Model processing method, device, equipment and storage medium
By adopting the idea of sorting learning in neural network models, we can build sample pairs by combining the object identifiers in the candidate object set and expanding the training samples, solving the low efficiency and overfitting problems caused by few training samples, and achieving faster convergence and more efficient training effects.
Patent Information
- Application Number
- CN202210875373.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-22
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-07-22
AI Technical Summary
In the advertising information push scenario, existing neural network models have a long training time, low efficiency, and are prone to overfitting the model.
Using the idea of sorting learning, we construct sample pairs by combining the object identifiers in the candidate object set, expand the training samples, and use position prediction functions to update model parameters to improve the convergence speed and training efficiency of the model.
It effectively improves the training efficiency and effect of the model, reduces the training time, improves the expressiveness and robustness of the model, can more accurately predict the behavioral parameters of the object, and optimizes the push decisions of information information.
Smart Images

Figure CN115130003B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence, and in particular to a model processing method, device, equipment and storage medium. Background Art
[0002] In the push scenario of information such as advertising information and event information, a neural network model can usually be trained to predict the behavior parameters of the object, and then the behavior parameters can be used to determine whether to push the information to the object. The behavior parameters can be used to measure the importance of a certain feedback behavior performed by the object on the information after receiving the pushed information. At present, the neural network model is usually trained by comparing the difference between the behavior parameters obtained by annotating the sample objects and the behavior parameters predicted by the neural network model for the sample objects; but since there are few sample objects with behavior parameters, this will result in fewer training samples for the neural network model, and fewer training samples will easily make it impossible for the neural network model to converge quickly during the training process, resulting in a longer model training time, which makes the model training efficiency lower. Summary of the invention
[0003] The embodiments of the present application provide a model processing method, apparatus, device, and storage medium, which can increase the convergence speed of the model, reduce the training time of the model, and improve the training efficiency of the model.
[0004] On the one hand, an embodiment of the present application provides a model processing method, including:
[0005] Obtain candidate sample pairs and corresponding sample annotation labels; the candidate sample pairs include two object identifiers in a candidate object set, the sample annotation labels are determined based on target behavior parameters of each object identifier in the candidate sample pairs, any target behavior parameter indicates the importance of feedback behavior under a target behavior type performed by the corresponding object on target information, and the sample annotation labels are used to mark the positional relationship between the two object identifiers in the candidate sample pairs in the candidate object set;
[0006] Calling the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs;
[0007] Using a position prediction function to predict the position relationship between two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information, to obtain a target position prediction result;
[0008] According to the sample annotation label and the target position prediction result, the model parameters of the target model are updated to obtain an updated target model corresponding to the target information; wherein the description information predicted by the updated target model corresponding to the target information information is used to decide whether to push the target information information to the corresponding object.
[0009] On the one hand, an embodiment of the present application provides a model processing device, including:
[0010] Acquire an original object set for updating the initial model; the original object set includes: object identifiers of multiple objects that have historically received the target information;
[0011] Select two object identifiers from the original object set to construct an original sample pair, and obtain an original annotation label based on the behavior label of each object identifier in the original sample pair; any behavior label indicates whether the corresponding object has performed feedback behavior on the target information; the original annotation label indicates the positional relationship between the two object identifiers in the original sample pair in the original object set;
[0012] The initial model is called to predict the behavior execution of the corresponding object according to the object features corresponding to each object identifier in the original sample pair, and obtain corresponding behavior prediction results; any behavior prediction result is used to indicate: the probability that the corresponding object performs feedback behavior on the target information;
[0013] Using the position prediction function to predict the position relationship between the two object identifiers in the original sample pair in the original object set according to the obtained behavior prediction results, to obtain the position prediction result corresponding to the original sample pair;
[0014] The initial model is updated based on the original annotation labels and the position prediction results corresponding to the original sample pairs to obtain a target model to be updated.
[0015] On the one hand, an embodiment of the present application provides an electronic device, characterized in that the electronic device includes an input interface and an output interface, and further includes:
[0016] a processor adapted to implement one or more instructions; and,
[0017] A computer storage medium storing one or more instructions, wherein the one or more instructions are suitable for being loaded by the processor and executing the above-mentioned model processing method.
[0018] On the one hand, an embodiment of the present application provides a computer storage medium, characterized in that computer program instructions are stored in the computer storage medium, and when the computer program instructions are executed by a processor, they are used to execute the above-mentioned model processing method.
[0019] On the one hand, an embodiment of the present application provides a computer program product, which includes a computer program stored in a computer storage medium; a processor of an electronic device reads the computer program from the computer storage medium, and the processor executes the computer program, so that the electronic device executes the above-mentioned model processing method.
[0020] The embodiment of the present application can construct a target model for predicting the description information of the behavior parameters of the object identifier of the object under the target behavior type. In the process of training the target model with the candidate object set, a candidate sample pair including two object identifiers in the candidate object set and a corresponding sample annotation label for marking the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set can be obtained, wherein the sample annotation label is determined based on the target behavior parameters of each object identifier in the candidate sample pair. Then, using the idea of ranking learning, the target model is called to predict the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the candidate sample pair, and the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set is predicted according to the predicted description information, and the target position prediction result is obtained, thereby updating the model parameters of the target model based on the sample annotation label and the target position prediction result. It can be seen that the embodiment of the present application utilizes the idea of ranking learning, so that the training samples can be expanded by combining each object identifier in the candidate object set in pairs, thereby increasing the number of training samples of the target model (one training sample is a sample pair) to a certain extent. This not only solves the problem of model overfitting caused by small training samples, but also enables the target model to learn more knowledge through richer training samples during the training process, thereby improving the convergence speed of the target model, reducing the training time of the target model, improving the training efficiency of the target model, and further improving the model training effect of the target model, so that the updated target model can have the advantages of good expressiveness and high robustness, and then the updated target model can more accurately predict the description information of the behavior parameters of the object identifier of any object under the target behavior type, so as to accurately decide whether to push the target information to the corresponding object based on the description information, thereby improving the push effect of the information information. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0022] Figure 1a It is a schematic diagram of a model processing solution jointly executed by a terminal device and a server provided in an embodiment of the present application;
[0023] Figure 1b It is a schematic diagram of another model processing solution jointly executed by a terminal device and a server provided in an embodiment of the present application;
[0024] Figure 2 It is a flow chart of a model processing method provided in an embodiment of the present application;
[0025] Figure 3 It is a flowchart of another model processing method provided in an embodiment of the present application;
[0026] Figure 4 is a schematic diagram of the structure of a target model provided in an embodiment of the present application;
[0027] Figure 5 It is a schematic diagram of processing intermediate features through a weight learning layer provided in an embodiment of the present application;
[0028] Figure 6 This is a schematic diagram of pushing target information to a target object provided by an embodiment of the present application;
[0029] Figure 7 This is a schematic diagram of pushing target information to multiple objects provided by an embodiment of the present application;
[0030] Figure 8 It is a flowchart of another model processing method provided in an embodiment of the present application;
[0031] Fig. 9 is a schematic diagram of obtaining a target model to be updated provided by an embodiment of the present application;
[0032] Fig.10 This is a schematic diagram of an embodiment of the present application that provides a method of expecting to push multiple information to a target object;
[0033] Fig.11 is a structural schematic diagram of a model processing device provided in an embodiment of the present application;
[0034] Fig.12It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0035] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0036] Artificial Intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.
[0037] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology (Computer Vision, CV), speech processing technology, natural language processing technology, and machine learning (Machine Learning, ML) / deep learning (Deep Learning, DL) and other major directions.
[0038] Among them, machine learning is a multi-disciplinary interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, supervised learning and other technologies.
[0039] The embodiment of the present application provides a model processing solution based on the machine learning technology in the field of artificial intelligence. The general principle of the model processing solution is as follows: first, a target model can be constructed to predict the descriptive information of the behavior parameters of the object identifier of the object under the target behavior type, and multiple sample pairs and corresponding annotation labels are obtained, wherein any sample pair includes two object identifiers in a candidate object set, and the candidate object set contains the object identifiers of objects that have historically performed feedback behaviors under the target behavior type on target information information. The annotation label of any sample pair is determined based on the target behavior parameters of each object identifier in the sample pair, and is used to mark the positional relationship between the two object identifiers in the sample pair in the candidate object set. Then, the idea of ranking learning can be used to call the target model to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to each object identifier in each sample pair, and predict the position relationship of the two object identifiers in each sample pair in the candidate object set according to the predicted description information, so as to obtain the position prediction results corresponding to each sample pair, and then update the model parameters of the target model based on the difference between the position prediction results corresponding to each sample pair and the corresponding annotation labels, so that the description information of the behavior parameters of the object identifier of any object under the target behavior type can be predicted based on the updated target model in the future, so as to decide whether to push the target information to the corresponding object based on the predicted description information.
[0040] Among them, the target information information mentioned above can be any information used to push to the object, for example, it can be advertising information, game activity information, etc. The target behavior type mentioned above can be any behavior type set according to specific needs, and there can be one or more feedback behaviors under one behavior type; the so-called feedback behavior can also be called conversion behavior, which can specifically refer to the behavior performed by the object after receiving the information information, such as click behavior, registration behavior, payment behavior, etc.; accordingly, the target behavior type can refer to: click behavior type, registration behavior type or payment behavior type, etc. In addition, the target behavior parameter of the object identifier mentioned above refers to the behavior parameter of the object identifier under the target behavior type, which is used to indicate the importance of the feedback behavior under the target behavior type performed by the object indicated by the object identifier on the target information information. Generally speaking, there are one or more feedback behaviors set according to specific needs under any behavior type, and each feedback behavior has different importance. The importance of any feedback behavior can be determined by the recognition degree of the specified object (such as the publisher of the target information information) to the feedback behavior. The higher the recognition degree of the specified object to the feedback behavior, the greater the value generated by the object performing the feedback behavior, and the higher the importance of the feedback behavior.
[0041] In a specific implementation, the model processing scheme proposed in the embodiment of the present application can be executed by an electronic device, wherein the electronic device can be a terminal device, which can include but is not limited to smart phones, tablet computers, laptops, desktop computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, smart wearable devices, etc.; or, the electronic device can also be a server, for example, it can be an independent physical server, or it can be a server cluster or distributed system composed of multiple physical servers, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (Content Delivery Network, CDN), and big data and artificial intelligence platforms.
[0042] Alternatively, the model processing solution proposed in the embodiment of the present application can also be executed by the terminal device and the server together. Figure 1a As shown, the terminal device can be responsible for obtaining multiple sample pairs and corresponding annotation labels; then, each sample pair and the corresponding annotation label are sent to the server, and the server uses the idea of ranking learning to perform corresponding steps according to each sample pair and the corresponding annotation label to update the model parameters of the target model, and apply the updated target model to the push scenario of the target information. For example, see Figure 1b As shown, the server can be responsible for obtaining multiple sample pairs and corresponding annotation labels, and then use the idea of ranking learning to perform corresponding steps to update the model parameters of the target model according to each sample pair and the corresponding annotation label, and then send the updated target model to the terminal device, so that the terminal device can predict the description information of the object identifier of any object according to the updated target model, and decide whether to push the target information to the corresponding object based on the predicted description information, so as to realize the application of the updated target model in the push scenario of the target information. It should be understood that Figure 1a and Figure 1b These are merely two schematic diagrams that exemplarily represent the two embodiments of the model processing scheme jointly executed by the terminal device and the server, and are not intended to be limiting.
[0043] It is particularly important to note that in the specific implementation of this application, object-related data, such as object characteristics, are involved. When the embodiments of this application are applied to specific products or technologies, the object's permission or consent must be obtained, and the collection, use and processing of relevant data must comply with local laws, regulations and standards.
[0044] Based on the above model processing solution, the present application embodiment provides a model processing method. Figure 2 , which is a flow chart of a model processing method provided in an embodiment of the present application. Figure 2 The model processing method shown can be executed by an electronic device, and can also be executed by any other device with computing power. The embodiment of the present application is described with reference to an electronic device. Figure 2 The model processing method shown may include the following steps:
[0045] S201, obtaining candidate sample pairs and corresponding sample annotation labels.
[0046] Among them, the candidate sample pair includes two object identifiers in the candidate object set, and the sample annotation label is determined based on the target behavior parameters of each object identifier in the candidate sample pair. Any target behavior parameter indicates the importance of the feedback behavior under the target behavior type performed by the corresponding object on the target information. The sample annotation label is used to mark the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set.
[0047] In one embodiment, the candidate object set includes: the object identification of the object that has performed the feedback behavior under the target behavior type on the target information information received historically, that is, the object identification of the object that has performed the feedback behavior under the target behavior type on the target information information can be added to the candidate object set from the objects that have pushed the target information historically (i.e., the objects that have received the target information historically); further, two object identifications can be selected from the candidate object set to construct a sample pair as a candidate sample pair. Specifically, two object identifiers can be randomly selected from the candidate object set to construct a candidate sample pair, that is, the two object identifiers selected in this way can be any two object identifiers in the candidate object set; for example, if the two object identifiers selected from the candidate object set are object A and object B, and a sample pair constructed based on object A and object B can be <object A, object B>, then the sample pair is used as a candidate sample pair, where object A can be called the first object identifier in the candidate sample pair, and object B can be called the second object identifier in the candidate sample pair; if another sample pair constructed based on object A and object B is <object B, object A>, at this time, the sample pair is used as a candidate sample pair, and the first object identifier in the candidate sample pair is object B, and the second object identifier in the candidate sample pair is object A. Optionally, two object identifiers can be selected from the candidate object set according to a preset object identifier selection strategy to construct a candidate sample pair, that is, the two object identifiers selected in this way must satisfy the object identifier selection strategy; for example, if the object identifier selection strategy indicates: object identifiers are selected in order of target behavior parameters from high to low, that is to say, one object identifier can be first selected in order of target behavior parameters from high to low, and when one object identifier is selected, another object identifier can be selected in order of target behavior parameters from high to low. For example, if the object identifiers included in the candidate object set are arranged from high to low according to the target behavior parameters, they are: object A, object B, object C and object D, if the first object identifier selected in order of target behavior parameters from high to low is object A, then when object A is selected, object B, object C and object D can be selected in sequence as another object identifier corresponding to object A.
[0048] In one embodiment, the sample annotation label can be used to mark the positional relationship between two object identifiers in the candidate sample pair in the candidate object set, and the sample annotation label can be determined based on the target behavior parameters of each object identifier in the candidate sample pair; wherein any target behavior parameter indicates the importance of the feedback behavior under the target behavior type performed by the corresponding object on the target information information, and optionally, the target behavior parameter can be obtained by judging the importance of the executed feedback behavior; generally speaking, when the object receives the target information information, the greater the importance of the feedback behavior under the target behavior type performed on the target information information, the greater the target behavior parameter; further, the target behavior parameter of the object identifier can be set by a specified object (for example, the delivery object of the target information information), or can be determined by the electronic device according to the target behavior parameter setting strategy set by the specified object; for example, when the target information information is advertising information, the delivery object of the target information information is the advertiser, for example, the target behavior parameter setting strategy can be set as follows: when the target behavior type is a paid behavior type, the target behavior parameter that generates a paid behavior and the payment amount is within different amount ranges can be set to different scores, for example, the paid behavior generated and the payment amount can be set to different scores. When the amount is (0, 200], the score of the target behavior parameter is set to 1, and a payment behavior will be generated. When the payment amount is (200, 500], the score of the target behavior parameter is set to 2, and a payment behavior will be generated. When the payment amount is (500, 1000], the score of the target behavior parameter is set to 3, and a payment behavior will be generated. When the payment amount is (1000, 2000], the score of the target behavior parameter is set to 4, and when the payment amount is (2000, ∞], the score of the target behavior parameter is set to 5. Further optionally, the electronic device generates a target behavior for the object identifier. The acquisition of parameters can be done by directly acquiring the target behavior parameters of the object identifier transmitted back by the specified object through the interactive interface output to the terminal device of the specified object, or by acquiring the target behavior parameter setting strategy transmitted back by the specified object, and then the electronic device generates the target behavior parameters of the corresponding object identifier based on the target behavior parameter setting strategy and the feedback behavior under the target behavior type performed by the object. Generally speaking, the arrangement position of the object identifier in the candidate sample set is related to the level of the target behavior parameter of the object identifier. The arrangement position of the object identifier with a high target behavior parameter in the candidate object set is located before the object identifier with a low target behavior parameter.
[0049] Furthermore, the target behavior parameter of any object identifier in the candidate sample pair used to determine the sample annotation label should be the target behavior parameter generated by the history of the object identifier, based on the target behavior parameter after deduplication at the object identifier * information granularity. That is to say, if the object identifier history generates multiple target behavior parameters, the maximum value of the multiple target behavior parameters should be used to determine the sample annotation label. It should be noted that the target behavior parameter of the object identifier mentioned in the embodiments of the present application should be the maximum value of the target behavior parameter generated by the history of the object identifier; for example, if the object indicated by the object identifier performs a first behavior type on the target information information after receiving the target information information The scores of the corresponding behavior parameters are 1, 2, and 3 respectively for multiple feedback behaviors of the second behavior type performed on the target information, and the scores of the corresponding behavior parameters are 3, 4, and 5 respectively. When the target behavior type is the first behavior type, the scores of the target behavior parameters generated by the object identification history of the object are 1, 2, and 3 respectively. The score of the target behavior parameter of the object identification used to determine the sample annotation label is 3. When the target behavior type is the second behavior type, the scores of the target behavior parameters generated by the object identification history of the object are 4, 5, and 6 respectively. The score of the target behavior parameter of the object identification used to determine the sample annotation label is 6.
[0050] S202, calling the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs.
[0051] In one embodiment, the target model to be updated can be any neural network model that can realize the descriptive information of the behavior parameters of the object identifier of the predicted object under the target behavior type and other related functions. It can be selected or designed according to specific needs. For example, the ranking learning model (i.e., LTR model) commonly used in the field of ranking learning (LTR) can be adopted, such as the ranknet model. In one embodiment, any description information can include: a prediction score for describing the behavior parameters under the target behavior type, that is, calling the target model to be updated according to the object characteristics corresponding to the object identifier can predict a score, which can be used to measure the target behavior parameters of the object identifier. Further, the score can also be used to measure the importance of the feedback behavior of the object indicated by the object identifier under the target behavior type performed on the target information information.
[0052] In one embodiment, the object characteristics may include at least one attribute characteristic related to the object, and the characteristic items of the attribute characteristics may be set according to specific needs. The embodiment of the present application does not limit the object characteristics. For example, the object characteristics may be set to include: attribute characteristics related to information such as the object's age, the object's gender, the object's permanent city of residence, and the terminal device held by the object.
[0053] S203, using a position prediction function to predict the position relationship between two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information, to obtain a target position prediction result.
[0054] In one embodiment, when any descriptive information includes a predicted score for describing a behavior parameter under a target behavior type, the electronic device uses a position prediction function to predict a target position prediction result based on each predicted descriptive information, which is related to the predicted scores of each object identifier in the candidate sample pair under the target behavior type. The object identifier with a high predicted score under the target behavior type is arranged in the candidate object set before the object identifier with a low predicted score under the target behavior type. The position prediction function can be any function that can realize the relevant function of predicting the positional relationship between two object identifiers in a candidate sample pair in the candidate object set based on each predicted descriptive information, and the embodiments of the present application are not limited thereto.
[0055] S204, updating the model parameters of the target model according to the sample annotation labels and the target position prediction results to obtain an updated target model corresponding to the target information.
[0056] Among them, the description information predicted by the updated target model corresponding to the target information is used to decide whether to push the target information to the corresponding object.
[0057] In one embodiment, the electronic device may update the target model by updating the model parameters in the target model in the direction of reducing the difference between the sample annotation label and the target position prediction result to obtain an updated target model corresponding to the target information.
[0058] The embodiment of the present application can construct a target model for predicting the description information of the behavior parameters of the object identifier of the object under the target behavior type. In the process of training the target model with the candidate object set, a candidate sample pair including two object identifiers in the candidate object set and a corresponding sample annotation label for marking the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set can be obtained, wherein the sample annotation label is determined based on the target behavior parameters of each object identifier in the candidate sample pair. Then, using the idea of ranking learning, the target model is called to predict the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the candidate sample pair, and the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set is predicted according to the predicted description information, and the target position prediction result is obtained, thereby updating the model parameters of the target model based on the sample annotation label and the target position prediction result. It can be seen that the embodiment of the present application utilizes the idea of ranking learning, so that the training samples can be expanded by combining each object identifier in the candidate object set in pairs, thereby increasing the number of training samples of the target model (one training sample is a sample pair) to a certain extent. This not only solves the problem of model overfitting caused by small training samples, but also enables the target model to learn more knowledge through richer training samples during the training process, thereby improving the convergence speed of the target model, reducing the training time of the target model, improving the training efficiency of the target model, and further improving the model training effect of the target model, so that the updated target model can have the advantages of good expressiveness and high robustness, and then the updated target model can more accurately predict the description information of the behavior parameters of the object identifier of any object under the target behavior type, so as to accurately decide whether to push the target information to the corresponding object based on the description information, thereby improving the push effect of the information information.
[0059] Based on the above-mentioned related embodiments of the model processing method, the present application embodiment provides another model processing method. Figure 3 , which is a flow chart of another model processing method provided in an embodiment of the present application. Figure 3 The model processing method shown can be executed by an electronic device, and can also be executed by any other device with computing power. The embodiment of the present application is described with reference to an electronic device. Figure 3 The model processing method shown may include the following steps:
[0060] S301, obtaining candidate sample pairs and corresponding sample annotation labels.
[0061] Among them, the candidate sample pair includes two object identifiers in the candidate object set, and the sample annotation label is determined based on the target behavior parameters of each object identifier in the candidate sample pair. Any target behavior parameter indicates the importance of the feedback behavior under the target behavior type performed by the corresponding object on the target information. The sample annotation label is used to mark the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set.
[0062] In a specific implementation, the arrangement position of the first object identifier in the candidate sample pair in the candidate object set can be called the first arrangement position, and the arrangement position of the second object identifier in the candidate sample pair in the candidate object set can be called the second arrangement position; the electronic device obtains the candidate sample pairs and the corresponding sample annotation labels, which may include: obtaining the candidate sample pairs, and the target behavior parameters of each object identifier in the candidate sample pairs; if the target behavior parameter of the first object identifier in the candidate sample pair is greater than the target behavior parameter of the second object identifier in the candidate sample pair, then the first label is determined as the sample annotation label, and the first label is used to indicate that: the first arrangement position is located before the second arrangement position; if the target behavior parameter of the first object identifier in the candidate sample pair is equal to the target behavior parameter of the second object identifier in the candidate sample pair, then the second label is determined as the sample annotation label, and the second label is used to indicate: the first arrangement position and the second arrangement position is the same position; if the target behavior parameter of the first object identifier in the candidate sample pair is less than the target behavior parameter of the second object identifier in the candidate sample pair, the third label is determined as the sample annotation label, and the third label is used to indicate that the first arrangement position is after the second arrangement position; wherein, the first label, the second label and the third label can be values set according to specific needs, and the first label, the second label and the third label should satisfy: the first label is greater than the second label, the second label is greater than the third label, and the distance from the second label to the first label should be equal to the distance from the second label to the third label, that is, the second label should be the average of the first label and the third label; optionally, the first label can be set to 1, the second label can be set to 0, and the third label can be set to -1, and the first label can also be set to 1, the second label can be set to 0.5, and the third label can be set to 0, then the sample annotation label satisfies the following formula 1:
[0063]
[0064] Among them, y represents the sample annotation label, weight1 represents the target behavior parameter of the first object identifier in the candidate sample pair, and weight2 represents the target behavior parameter of the second object identifier in the candidate sample pair.
[0065] S302, calling the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs.
[0066] In one embodiment, if the number of attribute features included in the object feature is N, that is, any object feature includes N attribute features, and N is a positive integer, the electronic device calls the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to each object identifier in the candidate sample pair, which may include: calling the target model to be updated, and performing feature dimensionality reduction processing on each attribute feature in the object feature corresponding to the i-th object identifier in the candidate sample pair, to obtain N intermediate features; i∈[1,2]; obtaining the feature weight of each intermediate feature in the N intermediate features; any feature weight is used to indicate the importance of the corresponding intermediate feature; based on each intermediate feature and the corresponding feature weight, predicting the description information of the behavior parameters of the i-th object identifier under the target behavior type; wherein any intermediate feature is a multidimensional vector.
[0067] In a specific implementation, the electronic device calls the target model to be updated, performs feature dimension reduction processing on each attribute feature in the object feature corresponding to the i-th object identifier in the candidate sample pair, and obtains N intermediate features, which can be achieved by performing feature embedding processing on each attribute feature. Further, the electronic device obtains the feature weight of each intermediate feature in the N intermediate features, which can include: performing feature squeezing processing on each intermediate feature in the N intermediate features to obtain N feature squeezing values, wherein any feature squeezing value is a one-dimensional value; determining the feature weight of the n-th intermediate feature according to the correlation between the feature squeezing value corresponding to the n-th intermediate feature in the N intermediate features and each feature squeezing value, n∈[1,N]. Further, the electronic device predicts the description information of the behavior parameters of the i-th object identifier under the target behavior type according to each intermediate feature and the corresponding feature weight, which can include: performing weighted processing on each intermediate feature and the corresponding feature weight to obtain N weighted features; performing splicing processing on the N weighted features to obtain spliced features, and then predicting the description information of the behavior parameters of the i-th object identifier under the target behavior type based on the spliced features.
[0068] See also Figure 4, is a schematic diagram of the structure of a target model provided in an embodiment of the present application, and the target model may include an embedding layer (i.e., EmbeddingLayer), a weight learning layer (which may be a Squeeze-and-ExcitationNetworksLayer, i.e., SENET Layer), a concatenation layer (i.e., Concat Layer), and a fully connected layer (i.e., DenseLayer). Among them, the embedding layer is mainly used to perform feature embedding processing on each attribute feature in the object feature corresponding to the i-th object identifier in the candidate sample pair, so as to achieve feature dimensionality reduction processing of each attribute feature, and obtain N intermediate features; among them, the dimension of the embedding layer can be set according to specific needs, for example, the dimension of the embedding layer can be set to 20, 10, etc., which is not limited in the embodiment of the present application. The embedding layer can be used to achieve dimensionality reduction of each attribute feature, which can reduce the subsequent processing data of the target model, save computing resources, and speed up the processing speed.
[0069] Furthermore, the weight learning layer can be used to learn the feature weight of each of the N intermediate features, and weight each intermediate feature and the corresponding feature weight to obtain N weighted features. Since any feature weight is used to indicate the importance of the corresponding intermediate feature, the introduction of the weight learning layer can realize the screening of intermediate features based on the importance of the intermediate features, so that the target model can focus on learning the intermediate features with high importance (i.e., high feature weight), thereby improving the expressiveness of the updated target model. Figure 5 , is a schematic diagram of processing intermediate features through a weight learning layer provided in an embodiment of the present application, as shown by the mark 501, there are N intermediate features, each of which is a multidimensional vector, as shown by the mark 502, there are N feature squeezing values obtained, each of which is a one-dimensional numerical value, as shown by the mark 503, there are N feature weights obtained, as shown by the mark 504, there are N weighted features obtained; wherein, through the weight learning layer, each of the N intermediate features is subjected to feature squeezing processing to obtain N feature squeezing values, which can be called a Squeeze operation, and this operation can be implemented by introducing relevant feature squeezing functions; through the weight learning layer, according to the correlation between the feature squeezing value corresponding to the nth intermediate feature of the N intermediate features and each feature squeezing value, determining the feature weight of the nth intermediate feature can be called an Excitation operation, and this operation can be implemented by introducing relevant feature weight determination functions; through the weight learning layer, each intermediate feature and the corresponding feature weight are weighted to obtain N weighted features, which can be called a Reweight operation, and this operation can be implemented by introducing relevant weighting functions.
[0070] Furthermore, the splicing layer is mainly used to splice N weighted features to obtain spliced features; the description information of the behavior parameters of the ith object identifier under the target behavior type can be predicted based on the spliced features through at least one fully connected layer, and the spliced features can be mapped to a score, which can be a predicted score for describing the behavior parameters under the target behavior type. At this time, it is not expected that the target model can directly predict the target behavior parameters of the object identifier, but only predict a score that can measure the target behavior parameters of the object identifier, and then the predicted score is used as the description information. Optionally, the score can be a predicted behavior parameter, that is, at this time, it is expected that the target model can directly predict the target behavior parameters of the object identifier, and then the predicted behavior parameters are used as the description information. Among them, the number of fully connected layers and the dimensions of each fully connected layer can be determined according to specific needs. For example, it can be set to include 3 fully connected layers, and the dimensions of the three fully connected layers are 128 dimensions, 64 dimensions, and 32 dimensions respectively.
[0071] In another feasible implementation, when the electronic device determines the feature weight of the nth intermediate feature based on the correlation between the feature squeezing value corresponding to the nth intermediate feature among the N intermediate features and each feature squeezing value, it can also sum the feature squeezing values corresponding to the intermediate features that are correlated with the nth intermediate feature among the N intermediate features to obtain the feature correlation value corresponding to the nth intermediate feature, and determine the feature weight of the nth intermediate feature based on the feature correlation values corresponding to the N intermediate features and the feature correlation value corresponding to the nth intermediate feature. Optionally, when the electronic device determines the feature weight of the nth intermediate feature based on the feature correlation values corresponding to the N intermediate features and the feature correlation value corresponding to the nth intermediate feature, it can use the result of normalizing the feature correlation value corresponding to the nth intermediate feature as the feature weight of the nth intermediate feature, and can also use the proportion of the feature correlation value corresponding to the nth intermediate feature in the total correlation value constituted by the feature correlation values corresponding to the N intermediate features as the feature weight of the nth intermediate feature.
[0072] S303, using a position prediction function to predict the position relationship between two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information, to obtain a target position prediction result.
[0073] In one embodiment, the electronic device uses a position prediction function to predict the position relationship between two object identifiers in a candidate sample pair in a candidate object set according to the predicted description information to obtain a target position prediction result, including: using a position prediction function to predict the probability that the first arrangement position is before the second arrangement position according to the predicted description information to obtain a predicted probability; and using the predicted probability as the target position prediction result; wherein, when the predicted probability is greater than a reference probability, it indicates that the predicted position relationship is: the first arrangement position is before the second arrangement position; when the predicted probability is equal to the reference probability, it indicates that the predicted position relationship is: the first arrangement position and the second arrangement position are the same position; when the predicted probability is less than the reference probability, it indicates that the predicted position relationship is: the first arrangement position is after the second arrangement position; wherein, the reference probability is a second label indicating that the first arrangement position and the second arrangement position are the same position, and if the sample annotation label is determined based on the above formula 1, the reference probability is 0.5.
[0074] In one embodiment, when any description information includes a predicted score for describing a behavior parameter under a target behavior type, the electronic device uses a position prediction function to predict the probability that the first arrangement position is before the second arrangement position based on the predicted description information to obtain the predicted probability, which may include: using the position prediction function to perform a difference operation on the predicted score of the first object identifier in the candidate sample pair under the target behavior type and the predicted score of the second object identifier under the target behavior type to obtain a difference operation result; mapping the difference operation result to a preset numerical interval to obtain the predicted probability; wherein the preset numerical interval is a preset numerical interval for determining the value of the sample annotation label, that is, a preset numerical interval for determining the first label, the second label and the third label of the sample annotation label. More specifically, the preset numerical interval is a numerical interval composed of the first label, the second label and the third label. When the first label is 1, the second label is 0.5 and the third label is 0, the preset numerical interval should be [0,1], that is, the electronic device can map the difference operation result to the preset numerical interval to obtain the predicted probability.
[0075] In a specific implementation, when setting the sample annotation label, when the target behavior parameter of the first object identifier in the candidate sample pair is greater than the target behavior parameter of the second object identifier, the first label is determined as the sample annotation label, and when the target behavior parameter of the first object identifier in the candidate sample pair is less than the target behavior parameter of the second object identifier, the third label is determined as the sample annotation label. Therefore, when the difference operation result is mapped to the preset numerical interval, the difference operation result can be mapped in direct proportion when the predicted probability is obtained; further optionally, since the difference operation result can be distributed throughout the (-∞, +∞) interval, the position prediction function can be implemented by an activation function that can realize the positive proportional mapping of real numbers in the (-∞, +∞) interval to the preset numerical interval. When the preset numerical interval is [0, 1], the activation function can use the sigmoid activation function, which can be specifically shown by the following formula 2:
[0076] p=Sigmoid(p1-p2) (2)
[0077] Among them, p represents the predicted probability, that is, the probability that the first arrangement position is before the second arrangement position, p1 represents the description information of the behavior parameters of the first object identifier in the candidate sample pair under the target behavior type, that is, the predicted score of the first object identifier in the candidate sample pair under the target behavior type, and p2 represents the description information of the behavior parameters of the second object identifier in the candidate sample pair under the target behavior type, that is, the predicted score of the second object identifier in the candidate sample pair under the target behavior type.
[0078] In one embodiment, when any description information includes a predicted behavior parameter under a target behavior type obtained by prediction of a target model, the electronic device uses a position prediction function to predict the probability that the first arrangement position is located before the second arrangement position based on the predicted description information to obtain a predicted probability, which may include: when the predicted behavior parameter of the first object identifier under the target behavior type in the candidate sample pair is greater than the predicted behavior parameter of the second object identifier under the target behavior type, determining the first predicted probability value as the predicted probability; when the predicted behavior parameter of the first object identifier under the target behavior type in the candidate sample pair is equal to the predicted behavior parameter of the second object identifier under the target behavior type, determining the second predicted probability value as the predicted probability. Probability measurement; when the predicted behavior parameter of the first object identifier under the target behavior type in the candidate sample pair is less than the predicted behavior parameter of the second object identifier under the target behavior type, the third predicted probability value is determined as the predicted probability; wherein the second predicted probability value is equal to the second label used to determine the sample annotation label, the first predicted probability value and the third predicted probability value can be set according to specific needs, and the first predicted probability value, the second predicted probability value and the third predicted probability value should satisfy: the first predicted probability value is greater than the second predicted probability value, and the second predicted probability value is greater than the third predicted probability value; optionally, when the first label is selected as the first predicted probability value and the third label is selected as the second predicted probability value, the training of the target model is more accurate.
[0079] S304, updating the model parameters of the target model according to the sample annotation labels and the target position prediction results to obtain an updated target model corresponding to the target information.
[0080] Among them, the description information predicted by the updated target model corresponding to the target information is used to decide whether to push the target information to the corresponding object.
[0081] In one embodiment, the electronic device can update the target model by updating the model parameters in the target model in the direction of reducing the difference between the sample annotation label and the target position prediction result, so as to obtain an updated target model corresponding to the target information information. When the target model is updated by using multiple sample pairs constructed based on the candidate object set, the electronic device can update the model parameters in the target model in the direction of reducing the difference between the annotation label corresponding to each sample pair and the corresponding position prediction result, and when the model convergence condition is reached, the updated target model corresponding to the target information information is obtained; wherein the difference between the annotation label corresponding to each sample pair and the corresponding position prediction result can be characterized by a target loss value, and the target loss value can be obtained by a target loss function, and the target loss function can be selected or designed according to specific needs. Optionally, the target loss function can be a cross entropy loss function, which can be specifically shown as the following formula 3:
[0082]
[0083] Among them, H is the number of multiple sample pairs constructed based on the candidate object set, h is the independent variable, and y h is the label corresponding to the hth sample pair among H sample pairs, p h The prediction result for the position corresponding to the h-th sample pair, that is, the probability that the arrangement position of the first object identifier in the h-th sample pair in the candidate object set is before the arrangement position of the second object identifier in the candidate object set, at this time, the electronic device can update the target model in the direction of reducing the target loss value. Further optionally, for each object identifier in the candidate object set, if each object identifier in the candidate object set is arranged in any arrangement order, the prediction probability of the sample pair constructed based on any two object identifiers in the candidate object set can be calculated through the prediction probability of the sample pair constructed by each adjacent two object identifiers indicated by the arrangement order; therefore, when predicting the prediction probability of all sample pairs constructed based on the objects in the candidate object set, only the prediction probability of the sample pair constructed by each adjacent two object identifiers can be predicted through the target model, and the prediction probabilities of other sample pairs are obtained by calculation, which can reduce the amount of calculation, save computing resources, and increase the model training speed.
[0084] In one embodiment, an update algorithm such as a stochastic gradient descent algorithm (i.e., SGD) or an adaptive random update algorithm (i.e., Adam) can be used to perform multiple rounds of updates on the target model based on the candidate object set to implement iterative updates of the model parameters of the target model, wherein the relevant update parameters used in the process of updating the target model can be set according to specific needs, for example, the number of update rounds (i.e., epochs), the initial learning rate, the number of samples in each batch when updating in batches (i.e., the number of sample pairs in each batch, i.e., batch size), the hyperparameters required for the update algorithm, etc. can be set according to specific needs; for example, in an optional implementation, an adaptive random update algorithm can be used, and the initial learning rate is set to 0.001, epoch=1000, batch=1000, and the initial learning rate is set to 0.001, epoch=1000, and batch=1000. size=512, the exponential decay rate of the first-order moment estimate in the hyperparameters required for the update algorithm is 0.975 (i.e., beta1=0.975), and the exponential decay rate of the second-order moment estimate is 0.999 (i.e., beta2=0.999). The target model is updated every 100 rounds at a decaying learning rate of 0.9, and the neuron dropout operation (i.e., dropout) and early stopping operation (i.e., early stoppping) commonly used in the model training process can be introduced to prevent the target model from overfitting. The neuron dropout probability in the neuron dropout operation can be set to 0.5, i.e., dropout rate=0.5.
[0085] In the process of updating the target model, the probability that the first object identifier in the candidate sample pair is located before the second object identifier in the candidate object set is used as the target position prediction result of the candidate sample pair, and then the target model is trained by comparing the difference between the sample annotation label and the target position prediction result. The prediction of the positional relationship between the first object identifier and the second object identifier in the candidate sample pair in the target sample set can be converted into a prediction of the probability that the first object identifier in the candidate sample pair is located before the second object identifier in the candidate object set. The pair method (i.e., Pair-wise) can be introduced on the basis of the ranking learning idea, and then the ranking problem can be solved from the perspective of probability on the basis of the ranking learning idea. The loss function used to update the target model can be designed from the perspective of probability loss, that is, the target model (i.e., the model corresponding to the ranking function in ranking learning) can be trained based on the perspective of probability loss, which can improve the model training effect of the target model and make the updated target model corresponding to the target information more expressive. On this basis, the present application proposes to train the target model in the direction of reducing the difference between the sample annotation labels and the target position prediction results, that is, in the direction of reducing the target loss function value. This is because in the process of training the target model based on multiple sample pairs constructed based on the candidate object set, the target model can be trained in the direction of reducing the number of sample pairs with incorrect predictions, where the sample pairs with incorrect predictions refer to: the position relationship indicated by the position prediction results of the sample pair is different from the position relationship indicated by the annotation labels of the sample pair.
[0086] S305, when it is intended to push target information to the target object and the target object is expected to perform a feedback behavior under a target behavior type on the target information, the object characteristics of the target object are obtained.
[0087] In one embodiment, the feature items of each attribute feature in the object feature of the target object should be the same as the feature items of each attribute feature in the object feature corresponding to the object identifier in the target sample set, and will not be described in detail here.
[0088] S306, calling the updated target model corresponding to the target information information, predicting the description information of the behavior parameters of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, and obtaining the target description information.
[0089] In one embodiment, the target description information includes a score. The electronic device calls the updated target model corresponding to the target information information, and predicts the description information of the behavior parameters of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, so as to obtain the target description information. The relevant process is similar to the relevant process of calling the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object characteristics corresponding to each object identifier in the candidate sample pair, and will not be repeated here.
[0090] S307, obtaining a score threshold corresponding to the target information, and pushing the target information to the target object when the target description information is greater than the score threshold corresponding to the target information.
[0091] In one embodiment, when any description information includes a predicted score for describing a behavior parameter under a target behavior type, or when any description information includes a predicted behavior parameter under a target behavior type predicted by a target model, the target description information is a score. In this case, the electronic device can obtain a score threshold corresponding to the target information information, and push the target information information to the target object when the target description information is greater than the score threshold corresponding to the target information information; wherein the score threshold corresponding to the target information information can be set based on specific needs.
[0092] In a feasible implementation, the calculation method of the score threshold corresponding to the target information information may include: selecting an object set from the candidate object set and the original object set as the test object set; calling the updated target model corresponding to the target information information, and predicting the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the test object set; and calculating the score threshold corresponding to the target information information based on the description information of the behavior parameters of each object identifier in the test object set under the target behavior type. Among them, each object identifier in the candidate object set may be selected from the original object set, and the original object set includes the object identifiers of multiple objects that have received the target information information in the past; the electronic device calls the updated target model corresponding to the target information information, and predicts the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the test object set. The process is similar to the process of calling the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the candidate sample pair, and will not be repeated here. Furthermore, when the electronic device calculates the score threshold corresponding to the target information information based on the descriptive information of the behavior parameters of each object identifier in the test object set under the target behavior type, it can average the descriptive information of the behavior parameters of each object identifier in the test object set under the target behavior type, and use the result of the averaging as the score threshold corresponding to the target information information. That is, when any descriptive information includes a predicted score for describing the behavior parameters under the target behavior type, the score threshold corresponding to the target information information may be the average of the predicted scores of each object identifier in the test object set under the target behavior type.
[0093] See also Figure 6 , which is a schematic diagram of pushing target information to a target object provided by an embodiment of the present application; taking the electronic device as a server as an example, the electronic device can obtain the object characteristics of the target object from the terminal device of the target object, and call the updated target model corresponding to the target information information, and predict the description information of the behavior parameters of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, and obtain the target description information, and further obtain the score threshold corresponding to the target information information, and when the target description information is greater than the score threshold corresponding to the target information information, push the target information information to the target object, specifically, push it to the terminal device of the target object.
[0094] In one embodiment, when it is desired to push target information to multiple objects, and it is expected that each of the multiple objects performs a feedback behavior under a target behavior type on the target information, the object features of each object can be obtained, and the updated target model corresponding to the target information can be called in sequence, and the description information of the behavior parameters of the object identifier of the corresponding object under the target behavior type can be predicted according to the object features of each object; further, based on the description information of the behavior parameters of each object under the target behavior type, the object identifiers of the multiple objects can be sorted in descending order to obtain a predicted object sequence, and the target information can be pushed to the object at a preset sequence position in the predicted object sequence; wherein the preset sequence position can be set according to specific needs. See Figure 7 , which is a schematic diagram of pushing target information to multiple objects provided by an embodiment of the present application; taking the electronic device as a server as an example, if there are an object indicated by object a, an object indicated by object b, an object indicated by object c, and an object indicated by object d, the object characteristics of the corresponding object can be obtained from the terminal device of each object, and the updated target model corresponding to the target information information is called in turn, and the description information of the behavior parameters of the object identifier of the corresponding object under the target behavior type is predicted according to the object characteristics of each object; further, based on the description information of the behavior parameters of each object under the target behavior type, the object identifiers of the multiple objects can be arranged in descending order to obtain a predicted object sequence. If the obtained predicted object sequence is object b, object c, object a, and object d, and the preset sequence positions are the first position and the second position, the electronic device pushes the target information to the terminal devices of the object indicated by object b and the object indicated by object c.
[0095] In an embodiment of the present application, by introducing the idea of ranking learning, when the target model is trained using the same candidate object set, the method used by the existing method of comparing the difference between the behavioral parameters obtained by object annotation and the behavioral parameters predicted for the object is converted into a method of comparing the difference between the sample annotation labels obtained by annotation for indicating the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set, and the target position prediction result obtained by prediction for indicating the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set. This can solve the problem of high difficulty in target modeling and poor training effect caused by the unbalanced score distribution of the target behavior parameters of the object identifiers in the candidate object set in the existing method. Moreover, by further taking the probability that the arrangement position of the first object identifier in the candidate sample pair in the candidate object set is before the arrangement position of the second object identifier in the candidate object set as the target position prediction result of the candidate sample pair, and then training the target model by comparing the difference between the sample annotation label and the target position prediction result, the prediction of the positional relationship between the first object identifier and the second object identifier in the candidate sample pair in the target sample set can be converted into a prediction of the probability that the arrangement position of the first object identifier in the candidate sample pair in the candidate object set is before the arrangement position of the second object identifier in the candidate object set. This can solve the sorting problem from a probability perspective based on the idea of sorting learning, and then design the loss function used to update the target model from the perspective of probability loss, that is, the target model can be trained based on the perspective of probability loss, which can further improve the model training effect of the target model and make the updated target model more expressive.
[0096] Based on the above-mentioned related embodiments of the model processing method, the present application embodiment provides another model processing method. Figure 8 , which is a flow chart of another model processing method provided in an embodiment of the present application. Figure 8 The model processing method shown can be executed by an electronic device, and can also be executed by any other device with computing power. The embodiment of the present application is described with reference to an electronic device. Figure 8 The model processing method shown may include the following steps:
[0097] S801: Acquire an original object set for updating an initial model.
[0098] Among them, the original object set may include: object identifiers of multiple objects that have historically received target information; the initial model and the target model to be updated have the same model structure but different model parameters, and the target model to be updated can be obtained by updating the initial model.
[0099] S802, selecting two object identifiers from the original object set to construct an original sample pair, and obtaining an original annotation label based on the behavior label of each object identifier in the original sample pair.
[0100] Among them, any behavior label indicates whether the corresponding object has performed feedback behavior on the target information; the original annotation label indicates the positional relationship between the two object identifiers in the original sample pair in the original object set. Optionally, the behavior label when the object has performed feedback behavior on the target information can be represented as 1, and the behavior label when the object has not performed feedback behavior on the target information can be represented as 0; further optionally, the feedback behavior parameter can be used to represent the importance of whether the object has performed feedback behavior on the target information. When the object has performed feedback behavior on the target information, the feedback behavior parameter of the object identifier of the object can be determined to be 1, indicating that the object has generated value for the target information. When the object has not performed feedback behavior on the target information, the feedback behavior parameter of the object identifier of the object can be determined to be 0, indicating that the object has not generated value for the target information.
[0101] In one embodiment, the arrangement position of the first object identifier in the original sample pair in the original object set is called the target arrangement position, and the arrangement position of the second object identifier in the original sample pair in the original object set is called the reference arrangement position; the electronic device obtains the original annotation label obtained based on the behavior labels of each object identifier in the original sample pair, which may include: if it is determined according to the behavior labels of each object identifier in the original sample pair that the object indicated by the first object identifier in the original sample pair has performed feedback behavior and the object indicated by the second object identifier has not performed feedback behavior, then a first value is determined as the original annotation label, and the first value is used to indicate that: the target arrangement position is located before the reference arrangement position; if the behavior labels of each object identifier in the original sample pair are the same, then a second value is determined as the original annotation label, and the second value is used to indicate that: the target arrangement position and the reference arrangement position are the same position; if it is determined according to the behavior labels of each object identifier in the original sample pair that the object indicated by the first object identifier in the original sample pair has not performed feedback behavior and the object indicated by the second object identifier has performed feedback behavior, then a third value is determined as the original annotation label, and the third value is used to indicate that: the target arrangement position is located after the reference arrangement position. Among them, the first value, the second value and the third value can be set according to specific needs, and the first value, the second value and the third value should satisfy: the first value is greater than the second value, the second value is greater than the third value, and the distance from the second value to the first value should be equal to the distance from the second value to the third value, that is, the second value should be the average of the first value and the third value; optionally, the first value can be set to 1, the second value can be set to 0, and the third value can be set to -1, or the first value can be set to 1, the second value can be set to 0.5, and the third value can be set to 0; when the first value is equal to the first label, the second value is equal to the second label, and the third value is equal to the third label, the condition satisfied by the original annotation label can also be as shown in the above formula 1, at this time, y represents the original annotation label, weight1 represents the behavior label or feedback behavior parameter of the first object identifier in the original sample pair, and weight2 represents the behavior label or feedback behavior parameter of the second object identifier in the original sample pair.
[0102] S803, calling the initial model to predict the behavior execution status of the corresponding objects according to the object features corresponding to each object identifier in the original sample pair, and obtaining corresponding behavior prediction results.
[0103] In one embodiment, the electronic device calls the initial model to predict the behavior execution of the corresponding object according to the object features corresponding to each object identifier in the original sample pair, and obtains the corresponding behavior prediction results. The process is similar to the above-mentioned electronic device calling the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the candidate sample pair, and will not be repeated here; that is, a score will be obtained by processing the object features through the initial model, and the obtained score can be used as the behavior prediction result of the corresponding object identifier to indicate the probability of the corresponding object performing feedback behavior on the target information information, that is, any behavior prediction result is used to indicate: the probability of the corresponding object performing feedback behavior on the target information information; when the concept of feedback behavior parameters is introduced, the obtained score can also be used to measure the feedback behavior parameters of the corresponding object identifier.
[0104] S804, using a position prediction function to predict the position relationship between two object identifiers in the original sample pair in the original object set according to each behavior prediction result obtained, to obtain a position prediction result corresponding to the original sample pair.
[0105] In one embodiment, the electronic device uses a position prediction function to predict the position relationship between two object identifiers in an original sample pair in an original object set based on the obtained behavior prediction results to obtain a related process of the position prediction result corresponding to the original sample pair. This process is similar to the electronic device using a position prediction function to predict the position relationship between two object identifiers in a candidate sample pair in a candidate object set based on the predicted description information to obtain a related process of the target position prediction result, and will not be repeated here.
[0106] S805, updating the initial model based on the original annotation labels and the position prediction results corresponding to the original sample pairs to obtain a target model to be updated.
[0107] In one embodiment, the updated initial model obtained by updating the original sample set and reaching the convergence condition can be directly used as the target model to be updated; optionally, some model parameters can also be selected from the updated initial model to initialize the corresponding model parameters of the target model to be updated. For example, the model parameters of the embedding layer in the updated initial model can be initialized as the model parameters in the embedding layer of the target model to be updated, and the model parameters of the embedding layer and the weight learning layer in the updated initial model can also be initialized as the model parameters of the embedding layer and the weight learning layer of the target model to be updated, and so on. In this case, the embodiment of the present application provides a schematic diagram of obtaining the target model to be updated, which can be specifically referred to. Fig. 9 .
[0108] In one embodiment, the process of updating the initial model based on the original annotation labels corresponding to the original sample pairs constructed based on the object identifiers in the original object set, and the corresponding position prediction results is similar to the process of updating the target model based on the sample annotation labels corresponding to the candidate sample pairs constructed based on the object identifiers in the candidate object set, and the corresponding target position prediction results, and will not be repeated here; but optionally, the relevant update parameters used in the process of updating the initial model can be set according to specific needs. For example, the batchsize in the update parameters used when updating the initial model can be set to 2048, and the remaining update parameters are the same as the settings of the update parameters used when updating the target model.
[0109] S806, obtaining candidate sample pairs and corresponding sample annotation labels.
[0110] Among them, the candidate sample pair includes two object identifiers in the candidate object set, and the sample annotation label is determined based on the target behavior parameters of each object identifier in the candidate sample pair. Any target behavior parameter indicates the importance of the feedback behavior under the target behavior type performed by the corresponding object on the target information. The sample annotation label is used to mark the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set.
[0111] S807, calling the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs.
[0112] S808, using a position prediction function to predict the position relationship between two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information, to obtain a target position prediction result.
[0113] S809, updating the model parameters of the target model according to the sample annotation labels and the target position prediction results to obtain an updated target model corresponding to the target information.
[0114] Among them, the description information predicted by the updated target model corresponding to the target information is used to decide whether to push the target information to the corresponding object; the relevant processes of steps S806 to S809 are similar to the relevant processes of steps S201 to S204 and steps S301 to S304 mentioned above, and will not be repeated here.
[0115] In one embodiment, the updated initial model and the updated target model can be evaluated using evaluation indicators commonly used in the ranking learning field. For example, evaluation indicators such as AUC, cumulative gain (CG), cumulative loss gain (DCG), and normalized cumulative loss gain (NDCG) can be used. Optionally, since the original object set used when updating the initial model includes object identifiers of objects that have historically received target information information, and only includes positive and negative sample objects, that is, object identifiers of objects that have performed feedback on the target information information, or object identifiers of objects that have not performed feedback on the target information information, the AUC indicator is more suitable for evaluating the updated initial model; and NDCG can weight the ranking position and distinguish the ranking differences between object identifiers of different target behavior parameters. Therefore, NDCG is more suitable for evaluating the updated target model than the AUC indicator; this application will be described later using the use of the AUC indicator to evaluate the updated initial model and the use of NDCG to evaluate the updated target model as examples.
[0116] Among them, AUC refers to the area under the receiver operating characteristic curve (ROC), which is used to characterize the probability that the positive example score is ranked before the negative example score. Its calculation method can be shown by the following formula 4.1:
[0117]
[0118] If the first evaluation object set is used to evaluate the updated initial model, G is the number of positive sample objects in the first evaluation object set, that is, the number of object identifiers of objects whose behavior labels indicate that they have performed feedback on the target information, and K is the number of negative sample objects in the first evaluation object set, that is, the number of object identifiers of objects whose behavior labels indicate that they have not performed feedback on the target information; positiveClass represents the set of positive sample objects in the first evaluation object set, and g is the independent variable; rank g is the rank value of the g-th positive sample object, where the sequence of descending order of the object identifiers in the first evaluation object set based on the behavior prediction results obtained by the initial model (i.e., the scores output by the initial model) is the target evaluation sequence. If the length of the target evaluation sequence is L, the rank value of the object at the first position in the target evaluation sequence is L, the rank value of the object at the second position is L-1, the rank value of the object at the Lth position is 1, and so on, to obtain the rank value of each object, where rank g Indicates the rank value corresponding to the g-th positive sample object.
[0119] If the updated target model is evaluated by the second evaluation object set, the object identifiers in the second evaluation object set are arranged in descending order based on the target behavior parameters to obtain a reference evaluation sequence. The arrangement sequence of object identifiers determined by the description information of the behavior parameters of each object identifier in the evaluation object set under the target behavior type predicted by the updated target model can be called a predicted evaluation sequence. The calculation of the cumulative gain, the cumulative loss gain, and the normalized cumulative loss gain can be realized based on the reference evaluation sequence and the predicted evaluation sequence; wherein the cumulative gain only considers the object relevance (i.e., the target behavior parameters of the object identifier), and does not consider the position of the object identifier in the reference evaluation sequence. The calculation method of the cumulative gain can be shown by the following formula 4.2:
[0120]
[0121] Among them, p can represent the pth position in the prediction evaluation sequence, j is the independent variable, and rel j It represents the correlation of the j-th object identification in the prediction evaluation sequence, that is, the target behavior parameter of the j-th object identification.
[0122] Furthermore, the cumulative impairment gain adds consideration of the position of the object identifier in the reference evaluation sequence on the basis of the cumulative gain. The calculation method of the cumulative impairment gain can be shown by the following formula 4.3:
[0123]
[0124] Furthermore, the normalized cumulative loss gain is the normalization of the best DCG value (i.e., IDCG) of DCG under ideal sorting. The position of the object identifier in the reference evaluation sequence is weighted so that the object identifier with a high target behavior parameter is given a greater weight than the object identifier with a low target behavior parameter during evaluation. This can be specifically shown by the following formula 4.4:
[0125]
[0126] Among them, when determining the best DCG value under the ideal sorting condition, each object identifier in the second evaluation object set can be arranged in descending order according to the object identifier relevance (i.e., the target behavior parameter of the object identifier), and the target behavior parameters of the p object identifiers with the highest relevance are selected to calculate the DCG, that is, the first p object identifiers in the reference evaluation sequence are selected to calculate the DCG. The DCG at this time is the best DCG value under the ideal sorting condition, which can be specifically shown by the following formula 4.5:
[0127]
[0128] Among them, |REL| represents the first p object identifiers in the reference evaluation sequence. Specifically, p can be set to the number of object identifiers in the evaluation object set. For a perfect sorting algorithm, DCG p Will be equal to IDCG p , the calculated nDCG value is 1, and all nDCG values fall in the interval [0,1].
[0129] In one embodiment, when there is an information information set to be pushed that includes M information information, and M is an integer greater than 1, the electronic device can take each of the M information information as the target information information in turn, and then can update a corresponding updated target model for each information information, which is used to predict the description information of the behavior parameters of the object identifier under the target behavior type when the object responds to the corresponding information information; further, the electronic device can determine which of the M information information to push to the object based on the updated target model corresponding to different information information and the description information of the behavior parameters of the object identifier under the target behavior type when the object responds to different information information. In a specific implementation, when M information information in an information information set is to be pushed to a target object, and the target object is expected to perform a feedback behavior under a target behavior type on the received information information, an updated target model corresponding to each of the M information information is obtained; M is an integer greater than 1; the updated target model corresponding to each information information is called respectively, and the description information of the behavior parameters of the object identifier of the target object under the target behavior type is predicted according to the object characteristics of the target object, to obtain M description information; one description information includes a score; based on the score threshold corresponding to each information information and the corresponding description information, a disturbance factor for each information information is determined; the disturbance factor is used to disturb the push order of the M information information; based on the disturbance factor of each information information, the information information to be pushed is selected from the M information information, and the selected information information is pushed to the target object. Among them, the electronic device calls the updated target model corresponding to each information information, predicts the description information of the behavior parameters of the target object's object identifier under the target behavior type according to the object characteristics of the target object, and obtains the relevant process of M description information, which is similar to the process shown in the above step S306 and will not be repeated here; the calculation method of the score threshold corresponding to each information information is similar to the calculation method of the score threshold corresponding to the target information information shown in the above step S307 and will not be repeated here.
[0130] In a specific implementation, determining the disturbance factor of each information information based on the score threshold and the corresponding description information corresponding to each information information may include: for the mth information information among the M information information, calculating the ratio between the description information corresponding to the mth information information and the score threshold corresponding to the mth information information; determining the calculated ratio as the disturbance factor of the mth information information, m∈[1,M]. Since the ratio between the description information corresponding to the mth information information and the score threshold corresponding to the mth information information represents the relative size between the description information corresponding to the mth information information and the score threshold corresponding to the mth information information, even if the description information corresponding to different information information among the M information information is incomparable, the disturbance factor determined based on the ratio between the description information and the corresponding score threshold is still comparable.
[0131] Further, when selecting information to be pushed from M information based on the disturbance factor of each information, the push order of the M information can be disturbed based on the disturbance factor of each information to obtain a disturbed information sorting queue, and the information at the target queue position in the disturbed information sorting queue is determined as the information to be pushed, wherein the target queue position can be set according to specific needs. In a feasible implementation, the M information can be sorted in descending order according to the disturbance factor of each information to obtain a disturbed information sorting queue, and when the target queue position is set to the first queue position, the information at the first queue position in the disturbed information sorting queue is determined as the information to be pushed. In another feasible implementation, the corresponding push revenue can be adjusted based on the disturbance factor of each information information to obtain the adjusted push revenue, and the M information information can be sorted in descending order according to the adjusted push revenue corresponding to each information information to obtain the disturbed information sorting queue; wherein, the push revenue (i.e., effective cost per mille, ECPM) of any information information refers to: the revenue obtained from every thousand displays of the corresponding information information.
[0132] For example, see Fig.10, is a schematic diagram of an embodiment of the present application in which a plurality of information information is expected to be pushed to a target object. If the information information set includes 4 information information, the 4 information information are arranged in the original push order as follows: information information 1, information information 2, information information 3, and information information 4; the electronic device can obtain the object characteristics of the target object from the terminal device of the target object, and call the updated target model corresponding to each information information respectively, and predict the description information of the behavior parameters of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, and obtain 4 description information, namely: description information 1, description information 2, description information 3, and description information 4; based on the corresponding information of each information information The score threshold and corresponding description information are used to determine the disturbance factors of each information information, which are: disturbance factor 1, disturbance factor 2, disturbance factor 3, and disturbance factor 4. Based on the disturbance factor of each information information, the push order of the four information information is disturbed to obtain a disturbed information information sorting queue, and the information information at the target queue position in the disturbed information information sorting queue is determined as the information information to be pushed, and then the selected information information is pushed to the terminal device of the target object; if the information information sorting queue after the disturbance is: information information 4, information information 2, information information 3, and information information 1, and the target queue position is the first position, the electronic device pushes information information 4 to the terminal device of the target object.
[0133] In an embodiment of the present application, the original sample pairs constructed based on the object identifiers in the original object set can be used to update the initial model to obtain the target model to be updated, and then the sample pairs constructed based on the object identifiers in the candidate object set can be used to further update the target model to be updated, that is, to fine-tune the target model to be updated; based on the idea of transfer learning, the knowledge learned in the original object set can be fully utilized to further expand the training samples, which can further solve the problem of model overfitting caused by the small number of training samples, further improve the model training effect of the target model, and make the updated target model more expressive and more robust.
[0134] Based on the above model processing method embodiment, the present application embodiment provides a model processing device. Fig.11 , is a structural diagram of a model processing device provided in an embodiment of the present application, and the model processing device may include an acquisition unit 1101 and a processing unit 1102. Fig.11 The model processing device shown can run the following units:
[0135] The acquisition unit 1101 is used to acquire a candidate sample pair and a corresponding sample annotation label; the candidate sample pair includes two object identifiers in a candidate object set, the sample annotation label is determined based on target behavior parameters of each object identifier in the candidate sample pair, any target behavior parameter indicates the importance of the feedback behavior under the target behavior type performed by the corresponding object on the target information information, and the sample annotation label is used to mark the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set;
[0136] Processing unit 1102 is used to call the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs;
[0137] The processing unit 1102 is further configured to predict the position relationship between the two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information using a position prediction function to obtain a target position prediction result;
[0138] The processing unit 1102 is also used to update the model parameters of the target model according to the sample annotation label and the target position prediction result to obtain an updated target model corresponding to the target information information; wherein the description information predicted by the updated target model corresponding to the target information information is used to decide whether to push the target information information to the corresponding object.
[0139] In one embodiment, the arrangement position of the first object identifier in the candidate sample pair in the candidate object set is called the first arrangement position, and the arrangement position of the second object identifier in the candidate sample pair in the candidate object set is called the second arrangement position;
[0140] When the acquisition unit 1101 acquires the candidate sample pairs and the corresponding sample annotation labels, it specifically performs the following operations:
[0141] Acquire the candidate sample pairs and target behavior parameters of each object identifier in the candidate sample pairs;
[0142] If the target behavior parameter of the first object identifier in the candidate sample pair is greater than the target behavior parameter of the second object identifier in the candidate sample pair, a first label is determined as the sample annotation label, and the first label is used to indicate that: the first arrangement position is located before the second arrangement position;
[0143] If the target behavior parameter of the first object identifier in the candidate sample pair is equal to the target behavior parameter of the second object identifier in the candidate sample pair, a second label is determined as the sample annotation label, and the second label is used to indicate that the first arrangement position and the second arrangement position are the same position;
[0144] If the target behavior parameter of the first object identifier in the candidate sample pair is less than the target behavior parameter of the second object identifier in the candidate sample pair, a third label is determined as the sample annotation label, and the third label is used to indicate that the first arrangement position is located after the second arrangement position.
[0145] In one embodiment, any object feature includes N attribute features, where N is a positive integer;
[0146] When the processing unit 1102 calls the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs, the following operations are specifically performed:
[0147] Call the target model to be updated, and perform feature dimension reduction processing on each attribute feature in the object feature corresponding to the i-th object identifier in the candidate sample pair to obtain N intermediate features; i∈[1,2];
[0148] Obtaining a feature weight of each of the N intermediate features; any feature weight is used to indicate the importance of the corresponding intermediate feature;
[0149] According to each intermediate feature and the corresponding feature weight, the description information of the behavior parameter of the i-th object identifier under the target behavior type is predicted.
[0150] In one embodiment, any intermediate feature is a multi-dimensional vector;
[0151] When the processing unit 1102 obtains the feature weight of each of the N intermediate features, it specifically performs the following operations:
[0152] Perform feature squeezing processing on each of the N intermediate features to obtain N feature squeezing values; any feature squeezing value is a one-dimensional value;
[0153] The feature weight of the nth intermediate feature is determined according to the correlation between the feature squeezing value corresponding to the nth intermediate feature among the N intermediate features and each feature squeezing value, n∈[1,N].
[0154] In one embodiment, the arrangement position of the first object identifier in the candidate sample pair in the candidate object set is called the first arrangement position, and the arrangement position of the second object identifier in the candidate sample pair in the candidate object set is called the second arrangement position;
[0155] The processing unit 1102 uses the position prediction function to predict the position relationship between the two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information, and when the target position prediction result is obtained, specifically performs the following operations:
[0156] Using the position prediction function to predict the probability that the first arrangement position is before the second arrangement position according to the predicted description information, to obtain a predicted probability; and using the predicted probability as the target position prediction result;
[0157] Among them, when the predicted probability is greater than the baseline probability, it indicates that the predicted position relationship is: the first arrangement position is before the second arrangement position; when the predicted probability is equal to the baseline probability, it indicates that the predicted position relationship is: the first arrangement position and the second arrangement position are the same position; when the predicted probability is less than the baseline probability, it indicates that the predicted position relationship is: the first arrangement position is after the second arrangement position.
[0158] In one embodiment, the sample annotation label includes a numerical value, and the numerical value is within a preset numerical range; any description information includes: a predicted score for describing a behavior parameter under a target behavior type;
[0159] The processing unit 1102 uses the position prediction function to predict the probability that the first arrangement position is before the second arrangement position according to the predicted description information, and when the predicted probability is obtained, specifically performs the following operations:
[0160] Using the position prediction function, performing a difference operation on the predicted score of the first object identifier in the candidate sample pair under the target behavior type and the predicted score of the second object identifier under the target behavior type to obtain a difference operation result;
[0161] The difference operation result is mapped to the preset numerical interval to obtain the predicted probability.
[0162] In one embodiment, the acquisition unit 1101 is further used to acquire an original object set for updating the initial model; the original object set includes: object identifiers of multiple objects that have historically received the target information;
[0163] The acquisition unit 1101 is further used to select two object identifiers from the original object set to construct an original sample pair, and obtain an original annotation label obtained based on the behavior label of each object identifier in the original sample pair; any behavior label indicates whether the corresponding object has performed feedback behavior on the target information; the original annotation label indicates the positional relationship between the two object identifiers in the original sample pair in the original object set;
[0164] The processing unit 1102 is further used to call the initial model to predict the behavior execution of the corresponding object according to the object features corresponding to each object identifier in the original sample pair, and obtain corresponding behavior prediction results; any behavior prediction result is used to indicate: the probability that the corresponding object performs feedback behavior on the target information;
[0165] The processing unit 1102 is further configured to use the position prediction function to predict the position relationship between the two object identifiers in the original sample pair in the original object set according to each behavior prediction result obtained, so as to obtain a position prediction result corresponding to the original sample pair;
[0166] The processing unit 1102 is further configured to update the initial model based on the original annotation labels and the position prediction results corresponding to the original sample pairs, so as to obtain a target model to be updated.
[0167] In one embodiment, the arrangement position of the first object identifier in the original sample pair in the original object set is called the target arrangement position, and the arrangement position of the second object identifier in the original sample pair in the original object set is called the reference arrangement position;
[0168] When the acquisition unit 1101 acquires the original annotation labels obtained based on the behavior labels of each object identifier in the original sample pair, the following operations are specifically performed:
[0169] If it is determined according to the behavior labels of the object identifiers in the original sample pair that the object indicated by the first object identifier in the original sample pair has performed a feedback behavior and the object indicated by the second object identifier has not performed a feedback behavior, a first value is determined as the original labeling label, and the first value is used to indicate that: the target arrangement position is located before the reference arrangement position;
[0170] If the behavior labels of the object identifiers in the original sample pair are the same, a second value is determined as the original label label, and the second value is used to indicate that: the target arrangement position and the reference arrangement position are the same position;
[0171] If it is determined based on the behavior labels of each object identifier in the original sample pair that the object indicated by the first object identifier in the original sample pair has not performed feedback behavior and the object indicated by the second object identifier has performed feedback behavior, then a third value is determined as the original annotation label, and the third value is used to indicate that: the target arrangement position is located after the reference arrangement position.
[0172] In one embodiment, the acquisition unit 1101 is further configured to acquire the object feature of the target object when the target information is to be pushed to the target object and the target object is expected to perform a feedback behavior under the target behavior type on the target information;
[0173] The processing unit 1102 is further configured to call the updated target model corresponding to the target information information, predict the description information of the behavior parameters of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, and obtain the target description information, wherein the target description information includes a score;
[0174] The processing unit 1101 is further configured to obtain a score threshold corresponding to the target information information, and push the target information information to the target object when the target description information is greater than the score threshold corresponding to the target information information.
[0175] In one embodiment, each object identifier in the candidate object set is selected from the original object set; when the processing unit 1102 is used to calculate the score threshold corresponding to the target information, the processing unit 1102 specifically performs the following operations:
[0176] Selecting an object set from the candidate object set and the original object set as a test object set;
[0177] Calling the updated target model corresponding to the target information information, and predicting the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the test object set;
[0178] Based on the description information of the behavior parameters of each object identifier in the test object set under the target behavior type, a score threshold corresponding to the target information information is calculated.
[0179] In one embodiment, the target information is an information in an information set, and any information in the information set has a corresponding updated target model;
[0180] The acquisition unit 1101 is further configured to acquire an updated target model corresponding to each of the M information information when M information information in the information information set is to be pushed to a target object and the target object is expected to perform a feedback behavior under the target behavior type on the received information information; M is an integer greater than 1;
[0181] The processing unit 1102 is further configured to respectively call the updated target model corresponding to each of the information information, and predict the description information of the behavior parameter of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, to obtain M description information; one description information includes a score;
[0182] The processing unit 1102 is further configured to determine a disturbance factor for each piece of information based on a score threshold and corresponding description information corresponding to each piece of information; the disturbance factor is used to disturb the order in which the M pieces of information are pushed;
[0183] The processing unit 1102 is further configured to select information to be pushed from the M information information based on the disturbance factor of each information information, and push the selected information information to the target object.
[0184] In one embodiment, when the processing unit 1102 determines the disturbance factor of each information information based on the score threshold and the corresponding description information corresponding to each information information, the processing unit 1102 specifically performs the following operations:
[0185] For an m-th information piece of information among the M pieces of information, calculating a ratio between the description information corresponding to the m-th information piece of information and a score threshold value corresponding to the m-th information piece of information;
[0186] The calculated ratio is determined as the disturbance factor of the m-th information, m∈[1,M].
[0187] According to one embodiment of the present application, Figure 2 , Figure 3 as well as Figure 8 The various steps involved in the model processing method shown can be Fig.11 The model processing device shown is executed by each unit. For example, Figure 2 Step S201 shown can be performed by Fig.11 The acquisition unit 1101 in the model processing device shown is executed; Figure 2 Steps S202 to S204 shown in FIG. Fig.11 The model processing device shown is executed by the processing unit 1102.
[0188] According to another embodiment of the present application, Fig.11 The various units in the model processing device shown can be separately or completely combined into one or several other units to constitute, or one (some) of the units can also be split into multiple smaller units in function to constitute, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above-mentioned units are divided based on logical functions. In practical applications, the function of one unit can also be implemented by multiple units, or the functions of multiple units can be implemented by one unit. In other embodiments of the present application, the model processing device based on logical function division can also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented by the collaboration of multiple units.
[0189] According to another embodiment of the present application, the program can be executed by running on a general computing device such as a computer including a central processing unit (CPU), a random access memory medium (RAM), a read-only memory medium (ROM), and other processing elements and storage elements. Figure 2 , Figure 3 as well as Figure 8 A computer program (including program code) for each step involved in the corresponding method shown in FIG. Fig.11 The model processing device shown in and the model processing method of the embodiment of the present application are implemented. The computer program can be recorded on, for example, a computer-readable storage medium, and loaded into the above-mentioned computing device through the computer-readable storage medium and run therein.
[0190] The embodiment of the present application can construct a target model for predicting the description information of the behavior parameters of the object identifier of the object under the target behavior type. In the process of training the target model with the candidate object set, a candidate sample pair including two object identifiers in the candidate object set and a corresponding sample annotation label for marking the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set can be obtained, wherein the sample annotation label is determined based on the target behavior parameters of each object identifier in the candidate sample pair. Then, using the idea of ranking learning, the target model is called to predict the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the candidate sample pair, and the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set is predicted according to the predicted description information, and the target position prediction result is obtained, thereby updating the model parameters of the target model based on the sample annotation label and the target position prediction result. It can be seen that the embodiment of the present application utilizes the idea of ranking learning, so that the training samples can be expanded by combining each object identifier in the candidate object set in pairs, thereby increasing the number of training samples of the target model (one training sample is a sample pair) to a certain extent. This not only solves the problem of model overfitting caused by small training samples, but also enables the target model to learn more knowledge through richer training samples during the training process, thereby improving the convergence speed of the target model, reducing the training time of the target model, improving the training efficiency of the target model, and further improving the model training effect of the target model, so that the updated target model can have the advantages of good expressiveness and high robustness, and then the updated target model can more accurately predict the description information of the behavior parameters of the object identifier of any object under the target behavior type, so as to accurately decide whether to push the target information to the corresponding object based on the description information, thereby improving the push effect of the information information.
[0191] Based on the above-mentioned model processing method embodiment and model processing device embodiment, the present application also provides an electronic device. Fig.12 , is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Fig.12 The electronic device shown may include at least a processor 1201, an input interface 1202, an output interface 1203, and a computer storage medium 1204. The processor 1201, the input interface 1202, the output interface 1203, and the computer storage medium 1204 may be connected via a bus or other means.
[0192] The computer storage medium 1204 can be stored in the memory of the electronic device. The computer storage medium 1204 is used to store a computer program. The computer program includes program instructions. The processor 1201 is used to execute the program instructions stored in the computer storage medium 1204. The processor 1201 (or CPU (Central Processing Unit)) is the computing core and control core of the electronic device, which is suitable for implementing one or more instructions, and is specifically suitable for loading and executing one or more instructions to implement the above-mentioned model processing method flow or corresponding functions.
[0193] The embodiment of the present application also provides a computer storage medium (Memory), which is a memory device in an electronic device for storing programs and data. It is understandable that the computer storage medium here can include both the built-in storage medium in the terminal and the extended storage medium supported by the terminal. The computer storage medium provides a storage space, which stores the operating system of the terminal. In addition, one or more instructions suitable for being loaded and executed by the processor 1201 are also stored in the storage space, and these instructions can be one or more computer programs (including program codes). It should be noted that the computer storage medium here can be a high-speed random access memory (RAM) memory, or a non-volatile memory (non-volatile memory), such as at least one disk storage; optionally, it can also be at least one computer storage medium located away from the aforementioned processor.
[0194] In one embodiment, the processor 1201 may load and execute one or more instructions stored in a computer storage medium to implement the above-mentioned Figure 2 , Figure 3 as well as Figure 8 The corresponding steps of the method in the model processing method embodiment, the processor 1201 is specifically used to execute:
[0195] Obtain candidate sample pairs and corresponding sample annotation labels; the candidate sample pairs include two object identifiers in a candidate object set, the sample annotation labels are determined based on target behavior parameters of each object identifier in the candidate sample pairs, any target behavior parameter indicates the importance of feedback behavior under a target behavior type performed by the corresponding object on target information, and the sample annotation labels are used to mark the positional relationship between the two object identifiers in the candidate sample pairs in the candidate object set;
[0196] Calling the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs;
[0197] Using a position prediction function to predict the position relationship between two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information, to obtain a target position prediction result;
[0198] According to the sample annotation label and the target position prediction result, the model parameters of the target model are updated to obtain an updated target model corresponding to the target information; wherein the description information predicted by the updated target model corresponding to the target information information is used to decide whether to push the target information information to the corresponding object.
[0199] In one embodiment, the arrangement position of the first object identifier in the candidate sample pair in the candidate object set is called the first arrangement position, and the arrangement position of the second object identifier in the candidate sample pair in the candidate object set is called the second arrangement position;
[0200] When the processor 1201 obtains the candidate sample pairs and the corresponding sample annotation labels, it specifically performs the following operations:
[0201] Acquire the candidate sample pairs and target behavior parameters of each object identifier in the candidate sample pairs;
[0202] If the target behavior parameter of the first object identifier in the candidate sample pair is greater than the target behavior parameter of the second object identifier in the candidate sample pair, a first label is determined as the sample annotation label, and the first label is used to indicate that: the first arrangement position is located before the second arrangement position;
[0203] If the target behavior parameter of the first object identifier in the candidate sample pair is equal to the target behavior parameter of the second object identifier in the candidate sample pair, a second label is determined as the sample annotation label, and the second label is used to indicate that the first arrangement position and the second arrangement position are the same position;
[0204] If the target behavior parameter of the first object identifier in the candidate sample pair is less than the target behavior parameter of the second object identifier in the candidate sample pair, a third label is determined as the sample annotation label, and the third label is used to indicate that the first arrangement position is located after the second arrangement position.
[0205] In one embodiment, any object feature includes N attribute features, where N is a positive integer;
[0206] When the processor 1201 calls the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs, the processor 1201 specifically performs the following operations:
[0207] Call the target model to be updated, and perform feature dimension reduction processing on each attribute feature in the object feature corresponding to the i-th object identifier in the candidate sample pair to obtain N intermediate features; i∈[1,2];
[0208] Obtaining a feature weight of each of the N intermediate features; any feature weight is used to indicate the importance of the corresponding intermediate feature;
[0209] According to each intermediate feature and the corresponding feature weight, the description information of the behavior parameter of the i-th object identifier under the target behavior type is predicted.
[0210] In one embodiment, any intermediate feature is a multi-dimensional vector;
[0211] When the processor 1201 obtains the feature weight of each of the N intermediate features, it specifically performs the following operations:
[0212] Perform feature squeezing processing on each of the N intermediate features to obtain N feature squeezing values; any feature squeezing value is a one-dimensional value;
[0213] The feature weight of the nth intermediate feature is determined according to the correlation between the feature squeezing value corresponding to the nth intermediate feature among the N intermediate features and each feature squeezing value, n∈[1,N].
[0214] In one embodiment, the arrangement position of the first object identifier in the candidate sample pair in the candidate object set is called the first arrangement position, and the arrangement position of the second object identifier in the candidate sample pair in the candidate object set is called the second arrangement position;
[0215] The processor 1201 predicts the position relationship of the two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information using the position prediction function, and when obtaining the target position prediction result, specifically performs the following operations:
[0216] Using the position prediction function to predict the probability that the first arrangement position is before the second arrangement position according to the predicted description information, to obtain a predicted probability; and using the predicted probability as the target position prediction result;
[0217] Among them, when the predicted probability is greater than the baseline probability, it indicates that the predicted position relationship is: the first arrangement position is before the second arrangement position; when the predicted probability is equal to the baseline probability, it indicates that the predicted position relationship is: the first arrangement position and the second arrangement position are the same position; when the predicted probability is less than the baseline probability, it indicates that the predicted position relationship is: the first arrangement position is after the second arrangement position.
[0218] In one embodiment, the sample annotation label includes a numerical value, and the numerical value is within a preset numerical range; any description information includes: a predicted score for describing a behavior parameter under a target behavior type;
[0219] The processor 1201 uses the position prediction function to predict the probability that the first arrangement position is before the second arrangement position according to the predicted description information, and when the predicted probability is obtained, specifically performs the following operations:
[0220] Using the position prediction function, performing a difference operation on the predicted score of the first object identifier in the candidate sample pair under the target behavior type and the predicted score of the second object identifier under the target behavior type to obtain a difference operation result;
[0221] The difference operation result is mapped to the preset numerical interval to obtain the predicted probability.
[0222] In one embodiment, the processor 1201 is further configured to:
[0223] Acquire an original object set for updating the initial model; the original object set includes: object identifiers of multiple objects that have historically received the target information;
[0224] Selecting two object identifiers from the original object set to construct an original sample pair, and obtaining an original annotation label based on the behavior label of each object identifier in the original sample pair; any behavior label indicates whether the corresponding object has performed feedback behavior on the target information; the original annotation label indicates the positional relationship between the two object identifiers in the original sample pair in the original object set;
[0225] The initial model is called to predict the behavior execution of the corresponding object according to the object features corresponding to each object identifier in the original sample pair, and obtain corresponding behavior prediction results; any behavior prediction result is used to indicate: the probability that the corresponding object performs feedback behavior on the target information;
[0226] Using the position prediction function to predict the position relationship between the two object identifiers in the original sample pair in the original object set according to the obtained behavior prediction results, to obtain the position prediction result corresponding to the original sample pair;
[0227] The initial model is updated based on the original annotation labels and the position prediction results corresponding to the original sample pairs to obtain a target model to be updated.
[0228] In one embodiment, the arrangement position of the first object identifier in the original sample pair in the original object set is called the target arrangement position, and the arrangement position of the second object identifier in the original sample pair in the original object set is called the reference arrangement position;
[0229] When the processor 1201 obtains the original annotation labels obtained based on the behavior labels of each object identifier in the original sample pair, the processor 1201 specifically performs the following operations:
[0230] If it is determined according to the behavior labels of the object identifiers in the original sample pair that the object indicated by the first object identifier in the original sample pair has performed a feedback behavior and the object indicated by the second object identifier has not performed a feedback behavior, a first value is determined as the original labeling label, and the first value is used to indicate that: the target arrangement position is located before the reference arrangement position;
[0231] If the behavior labels of the object identifiers in the original sample pair are the same, a second value is determined as the original label label, and the second value is used to indicate that: the target arrangement position and the reference arrangement position are the same position;
[0232] If it is determined based on the behavior labels of each object identifier in the original sample pair that the object indicated by the first object identifier in the original sample pair has not performed feedback behavior and the object indicated by the second object identifier has performed feedback behavior, then a third value is determined as the original annotation label, and the third value is used to indicate that: the target arrangement position is located after the reference arrangement position.
[0233] In one embodiment, the processor 1201 is further configured to:
[0234] When the target information is to be pushed to the target object, and the target object is expected to perform a feedback behavior under the target behavior type on the target information, obtaining the object characteristics of the target object;
[0235] Calling the updated target model corresponding to the target information information, predicting the description information of the behavior parameters of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, and obtaining target description information, wherein the target description information includes a score;
[0236] A score threshold corresponding to the target information is obtained, and when the target description information is greater than the score threshold corresponding to the target information information, the target information is pushed to the target object.
[0237] In one embodiment, each object identifier in the candidate object set is selected from the original object set;
[0238] When the processor 1201 is used to calculate the score threshold corresponding to the target information, the processor 1201 specifically performs the following operations:
[0239] Selecting an object set from the candidate object set and the original object set as a test object set;
[0240] Calling the updated target model corresponding to the target information information, and predicting the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the test object set;
[0241] Based on the description information of the behavior parameters of each object identifier in the test object set under the target behavior type, a score threshold corresponding to the target information information is calculated.
[0242] In one embodiment, the target information is an information in an information set, and any information in the information set has a corresponding updated target model;
[0243] The processor 1201 is further configured to:
[0244] When M information in the information set is to be pushed to a target object, and the target object is expected to perform a feedback behavior under the target behavior type on the received information, an updated target model corresponding to each of the M information is obtained; M is an integer greater than 1;
[0245] Respectively calling the updated target model corresponding to each of the information information, predicting the description information of the behavior parameter of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, and obtaining M description information; one description information includes a score;
[0246] Determine a disturbance factor for each piece of information based on a score threshold and corresponding description information corresponding to each piece of information; the disturbance factor is used to disturb the order in which the M pieces of information are pushed;
[0247] Based on the disturbance factor of each information information, information information to be pushed is selected from the M information information, and the selected information information is pushed to the target object.
[0248] In one embodiment, when the processor 1201 determines the disturbance factor of each information information based on the score threshold and the corresponding description information corresponding to each information information, the processor 1201 specifically performs the following operations:
[0249] For an m-th information piece of information among the M pieces of information, calculating a ratio between the description information corresponding to the m-th information piece of information and a score threshold value corresponding to the m-th information piece of information;
[0250] The calculated ratio is determined as the disturbance factor of the m-th information, m∈[1,M].
[0251] The present application provides a computer program product, which includes a computer program stored in a computer storage medium; a processor of an electronic device reads the computer program from the computer storage medium, and the processor executes the computer program, so that the electronic device performs the above-mentioned Figure 2 , Figure 3 as well as Figure 8 The computer-readable storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).
[0252] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A model processing method, It is characterized in that include: Obtain candidate sample pairs and corresponding sample annotation labels; the candidate sample pairs include two object identifiers in a candidate object set, the sample annotation labels are determined based on target behavior parameters of each object identifier in the candidate sample pairs, any target behavior parameter indicates the importance of feedback behavior under a target behavior type performed by the corresponding object on target information, and the sample annotation labels are used to mark the positional relationship between the two object identifiers in the candidate sample pairs in the candidate object set; Calling the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs; Using a position prediction function to predict the position relationship between two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information, to obtain a target position prediction result; According to the sample annotation label and the target position prediction result, the model parameters of the target model are updated to obtain an updated target model corresponding to the target information; wherein the description information predicted by the updated target model corresponding to the target information information is used to decide whether to push the target information information to the corresponding object.
2. The method according to claim 1, It is characterized in that The arrangement position of the first object identifier in the candidate sample pair in the candidate object set is called the first arrangement position, and the arrangement position of the second object identifier in the candidate sample pair in the candidate object set is called the second arrangement position; The step of obtaining candidate sample pairs and corresponding sample annotation labels includes: Acquire the candidate sample pairs and target behavior parameters of each object identifier in the candidate sample pairs; If the target behavior parameter of the first object identifier in the candidate sample pair is greater than the target behavior parameter of the second object identifier in the candidate sample pair, a first label is determined as the sample annotation label, and the first label is used to indicate that: the first arrangement position is located before the second arrangement position; If the target behavior parameter of the first object identifier in the candidate sample pair is equal to the target behavior parameter of the second object identifier in the candidate sample pair, a second label is determined as the sample annotation label, and the second label is used to indicate that the first arrangement position and the second arrangement position are the same position; If the target behavior parameter of the first object identifier in the candidate sample pair is less than the target behavior parameter of the second object identifier in the candidate sample pair, a third label is determined as the sample annotation label, and the third label is used to indicate that the first arrangement position is located after the second arrangement position.
3. The method according to claim 1 or 2, It is characterized in that Any object feature includes N attribute features, where N is a positive integer; the calling of the target model to be updated predicts the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the candidate sample pair, including: Call the target model to be updated, and perform feature dimension reduction processing on each attribute feature in the object feature corresponding to the i-th object identifier in the candidate sample pair to obtain N intermediate features; i∈[1,2]; Obtaining a feature weight of each of the N intermediate features; any feature weight is used to indicate the importance of the corresponding intermediate feature; According to each intermediate feature and the corresponding feature weight, the description information of the behavior parameter of the i-th object identifier under the target behavior type is predicted.
4. The method according to claim 3, It is characterized in that Any intermediate feature is a multidimensional vector; The obtaining of the feature weight of each of the N intermediate features includes: Perform feature squeezing processing on each of the N intermediate features to obtain N feature squeezing values; any feature squeezing value is a one-dimensional value; The feature weight of the nth intermediate feature is determined according to the correlation between the feature squeezing value corresponding to the nth intermediate feature among the N intermediate features and each feature squeezing value, n∈[1,N].
5. The method according to claim 1, It is characterized in that The arrangement position of the first object identifier in the candidate sample pair in the candidate object set is called the first arrangement position, and the arrangement position of the second object identifier in the candidate sample pair in the candidate object set is called the second arrangement position; The position prediction function is used to predict the position relationship between the two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information to obtain the target position prediction result, including: Using the position prediction function to predict the probability that the first arrangement position is before the second arrangement position according to the predicted description information, to obtain a predicted probability; and using the predicted probability as the target position prediction result; Among them, when the predicted probability is greater than the baseline probability, it indicates that the predicted position relationship is: the first arrangement position is before the second arrangement position; when the predicted probability is equal to the baseline probability, it indicates that the predicted position relationship is: the first arrangement position and the second arrangement position are the same position; when the predicted probability is less than the baseline probability, it indicates that the predicted position relationship is: the first arrangement position is after the second arrangement position.
6. The method according to claim 5, It is characterized in that The sample annotation label includes a numerical value, and the numerical value is within a preset numerical range; any description information includes: a predicted score for describing a behavior parameter under a target behavior type; The using the position prediction function to predict the probability that the first arrangement position is before the second arrangement position according to the predicted description information to obtain the predicted probability includes: Using the position prediction function, performing a difference operation on the predicted score of the first object identifier in the candidate sample pair under the target behavior type and the predicted score of the second object identifier under the target behavior type to obtain a difference operation result; The difference operation result is mapped to the preset numerical interval to obtain the predicted probability.
7. The method according to claim 1, It is characterized in that The method further comprises: Acquire an original object set for updating the initial model; the original object set includes: object identifiers of multiple objects that have historically received the target information; Selecting two object identifiers from the original object set to construct an original sample pair, and obtaining an original annotation label based on the behavior label of each object identifier in the original sample pair; any behavior label indicates whether the corresponding object has performed feedback behavior on the target information; the original annotation label indicates the positional relationship between the two object identifiers in the original sample pair in the original object set; The initial model is called to predict the behavior execution of the corresponding object according to the object features corresponding to each object identifier in the original sample pair, and obtain corresponding behavior prediction results; any behavior prediction result is used to indicate: the probability that the corresponding object performs feedback behavior on the target information; Using the position prediction function to predict the position relationship between the two object identifiers in the original sample pair in the original object set according to the obtained behavior prediction results, to obtain the position prediction result corresponding to the original sample pair; The initial model is updated based on the original annotation labels and the position prediction results corresponding to the original sample pairs to obtain a target model to be updated.
8. The method according to claim 7, It is characterized in that The arrangement position of the first object identifier in the original sample pair in the original object set is called the target arrangement position, and the arrangement position of the second object identifier in the original sample pair in the original object set is called the reference arrangement position; The obtaining of original annotation labels based on the behavior labels of each object identifier in the original sample pair includes: If it is determined according to the behavior labels of the object identifiers in the original sample pair that the object indicated by the first object identifier in the original sample pair has performed a feedback behavior and the object indicated by the second object identifier has not performed a feedback behavior, a first value is determined as the original labeling label, and the first value is used to indicate that: the target arrangement position is located before the reference arrangement position; If the behavior labels of the object identifiers in the original sample pair are the same, a second value is determined as the original label label, and the second value is used to indicate that: the target arrangement position and the reference arrangement position are the same position; If it is determined based on the behavior labels of each object identifier in the original sample pair that the object indicated by the first object identifier in the original sample pair has not performed feedback behavior and the object indicated by the second object identifier has performed feedback behavior, then a third value is determined as the original annotation label, and the third value is used to indicate that: the target arrangement position is located after the reference arrangement position.
9. The method according to claim 1, It is characterized in that The method further comprises: When the target information is to be pushed to the target object, and the target object is expected to perform a feedback behavior under the target behavior type on the target information, obtaining the object characteristics of the target object; Calling the updated target model corresponding to the target information information, predicting the description information of the behavior parameters of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, and obtaining target description information, wherein the target description information includes a score; A score threshold corresponding to the target information is obtained, and when the target description information is greater than the score threshold corresponding to the target information information, the target information is pushed to the target object.
10. The method according to claim 9, It is characterized in that Each object identifier in the candidate object set is selected from the original object set; the calculation method of the score threshold corresponding to the target information includes: Selecting an object set from the candidate object set and the original object set as a test object set; Calling the updated target model corresponding to the target information information, and predicting the description information of the behavior parameters of the corresponding object identifier under the target behavior type according to the object features corresponding to each object identifier in the test object set; Based on the description information of the behavior parameters of each object identifier in the test object set under the target behavior type, a score threshold corresponding to the target information information is calculated.
11. The method according to claim 1, It is characterized in that The target information is an information in an information set, and any information in the information set has a corresponding updated target model; The method further comprises: When M information in the information set is to be pushed to a target object, and the target object is expected to perform a feedback behavior under the target behavior type on the received information, an updated target model corresponding to each of the M information is obtained; M is an integer greater than 1; Respectively calling the updated target model corresponding to each of the information information, predicting the description information of the behavior parameter of the object identifier of the target object under the target behavior type according to the object characteristics of the target object, and obtaining M description information; one description information includes a score; Determine a disturbance factor for each piece of information based on a score threshold and corresponding description information corresponding to each piece of information; the disturbance factor is used to disturb the order in which the M pieces of information are pushed; Based on the disturbance factor of each information information, information information to be pushed is selected from the M information information, and the selected information information is pushed to the target object.
12. The method according to claim 11, It is characterized in that The determining the disturbance factor of each information information based on the score threshold corresponding to each information information and the corresponding description information includes: For an m-th information piece of information among the M pieces of information, calculating a ratio between the description information corresponding to the m-th information piece of information and a score threshold value corresponding to the m-th information piece of information; The calculated ratio is determined as the disturbance factor of the m-th information, m∈[1,M].
13. A model processing device, It is characterized in that include: An acquisition unit, used to acquire candidate sample pairs and corresponding sample annotation labels; The candidate sample pair includes two object identifiers in the candidate object set, the sample annotation label is determined based on the target behavior parameters of each object identifier in the candidate sample pair, any target behavior parameter indicates the importance of the feedback behavior under the target behavior type performed by the corresponding object on the target information information, and the sample annotation label is used to mark the positional relationship between the two object identifiers in the candidate sample pair in the candidate object set; A processing unit, configured to call the target model to be updated to predict the description information of the behavior parameters of the corresponding object identifiers under the target behavior type according to the object features corresponding to the object identifiers in the candidate sample pairs; The processing unit is further configured to predict the position relationship between the two object identifiers in the candidate sample pair in the candidate object set according to the predicted description information using a position prediction function to obtain a target position prediction result; The processing unit is also used to update the model parameters of the target model according to the sample annotation label and the target position prediction result to obtain an updated target model corresponding to the target information information; wherein the description information predicted by the updated target model corresponding to the target information information is used to decide whether to push the target information information to the corresponding object.
14. An electronic device, It is characterized in that The electronic device comprises an input interface and an output interface, and further comprises: a processor adapted to implement one or more instructions; and, A computer storage medium storing one or more instructions, wherein the one or more instructions are suitable for being loaded by the processor and executing the model processing method according to any one of claims 1 to 12.
15. A computer storage medium, It is characterized in that The computer storage medium stores computer program instructions, which, when executed by a processor, are used to execute the model processing method according to any one of claims 1 to 12.
16. A computer program product, It is characterized in that The computer program product comprises a computer program, which, when executed by a processor, is used to load and execute the model processing method according to any one of claims 1 to 12.
