can extract the login object identifier from the login request and use the login object identifier as the object identifier in an online state, that is, use the login object identifier as the target object; or, after detecting that the object has performed a login operation, the terminal verifies the login information carried in the login request. If the verification passes, the terminal extracts the login object identifier from the login request and uses the login object identifier as the object identifier in an online state, that is, uses the login object identifier as the target object.
[0064] Among them, the object information of the target object may include information used to describe the characteristics of the target object, the object information of the target object may include object attribute information and data interaction information, the object attribute information may include a unique identification code of the target object in the target application; the data interaction information may include a data sequence composed of data on which the target object has performed interactive operations.
[0065] The data pool can be used to store data uploaded by all objects through the target application. In a specific implementation, the target application provides a data upload portal, through which the object can upload data. Each data stored in the data pool can be used as data to be processed, and the data information corresponding to each data to be processed can be obtained.
[0066] Among them, the data information corresponding to the data to be processed may include information used to describe the characteristics of the data to be processed, and the data information corresponding to the data to be processed may include data attribute information and object interaction information. The data attribute information may include a unique identification code of the data to be processed in the target application, the number of times the data to be processed has been interacted, etc.; the object interaction information may include an object sequence consisting of objects that have performed interactive operations on the data to be processed.
[0067] Step 204 , based on the object information and the data information corresponding to each of the to-be-processed data, obtain preliminary feature data of each of the to-be-processed data on at least two interactive prediction targets, wherein the at least two interactive prediction targets have a sequential dependency relationship.
[0068] Among them, for each data to be processed, the server can perform feature mapping processing on the object information of the target object and the data information of the data to be processed to obtain various mapping features; splice the various mapping features to obtain spliced features, and use the preliminary feature extraction network in the data sorting model to extract features from the spliced features to obtain preliminary feature data of the data to be processed on at least two interactive prediction targets respectively.
[0069] Among them, for the discrete information in the object information of the target object and the discrete information in the data information of the data to be processed, the server can perform vector mapping on them to obtain corresponding mapping features. The discrete information in the object information of the target object may include, for example, a unique identification code of the target object in the target application, and the discrete information in the data information of the data to be processed may include, for example, a unique identification code of the data to be processed in the target application.
[0070] Among them, for the continuous information in the object information of the target object and the continuous information in the data information of the data to be processed, the server can obtain the corresponding mapping features through value transformation. The continuous information in the object information of the target object, for example, object attribute information with digital values, and the continuous information in the data information of the data to be processed, for example, the number of times the data to be processed is interacted.
[0071] Among them, for the indefinite-length information in the object information of the target object and the indefinite-length information in the data information of the data to be processed, the server can sample and process the indefinite-length information to obtain a fixed-length feature, and the fixed-length feature can be used as the corresponding mapping feature. The indefinite-length information in the object information of the target object, for example, includes a data sequence composed of data on which the target object has performed interactive operations, and the indefinite-length information in the data information of the data to be processed, for example, includes an object sequence composed of objects on which interactive operations have been performed on the data to be processed.
[0072] Among them, for each data to be processed, after obtaining the mapping features in the above manner, the server can splice all the mapping features to obtain spliced features, and subsequently obtain preliminary feature data based on the spliced features.
[0073] Among them, at least two interactive prediction targets have a sequential dependency relationship, that is, the occurrence of at least two interactive prediction targets has a fixed path. Exemplarily, the at least two interactive prediction targets include target l, target k and target p. When target p occurs, target k may occur, and when target k occurs, target l may occur. The dependency relationship is: the occurrence of target l depends on the occurrence of target k, and the occurrence of target k depends on the occurrence of target p.
[0074] Step 206, for each data to be processed, the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion feature corresponding to each interactive prediction target, and the forward fusion feature is the fusion feature corresponding to the interactive prediction target on which the corresponding interactive prediction target depends; based on the fusion feature corresponding to each interactive prediction target, the prediction result corresponding to each interactive prediction target is determined; based on the prediction result corresponding to each interactive prediction target, the sorting reference information of the current data to be processed is determined.
[0075] Among them, for each data to be processed, since there are at least two interactive prediction targets, for each interactive prediction target, the server can fuse the preliminary feature data of the current data to be processed on the current interactive prediction target and the forward fusion feature to obtain the fusion feature corresponding to the current interactive prediction target, where the forward fusion feature is the fusion feature corresponding to the interactive prediction target on which the current interactive prediction target depends. The above processing is performed for each interactive prediction target to obtain the fusion feature corresponding to each interactive prediction target.
[0076] Exemplarily, at least two interactive prediction targets include target l, target k and target p, the occurrence of target l depends on the occurrence of target k, and the occurrence of target k depends on the occurrence of target p. The preliminary feature data of the current data to be processed on target p and the feature reference data can be fused to obtain the fused features corresponding to target p, and the fused features corresponding to target p are the forward fusion features of target k. The preliminary feature data of the current data to be processed on target k and the fused features corresponding to target p can be fused to obtain the fused features corresponding to target k, and the fused features corresponding to target k are the forward fusion features of target l. The preliminary feature data of the current data to be processed on target l and the fused features corresponding to target k can be fused to obtain the fused features corresponding to target l. Thus, the fused features corresponding to targets l, k and p can be obtained.
[0077] Among them, for each data to be processed, each interactive prediction target corresponds to a prediction network. After obtaining the fusion features corresponding to each interactive prediction target, the server can input each fusion feature into the corresponding prediction network and use the output of the prediction network as the prediction result corresponding to the corresponding interactive prediction target.
[0078] Among them, for each data to be processed, after obtaining the prediction results corresponding to each interactive prediction target, the server can combine the prediction results corresponding to each interactive prediction target to obtain the sorting reference information of the current data to be processed.
[0079] The server may traverse the data to be processed in the data pool, and use the currently traversed data to be processed as the current data to be processed. By performing the above operation, the sorting reference information of the current data to be processed may be obtained. After the traversal is completed, the sorting reference information of each data to be processed may be obtained. The sorting reference information is used to indicate the score of the push priority.
[0080] Step 208: Determine the push data corresponding to the target object based on the sorting reference information corresponding to each piece of data to be processed in the data pool.
[0081] Among them, after obtaining the sorting reference information corresponding to each of the to-be-processed data in the data pool, the server can use the to-be-processed data whose sorting reference information meets the preset conditions as the push data corresponding to the target object. Alternatively, in the embodiment of the present application, the data in the data pool is the data that is ranked in front of the preset number in the rough sorting result. After obtaining the sorting reference information corresponding to each of the to-be-processed data in the data pool, the server can find the to-be-processed data whose sorting reference information meets the preset conditions as the refined sorting result, and can adopt the preset strategy to further screen the refined sorting result to obtain the push data corresponding to the target object.
[0082] In the above embodiment, after obtaining the object information of the target object and the data information corresponding to each of the to-be-processed data in the data pool, the preliminary feature data of each to-be-processed data on at least two interactive prediction targets are first obtained based on the object information and the data information corresponding to each of the to-be-processed data; then, for each to-be-processed data, the preliminary feature data of the current to-be-processed data on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion feature corresponding to each of the interactive prediction targets; then, based on the fusion feature corresponding to each of the interactive prediction targets, the prediction result corresponding to each of the interactive prediction targets is determined; this method of fusion processing the forward fusion feature and the preliminary feature data corresponding to the current interactive prediction target makes the prediction result of the current interactive prediction target more based on the amount of information, greatly improving the accuracy of the prediction result corresponding to the current interactive prediction target; based on the prediction result corresponding to each of the interactive prediction targets, the sorting reference information of the current to-be-processed data is determined, and the improvement of the accuracy of the prediction result can further promote the improvement of the accuracy of the sorting reference information; based on the sorting reference information corresponding to each of the to-be-processed data in the data pool, the push data corresponding to the target object is determined, and the improvement of the accuracy of the sorting reference information makes it more likely that the target object is interested in the push data, thereby improving the accuracy of data push.
[0083] In some embodiments, based on the object information and the data information corresponding to each data to be processed, preliminary feature data of each data to be processed on at least two interactive prediction targets are obtained, including: for each data to be processed, feature mapping processing is performed on the object information, the data information corresponding to the current data to be processed, and the interactive description information to obtain each mapping feature; each mapping feature is spliced to obtain a spliced feature; feature extraction is performed on the spliced feature through each feature extraction network to obtain the feature extraction results corresponding to each feature extraction network; weight information corresponding to each feature extraction network is obtained; weighted features corresponding to each interactive prediction target are determined according to the target feature extraction results corresponding to each target feature extraction network corresponding to each interactive prediction target and the target weight information corresponding to each target feature extraction network; nonlinear transformation is performed on the weighted features corresponding to each interactive prediction target to obtain preliminary feature data of the current data to be processed on each interactive prediction target.
[0084] For each data to be processed, the server can perform vector mapping on the discrete information in the object information of the target object and the discrete information in the data information of the data to be processed to obtain the corresponding mapping features. The continuous information in the object information of the target object and the continuous information in the data information of the data to be processed are transformed to obtain the corresponding mapping features; the variable-length information in the object information of the target object and the variable-length information in the data information of the data to be processed are converted to fixed length to obtain the corresponding mapping features. After obtaining each mapping feature, the server splices all the mapping features to obtain the spliced features.
[0085] In the specific implementation, assume that there are K mapping features in total, and the kth mapping feature is recorded as F k , the K mapping features can be concatenated by the following formula to obtain the concatenated features:
[0086] X=concat(...,F k ,...)
[0087] Among them, X is the splicing feature, F k is the kth mapping feature.
[0088] There may be multiple feature extraction networks, wherein the multiple feature extraction networks include shared networks and private networks. The shared network refers to a network shared by at least two interactive prediction targets, and the private network refers to a network dedicated to the interactive prediction target.
[0089] Among them, the server can input the splicing features into each feature extraction network respectively, and each feature extraction network can extract features from the splicing features respectively to obtain the feature extraction results corresponding to each feature extraction network.
[0090] Among them, the server can obtain the weight information corresponding to each feature extraction network according to the splicing features and the weight parameter matrix. For each interactive prediction target, the feature extraction network corresponding to the interactive prediction target includes a shared network and a network dedicated to the interactive prediction target. Each shared network and each network dedicated to the interactive prediction target can be called a target feature extraction network corresponding to the interactive prediction target. The server can determine the weighted features corresponding to each interactive prediction target according to the target feature extraction results corresponding to each target feature extraction network corresponding to each interactive prediction target and the corresponding target weight information.
[0091] Among them, after obtaining the weighted features corresponding to each interactive prediction target, the server can perform nonlinear transformation on the weighted features corresponding to each interactive prediction target through a nonlinear transformation network, thereby obtaining preliminary feature data on each interactive prediction target of the corresponding data to be processed.
[0092] Exemplarily, the network structure of the feature extraction network and the nonlinear transformation network can adopt a multi-layer perceptron (MLP for short).
[0093] In the above embodiment, for each data to be processed, first, feature mapping is performed on the object information and the data information corresponding to the current data to be processed to obtain each mapping feature; each mapping feature is spliced to obtain a spliced feature; then, feature extraction is performed on the spliced feature through each feature extraction network to obtain the feature extraction results corresponding to each feature extraction network; and the weight information corresponding to each feature extraction network is obtained; then, according to the target feature extraction results corresponding to each target feature extraction network corresponding to each interactive prediction target and the target weight information corresponding to each target feature extraction network, the weighted features corresponding to each interactive prediction target are determined; finally, the weighted features corresponding to each interactive prediction target are nonlinearly transformed to obtain the preliminary feature data on each interactive prediction target of the current data to be processed. The preliminary feature data can be used for subsequent fusion processing, which makes the prediction result of the interactive prediction target more based on the amount of information, greatly improving the accuracy of the prediction result corresponding to the interactive prediction target.
[0094] In some embodiments, the weight information corresponding to each feature extraction network is obtained, including: performing dimension transformation processing on the weight parameter matrix based on the splicing features to obtain a target vector whose dimension matches the number of feature extraction networks; and performing activation processing on the target vector to obtain the weight information corresponding to each feature extraction network.
[0095] The server may multiply the concatenated features and the weight parameter matrix to achieve dimensional transformation of the weight parameter matrix, thereby obtaining a target vector whose dimension matches the number of feature extraction networks. After obtaining the target vector whose dimension matches the number of feature extraction networks, the server may use the first activation function to activate the target vector, thereby obtaining weight information corresponding to each feature extraction network.
[0096] In the specific implementation, the following formula can be used to determine the weight information corresponding to each feature extraction network:
[0097] g l (X) = soft max(W l ·X T )
[0098] Among them, W l is the weight parameter matrix, X T is the transpose of the concatenated features, soft max() is the first activation function, g l (X) is the weight vector. The number of elements in the weight vector is consistent with the number of feature extraction networks. Each element in the weight vector can be used as the weight information corresponding to each feature extraction network. The dimension of the weight parameter matrix is (ms+mp)×dx, ms is the number of shared networks in each feature extraction network, mp is the number of private networks in each feature extraction network, and dx is the number of elements in the spliced feature. Exemplarily, the spliced feature is a 1×50 vector, the transpose of the spliced feature is a 50×1 vector, ms is 2, mp is 2, the dimension of the weight parameter matrix is 4×50, and the dimension of the obtained weight vector is 4×1.
[0099] It should be noted that the weight information corresponding to each feature extraction network in the embodiment of the present application can be understood as the weight corresponding to each feature extraction network.
[0100] In the above embodiment, the weight parameter matrix and the splicing features can be combined to determine the weight information corresponding to each feature extraction network. Different interactive prediction targets correspond to network combinations composed of different feature extraction networks. After obtaining the weight information corresponding to each feature extraction network, it can be used to calculate the weighted features corresponding to each interactive prediction target, so that the weighted features are determined by combining the outputs of multiple feature extraction networks, thereby improving the feature accuracy.
[0101] In some embodiments, the weighted features corresponding to each interactive prediction target are determined based on the target feature extraction results and target weight information corresponding to each target feature extraction network corresponding to each interactive prediction target, including: for each interactive prediction target, linearly processing the target feature extraction results and corresponding target weight information of each target feature extraction network corresponding to the current interactive prediction target to obtain the linear processing results of each target feature extraction network corresponding to the current interactive prediction target; and statistically processing the linear processing results of each target feature extraction network corresponding to the current interactive prediction target to obtain the weighted features corresponding to the current interactive prediction target.
[0102] Among them, each interactive prediction target corresponds to a network combination composed of at least two feature extraction networks. For each interactive prediction target, each feature extraction network in the network combination corresponding to the interactive prediction target can be used as the target feature extraction network, the feature extraction result corresponding to the target feature extraction network can be used as the target feature extraction result, and the weight information corresponding to the target feature extraction network can be used as the target weight information.
[0103] Among them, for each interactive prediction target, the server can multiply the target feature extraction result of each target feature extraction network corresponding to the current interactive prediction target and the corresponding target weight information to obtain the linear processing result of each target feature extraction network corresponding to the current interactive prediction target; then add the linear processing results of each target feature extraction network corresponding to the current interactive prediction target to obtain the weighted feature corresponding to the current interactive prediction target. The weighted feature corresponding to each interactive prediction target can be obtained in the above manner.
[0104] In the specific implementation, the following formula can be used to calculate the weighted features corresponding to the current interactive prediction target:
[0105]
[0106] in, is the weighted feature corresponding to the current interactive prediction target, g l (X) i Extract the target weight information of the network for the i-th target feature corresponding to the current interactive prediction target, e i (X) is the target feature extraction result of the i-th target feature extraction network corresponding to the current interaction prediction target, and n is the number of target feature extraction networks corresponding to the current interaction prediction target.
[0107] In the above embodiment, different interactive prediction targets correspond to network combinations composed of different feature extraction networks. After obtaining the weight information corresponding to each feature extraction network, for each interactive prediction target, the weighted features corresponding to the interactive prediction target can be determined based on the weight information and feature extraction results of each feature extraction network in the network combination corresponding to the interactive prediction target, so that the weighted features are determined by combining the outputs of multiple feature extraction networks, thereby improving the feature accuracy.
[0108] In some embodiments, preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion features are fused to obtain the fusion features corresponding to each interactive prediction target, including: for each interactive prediction target, information filtering parameters are calculated based on the preliminary feature data of the current data to be processed on the current interactive prediction target and the corresponding forward fusion features; candidate output features are calculated based on the preliminary feature data of the current data to be processed on the current interactive prediction target, the corresponding forward fusion features and the information filtering parameters; fusion features corresponding to the current interactive prediction target are calculated based on the information filtering parameters, the preliminary feature data of the current data to be processed on the current interactive prediction target and the candidate output features.
[0109] Among them, for each interactive prediction target, the server can splice the preliminary feature data of the current data to be processed on the current interactive prediction target and the corresponding forward fusion features to obtain a first splicing result, multiply the first splicing result and the first information filtering matrix to obtain a first matrix multiplication result, add the first matrix multiplication result and the first bias parameter vector to obtain a vector to be standardized, and use a standardization function to process the vector to be standardized to obtain an information filtering parameter.
[0110] In a specific implementation, the following formula can be used to calculate the information filtering parameters:
[0111] f t =σ(W f ×[h t-1 ,x t ]+b f )
[0112] Among them, f t is the information filtering parameter, σ() is the normalization function, W f is the first information filtering matrix, h t-1 is the forward fusion feature, x t is the preliminary feature data of the current data to be processed on the current interactive prediction target, [h t-1 ,x t ] is the first splicing result obtained by splicing the preliminary feature data of the current data to be processed on the current interactive prediction target and the forward fusion feature, bf is the first bias parameter vector. t-1 and x t is a column vector, h t-1 The length is d h , x t The length is d x , concatenating the two column vectors to obtain [h t-1 ,x t ] is the length d w Column vector of w =d h +d x , W f The dimension is d h ×d w , b f The length is d h A column vector of . t The length is d h Column vector of .
[0113] Among them, after obtaining the information filtering parameters, the server can perform vector multiplication operations on the information filtering parameters and the forward fusion features to obtain a first vector multiplication result, splice the first vector multiplication result and the preliminary feature data of the current data to be processed on the current interactive prediction target to obtain a second splicing result, multiply the second splicing result with the second information filtering matrix to obtain a second matrix multiplication result, add the second matrix multiplication result and the second bias parameter vector to obtain a vector to be activated, use a second activation function to process the vector to be activated, and obtain candidate output features.
[0114] In the specific implementation, the following formula can be used to calculate the candidate output features:
[0115]
[0116] in, is the candidate output feature, tanh() is the second activation function, W h is the second information filtering matrix, f t is the information filtering parameter, ⊙ is the vector multiplication operation (element wise), h t-1 is the forward fusion feature, x t is the preliminary feature data of the current data to be processed on the current interactive prediction target, f t ⊙h t-1 is the first vector multiplication result obtained by performing vector multiplication on the information filtering parameter and the forward fusion feature, [f t ⊙h t-1 ,x t] is a second concatenation result obtained by concatenating the first vector multiplication result and the preliminary feature data of the current data to be processed on the current interactive prediction target, b f is the second bias parameter vector. t-1 and x t is a column vector, h t-1 The length is d h , x t The length is d x , f t The length is d h Column vector of t ⊙h t-1 The length is d h Column vector of h The length is d h The column vector, W h The dimension is d h ×d w , d w =d h +d x .
[0117] After obtaining the information filtering parameters and the candidate output features, the server may perform vector multiplication on the information filtering parameters and the candidate output features to obtain a second vector multiplication result; calculate the difference between the preset value and the information filtering parameter, perform vector multiplication on the difference and the forward fusion feature to obtain a third vector multiplication result; add the second vector multiplication result and the third vector multiplication result to obtain the fusion feature corresponding to the current interactive prediction target. Exemplarily, the preset value may be 1.
[0118] In the specific implementation, the following formula can be used to calculate the fusion feature corresponding to the current interactive prediction target:
[0119]
[0120] Among them, h t is the fusion feature corresponding to the current interactive prediction target, f t is the information filtering parameter, h t-1 is the forward fusion feature, are candidate output features.
[0121] In the above embodiment, for each data to be processed, when calculating the prediction result corresponding to each interactive prediction target, after obtaining the initial feature data corresponding to the original information such as the object information of the target object and the data information of the data to be processed, considering that the occurrence of the current interactive prediction target depends on the occurrence of another interactive prediction target, the feature corresponding to the other interactive prediction target is a very important consideration for the prediction of the current interactive prediction target, which greatly affects the accuracy of the prediction result corresponding to the current interactive prediction target. Therefore, the initial feature data corresponding to the current interactive prediction target and the fused feature corresponding to the dependent interactive prediction target are fused to obtain the fused feature corresponding to the current interactive prediction target, and the prediction result corresponding to the current interactive prediction target is determined based on the fused feature. The above fusion processing process expands the feature richness and improves the accuracy of the prediction result.
[0122] In some embodiments, based on the prediction results corresponding to each interactive prediction target, the sorting reference information of the current data to be processed is determined, including: aggregating the prediction results corresponding to each interactive prediction target to obtain the sorting reference information of the current data to be processed.
[0123] The aggregation processing may be multiplication, addition, or weighted summation, etc., which is not limited in the embodiments of the present application. Exemplarily, the aggregation processing is multiplication. After obtaining the prediction results corresponding to each interactive prediction target, the server multiplies the prediction results corresponding to each interactive prediction target, and uses the multiplication result as the sorting reference information of the current data to be processed.
[0124] In the above embodiment, after obtaining the prediction results corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are aggregated to obtain the sorting reference information of the current data to be processed. Compared with the single target prediction method, the above method combines the prediction results corresponding to multiple interactive prediction targets, so that the final sorting reference information is more accurate.
[0125] In some embodiments, based on the object information and the data information corresponding to each data to be processed, preliminary feature data of each data to be processed on at least two interactive prediction targets are obtained, including: for each data to be processed, based on the object information and the data information corresponding to the current data to be processed, the preliminary feature data of the current data to be processed on at least two interactive prediction targets are obtained through the preliminary feature extraction network in the data sorting model; for each data to be processed, the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion features are fused to obtain the fused features corresponding to each interactive prediction target, including: for each interactive prediction target, the preliminary feature data of the current data to be processed on the current interactive prediction target and the corresponding forward fusion features are fused through the state transfer network corresponding to the current interactive prediction target in the data sorting model to obtain the fused features corresponding to the current interactive prediction target.
[0126] Among them, the data sorting model includes a preliminary feature extraction network, a state transfer network corresponding to each interactive prediction target, and a prediction network corresponding to each interactive prediction target. The state transfer network corresponding to each interactive prediction target constitutes a state transfer network layer, and the prediction network corresponding to each interactive prediction target constitutes a prediction layer. The preliminary feature extraction network includes a feature mapping layer, a feature splicing layer, a feature extraction layer, a feature weighting layer, and a nonlinear transformation layer. The feature extraction layer includes multiple feature extraction networks, which can also be called expert networks. The multiple feature extraction networks include shared networks and private networks. The shared network refers to a network shared by all interactive prediction targets, and the private network refers to a network exclusive to the interactive prediction target. The network structure of multiple feature extraction networks can all be deep neural networks (Deep Neural Networks, referred to as DNN). The feature weighting layer includes a weighted processing network corresponding to each interactive prediction target. The nonlinear transformation layer includes a nonlinear transformation network corresponding to each interactive prediction target.
[0127] Among them, for each data to be processed, the server can input the object information of the target object, the data information of the data to be processed, and the context information into the feature mapping layer, and the feature mapping layer will output multiple mapping features. The context information includes the date and time point of the current moment, etc. The multiple mapping features can be input into the feature splicing layer, and the feature splicing layer will splice the multiple mapping features to obtain splicing features. The splicing features can be respectively input into each feature extraction network, and each feature extraction network outputs the corresponding feature extraction results. Each interactive prediction target corresponds to a network combination consisting of at least two feature extraction networks. For each interactive prediction target, each feature extraction network in the network combination corresponding to the interactive prediction target can be used as a target feature extraction network, and the feature extraction result corresponding to the target feature extraction network is used as the target feature extraction result. The feature extraction results corresponding to each target feature extraction network can be input into the weighted processing network corresponding to the interactive prediction target, and the weighted processing network outputs the corresponding weighted features. The weighted feature number is input into the nonlinear transformation network corresponding to the interactive prediction target, and the nonlinear transformation network outputs the preliminary feature data of the data to be processed on the interactive prediction target. The above processing is performed on each interactive prediction target, and preliminary feature data of the data to be processed on each interactive prediction target can be obtained.
[0128] Among them, after the server obtains the preliminary feature data of the data to be processed on each interactive prediction target, for each interactive prediction target, the server inputs the preliminary feature data of the data to be processed on the interactive prediction target and the fused features corresponding to the interactive prediction target on which the interactive prediction target depends, into the state transfer network corresponding to the interactive prediction target, and the state transfer network will output the fused features corresponding to the interactive prediction target. The fused features are input into the prediction network corresponding to the interactive prediction target, and the prediction network will output the prediction result corresponding to the interactive prediction target.
[0129] Among them, the structure of the state transfer network can adopt a recurrent neural network (Recurrent Neural Network, referred to as RNN). RNN has various forms, and a recurrent neural network in the form of a minimal gated unit (MGU) can be used.
[0130] For example, see Figure 3As shown, at least two interactive prediction targets include target l and target k. The occurrence of target l depends on the occurrence of target k. The feature extraction layer includes 4 feature extraction networks. The network structures of the 4 feature extraction networks are all DNNs. The 4 feature extraction networks are represented by DNN1, DNN2, DNN3, and DNN4, respectively, where DNN1 and DNN2 are shared networks, DNN3 is a network dedicated to target k, and DNN4 is a network dedicated to target l. For each data to be processed, the object information, the data information of the data to be processed, and the context information are input into the feature mapping layer, and the feature mapping layer will output multiple mapping features. The feature concatenation layer concatenates multiple mapping features to obtain concatenated features. DNN1, DNN2, DNN3, and DNN4 extract features from the concatenated features respectively. DNN1, DNN2, and DNN3 input the feature extraction results into the weighted processing network corresponding to target k. Figure 3 G(k) is used to represent that DNN1, DNN2, and DNN4 input the feature extraction results into the weighted processing network corresponding to target l. Figure 3 G(l) is used to represent it. The weighted features output by G(k) enter the nonlinear transformation network k corresponding to target k, and the nonlinear transformation network k outputs the preliminary feature data of the data to be processed on target k. The weighted features output by G(l) enter the nonlinear transformation network l corresponding to target l, and the nonlinear transformation network l outputs the preliminary feature data of the data to be processed on target l. The state transfer network corresponding to target l is represented by MGUl, and the state transfer network corresponding to target k is represented by MGUk. Corresponding to target l, MGUl fuses the preliminary feature data of the data to be processed on target l and the fused features output by MGUk to obtain the fused features corresponding to target l. The fused features corresponding to target l are input into the prediction network l corresponding to target l, and the prediction network l will output the prediction result corresponding to target l. The fused features output by MGUk are input into the prediction network k corresponding to target k, and the prediction network k will output the prediction result corresponding to target k.
[0131] It should be noted that: Figure 3 In the model structure shown, there are only two interactive prediction targets: target k and target l. The occurrence of target l depends on the occurrence of target k. Therefore, MGUl can fuse the preliminary feature data of the processed data on target l and the fused features output by MGUk to obtain the fused features corresponding to target l. However, target k has no dependent targets. MGUk can fuse the preliminary feature data of the processed data on target k with the preset data to obtain the fused features corresponding to target k, and output the fused features corresponding to target k as the forward fused features of target l to MGUl, so that MGUl can perform the above fusion processing to obtain the fused features corresponding to target l. Among them, the preset data can be continuously updated as part of the model parameters during the model training process.
[0132] See also Figure 4 As shown, the computational framework of the state transfer network can be shown as the dotted box, where h t-1 is the forward fusion feature, x t is the preliminary feature data of the data to be processed on the current interactive prediction target, h t is the fusion feature corresponding to the current interactive prediction target. Figure 3 For example, the target l in h t-1 is the fusion feature corresponding to target k, x t is the preliminary feature data of the data to be processed on the target l, h t is the fusion feature corresponding to target l. Figure 3 For example, h t-1 is the preset data, x t is the preliminary feature data of the data to be processed on target k, h t is the fusion feature corresponding to the target k. For the introduction of the calculation process in the calculation framework, please refer to the description of the above embodiment.
[0133] In the above embodiment, preliminary feature data of the data to be processed on at least two interactive prediction targets are obtained through the preliminary feature extraction network in the data sorting model, and the preliminary feature data and the forward fusion feature are fused through the state transition network corresponding to each interactive prediction target in the data sorting model. This method of performing feature fusion processing between targets with dependencies makes the prediction basis of the interactive prediction target more information, greatly improves the accuracy of the prediction results corresponding to the interactive prediction target, the sorting reference information of the data to be processed is determined according to the prediction results corresponding to each interactive prediction target, and the sorting reference information is also improved accordingly, and the push data corresponding to the target object is determined based on the sorting reference information of each data to be processed, and the target object is more likely to be interested in the pushed data.
[0134] In some embodiments, the training step of the data sorting model includes: obtaining training samples, the training samples include sample object information of the sample object and sample data information of the sample data; based on the sample object information and the sample data information, obtaining sample preliminary feature data of the sample data on at least two interactive prediction targets respectively through the preliminary feature extraction network in the data sorting model to be trained, and there is a sequential dependency relationship between the at least two interactive prediction targets; for each interactive prediction target, the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion features are fused through the state transition network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion features corresponding to the current interactive prediction target; based on the sample fusion features corresponding to each interactive prediction target, determining the prediction results corresponding to each interactive prediction target; based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target, determining the model loss; based on the model loss, adjusting the parameters of the data sorting model to be trained until the stopping condition of the model training is met, and obtaining the trained data sorting model.
[0135] Among them, different applications (Application, APP for short) can be installed on the terminal, and the APP corresponding to the data push method provided in the embodiment of the present application can be called the target application. Each object that has completed registration on the target application may or may not use the data display function in the process of using the target application. In the process of using the data display function, the object needs to first click on the data display entrance on the target application. At this time, the target application will jump to the data display interface and start playing the data. For each data played by the target application, the object may have different interactive operations on it, such as watching for a few seconds and then swiping away, liking, forwarding, etc. The object's interactive operations on each data can be saved as historical data.
[0136] The sample objects can be obtained from the objects registered on the target application, the objects registered on the target application can upload data through the target application, and the sample data can be obtained from all the uploaded data. The obtained sample objects and sample data can constitute training samples.
[0137] Among them, at least two interaction prediction targets include: executing a trigger operation on the data display portal and satisfying a preset condition for data interaction, and satisfying the preset condition for data interaction depends on the occurrence of executing a trigger operation on the data display portal. For sample objects and sample data in the training samples, it can be determined whether the sample objects have executed a trigger operation on the data display portal based on the historical data corresponding to the sample objects. If the sample objects have not executed a trigger operation on the data display portal, it can be determined that there is no interaction with the sample data. At this time, it can be determined that the label corresponding to the target of "executing a trigger operation on the data display portal" is 0, and it can also be determined that the label corresponding to the target of "data interaction satisfies the preset condition" is also 0. In the case where the sample object has performed a trigger operation on the data display entrance, it can be determined that the label corresponding to the goal of "performing a trigger operation on the data display entrance" is 1, and it can be further determined based on the historical data corresponding to the sample object whether the data interaction between the sample object and the sample data meets the preset conditions. If the preset conditions are met, it is determined that the label corresponding to the goal of "data interaction meets the preset conditions" is 1; if the preset conditions are not met, the label corresponding to the goal of "data interaction meets the preset conditions" is 0. The above use of 0 and 1 as labels is only an example, and other numerical values or characters may also be used, which is not limited in the embodiments of the present application.
[0138] Similar to the object information of the target object, the sample object information of the sample object may include information used to describe the characteristics of the sample object. For detailed description, please refer to the relevant description of the object information of the target object. Similar to the data information of the data to be processed, the sample data information of the sample data may include information used to describe the characteristics of the sample data. For detailed description, please refer to the relevant description of the data information of the data to be processed.
[0139] Among them, the data sorting model to be trained includes a preliminary feature extraction network, a state transfer network layer and a prediction layer. The state transfer network layer includes a state transfer network corresponding to each interactive prediction target, and the prediction layer includes a prediction network corresponding to each interactive prediction target.
[0140] Among them, at least two interactive prediction targets have a sequential dependency relationship, that is, the occurrence of at least two interactive prediction targets has a fixed path.
[0141] Among them, sample object information, sample data information and context information can be input into the preliminary feature extraction network, and the preliminary feature extraction network processes the input information to obtain sample preliminary feature data of the sample data on at least two interactive prediction targets.
[0142] In the specific implementation, the preliminary feature extraction network includes a feature mapping layer, a feature concatenation layer, a feature extraction layer, a feature weighting layer, and a nonlinear transformation layer. The feature extraction layer includes multiple feature extraction networks, which can also be called expert networks. The multiple feature extraction networks include shared networks and private networks. The shared network refers to the network shared by all interactive prediction targets, and the private network refers to the network exclusive to the interactive prediction target. The feature weighting layer includes a weighted processing network corresponding to each interactive prediction target. The nonlinear transformation layer includes a nonlinear transformation network corresponding to each interactive prediction target.
[0143] Among them, the server can input the sample object information of the sample object, the sample data information of the sample data and the sample context information into the feature mapping layer, and the feature mapping layer will output multiple sample mapping features. The multiple sample mapping features can be input into the feature splicing layer, and the feature splicing layer will splice the multiple sample mapping features to obtain sample splicing features. The sample splicing features can be respectively input into each feature extraction network, and each feature extraction network outputs the corresponding sample feature extraction results. Each interactive prediction target corresponds to a network combination consisting of at least two feature extraction networks. For each interactive prediction target, each feature extraction network in the network combination corresponding to the interactive prediction target can be used as a target feature extraction network. The sample feature extraction results corresponding to each target feature extraction network can be input into the weighted processing network corresponding to the interactive prediction target, and the weighted processing network outputs the corresponding sample weighted features. The sample weighted feature number is input into the nonlinear conversion network corresponding to the interactive prediction target, and the nonlinear conversion network outputs the sample preliminary feature data of the sample data on the interactive prediction target. The above processing is performed on each interactive prediction target to obtain the sample preliminary feature data of the sample data on each interactive prediction target.
[0144] Among them, after the server obtains the sample preliminary feature data of the sample data on each interactive prediction target, for each interactive prediction target, the server inputs the sample preliminary feature data of the sample data on the interactive prediction target and the sample fusion features corresponding to the interactive prediction target on which the interactive prediction target depends, into the state transfer network corresponding to the interactive prediction target, and the state transfer network will output the sample fusion features corresponding to the interactive prediction target. The sample fusion features are input into the prediction network corresponding to the interactive prediction target, and the prediction network will output the prediction result corresponding to the interactive prediction target.
[0145] Among them, for each interactive prediction target, the prediction result and label corresponding to the interactive prediction target can be input into the loss function corresponding to the interactive prediction target to obtain the loss value corresponding to the interactive prediction target, and the loss values corresponding to each interactive prediction target are added to obtain the model loss. Based on the model loss, the parameters of the data sorting model to be trained are adjusted, and the training continues until the stopping condition of the model training is met, thereby obtaining the trained data sorting model.
[0146] In the above embodiment, a training step of a data sorting model is provided. First, a training sample is obtained, and the training sample includes sample object information of a sample object and sample data information of a sample data. Then, based on the sample object information and the sample data information, sample preliminary feature data of the sample data on at least two interactive prediction targets are obtained through a preliminary feature extraction network in the data sorting model to be trained. There is a sequential dependency relationship between the at least two interactive prediction targets. Then, for each interactive prediction target, the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion features are fused through a state transition network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion features corresponding to the current interactive prediction target. Based on the sample fusion features corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are determined. This method of performing feature fusion processing between targets with a dependency relationship makes the amount of information based on which the interactive prediction target is predicted increase, greatly improving the accuracy of the prediction results corresponding to the interactive prediction target. Finally, the model loss is determined based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target. The parameters of the data sorting model to be trained are adjusted based on the model loss until the stopping condition of the model training is met, and the trained data sorting model is obtained. During the above model training process, the fused features corresponding to the interactive prediction targets that the current interactive prediction targets depend on will participate in the prediction process of the current interactive prediction targets. This modeling method expands the feature richness of the interactive prediction targets and improves the prediction accuracy of the data sorting model.
[0147] In some embodiments, before obtaining the object information of the target object and the data information corresponding to each piece of data to be processed in the data pool, it also includes: in response to a login request of the target application, extracting a login object identifier from the login request, and using the login object identifier as the target object; after determining the push data corresponding to the target object based on the sorting reference information corresponding to each piece of data to be processed in the data pool, it also includes: in response to a trigger operation on the data display entrance in the target application, displaying the push data corresponding to the target object in sequence.
[0148] Among them, the APP corresponding to the data push method provided in the embodiment of the present application can be called the target application. The target application can be a social APP, a video APP, or a news APP, etc., which is not limited in the embodiment of the present application.
[0149] Among them, the object can perform a trigger operation on the target application on the terminal. In response to the trigger operation, the terminal can display the target application login interface. The target application login interface includes an identification input box, a password input box, a login control and a cancel control. The object can input the login object identification through the identification input box, and input the password corresponding to the login object identification through the password input box, and then perform a trigger operation on the login control. The terminal responds to the trigger operation, generates a login request, and sends the login request to the server. After receiving the login request, the server extracts the login object identification from the login request, uses the login object identification as the target object, and then starts to execute the steps in the above embodiment: obtaining the object information of the target object and the data information corresponding to each data to be processed in the data pool, obtaining preliminary feature data, fusion processing, determining the push data corresponding to the target object, etc. The detailed process can be found in the description of the above embodiment.
[0150] After determining the push data corresponding to the target object, the server can save the data identifier of the push data corresponding to the target object. The target application provides a data display entry. When the object performs a trigger operation on the data display entry, the terminal can send a push request to the server. After receiving the push request, the server sends the push data corresponding to each of the saved data identifiers to the terminal in sequence, and the terminal displays the received push data in sequence.
[0151] Wherein, there are multiple push data corresponding to the target object, and the multiple push data have sorting reference information, and the server can send the push data to the terminal in descending order according to the sorting reference information. When the data in the data push method provided in the embodiment of the present application is a video, the terminal plays each received video in sequence according to the order in which the video is received.
[0152] It should be noted that: while watching a video played by the terminal, the subject may not be interested in the video and thus swipe away the video. In response to the swipe away operation, the terminal may directly start playing the next video.
[0153] In the above embodiment, in response to the login request of the target application, the login object identifier is extracted from the login request, and the login object identifier is used as the target object. Then, the push data corresponding to the target object is determined by the method provided in the embodiment of the present application, and in response to the triggering operation of the data display entrance in the target application, the push data corresponding to the target object is displayed in sequence. This method of embedding the data display entrance in the target application makes the functions of the target application more diversified and improves the object's experience of using the target application.
[0154] In some embodiments, the at least two interaction prediction targets include: executing a trigger operation on a data display portal and data interaction satisfying a preset condition, wherein the satisfying of the preset condition by the data interaction depends on the occurrence of executing a trigger operation on the data display portal.
[0155] The APP corresponding to the data push method provided in the embodiment of the present application can be called the target application. The data display entry can be an entry in the target application, and the trigger operation performed on the data display entry can be any operation that can jump to the data display interface, such as a click operation, a double-click operation, a slide operation, etc.
[0156] The data interaction meeting the preset conditions may be that the object has viewed the data for a preset time, the object has liked the data, the object has forwarded the data, or any other goal that can indicate that the object is interested in the data.
[0157] For example, the data in the data push method provided in the embodiment of the present application is a video, and the target application may be a social application. Figure 5 It is an interface in the target application, which includes two entrances, namely, entrance 1 and entrance 2. Entrance 1 can be a display entrance of a list of objects that have been socialized recently, and entrance 2 can be a video display entrance mentioned in the embodiment of the present application. After the object performs a trigger operation on the video display entrance, the terminal will jump to the video display interface, and the object can watch the video on the video display interface. After the video display interface plays a video, it will continue to play the next video. In the process of watching the video, the object may finish watching the video, and may perform a swipe operation after watching for a period of time. The object can watch the video for a preset time as an interactive prediction target. Obviously, the occurrence of the goal of "the object watches the video for a preset time" depends on the occurrence of the goal of "performing a trigger operation on the video display entrance". For each video to be processed in the video pool, it can be predicted whether the object will perform a trigger operation on the above-mentioned video display entrance, and the prediction result corresponding to the goal of "performing a trigger operation on the video display entrance" can be obtained. It can also be predicted whether the object can watch the video to be processed. The duration of the preset duration is obtained, and the prediction result corresponding to the goal of "watching the video for a preset duration" is obtained. Based on these two prediction results, the sorting reference information of the video to be processed is obtained.
[0158] In the above embodiment, a scenario is provided in which there are at least two interactive prediction targets with a dependency relationship. In this scenario, the pushed video object obtained using the data pushing method provided in the embodiment of the present application is more likely to be of interest.
[0159] In some embodiments, a method for training a data sorting model is provided. The method can be executed by a terminal or a server alone, or by a system consisting of a terminal and a server. For example, the method can be executed by a server alone. Figure 7 As shown, the method comprises the following steps:
[0160] Step 702: Acquire training samples, where the training samples include sample object information of sample objects and sample data information of sample data.
[0161] Among them, different applications (Application, referred to as APP) can be installed on the terminal, and the APP corresponding to the data push method provided in the embodiment of the present application can be called the target application. The server can obtain sample objects from the objects registered on the target application, and the objects registered on the target application can upload data through the target application, and the sample data can be obtained from all the uploaded data. The obtained sample objects and sample data can constitute training samples. For the detailed process, please refer to the description of the aforementioned embodiment.
[0162] Step 704, based on the sample object information and the sample data information, obtain sample preliminary feature data of the sample data on at least two interactive prediction targets respectively through the preliminary feature extraction network in the data sorting model to be trained, and at least two interactive prediction targets have a sequential dependency relationship.
[0163] The data sorting model to be trained includes a preliminary feature extraction network, a state transition network layer and a prediction layer, the state transition network layer includes a state transition network corresponding to each interactive prediction target, and the prediction layer includes a prediction network corresponding to each interactive prediction target. The method of obtaining sample preliminary feature data of sample data on at least two interactive prediction targets respectively through the preliminary feature extraction network can refer to the description of the above embodiment.
[0164] Step 706, for each interactive prediction target, the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion features are fused through the state transition network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion features corresponding to the current interactive prediction target.
[0165] Step 708: Determine the prediction results corresponding to each interactive prediction target based on the sample fusion features corresponding to each interactive prediction target.
[0166] Among them, after the server obtains the sample preliminary feature data of the sample data on each interactive prediction target, for each interactive prediction target, the sample preliminary feature data of the sample data on the interactive prediction target and the sample fusion feature corresponding to the interactive prediction target on which the interactive prediction target depends are input into the state transfer network corresponding to the interactive prediction target, and the state transfer network will output the sample fusion feature corresponding to the interactive prediction target, and the sample fusion feature is input into the prediction network corresponding to the interactive prediction target, and the prediction network will output the prediction result corresponding to the interactive prediction target. For the detailed fusion process, please refer to the description of fusing the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature, which will not be repeated in the embodiments of the present application.
[0167] Step 710, based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target, determine the model loss; based on the model loss, adjust the parameters of the data sorting model to be trained until the stopping condition of the model training is met, and obtain the trained data sorting model.
[0168] Among them, for each interactive prediction target, the prediction result and label corresponding to the interactive prediction target can be input into the loss function corresponding to the interactive prediction target to obtain the loss value corresponding to the interactive prediction target, and the loss values corresponding to each interactive prediction target are added to obtain the model loss. Based on the model loss, the parameters of the data sorting model to be trained are adjusted, and the training continues until the stopping condition of the model training is met, thereby obtaining the trained data sorting model.
[0169] In the above embodiment, a training sample is first obtained, and the training sample includes sample object information of the sample object and sample data information of the sample data. Then, based on the sample object information and the sample data information, the sample preliminary feature data of the sample data on at least two interactive prediction targets are obtained through the preliminary feature extraction network in the data sorting model to be trained, and there is a chronological dependency relationship between the at least two interactive prediction targets. Then, for each interactive prediction target, the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion features are fused through the state transition network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion features corresponding to the current interactive prediction target. Based on the sample fusion features corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are determined. This method of performing feature fusion processing between targets with a dependency relationship makes the amount of information based on which the interactive prediction target is predicted increase, greatly improving the accuracy of the prediction results corresponding to the interactive prediction target. Finally, the model loss is determined based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target. The parameters of the data sorting model to be trained are adjusted based on the model loss until the stopping condition of the model training is met, and the trained data sorting model is obtained. During the above model training process, the fused features corresponding to the interactive prediction targets that the current interactive prediction targets depend on will participate in the prediction process of the current interactive prediction targets. This modeling method expands the feature richness of the interactive prediction targets and improves the prediction accuracy of the data sorting model.
[0170] In some embodiments, a data push method is provided, which is described by taking the method executed by a server as an example, and the method includes:
[0171] In response to a login request from a target application, extract a login object identifier from the login request, and use the login object identifier as a target object; obtain object information of the target object and data information corresponding to each piece of data to be processed in a data pool;
[0172] For each data to be processed, feature mapping is performed on the object information, the data information corresponding to the current data to be processed, and the interaction description information to obtain various mapping features; each mapping feature is spliced to obtain a spliced feature; the spliced features are respectively extracted through each feature extraction network to obtain the feature extraction results corresponding to each feature extraction network; the weight parameter matrix is dimensionally transformed based on the spliced features to obtain a target vector whose dimension matches the number of feature extraction networks; the target vector is activated to obtain the weight information corresponding to each feature extraction network.
[0173] For each interactive prediction target, the target feature extraction results and corresponding target weight information of each target feature extraction network corresponding to the current interactive prediction target are linearly processed to obtain the linear processing results of each target feature extraction network corresponding to the current interactive prediction target; the linear processing results of each target feature extraction network corresponding to the current interactive prediction target are statistically processed to obtain the weighted features corresponding to the current interactive prediction target. The weighted features corresponding to each interactive prediction target are nonlinearly transformed to obtain the preliminary feature data of the current data to be processed on each interactive prediction target.
[0174] For each interactive prediction target, the information filtering parameters are calculated based on the preliminary feature data of the current data to be processed on the current interactive prediction target and the corresponding forward fusion features; the candidate output features are calculated based on the preliminary feature data of the current data to be processed on the current interactive prediction target, the corresponding forward fusion features and the information filtering parameters; the fusion features corresponding to the current interactive prediction target are calculated based on the information filtering parameters, the preliminary feature data of the current data to be processed on the current interactive prediction target and the candidate output features. The prediction results corresponding to each interactive prediction target are aggregated to obtain the sorting reference information of the current data to be processed.
[0175] Based on the sorting reference information corresponding to each piece of data to be processed in the data pool, the push data corresponding to the target object is determined.
[0176] In the above embodiment, after obtaining the object information of the target object and the data information corresponding to each of the to-be-processed data in the data pool, the preliminary feature data of each to-be-processed data on at least two interactive prediction targets are first obtained based on the object information and the data information corresponding to each of the to-be-processed data; then, for each to-be-processed data, the preliminary feature data of the current to-be-processed data on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion feature corresponding to each of the interactive prediction targets; then, based on the fusion feature corresponding to each of the interactive prediction targets, the prediction result corresponding to each of the interactive prediction targets is determined; this method of fusion processing the forward fusion feature and the preliminary feature data corresponding to the current interactive prediction target makes the prediction result of the current interactive prediction target more based on the amount of information, greatly improving the accuracy of the prediction result corresponding to the current interactive prediction target; based on the prediction result corresponding to each of the interactive prediction targets, the sorting reference information of the current to-be-processed data is determined, and the improvement of the accuracy of the prediction result can further promote the improvement of the accuracy of the sorting reference information; based on the sorting reference information corresponding to each of the to-be-processed data in the data pool, the push data corresponding to the target object is determined, and the improvement of the accuracy of the sorting reference information makes it more likely that the target object is interested in the push data, thereby improving the accuracy of data push.
[0177] In some embodiments, the data push method provided by the embodiment of the present application can be applied in a video push scenario. A target application corresponding to the data push method provided by the embodiment of the present application is installed on the terminal. After detecting that the object has performed a login operation on the target application, the terminal can generate a login request and send it to the server. The server extracts the login object identifier from the login request, uses the login object identifier as the target object, and obtains the object information of the target object. Based on the object information, thousands of videos are recalled from millions of candidate videos, and then hundreds of videos are screened from thousands of videos in a rough sorting manner. The hundreds of videos are stored in a video pool, and each video in the video pool is used as For the videos to be processed, the video information corresponding to each video to be processed can be obtained. Based on the object information and the video information corresponding to each video to be processed, the preliminary feature video of each video to be processed on at least two interactive prediction targets can be obtained; for each video to be processed, the preliminary feature video of the current video to be processed on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion features corresponding to each interactive prediction target; based on the fusion features corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are determined; based on the prediction results corresponding to each interactive prediction target, the sorting reference information of the current video to be processed is determined. The videos to be processed whose sorting reference information meets the preset conditions are taken as the refined sorting results, and the pre-set strategies are used to further screen the pushed videos corresponding to the target objects from the refined sorting results. After the object performs a trigger operation on the video display entrance on the target application, the terminal sends a push request to the server. After receiving the push request, the server sends the pushed videos corresponding to the target object to the terminal in sequence, and the terminal plays each pushed video in sequence according to the order of reception.
[0178] It should be noted that the application scenarios provided above are only used to illustrate the data push method of this application, and the application of the data push method provided by this application is not limited to the application scenarios provided above. For example, it can also be applied to product push scenarios, news push scenarios, etc., which are not limited in this embodiment of the application.
[0179] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.
[0180] Based on the same inventive concept, the embodiment of the present application also provides a data push device for implementing the data push method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more data push device embodiments provided below can refer to the limitations on the data push method above, and will not be repeated here.
[0181] In one embodiment, Fig. 9 As shown, a data push device is provided, comprising:
[0182] The first acquisition module 901 is used to acquire object information of the target object and data information corresponding to each piece of data to be processed in the data pool;
[0183] The second acquisition module 902 is used to acquire preliminary feature data of each to-be-processed data on at least two interactive prediction targets based on the object information and the data information corresponding to each to-be-processed data, wherein the at least two interactive prediction targets have a sequential dependency relationship;
[0184] The fusion module 903 is used to fuse the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature for each data to be processed, so as to obtain the fusion feature corresponding to each interactive prediction target, wherein the forward fusion feature is the fusion feature corresponding to the interactive prediction target on which the corresponding interactive prediction target depends; based on the fusion feature corresponding to each interactive prediction target, determine the prediction result corresponding to each interactive prediction target; based on the prediction result corresponding to each interactive prediction target, determine the sorting reference information of the current data to be processed;
[0185] The determination module 904 is used to determine the push data corresponding to the target object based on the sorting reference information corresponding to each piece of data to be processed in the data pool.
[0186] In some embodiments, the second acquisition module 902 is specifically used to perform feature mapping processing on the object information, the data information corresponding to the current data to be processed, and the interaction description information for each data to be processed to obtain each mapping feature; perform splicing processing on each mapping feature to obtain a spliced feature; perform feature extraction on the spliced feature through each feature extraction network to obtain the feature extraction results corresponding to each feature extraction network; obtain the weight information corresponding to each feature extraction network; determine the weighted features corresponding to each interaction prediction target according to the target feature extraction results corresponding to each target feature extraction network corresponding to each interaction prediction target and the target weight information corresponding to each target feature extraction network; perform nonlinear transformation on the weighted features corresponding to each interaction prediction target to obtain preliminary feature data of the current data to be processed on each interaction prediction target.
[0187] In some embodiments, the second acquisition module 902 is specifically used to perform dimension transformation processing on the weight parameter matrix based on the splicing features to obtain a target vector whose dimension matches the number of feature extraction networks; and activate the target vector to obtain the weight information corresponding to each feature extraction network.
[0188] In some embodiments, the second acquisition module 902 is specifically used to linearly process the target feature extraction results and corresponding target weight information of each target feature extraction network corresponding to the current interaction prediction target for each interaction prediction target, to obtain the linear processing results of each target feature extraction network corresponding to the current interaction prediction target; and perform statistical processing on the linear processing results of each target feature extraction network corresponding to the current interaction prediction target to obtain the weighted features corresponding to the current interaction prediction target.
[0189] In some embodiments, the fusion module 903 is specifically used to calculate, for each interactive prediction target, information filtering parameters based on preliminary feature data of the current data to be processed on the current interactive prediction target and the corresponding forward fusion features; calculate candidate output features based on preliminary feature data of the current data to be processed on the current interactive prediction target, the corresponding forward fusion features and information filtering parameters; calculate the fusion features corresponding to the current interactive prediction target based on the information filtering parameters, preliminary feature data of the current data to be processed on the current interactive prediction target and the candidate output features.
[0190] In some embodiments, the fusion module 903 is specifically used to aggregate the prediction results corresponding to each interactive prediction target to obtain the sorting reference information of the current data to be processed.
[0191] In some embodiments, the second acquisition module 902 is specifically used to, for each data to be processed, based on the object information and the data information corresponding to the current data to be processed, obtain the preliminary feature data of the current data to be processed on at least two interaction prediction targets respectively through the preliminary feature extraction network in the data sorting model; the fusion module is specifically used to, for each interaction prediction target, fuse the preliminary feature data of the current data to be processed on the current interaction prediction target and the corresponding forward fusion feature through the state transition network corresponding to the current interaction prediction target in the data sorting model to obtain the fusion feature corresponding to the current interaction prediction target.
[0192] In some embodiments, the data push device also includes a training module, which is used to obtain training samples, and the training samples include sample object information of the sample object and sample data information of the sample data; based on the sample object information and the sample data information, the sample preliminary feature data of the sample data on at least two interactive prediction targets are obtained through the preliminary feature extraction network in the data sorting model to be trained, and there is a chronological dependency relationship between the at least two interactive prediction targets; for each interactive prediction target, the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion features are fused through the state transition network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion features corresponding to the current interactive prediction target; based on the sample fusion features corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are determined; based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target, the model loss is determined; based on the model loss, the parameters of the data sorting model to be trained are adjusted until the stopping condition of the model training is met, and the trained data sorting model is obtained.
[0193] In some embodiments, the first acquisition module 901 is also used to respond to a login request from the target application, extract a login object identifier from the login request, and use the login object identifier as the target object; the determination module is also used to respond to a trigger operation on the data display entrance in the target application, and display the push data corresponding to the target object in sequence.
[0194] In some embodiments, the at least two interaction prediction targets include: executing a trigger operation on a data display portal and data interaction satisfying a preset condition, wherein the satisfying of the preset condition by the data interaction depends on the occurrence of executing a trigger operation on the data display portal.
[0195] In one embodiment, Fig.10 As shown, a training device for a data sorting model is provided, comprising:
[0196] The third acquisition module 1001 is used to acquire a training sample, where the training sample includes sample object information of a sample object and sample data information of a sample data;
[0197] A fourth acquisition module 1002 is used to acquire sample preliminary feature data of the sample data on at least two interactive prediction targets respectively based on the sample object information and the sample data information through a preliminary feature extraction network in the data sorting model to be trained, wherein the at least two interactive prediction targets have a sequential dependency relationship;
[0198] The second fusion module 1003 is used to fuse the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion feature for each interactive prediction target through the state transition network corresponding to the current interactive prediction target in the data sorting model to be trained, so as to obtain the sample fusion feature corresponding to the current interactive prediction target;
[0199] The second determination module 1004 is used to determine the prediction results corresponding to each interactive prediction target based on the sample fusion features corresponding to each interactive prediction target;
[0200] The training module 1005 is used to determine the model loss based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target; adjust the parameters of the data sorting model to be trained based on the model loss until the stopping condition of the model training is met, and obtain the trained data sorting model, which is used to determine the push data corresponding to the target object.
[0201] Each module in the above-mentioned data push device and data sorting model training device can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.
[0202] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig.11As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. Among them, the processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used 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 the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store object information and data information. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a data push method or a data sorting model training method is implemented.
[0203] Those skilled in the art will understand that Fig.11 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may 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 further provided, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above method embodiments when executing the computer program.
[0205] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.
[0206] In one embodiment, a computer program product is provided, including a computer program, which implements the steps in the above method embodiments when executed by a processor.
[0207] It should be noted that the object information (including but not limited to object attribute information, object behavior information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.
[0208] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed 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 the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. 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). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.
[0209] The technical features of the above embodiments may be arbitrarily combined. 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.
[0210] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the present application. It should be noted that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.
Claims
1. A data push method, characterized in that: The method comprises: Obtain object information of the target object and data information corresponding to each piece of data to be processed in the data pool; Based on the object information and data information corresponding to each of the to-be-processed data, preliminary feature data of each of the to-be-processed data on at least two interactive prediction targets are obtained, wherein the at least two interactive prediction targets have a sequential dependency relationship; For each data to be processed, the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion feature corresponding to each interactive prediction target, wherein the forward fusion feature is the fusion feature corresponding to the interactive prediction target on which the corresponding interactive prediction target depends; based on the fusion feature corresponding to each interactive prediction target, the prediction result corresponding to each interactive prediction target is determined; based on the prediction result corresponding to each interactive prediction target, the sorting reference information of the current data to be processed is determined; Based on the sorting reference information corresponding to each piece of to-be-processed data in the data pool, the push data corresponding to the target object is determined.
2. The method according to claim 1, characterized in that: The obtaining, based on the object information and the data information corresponding to each of the to-be-processed data, preliminary feature data of each to-be-processed data on at least two interactive prediction targets respectively includes: For each data to be processed, feature mapping is performed on the object information, data information corresponding to the current data to be processed, and interaction description information to obtain various mapping features; The mapping features are spliced to obtain spliced features; the spliced features are extracted by each feature extraction network to obtain feature extraction results corresponding to each feature extraction network; and weight information corresponding to each feature extraction network is obtained; Determine the weighted features corresponding to each interactive prediction target according to the target feature extraction results corresponding to each target feature extraction network corresponding to each interactive prediction target and the target weight information corresponding to each target feature extraction network; The weighted features corresponding to each interactive prediction target are transformed nonlinearly to obtain preliminary feature data of the current data to be processed on each interactive prediction target.
3. The method according to claim 2, characterized in that The obtaining of weight information corresponding to each feature extraction network includes: Performing dimension transformation processing on the weight parameter matrix based on the splicing features to obtain a target vector whose dimension matches the number of feature extraction networks; The target vector is activated to obtain weight information corresponding to each feature extraction network.
4. The method according to claim 2, characterized in that: The step of determining the weighted features corresponding to each interactive prediction target according to the target feature extraction results corresponding to each target feature extraction network corresponding to each interactive prediction target and the target weight information corresponding to each target feature extraction network includes: For each interactive prediction target, linearly process the target feature extraction result of each target feature extraction network corresponding to the current interactive prediction target and the corresponding target weight information to obtain the linear processing result of each target feature extraction network corresponding to the current interactive prediction target; The linear processing results of each target feature extraction network corresponding to the current interactive prediction target are statistically processed to obtain the weighted features corresponding to the current interactive prediction target.
5. The method according to claim 1, characterized in that: The preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion features corresponding to each interactive prediction target, including: For each interactive prediction target, information filtering parameters are calculated based on the preliminary feature data of the current data to be processed on the current interactive prediction target and the corresponding forward fusion features; Based on the preliminary feature data of the current data to be processed on the current interactive prediction target, the corresponding forward fusion features and the information filtering parameters, the candidate output features are calculated; Based on the information filtering parameters, preliminary feature data of the current data to be processed on the current interactive prediction target and the candidate output features, a fusion feature corresponding to the current interactive prediction target is calculated.
6. The method according to claim 1, characterized in that The step of determining the sorting reference information of the current data to be processed based on the prediction results corresponding to each interactive prediction target includes: The prediction results corresponding to each interactive prediction target are aggregated to obtain the sorting reference information of the current data to be processed.
7. The method according to claim 1, characterized in that The obtaining, based on the object information and the data information corresponding to each of the to-be-processed data, preliminary feature data of each to-be-processed data on at least two interactive prediction targets respectively includes: For each data to be processed, based on the object information and data information corresponding to the current data to be processed, preliminary feature data of the current data to be processed on at least two interactive prediction targets are obtained through a preliminary feature extraction network in a data sorting model; For each data to be processed, the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature are fused to obtain the fusion features corresponding to each interactive prediction target, including: For each interactive prediction target, the preliminary feature data of the current data to be processed on the current interactive prediction target and the corresponding forward fusion features are fused through the state transition network corresponding to the current interactive prediction target in the data sorting model to obtain the fusion features corresponding to the current interactive prediction target.
8. The method according to claim 7, characterized in that The training steps of the data sorting model include: Acquire a training sample, wherein the training sample includes sample object information of a sample object and sample data information of a sample data; Based on the sample object information and the sample data information, obtaining sample preliminary feature data of the sample data on at least two interactive prediction targets respectively through a preliminary feature extraction network in a data sorting model to be trained, wherein the at least two interactive prediction targets have a sequential dependency relationship; For each interactive prediction target, the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion feature are fused through the state transfer network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion feature corresponding to the current interactive prediction target; Based on the sample fusion features corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are determined; Based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target, the model loss is determined; based on the model loss, the parameters of the data sorting model to be trained are adjusted until the stopping condition of the model training is met, so as to obtain a trained data sorting model.
9. The method according to claim 1, characterized in that: Before obtaining the object information of the target object and the data information corresponding to each of the to-be-processed data in the data pool, the method further includes: In response to a login request of a target application, extracting a login object identifier from the login request, and using the login object identifier as a target object; After determining the push data corresponding to the target object based on the sorting reference information corresponding to each of the to-be-processed data in the data pool, the method further includes: In response to a triggering operation on a data display entry in the target application, push data corresponding to the target object is displayed in sequence.
10. The method according to any one of claims 1 to 9, characterized in that: The at least two interaction prediction targets include: executing a trigger operation on a data display portal and data interaction satisfying a preset condition, wherein the satisfying of the preset condition by the data interaction depends on the occurrence of executing a trigger operation on the data display portal.
11. A method for training a data sorting model, characterized in that: The method comprises: Acquire a training sample, wherein the training sample includes sample object information of a sample object and sample data information of a sample data; Based on the sample object information and the sample data information, obtaining sample preliminary feature data of the sample data on at least two interactive prediction targets respectively through a preliminary feature extraction network in a data sorting model to be trained, wherein the at least two interactive prediction targets have a sequential dependency relationship; For each interactive prediction target, the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion feature are fused through the state transfer network corresponding to the current interactive prediction target in the data sorting model to be trained to obtain the sample fusion feature corresponding to the current interactive prediction target; Based on the sample fusion features corresponding to each interactive prediction target, the prediction results corresponding to each interactive prediction target are determined; Based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target, the model loss is determined; based on the model loss, the parameters of the data sorting model to be trained are adjusted until the stopping condition of the model training is met, and the trained data sorting model is obtained, and the data sorting model is used to determine the pushed data corresponding to the target object.
12. A data push device, characterized in that: The device comprises: A first acquisition module is used to acquire object information of a target object and data information corresponding to each piece of data to be processed in a data pool; A second acquisition module is used to acquire preliminary feature data of each to-be-processed data on at least two interactive prediction targets based on the object information and data information corresponding to each to-be-processed data, wherein the at least two interactive prediction targets have a sequential dependency relationship; A fusion module is used to fuse the preliminary feature data of the current data to be processed on each interactive prediction target and the corresponding forward fusion feature for each data to be processed, so as to obtain the fusion feature corresponding to each interactive prediction target, wherein the forward fusion feature is the fusion feature corresponding to the interactive prediction target on which the corresponding interactive prediction target depends; based on the fusion feature corresponding to each interactive prediction target, determine the prediction result corresponding to each interactive prediction target; based on the prediction result corresponding to each interactive prediction target, determine the sorting reference information of the current data to be processed; The determination module is used to determine the push data corresponding to the target object based on the sorting reference information corresponding to each piece of to-be-processed data in the data pool.
13. A training device for a data sorting model, characterized in that: The device comprises: A third acquisition module is used to acquire a training sample, where the training sample includes sample object information of the sample object and sample data information of the sample data; A fourth acquisition module is used to acquire sample preliminary feature data of the sample data on at least two interactive prediction targets respectively based on the sample object information and the sample data information through a preliminary feature extraction network in a data sorting model to be trained, wherein the at least two interactive prediction targets have a sequential dependency relationship; The second fusion module is used to fuse the sample preliminary feature data of the sample data on the current interactive prediction target and the corresponding forward sample fusion feature for each interactive prediction target through the state transition network corresponding to the current interactive prediction target in the data sorting model to be trained, so as to obtain the sample fusion feature corresponding to the current interactive prediction target; A second determination module is used to determine the prediction results corresponding to each interactive prediction target based on the sample fusion features corresponding to each interactive prediction target; The training module is used to determine the model loss based on the prediction results corresponding to each interactive prediction target and the labels corresponding to each interactive prediction target; adjust the parameters of the data sorting model to be trained based on the model loss until the stop condition of the model training is met, and obtain the trained data sorting model, which is used to determine the push data corresponding to the target object.
14. 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.
15. A computer-readable storage medium having a computer program stored thereon, 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.
16. A computer program product comprising 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.
