Object recognition method, apparatus, computer device, and storage medium

Abstract

The application relates to an object recognition method and device, computer equipment and a storage medium. The method comprises the following steps: inputting a training sample into a shared feature extraction network in an object recognition model to perform feature extraction, obtaining sample extraction features, inputting sample extraction features of a target sample type into a push object recognition network in the object recognition model, obtaining a push object content conversion degree, inputting sample extraction features of a target sample type into a reference object recognition network in the object recognition model, obtaining a reference object content conversion degree, obtaining a push recognition loss value based on the push object content conversion degree, obtaining a reference recognition loss value based on the reference object content degree, and adjusting parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value to obtain a trained object recognition model. The method can improve the accuracy of object recognition.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an object recognition method, apparatus, computer device, and storage medium. Background Technology

[0002] With the development of computer and internet technologies, more and more content is being pushed to users via the internet, such as advertising or coupons.

[0003] In many cases, object recognition is required to identify the recipients of content. For example, different user groups may react differently to the same content, so it is necessary to identify users who match the content and then push the content to them to improve efficiency. However, traditional object recognition methods suffer from low accuracy. Summary of the Invention

[0004] Therefore, it is necessary to provide an object recognition method, apparatus, computer device, and storage medium that can improve the accuracy of object recognition in response to the above-mentioned technical problems.

[0005] An object recognition method includes: acquiring a training sample set corresponding to an object recognition model to be trained, the training sample set including training samples corresponding to content push objects and training samples corresponding to reference objects; inputting the training samples in the training sample set into a shared feature extraction network in the object recognition model for feature extraction to obtain sample extraction features corresponding to the training samples; determining the target sample type corresponding to the sample extraction features; inputting the sample extraction features of the target sample type being a push sample type into a push object recognition network in the object recognition model for recognition to obtain the push object content conversion degree corresponding to the content push object; inputting the sample extraction features of the target sample type being a reference sample type into a reference object recognition network in the object recognition model for recognition to obtain the reference object content conversion degree corresponding to the reference object; obtaining a push recognition loss value based on the push object content conversion degree corresponding to the content push object, and obtaining a reference recognition loss value based on the reference object content conversion degree corresponding to the reference object; adjusting the parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining a trained object recognition model based on the adjusted shared feature extraction network.

[0006] An object recognition device includes: a training sample set acquisition module, configured to acquire a training sample set corresponding to an object recognition model to be trained, the training sample set including training samples corresponding to content push objects and training samples corresponding to reference objects; a sample feature extraction module, configured to input the training samples in the training sample set into a shared feature extraction network in the object recognition model for feature extraction, thereby obtaining sample extraction features corresponding to the training samples; a target sample type determination module, configured to determine the target sample type corresponding to the sample extraction features; and a push object content conversion degree determination module, configured to input the sample extraction features whose target sample type is a push sample type into the push object recognition network in the object recognition model for recognition, thereby obtaining the target sample type; and a target sample type determination module, configured to determine the target sample type corresponding to the target sample type; and a push object content conversion degree determination module, configured to input the sample extraction features whose target sample type is a push sample type into the push object recognition network in the object recognition model for recognition, thereby obtaining the target sample type corresponding to the target sample type; and a target sample type determination module, configured to determine ... The content conversion degree of the push object corresponding to the content push object; the reference object content conversion degree acquisition module is used to input the sample extraction features of the target sample type as the reference sample type into the reference object recognition network in the object recognition model for recognition, and to obtain the reference object content conversion degree corresponding to the reference object; the recognition loss value acquisition module is used to obtain the push recognition loss value based on the push object content conversion degree corresponding to the content push object, and to obtain the reference recognition loss value based on the reference object content conversion degree corresponding to the reference object; the trained object recognition model acquisition module is used to adjust the parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and to obtain the trained object recognition model based on the adjusted shared feature extraction network.

[0007] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the object recognition method described above.

[0008] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the object recognition method described above.

[0009] The aforementioned object recognition method, apparatus, computer equipment, and storage medium, because the sample extraction features corresponding to the training samples of the content push object are extracted by a shared feature extraction network, and the sample extraction features corresponding to the training samples of the reference object are also extracted by a shared feature extraction network, the training samples of the content push object and the training samples of the reference object share a single shared feature extraction network. This allows the feature extraction network to be trained using both types of samples, even when the number of samples is unbalanced. This enables the shared feature extraction network to learn the ability to distinguish between the two types of samples, improving the model training effect and thus improving the accuracy of the trained object recognition model and the accuracy of object recognition.

[0010] An object recognition method includes: acquiring a candidate object set corresponding to target push content; the candidate object set includes multiple candidate objects; inputting object samples of the candidate objects into a shared feature extraction network of a trained object recognition model for feature extraction to obtain object extraction features corresponding to the candidate objects; inputting the object extraction features into a push object recognition network in the object recognition model for recognition, and obtaining a first content conversion probability corresponding to the candidate objects based on the recognition conversion probability; inputting the object extraction features into a reference object recognition network in the object recognition model for recognition, and obtaining a second content conversion probability corresponding to the candidate objects based on the recognition conversion probability; acquiring the probability difference between the first content conversion probability and the second content conversion probability corresponding to the candidate objects; and selecting candidate objects that meet the difference conditions from the candidate object set based on the probability difference, and using them as push objects corresponding to the target push content; the difference conditions include at least one of the following: the probability difference is ranked before a preset ranking or the probability difference is greater than a difference threshold.

[0011] An object recognition device, comprising: a candidate object set acquisition module, configured to acquire a candidate object set corresponding to target push content; the candidate object set includes multiple candidate objects; an object extraction feature acquisition module, configured to input object samples of the candidate objects into a shared feature extraction network of a trained object recognition model for feature extraction, thereby obtaining object extraction features corresponding to the candidate objects; a first content conversion probability acquisition module, configured to input the object extraction features into a push object recognition network in the object recognition model for recognition, and obtain a first content conversion probability corresponding to the candidate objects based on the recognition-obtained conversion probability; and a second content conversion probability acquisition module. The module is used to input the extracted features of the object into the reference object recognition network in the object recognition model for recognition, and to obtain the second content conversion probability corresponding to the candidate object based on the conversion probability obtained by recognition; the probability difference obtaining module is used to obtain the probability difference between the first content conversion probability corresponding to the candidate object and the second content conversion probability; the push object determining module is used to filter the candidate objects that meet the difference conditions from the candidate object set based on the probability difference corresponding to the candidate objects, and use them as the push objects corresponding to the target push content; the difference conditions include at least one of the following: the probability difference is ranked before a preset ranking or the probability difference is greater than a difference threshold.

[0012] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the object recognition method described above.

[0013] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the object recognition method described above.

[0014] In some embodiments, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above-described method embodiments.

[0015] The aforementioned object recognition method, apparatus, computer equipment, and storage medium acquire a candidate object set corresponding to the target push content. The candidate object set includes multiple candidate objects. Object samples of the candidate objects are input into a shared feature extraction network of a trained object recognition model for feature extraction, resulting in extracted object features corresponding to the candidate objects. The extracted object features are then input into the push object recognition network of the object recognition model for recognition. Based on the recognition-derived conversion probability, a first content conversion probability corresponding to the candidate object is obtained. The extracted object features are input into a reference object recognition network of the object recognition model for recognition, resulting in a second content conversion probability corresponding to the candidate object. The probability difference between the first and second content conversion probabilities corresponding to the candidate objects is obtained. Based on the probability difference, candidate objects that meet the difference conditions are selected from the candidate object set as the push objects corresponding to the target push content. The difference conditions include at least one of the following: the probability difference ranking is before a preset ranking or the probability difference is greater than a difference threshold. Thus, the trained object recognition model quickly and conveniently identifies the push objects corresponding to the target push content, improving the accuracy of object recognition. Attached Figure Description

[0016] Figure 1 This is a diagram illustrating the application environment of the object recognition method in some embodiments;

[0017] Figure 2 This is a flowchart illustrating the object recognition method in some embodiments;

[0018] Figure 3 The diagram shows the model structure of the delayed feedback model in some embodiments;

[0019] Figure 4 Here are structural diagrams of the object recognition model in some embodiments;

[0020] Figure 5 This is a flowchart illustrating the object recognition method in some embodiments;

[0021] Figure 6 Here are flowcharts illustrating the training of two models in some embodiments;

[0022] Figure 7 Here is a flowchart of a process for predicting gain values ​​using two trained models in some embodiments.

[0023] Figure 8 Here are model structure diagrams for a single model in some embodiments;

[0024] Figure 9 Here are model structure diagrams of the category conversion methods in some embodiments;

[0025] Figure 10 This is a structural block diagram of the object recognition device in some embodiments;

[0026] Figure 11 This is a structural block diagram of the object recognition device in some embodiments;

[0027] Figure 12 These are internal structural diagrams of the computer device in some embodiments;

[0028] Figure 13 This is a diagram showing the internal structure of a computer device in some embodiments. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0030] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0031] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and intelligent transportation.

[0032] Computer vision (CV) is a science that studies how to enable machines to "see." More specifically, it refers to machine vision, which uses cameras and computers to replace human eyes in tasks such as target recognition, tracking, and measurement, and further performs image processing to create images more suitable for human observation or transmission to instruments. As a scientific discipline, computer vision studies related theories and technologies, attempting to build artificial intelligence systems capable of extracting information from images or multidimensional data. Computer vision technologies typically include image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, 3D object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping (SLAM), autonomous driving, intelligent transportation, and other technologies, as well as common biometric recognition technologies such as facial recognition and fingerprint recognition.

[0033] Machine Learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or 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 endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0034] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, smart customer service, vehicle networking, and intelligent transportation. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0035] The solutions provided in this application involve technologies such as machine learning in artificial intelligence, and are specifically illustrated through the following embodiments:

[0036] The object recognition method provided in this application can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network.

[0037] Server 104 can be a server for pushing content, such as a server that pushes content to an application, like a server that pushes advertisements. The application may include at least one of an instant messaging application or a shopping application. Terminal 102 may have a client corresponding to the application installed. Server 104 can push content to the client corresponding to the application in terminal 102. Terminal 102 can display the pushed content in the client. For example, the client can display the pushed content in a push content display area. When the pushed content is an advertisement, the push content display area can be an advertisement slot.

[0038] Specifically, server 104 can obtain the training sample set corresponding to the object recognition model to be trained. The training sample set includes training samples corresponding to content push objects and training samples corresponding to reference objects. The training samples in the training sample set are input into the shared feature extraction network in the object recognition model for feature extraction to obtain the sample extraction features corresponding to the training samples. The target sample type corresponding to the sample extraction features is determined. The sample extraction features of the target sample type being push sample type are input into the push object recognition network in the object recognition model for recognition to obtain the push object content conversion degree of the content push object. The sample extraction features of the target sample type being reference sample type are input into the reference object recognition network in the object recognition model for recognition to obtain the reference object content conversion degree of the reference object. The push recognition loss value is obtained based on the push object content conversion degree, and the reference recognition loss value is obtained based on the reference content conversion degree. The parameters of the shared feature extraction network are adjusted based on the push recognition loss value and the reference recognition loss value. The trained object recognition model is obtained based on the adjusted shared feature extraction network. Here, the content push object refers to the object that has been pushed training push content, and the reference object refers to the object that has not been pushed training push content.

[0039] Server 104 can respond to a content push request for a target push content by obtaining a candidate object set, which includes multiple candidate objects. Object samples of the candidate objects are input into a shared feature extraction network of a trained object recognition model for feature extraction, resulting in extracted object features. These extracted features are then input into a push object recognition network within the object recognition model for recognition. Based on the recognition's conversion probability, a first content conversion probability is obtained for each candidate object. The extracted object features are then input into a reference object recognition network within the object recognition model for recognition, resulting in a second content conversion probability for each candidate object. The probability difference between the first and second content conversion probabilities for each candidate object is obtained. Based on this probability difference, candidate objects that meet the difference criteria are selected from the candidate object set and used as the push objects for the target push content. The difference criteria include at least one of the following: the probability difference ranking is before a preset ranking, or the probability difference is greater than a difference threshold. Server 104 can then push the target push content to the push object. The target push content can be the same as or different from the training push content.

[0040] The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server 104 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud storage, network services, cloud communication, big data, and artificial intelligence platforms. The terminal and the server can be connected directly or indirectly through wired or wireless communication, which is not limited herein.

[0041] It is understood that the above application scenarios are merely examples and do not constitute a limitation on the object recognition provided in the embodiments of this application. The method provided in the embodiments of this application can also be applied in other application scenarios. For example, the object recognition provided in this application can be performed by terminal 102. Terminal 102 can upload the obtained trained object recognition model to server 104. Server 104 can store the trained object recognition model or forward the trained object recognition model to other devices.

[0042] As provided in this application, the object recognition method includes a server 104, which can be a node in a blockchain. The server 104 can store the trained object recognition model in the blockchain. The server 104 can also filter the push objects corresponding to the push content from the candidate object set, establish a correspondence between the push content and the push objects, and store this correspondence in the blockchain. The server 104 can then query the blockchain to obtain the push object corresponding to the push content and push the push content to that object.

[0043] In some embodiments, such as Figure 2 As shown, an object recognition method is provided. This method can be executed by a terminal or a server, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the following steps are included:

[0044] S202, Obtain the training sample set corresponding to the object recognition model to be trained. The training sample set includes the training samples corresponding to the content push object and the training samples corresponding to the reference object.

[0045] The target audience can be natural persons, such as users of an application. Applications can include shopping apps or financial management apps. The content push target refers to the object to which content was pushed. Content can include virtual things that can be pushed via the network, such as electronic coupons or advertisements, or real things, such as paper coupons. The reference target is the opposite of the content push target; it refers to the object to which content was not pushed. The content pushed to the content push target and the content not pushed to the reference target can be the same. For example, for a coupon, if the coupon is pushed to object A, then object A is the content push target; if the coupon is not pushed to object B, then object B is the reference target. Of course, the content can also be different.

[0046] Content is pushed to an audience to incentivize conversion. The conversion rate depends on the purpose of the content. If the purpose of the content is to incentivize a user to click, and the audience clicks on the content, a conversion is considered complete. If the purpose of the content is to incentivize a user to purchase a product related to the content, and the audience purchases that product, a conversion is considered complete. The audience to whom content is pushed can also be called the intervention audience, and the reference audience can also be called the reference audience.

[0047] The training sample set includes multiple training samples used to train the object recognition model. Training samples can be determined based on object attribute information, such as age, occupation, education level, hobbies, region, or gender. Training samples for content push objects can be determined based on the corresponding object attribute information, and training samples for reference objects can be determined based on the corresponding object attribute information. Training samples for content push objects can be called intervention samples, and training samples for reference objects can be called control samples. Here, intervention (treatment) indicates that content push has been performed; for example, issuing coupons to an object is an intervention, as is pushing advertisements. Training samples for content push objects can also be called push training samples, and training samples for reference objects can also be called reference training samples.

[0048] The training samples can include both positive and negative samples. The training samples for the content push object can include both positive and negative samples, and the training samples for the reference object can include both positive and negative samples.

[0049] Positive samples refer to training samples corresponding to objects that have undergone conversion within the observation period, while negative samples refer to training samples corresponding to objects that have not undergone conversion within the observation period. The observation period refers to the time between the content push time and the reference time for dividing positive and negative samples; this reference time can also be called the observation day. The content push time refers to the time when content is pushed to the content push object. Because the same object may not have converted at one time but may have converted at another, meaning the conversion result of the same object may be different at different times—that is, the conversion result changes over time and can be either converted or not converted—it is necessary to determine whether the training samples are positive or negative samples based on time. The reference time for dividing positive and negative samples refers to the time used to classify training samples as positive or negative samples. For example, if the content was pushed on January 1st and the reference time for dividing positive and negative samples is January 10th, then the observation period is from January 1st to January 10th. If the object converted during the period from January 1st to January 10th, then the training sample corresponding to that object is a positive sample; if the object did not convert during the same period, then the training sample corresponding to that object is a negative sample. The reference time for dividing positive and negative samples can be preset or set as needed.

[0050] The number of positive samples and the number of negative samples can be the same or different. For example, in a financial scenario, since the reference object is not intervened in, the transformation of the reference object relies on the natural transformation of the object, resulting in a low transformation rate. Consequently, the number of positive samples corresponding to the reference object is less than the number of negative samples corresponding to the reference object.

[0051] The object recognition model can be a neural network model based on artificial intelligence. It is used to identify object types, which can include at least one of the following: marketing-sensitive objects, naturally converting objects, indifferent objects, or reactive objects. The object recognition model to be trained can be an untrained model or a trained model that requires further training. Marketing-sensitive objects are those that convert only when content is pushed to them; naturally converting objects convert regardless of whether content is pushed to them; indifferent objects do not convert regardless of whether content is pushed to them; and reactive objects are those that are averse to marketing activities, converting only when content is pushed to them. For example, marketing-sensitive objects are those that will only buy when coupons are issued; naturally converting objects will buy regardless of coupons; indifferent objects will not buy regardless of coupons; and reactive objects are those that are averse to marketing activities, making purchases when coupons are not issued but not after coupons are issued. By using object recognition models, marketing-sensitive targets can be identified, allowing content to be pushed to them, incentivizing conversions, improving the accuracy of content delivery, and reducing the cost of content delivery.

[0052] Specifically, the server can obtain a set of objects, which includes multiple objects, determine the training content to be pushed (the training content can be any content), randomly select multiple first objects from the object set, and use the objects in the object set other than the first objects as second objects. The training content is pushed to each of the first objects, and the first objects that have received the training content are designated as content push objects. Each second object is used as a reference object. The number of first objects and the number of second objects can be the same or different; that is, the number of push content objects and the number of reference objects can be the same or different.

[0053] In some embodiments, the number of training samples corresponding to the content push object is less than the number of training samples corresponding to the reference object. For example, in a financial scenario, due to the low conversion rate and low frequency of object behavior, it is difficult to reserve a large number of reference objects. Reference objects can also be called unbiased control objects, and the training samples corresponding to the reference objects can also be called unbiased control samples. That is, unbiased control samples are scarce in a financial scenario, and users have high decision-making costs and difficulty in conversion, which leads to the number of samples of the content push object being less than the number of samples of the reference object.

[0054] In some embodiments, the number of training samples corresponding to the content push object is less than the number of training samples corresponding to the reference object, and the number of positive samples in the training samples corresponding to the reference object is less than the number of negative samples in the training samples corresponding to the reference object.

[0055] S204, input the training samples in the training sample set into the shared feature extraction network in the object recognition model to extract features, and obtain the sample extraction features corresponding to the training samples.

[0056] The object recognition model may include a shared feature extraction network, where the extracted features are the features obtained by the shared feature extraction network from the training samples.

[0057] Specifically, the server can input the training samples corresponding to the content push objects into a shared feature extraction network for feature extraction, obtaining the extracted features of the training samples corresponding to the content push objects. Similarly, it can input the training samples corresponding to the reference objects into the shared feature extraction network for feature extraction, obtaining the extracted features of the training samples corresponding to the reference objects. When training the object recognition model, one training sample can be input into the model at a time, meaning one training sample is used per training session. Alternatively, multiple training samples can be used per training session, such as all training samples.

[0058] In some embodiments, the shared feature extraction network may include an embedding layer, a concatenation layer, and a feature extraction layer. Training samples include multiple object attribute information, such as age, occupation, and gender. The server can use the embedding layer in the shared feature extraction network to transform the word vectors of each object attribute information in the training samples into word vectors, obtaining word vectors corresponding to each object attribute information. These word vectors are then input into the concatenation layer in the shared feature extraction network for aggregation, yielding an aggregation result. This aggregation result is then input into the feature extraction layer for further feature extraction, obtaining the extracted features corresponding to the training samples. The aggregation can be achieved by concatenating the word vectors of each object attribute information or by weighting the word vectors of each object attribute information. The feature extraction layer may include at least one of a convolutional layer or an activation layer.

[0059] S206, Determine the target sample type corresponding to the extracted sample features.

[0060] The sample type distinguishes between training samples corresponding to the content push object and training samples corresponding to the reference object. The sample type can include both push sample type and reference sample type, and the target sample type can be either push sample type or reference sample type. The target sample type corresponding to the extracted feature refers to the sample type of the training sample to which the extracted feature belongs. If the target sample type is a push sample type, the extracted feature is the feature obtained by extracting features from the training sample of the content push object; if the target sample type is a reference sample type, the extracted feature is the feature obtained by extracting features from the training sample of the reference object.

[0061] Specifically, when the server obtains the sample extraction features corresponding to the training samples, it can determine the sample type corresponding to the sample extraction features. For example, the sample type can be determined based on the sample label value of the training samples. The sample label value is used to distinguish sample types; different sample label values ​​correspond to different sample types. The sample label value can be preset or set as needed. For example, the sample label value corresponding to the push sample type can be set to 1, and the sample label value corresponding to the reference sample type can be set to 0.

[0062] S208, the sample extraction features of the target sample type are input into the push object recognition network in the object recognition model for recognition, and the content conversion degree of the push object corresponding to the content push object is obtained.

[0063] The object recognition model can also include a push object recognition network. The input to the push object recognition network is the sample extraction features of the training samples corresponding to the content push object.

[0064] If the training samples corresponding to the content push target are positive samples, then the content conversion rate refers to the probability of conversion of the content push target at the reference time when the positive and negative samples are divided. For example, the probability of the content push target clicking on the pushed content at the reference time, or the probability of the content push target purchasing a product related to the pushed content at the reference time. For instance, if the pushed content is a computer coupon, the conversion probability can be the probability that a user who receives the coupon will purchase the computer. For example, if the observation period is January 1st to January 10th, and the reference time for dividing the positive and negative samples is January 10th, then the content conversion rate reflects the probability of conversion of the content push target on January 10th. The content conversion rate is positively correlated with the probability of conversion of the content push target at the reference time when the positive and negative samples are divided.

[0065] If the training samples corresponding to the content push target are negative samples, then the content conversion rate of the push target refers to the probability that the content push target will not convert at the reference time when the positive and negative samples are divided. The content conversion rate of the push target is positively correlated with the probability that the content push target will not convert at the reference time when the positive and negative samples are divided.

[0066] A positive correlation refers to a relationship where, all other things being equal, two variables change in the same direction; when one variable decreases, the other also decreases. It's important to understand that a positive correlation means the direction of change is consistent, but it doesn't require that a slight change in one variable necessarily means a change in the other. For example, we can set variable b to 100 when variable a is between 10 and 20, and to 120 when variable a is between 20 and 30. In this case, the direction of change for both a and b is that as a increases, b also increases. However, when a is between 10 and 20, b may remain unchanged.

[0067] Specifically, the server can determine the sample type corresponding to the extracted sample features. If the sample type corresponding to the extracted sample features is a push sample type, the extracted sample features are input into the push object recognition network for recognition.

[0068] In some embodiments, the push object identification network may include a conversion probability identification network and a probability decay factor identification network. The server can input the sample extracted features into the conversion probability identification network to identify the conversion probability and obtain the total conversion probability of the content push object. The sample extracted features can be input into the probability decay factor identification network to identify the decay factor and obtain the probability decay factor corresponding to the content push object. Based on the probability decay factor, the sub-conversion probability of the content push object is determined. The probability decay factor refers to the parameter of the distribution function that the probability of the object converting decreases over time, that is, the parameter of the distribution function that the probability of the object converting decreases over time. For example, when the probability of the object converting decreases over time according to an exponential distribution, the distribution function can be an exponential distribution function. The sub-conversion probability refers to the probability that an object that will eventually convert will convert at a certain time, such as the probability that an object that will eventually convert will convert on the dth day after the content push. Total conversion probability refers to the likelihood that an object will eventually convert, such as the probability that a user will eventually purchase a product from a pushed advertisement. The server can calculate the content conversion rate of the content push object based on the total conversion probability of the content push object and the sub-conversion probability of the content push object.

[0069] The first conversion probability identification network mentioned below refers to the conversion probability identification network in the push object identification network, and the first probability decay factor identification network refers to the probability decay factor identification network in the push object identification network.

[0070] S210, the sample extracted features of the target sample type are input into the reference object recognition network in the object recognition model for recognition, and the reference object content conversion degree corresponding to the reference object is obtained.

[0071] The object recognition model may also include a reference object recognition network. The input to the reference object recognition network is the sample extraction features of the training samples corresponding to the reference object. If the training samples corresponding to the reference object are positive samples, then the reference object content conversion degree refers to the probability that the reference object will convert at the reference time when the positive and negative samples are divided, such as the probability of conversion on the observation day. The reference object content conversion degree is positively correlated with the probability that the reference object will convert at the reference time when the positive and negative samples are divided. If the training samples corresponding to the reference object are negative samples, then the reference object content conversion degree refers to the probability that the reference object will not convert at the reference time when the positive and negative samples are divided. The reference object content conversion degree is positively correlated with the probability that the reference object will not convert at the reference time when the positive and negative samples are divided.

[0072] Specifically, the server can determine the sample type corresponding to the extracted sample features. If the sample type corresponding to the extracted sample features is a reference sample type, the extracted sample features are input into the reference object recognition network for recognition.

[0073] In some embodiments, the reference object identification network may include a conversion probability identification network and a probability decay factor identification network. The conversion probability identification network takes sample extraction features as input and outputs the total conversion probability. The probability decay factor identification network takes sample extraction features as input and outputs a probability decay factor. The server can input the sample extraction features into the conversion probability identification network to identify the conversion probability and obtain the total conversion probability of the reference object. It can also input the sample extraction features into the probability decay factor identification network to identify the decay factor and obtain the probability decay factor corresponding to the reference object. Based on the probability decay factor and the total conversion probability of the reference object, the server can calculate the reference object content conversion degree corresponding to the reference object.

[0074] The second transformation probability identification network mentioned below refers to the transformation probability identification network in the reference object identification network, and the second probability attenuation factor identification network refers to the probability attenuation factor identification network in the reference object identification network.

[0075] In some embodiments, the server processes the sample extraction features using the network parameters of the second conversion probability recognition network to obtain processed sample extraction features. The processed sample extraction features are then normalized, and the normalized result is used as the total conversion probability of the reference object. For example, the server can calculate the total conversion probability using formula (1). Where X... i The sample features corresponding to object i are represented by C, which indicates whether the object will eventually convert (C=1 indicates that the user eventually converts, C=0 indicates that the user eventually does not convert), and X represents the user feature set, including the user features X corresponding to multiple users. i w c The network parameters representing the transformation probability identification network are: P(X) i ) represents the total conversion probability.

[0076]

[0077] In some embodiments, the training samples corresponding to the reference object are positive samples. The server can obtain the conversion duration corresponding to the reference object and calculate the sub-conversion probability based on the conversion duration and the probability decay factor. Specifically, when the conversion probability of an object gradually decays over time and the decay follows an exponential distribution, formula (2) can be used to calculate the sub-conversion probability. P(D=d|X=X i,C=1)=λ(X i )·exp(-λ(X i )d) (2). Wherein, λ(X) i P(D=d|X=X) represents the probability decay factor, and d represents the duration of the transformation. i C = 1) represents the sub-conversion probability, which also represents the probability that the object to be converted will be converted on day d. The server can multiply the probability decay coefficient with the total conversion probability of the reference object, and use the result of the multiplication operation as the reference object content conversion degree. For example, the reference object content conversion degree can be calculated using formula (3). P(Y=1,D=d i |X=X i E = e i )=P(X i )·λ(X i )·exp(-λ(X i )d i (3)

[0078] Where P(Y=1,D=d) i |X=X i E = e i The conversion rate of the reference object content is represented by P(X). Specifically, the conversion rate of the reference object content for positive samples can be expressed as P(X). i )·λ(X i )·exp(-λ(X i )d i ), d i Let D represent the conversion time for object i. D indicates the conversion time, such as the time interval from the initial delivery to the final conversion by the user. Y indicates whether the user has converted by the observation time (which can be understood as the reference time for dividing positive and negative samples as mentioned above). Y = 1 indicates that the user has converted by the observation time, and Y = 0 indicates that the user has not converted by the observation time. E represents the observation time, such as the time interval from the initial delivery to the current waiting period. For example, suppose a user was delivered a red dot on May 1st and converted on May 5th. If the user had not converted by the observation time on May 3rd, the values ​​corresponding to the above symbols would be: Y = 0, C = 1, D = 5, E = 3.

[0079] In some embodiments, the training samples corresponding to the reference object are negative samples. The server can obtain the observation duration, which refers to the length of time corresponding to the observation period. The server can calculate the probability adjustment value of the reference object based on the observation duration, the total conversion probability of the reference object, and the probability decay factor. Based on the probability adjustment value of the reference object and the total conversion probability, the server can calculate the content conversion degree of the reference object. For example, the server can use formula (4) to calculate the content conversion degree of the reference object. Wherein, P(Y=0|X=X i E = e i The value 1-P(X) represents the conversion degree of the reference object content, i.e., the conversion degree of the reference object content corresponding to the negative sample is 1-P(X). i )+P(X i )exp(-λ(X i )e i ), P(X i )exp(-λ(X i )e i ) indicates the possible adjustment value.

[0080]

[0081] S212, obtain the push recognition loss value based on the content conversion degree of the push object corresponding to the content push object, and obtain the reference recognition loss value based on the content conversion degree of the reference object.

[0082] The push recognition loss value is calculated based on the content conversion rate of the push object. It can be calculated based on the conversion rate of a single push object (i.e., one conversion rate corresponds to one push recognition loss value), or it can be calculated based on the conversion rates of multiple push objects (at least two). The push recognition loss value is negatively correlated with the content conversion rate of the push object.

[0083] The reference recognition loss value is calculated based on the reference content conversion degree of a reference object. It can be calculated based on the reference content conversion degree of a single reference object (i.e., one reference content conversion degree corresponds to one reference recognition loss value), or it can be calculated based on the reference content conversion degrees of multiple reference objects. The reference recognition loss value is negatively correlated with the reference content conversion degree.

[0084] A negative correlation refers to a situation where, all other things being equal, two variables change in opposite directions; when one variable decreases, the other increases. It's important to understand that a negative correlation means the changes are in opposite directions, but it doesn't mean that a slight change in one variable necessarily requires a change in the other.

[0085] Specifically, the server can perform statistical calculations on the content conversion rates of each content push object to obtain a push statistical conversion rate. For example, the server can sum the conversion rates of each push object and use the sum as the push statistical conversion rate; alternatively, the server can multiply the conversion rates of each push object and use the result of the product operation as the push statistical conversion rate; or the server can multiply the conversion rates of each push object, perform a logarithmic operation on the result of the product operation, and use the result of the logarithmic operation as the push statistical conversion rate. The server can obtain a push recognition loss value based on the push statistical conversion rate, and the push recognition loss value is negatively correlated with the push statistical conversion rate.

[0086] In some embodiments, the server can perform statistical calculations on the content conversion rates of each reference object to obtain a reference statistical conversion rate. For example, the server can sum the content conversion rates of each reference object and use the sum as the reference statistical conversion rate. Alternatively, the server can multiply the content conversion rates of each reference object and use the result of the multiplication as the reference statistical conversion rate. Or, the server can multiply the content conversion rates of each reference object, perform a logarithmic operation on the result of the multiplication, and use the result of the logarithmic operation as the reference statistical conversion rate. The server can obtain a push recognition loss value based on the reference statistical conversion rate, and the push recognition loss value is negatively correlated with the reference statistical conversion rate. For example, the server can calculate the push recognition loss value using formula (5), where L DFM This indicates the push recognition loss value.

[0087] L DFM =-∑ y＝1 logP(X i )+logλ(X i )-λ(X i )d i -∑ y＝0 log(1-P(X i )+P(X i )exp(-λ(X i )e i ))(5)

[0088] S214, The parameters of the shared feature extraction network are adjusted based on the push recognition loss value and the reference recognition loss value, and the trained object recognition model is obtained based on the adjusted shared feature extraction network.

[0089] In this context, the shared feature extraction network can have multiple parameters. The term "shared feature extraction network" refers to the variable parameters within the network. For neural networks, these parameters can also be called network weights. The trained object recognition model can be obtained through one or more training iterations.

[0090] Specifically, the server can adjust the parameters of the shared feature extraction network in the direction of decreasing the push recognition loss value and in the direction of decreasing the reference recognition loss value. This can be done through multiple iterations of training. Training stops when a convergence condition is met, resulting in a trained object recognition model. The convergence condition can include the reference recognition loss value being less than a reference loss value threshold, the push recognition loss value being less than a push loss value threshold, or the sum of the reference recognition loss value and the push recognition loss value being less than an exponent of a total loss value threshold. The convergence condition can also include the parameter change being less than a parameter change threshold. The reference loss value threshold, push loss value threshold, total loss value threshold, and parameter change threshold can be preset or set as needed.

[0091] In some embodiments, the server can sequentially adjust the parameters of the shared feature extraction network using the push recognition loss value and the reference recognition loss value. For example, it can first adjust the parameters of the shared feature extraction network using the push recognition loss value to obtain a shared feature extraction network with adjusted push recognition loss value, and then adjust the parameters of the shared feature extraction network with adjusted push recognition loss value using the reference recognition loss value to obtain a shared feature extraction network with adjusted reference recognition loss value. Alternatively, it can first adjust the parameters of the shared feature extraction network using the reference recognition loss value to obtain a shared feature extraction network with adjusted reference recognition loss value, and then adjust the parameters of the shared feature extraction network with adjusted push recognition loss value using the push recognition loss value to obtain a shared feature extraction network with adjusted push recognition loss value. The trained object recognition model is then obtained based on the adjusted shared feature extraction network.

[0092] In some embodiments, there are multiple push recognition loss values. For example, a push recognition loss value can be calculated for each push object content conversion degree. There are also multiple reference recognition loss values. For example, a reference recognition loss value can be calculated for each reference object content conversion degree. The server can use multiple push recognition loss values ​​and multiple reference recognition loss values ​​to adjust the parameters of the shared feature extraction network in sequence, and obtain the trained object recognition model based on the adjusted shared feature extraction network.

[0093] In some embodiments, the server can perform statistical calculations on the reference recognition loss value and the push recognition loss value to obtain a statistical loss value, use the statistical loss value to adjust the parameters of the shared feature extraction network, and obtain a trained object recognition model based on the adjusted shared feature extraction network.

[0094] In the above object recognition method, since the sample extraction features corresponding to the training samples of the content push object are extracted by a shared feature extraction network, and the sample extraction features corresponding to the training samples of the reference object are also extracted by a shared feature extraction network, the training samples of the content push object and the training samples of the reference object share the same shared feature extraction network. This allows the shared feature extraction network to be trained using two different samples, namely the training samples of the content push object and the training samples of the reference object, even when the number of samples is unbalanced. This enables the shared feature extraction network to learn the ability to distinguish between the two types of samples, improving the model training effect and thus improving the accuracy of the trained object recognition model and the accuracy of object recognition.

[0095] In some embodiments, determining the target sample type corresponding to the sample extraction feature includes: obtaining the sample label value of the training sample corresponding to the sample extraction feature; and determining the target sample type corresponding to the sample extraction feature based on the sample label value.

[0096] Each training sample can have a corresponding sample label value. The sample label value is used to distinguish the sample type; different sample label values ​​correspond to different sample types. The sample label value corresponding to the push sample type can be a value greater than or equal to 1, for example, it can be 1. The sample label value corresponding to the reference sample type can be a value less than or equal to 0, for example, it can be 0.

[0097] Specifically, the server can input training samples and their corresponding sample label values ​​into the object recognition model. When the sample extraction features corresponding to the training samples are obtained, the sample type of the training samples is determined based on the sample label values. If the sample type is a push sample type, the sample extraction features are input into the push object recognition network in the object recognition model for recognition. If the sample type is a reference sample type, the sample extraction features are input into the reference object recognition network in the object recognition model for recognition.

[0098] In this embodiment, the sample label value of the training sample corresponding to the sample extraction feature is obtained, and the target sample type corresponding to the sample extraction feature is determined based on the sample label value. This can distinguish between the push training sample and the reference training sample in the training sample set, so that although the push training sample and the reference training sample share a shared feature extraction network, the object recognition model can still distinguish between the push training sample and the reference training sample.

[0099] In some embodiments, the content conversion degree of the push object is used to represent the conversion probability of the content push object for the push content. The push recognition loss value is obtained based on the content conversion degree of the push object corresponding to the content push object. This includes: multiplying the sample label value with the content conversion degree of the push object to obtain the object recognition loss value corresponding to the content push object; and statistically analyzing the object recognition loss values ​​of multiple content push objects to obtain the push recognition loss value.

[0100] Conversion probability refers to the likelihood of an object converting, such as the likelihood of an object purchasing a product from a pushed advertisement, or the likelihood of an object clicking on a pushed advertisement. When the training samples of the content push object are positive samples, the content conversion rate of the push object represents the probability of the content push object converting in response to the pushed content; that is, the content conversion rate of the push object represents the probability of the content push object converting at the reference time when positive and negative samples are separated. The higher the conversion probability, the greater the probability that the user will convert.

[0101] Specifically, the server can multiply the sample label value with the content conversion degree of the push object to obtain the multiplied content conversion degree of the push object. Based on the multiplied content conversion degree of the push object, the server can obtain the object recognition loss value of the content push object. The object recognition loss value is negatively correlated with the multiplied content conversion degree of the push object. The server can sum the object recognition loss values ​​of multiple content push objects and use the summation result as the push recognition loss value.

[0102] In some embodiments, the server can determine the first initial object loss value corresponding to the content push object based on the content conversion degree of the push object. The first initial object loss value is negatively correlated with the content conversion degree of the push object. The first initial identification loss value is weighted using sample label values, and the result of the weighted calculation is used as the push identification loss value. For example, the push identification loss value can be calculated using formula (6), where L T S represents the push recognition loss value. T Let S represent the set of content push objects, and k represent the training sample set. i k represents the sample label value of the training sample corresponding to object i. i =1 indicates that the sample type is a push sample type, k i =0 indicates that the sample type is the reference sample type. K includes 0 and 1. S represents the first initial object loss value for object i. i This represents the training sample corresponding to object i in the training sample set S.

[0103]

[0104] In some embodiments, the server can determine the second initial object loss value corresponding to the reference object based on the content conversion degree of the reference object. The second initial object loss value is negatively correlated with the content conversion degree of the reference object. The server can calculate the reference recognition loss value using the sample label value and the second initial object loss value. For example, the server can calculate the reference recognition loss value using formula (7), where L C S represents the reference recognition loss value. C S represents a set of reference objects, where S includes S0. T and S C .

[0105]

[0106] In this embodiment, the sample label value is multiplied by the conversion degree of the content of the push object to obtain the object recognition loss value corresponding to the content push object. The object recognition loss values ​​of multiple content push objects are statistically analyzed to obtain the push recognition loss value, which improves the efficiency of calculating the loss value.

[0107] In some embodiments, adjusting the parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining the trained object recognition model based on the adjusted shared feature extraction network includes: adjusting the parameters of the shared feature extraction network based on the push recognition loss value to obtain a shared feature extraction network with adjusted push recognition loss value; adjusting the parameters of the shared feature extraction network with adjusted push recognition loss value based on the reference recognition loss value to obtain a shared feature extraction network with both reference recognition loss value and push recognition loss value adjusted; and obtaining the trained object recognition model based on the shared feature extraction network with both reference recognition loss value and push recognition loss value adjusted.

[0108] Specifically, the server can input training samples into the object recognition model to obtain the content conversion rate of the pushed objects. Based on the content conversion rate, a push recognition loss value is obtained. The parameters of the shared feature extraction network are adjusted based on this loss value, resulting in an adjusted shared feature extraction network. Reference training samples are then input into the object recognition model to obtain the reference content conversion rate of the reference objects. A reference recognition loss value is obtained based on this reference conversion rate. The parameters of the adjusted shared feature extraction network are then adjusted based on this reference loss value, resulting in both the reference loss value and the adjusted shared feature extraction network. Finally, the trained object recognition model is obtained. The shared feature extraction network can be adjusted multiple times using the push recognition loss value or multiple times using the reference loss value. The order of parameter adjustment using the push recognition loss value and the reference loss value can be set as needed without significant restrictions.

[0109] In this embodiment, by adjusting the parameters of the shared feature extraction network using the push recognition loss value and by adjusting the parameters of the shared feature extraction network using the reference recognition loss value, the trained shared feature extraction network can learn both the features of the content push object and the corresponding features of the reference, thereby improving the training accuracy.

[0110] In some embodiments, adjusting the parameters of the shared feature extraction network based on the push recognition loss value to obtain the shared feature extraction network after push recognition loss value adjustment includes: determining the push parameter adjustment related value corresponding to the push object recognition network based on the push recognition loss value; adjusting the parameters of the push object recognition network based on the push parameter adjustment related value to obtain the parameter-adjusted push object recognition network; determining the first feature parameter adjustment related value corresponding to the shared feature extraction network based on the push recognition loss value and the first loss value adjustment direction; and adjusting the parameters of the shared feature extraction network based on the first feature parameter adjustment related value to obtain the shared feature extraction network after push recognition loss value adjustment.

[0111] Among them, the push parameter adjustment related values ​​refer to the values ​​used to adjust the parameters of the push object recognition network, and the push recognition loss value corresponds to the push recognition loss function, which has specific values. The first loss value adjustment direction can be the direction of gradient descent. The first feature parameter adjustment related values ​​refer to the values ​​used to adjust the parameters of the shared feature extraction network.

[0112] Specifically, the server can adjust the parameters in the shared feature extraction network in the direction of reducing the push recognition loss value, and adjust the parameters in the push object recognition network in the direction of reducing the push recognition loss value, to obtain the adjusted shared feature extraction network and the adjusted push object recognition network. Based on the adjusted shared feature extraction network and the adjusted push object recognition network, a trained object recognition model is obtained.

[0113] In some embodiments, the server can use the push recognition loss function to perform gradient calculation on the parameters of the push object recognition network to obtain a first gradient, and determine the push parameter adjustment related value based on the first gradient. In this case, adjusting the parameters of the push object recognition network using the push parameter adjustment related value can make the first gradient change in a decreasing direction.

[0114] In some embodiments, the server can use the push recognition loss value to perform gradient calculation on the parameters of the shared feature extraction network to obtain a second gradient. The adjustment direction of the first loss value is the descent direction of the second gradient. The server can determine the adjustment correlation value of the first feature parameter based on the second gradient. Adjusting the parameters of the shared feature extraction network using the adjustment correlation value of the first feature parameter can cause the second gradient to change in the descent direction. The server can adjust the parameters of the shared feature extraction network in the direction that causes the second gradient to decrease, resulting in a shared feature extraction network with adjusted push recognition loss value.

[0115] In this embodiment, the push object recognition network and the shared feature extraction network were adjusted using the push recognition loss value. One loss value adjusted both networks, which improved the efficiency of parameter adjustment.

[0116] In some embodiments, adjusting the parameters of the shared feature extraction network after adjusting the push recognition loss value based on the reference recognition loss value to obtain the reference recognition loss value and the shared feature extraction network after adjusting the push recognition loss value includes:

[0117] Based on the reference recognition loss value, determine the reference parameter adjustment correlation value corresponding to the reference object recognition network; adjust the parameters of the reference object recognition network based on the reference parameter adjustment correlation value to obtain the parameter-adjusted reference object recognition network; based on the reference recognition loss value and the adjustment direction of the second loss value, determine the second feature parameter adjustment correlation value corresponding to the shared feature extraction network after the push recognition loss value adjustment; adjust the parameters of the shared feature extraction network after the push recognition loss value adjustment based on the second feature parameter adjustment correlation value to obtain the reference recognition loss value and the shared feature extraction network after the push recognition loss value adjustment.

[0118] Among them, the reference parameter adjustment related value is a numerical value used to adjust the parameters of the reference object recognition network, the reference recognition loss value corresponds to the reference recognition loss function, and the reference recognition loss value is the specific value of the reference recognition loss function. The direction of the second loss value adjustment can be the direction of gradient descent. The second feature parameter adjustment related value refers to a numerical value used to adjust the parameters of the shared feature extraction network.

[0119] Specifically, the server can adjust the parameters in the shared feature extraction network in the direction of reducing the reference recognition loss value, and adjust the parameters in the reference object recognition network in the direction of reducing the reference recognition loss value, to obtain the adjusted shared feature extraction network and the adjusted reference object recognition network. Based on the adjusted shared feature extraction network and the adjusted reference object recognition network, a trained object recognition model is obtained.

[0120] In some embodiments, the server can use the reference recognition loss function to perform gradient calculation on the parameters of the reference object recognition network to obtain the third gradient, and determine the reference parameter adjustment correlation value based on the third gradient. In this case, adjusting the parameters of the reference object recognition network using the reference parameter adjustment correlation value can make the third gradient change in the descent direction.

[0121] In some embodiments, after adjusting the shared feature extraction network using the push recognition loss value, the server can perform gradient calculation on the parameters of the shared feature extraction network using the reference recognition loss value to obtain a fourth gradient. The adjustment direction of the second loss value is the descent direction of the fourth gradient. The server can determine the adjustment correlation value of the second feature parameter based on the fourth gradient. Adjusting the parameters of the shared feature extraction network using the adjustment correlation value of the second feature parameter can cause the fourth gradient to change in the descent direction. The server can adjust the parameters of the shared feature extraction network in the direction that causes the fourth gradient to decrease, obtaining the reference recognition loss value and the shared feature extraction network adjusted by the push recognition loss value.

[0122] In this embodiment, the reference object recognition network and the shared feature extraction network were adjusted using a reference recognition loss value. One loss value was used to adjust the two networks, which improved the efficiency of parameter adjustment.

[0123] In some embodiments, the push object identification network includes a conversion probability identification network and a probability decay factor identification network. The process involves inputting the sample extraction features (target sample type is push sample type) into the push object identification network of the object identification model for identification, and identifying the push object content conversion degree corresponding to the content push object. This includes: inputting the sample extraction features (target sample type is push sample type) into the conversion probability identification network of the push object identification network for conversion probability identification to obtain the total conversion probability; inputting the sample extraction features (target sample type is push sample type) into the probability decay factor identification network of the push object identification network for decay factor identification to obtain the probability decay factor; and determining the push object content conversion degree of the content push object based on the total conversion probability and the probability decay factor.

[0124] Specifically, the server can input the extracted features from the samples into the probability decay factor recognition network, and process the extracted features using the network parameters of the probability decay factor recognition network to obtain the probability decay factor of the content push object. The probability decay factor can be, for example,... Among them, w d Identify network parameters for the probability decay factor of the network.

[0125] In some embodiments, the training samples for the content push objects are positive samples. The server can obtain the conversion occurrence duration, which refers to the length of time between the content push time and the conversion occurrence time. The conversion occurrence time refers to the time when the object undergoes a conversion. The server can calculate the sub-conversion probability of the content push object based on the observation duration and the probability decay factor. Based on the sub-conversion probability and the total conversion probability of the content push object, the server can calculate the content conversion degree of the content push object. For example, the server can multiply the sub-conversion probability and the total conversion probability, and use the result of the product as the content conversion degree of the push object.

[0126] For example, if the server pushes content to the target audience on January 1st, and the reference time for dividing positive and negative samples is January 10th, if the target audience converts on January 6th, then the target audience is a positive sample. If the sub-conversion probability of the target audience on January 10th is 0.6, and the total conversion probability of the target audience is 0.8, then the conversion probability of the target audience is 0.6 × 0.8 = 0.48.

[0127] In some embodiments, the training samples for the content push object are negative samples. The server can obtain the observation duration, which refers to the length of time corresponding to the observation period. The server can calculate the probability adjustment value of the content push object based on the observation duration, the total conversion probability of the content push object, and the probability decay factor. Based on the probability adjustment value of the content push object and the total conversion probability, the server can calculate the content conversion degree of the push object.

[0128] In some embodiments, the push object recognition network can be implemented using a delayed feedback model (DFM). A reference object recognition network can also employ a delayed feedback model. For example... Figure 3 The diagram shows the model structure of the delayed feedback model, which includes an embedding layer, a concatenation layer, a conversion rate layer (Cvr Layer), and a parameter layer (Lambda Layer). Figure 4 The diagram shown is a structural diagram of an object recognition model in some embodiments. The object recognition model includes a shared feature extraction network, a push object recognition network, and a reference object recognition network. The shared feature extraction network includes an embedding layer, an aggregation layer, and a feature extraction layer. The push object recognition network includes a first conversion probability recognition network and a first probability decay factor recognition network. The reference object recognition network includes a second conversion probability recognition network and a second probability decay factor recognition network.

[0129] In this embodiment, the conversion rate of the pushed object content is obtained by using the probability decay factor and the total conversion probability. Since the probability decay factor can reflect the decay of the probability of conversion of the object over time, the calculated conversion rate of the pushed object content also reflects the decay of the probability of conversion over time, thus improving the accuracy of the conversion rate of the pushed object content.

[0130] In some embodiments, determining the content conversion degree of a content push object based on the total conversion probability and the probability decay factor includes: obtaining the probability decay duration; determining the sub-conversion probability of the content push object based on the probability decay duration and the probability decay factor, wherein the sub-conversion probability reflects the probability of conversion of the object corresponding to the time interval between the content push time and the probability decay duration; and determining the content conversion degree of the content push object based on the sub-conversion probability and the total conversion probability.

[0131] The probability decay duration can be either the observation duration or the conversion occurrence duration. When the training sample is a positive sample, the probability decay duration is the conversion occurrence duration; when the training sample is a negative sample, the probability decay duration is the observation duration.

[0132] Specifically, when the training samples are positive samples, the server can obtain the conversion occurrence duration corresponding to the content push object, and use the conversion occurrence duration as the probability decay duration. Based on the conversion occurrence duration and the probability decay factor, the server can determine the sub-conversion probability of the content push object. The sub-conversion probability represents the probability that the content push object will convert within a time interval equal to the conversion occurrence duration between the content push time and the content push time. For example, if the content push time is January 1st and the conversion occurrence duration is 4 days, then the sub-conversion probability can represent the probability that the content push object will convert on January 5th. The server can multiply the sub-conversion probability of the content push object by the total conversion probability, and use the result as the content conversion degree of the content push object. In some embodiments, when the training sample is a negative sample, the server can obtain the observation duration and use it as the probability decay duration. For objects that have not been converted by the reference time for dividing positive and negative samples, i.e., for negative samples, there are two possibilities: one is that the object will not be converted in the end, i.e., C=0; the other is that the object will be converted in the end, i.e., C=1. However, since the observation duration is less than the conversion time of the object, no conversion has been found by the observation date. Here, the observation date refers to the reference time for dividing positive and negative samples. Therefore, when the training sample is a negative sample, the probability of the content push object being converted on the observation date, i.e., the content conversion degree of the content push object, can be represented by formula (4).

[0133] In this embodiment, the calculated conversion rate of the push object content also reflects the decay of the probability of conversion over time, thus improving the accuracy of the conversion rate of the push object content.

[0134] In some embodiments, obtaining the training sample set corresponding to the object recognition model to be trained includes: obtaining training push content; dividing the training object set to obtain a content push object set and a reference object set; pushing the training push content to each content push object in the content push object set, and obtaining the content push time of the training push content, wherein the reference objects in the reference object set block the automatic push of training push content; obtaining the backward operation behavior record of the content push object for the training push content; determining the content conversion result of the content push object in the content push object based on the backward operation behavior record; for the reference object in the reference object set, obtaining the content conversion result of the reference object after the content push time; taking the training sample corresponding to the object whose content conversion result is converted as the positive sample in the training sample set, and taking the training sample corresponding to the object whose content conversion result is not converted as the negative sample in the training sample set.

[0135] The training push content refers to the content pushed to the content push recipients. The training recipient set includes multiple objects, which can be users on the same platform, such as users on the same shopping platform. The content push time refers to the time when the training push content is pushed to the content push recipients. The subsequent action records for the training push content refer to the actions of the content push recipients towards the training push content after the content push time and before the positive / negative sample partitioning reference time (which may include the positive / negative sample partitioning reference time). Examples include clicking on the training push content or purchasing products related to the training push content. The content conversion result can be either converted or not converted.

[0136] Specifically, the server can divide the training object set into a content push object set and a reference object set. The number of objects in the content push object set can be the same as or different from the number of objects in the reference object set; for example, the number of objects in the content push object set can be less than the number of objects in the reference object set. The server proactively pushes training content to the content push objects but not to the reference objects. The server can determine whether the subsequent operation behavior record includes a conversion behavior. If a conversion behavior is included, the content push object's conversion result is determined to be converted; if no conversion behavior is included, the content push object's conversion result is determined to be unconverted. The conversion behavior is used to determine whether a conversion has occurred. For example, if a user makes a purchase, it is determined that the user has converted, and the conversion behavior can be "purchase".

[0137] In some embodiments, the server can obtain the object behavior of the reference object after the content push time. When the object behavior includes conversion behavior, the content conversion result of the reference object is determined to be converted. When the object behavior does not include conversion behavior, the content conversion result of the reference object is determined to be unconverted. For example, when the training push content is an advertisement for computer A, if the user buys computer A, it means that the user has converted. Therefore, when the user buys computer A, the content conversion result of the user is determined to be converted.

[0138] In some embodiments, the server may use training samples corresponding to objects whose content conversion result is "converted" before the positive / negative sample division reference time as positive samples in the training sample set, and training samples corresponding to objects whose content conversion result is "unconverted" before the positive / negative sample division reference time as negative samples in the training sample set. The positive / negative sample division reference time is after the content push.

[0139] In this embodiment, positive and negative samples are determined by the user's actions after the content is pushed to them, which improves the efficiency and accuracy of obtaining training samples.

[0140] In some embodiments, such as Figure 5 As shown, an object recognition method is provided, which uses a trained object recognition model generated in the above embodiments to identify objects. This method is applied to... Figure 1 Taking server 104 as an example, the following steps are included:

[0141] S502, obtain the set of candidate objects corresponding to the target push content; the set of candidate objects includes multiple candidate objects.

[0142] The target push content can be any content, either pre-set or determined as needed, such as coupons. The target push content can be the same as or different from the training push content. The candidate object set can be pre-stored on the server. Content push requests can be automatically triggered periodically by the server or retrieved from other devices.

[0143] Specifically, in response to a content push request targeting specific content, the server can obtain a set of candidate objects. The content push request is used to request the push of the target content to the object. The server can determine the push object from the candidate object set and push the target content to the push object. The push object includes marketing-sensitive objects, that is, objects that will not convert without content push but will convert with content push.

[0144] S504, input the object sample of the candidate object into the shared feature extraction network of the trained object recognition model for feature extraction, and obtain the object extraction features corresponding to the candidate object.

[0145] Here, object samples refer to the samples corresponding to candidate objects. The server can generate object samples based on the object attribute information of the candidate objects. Object extraction features are the features obtained by extracting features from object samples using a shared feature extraction network.

[0146] S506, the extracted features of the object are input into the push object recognition network in the object recognition model for recognition, and the first content conversion probability corresponding to the candidate object is obtained based on the total conversion probability obtained from the recognition.

[0147] Here, the first content conversion probability represents the probability that a candidate object will convert within the target probability decay time, and the first content conversion probability is calculated based on the results obtained by the push object recognition network through feature extraction of the object. The target probability decay time can be set as needed, for example, it can be 10 days.

[0148] Specifically, the push object identification network includes a first conversion probability identification network and a first probability decay factor identification network. The server can use the first conversion probability identification network to process the extracted features of the object to obtain the first total conversion probability of the candidate object. The server can use the first probability decay factor identification network to process the extracted features of the object to obtain the first probability decay factor of the candidate object. The server can obtain the target probability decay time and calculate the first content conversion probability corresponding to the candidate object based on the target probability decay time, the first probability decay factor, and the first total conversion probability. For example, the server can use formula (8) to calculate the first content conversion probability corresponding to the candidate object.

[0149]

[0150] Where d represents the target probability decay time, for example, 5 days; k is the time within the target probability decay time, for example, k=1 represents the first day of 5 days; pcvr represents the first total conversion probability; λ(X) i X represents the first probability decay factor. i This represents the extracted features of the object. pCVR_d represents the first content conversion probability.

[0151] S508, the extracted features of the object are input into the reference object recognition network in the object recognition model for recognition, and the second content conversion probability corresponding to the candidate object is obtained based on the conversion probability obtained from the recognition.

[0152] The second content conversion probability represents the probability that the candidate object will be converted within the target probability decay time, and the second content conversion probability is calculated based on the result of the object feature extraction and recognition by the reference object recognition network.

[0153] Specifically, the reference object recognition network includes a second conversion probability recognition network and a second probability decay factor recognition network. The server can use the second conversion probability recognition network to process the extracted features of the object to obtain the second total conversion probability of the candidate object, and use the second probability decay factor recognition network to process the extracted features of the object to obtain the second probability decay factor of the candidate object. The server can obtain the target probability decay time, and calculate the second content conversion probability of the candidate object based on the target probability decay time, the second probability decay factor, and the second total conversion probability. For example, the server can use formula (8) to calculate the second content conversion probability of the candidate object.

[0154] S510, obtain the difference in probability between the first content conversion probability and the second content conversion probability corresponding to the candidate object.

[0155] The probability difference refers to the difference between the probability of conversion of the first content and the probability of conversion of the second content. The server can use the result of subtracting the probability of conversion of the second content from the probability of conversion of the first content as the probability difference.

[0156] In some embodiments, the structure diagram of the trained object recognition model is as follows: Figure 4 As shown, Figure 4 As shown, the server inputs the extracted object features into the first conversion probability recognition network to obtain the first total conversion probability, and inputs the extracted object features into the first probability decay factor recognition network to obtain the first probability decay factor. Based on the first total conversion probability, the first probability decay factor, and the target probability decay time, the first content conversion probability is calculated. G is used in the figure. T (Y i |X i The first content conversion probability is represented by ). Similarly, the server inputs the extracted object features into the second conversion probability recognition network to obtain the second total conversion probability, and inputs the extracted object features into the second probability decay factor recognition network to obtain the second probability decay factor. Based on the second total conversion probability, the second probability decay factor, and the target probability decay time, the second content conversion probability is calculated, denoted as G in the figure. C (Y i |X i The first content conversion probability is represented by τ(X). The server calculates the difference between the first and second content conversion probabilities to obtain the probability difference, which is denoted as τ(X) in the figure. i ) = G T (Y i |X i )-G C (Y i |X i ) represents the difference in probability, τ(X) i ) indicates the degree of difference, X i This indicates that features are extracted from the object.

[0157] S512, Based on the probability difference corresponding to the candidate object, select the candidate object that meets the difference condition from the candidate object set, and use it as the push object corresponding to the target push content; the difference condition includes at least one of the following: the probability difference is ranked before the preset ranking or the probability difference is greater than the difference threshold.

[0158] Here, ranking refers to the order of candidate objects within a candidate object sequence. The candidate object sequence is obtained by arranging the candidate objects in the candidate object set in descending order of their probability differences. The greater the probability difference, the higher the candidate object ranks in the candidate object sequence. The preset ranking can be set as needed or pre-defined, for example, it can be 10. The difference threshold can be preset or set as needed.

[0159] Specifically, the server can compare the probability differences of candidate objects with a difference threshold. If the probability difference is greater than the difference threshold, the candidate object is selected as the push target. Alternatively, the server can select candidate objects from the candidate object sequence that are ranked before a preset order as push targets. The server can then push the target content to the corresponding terminal.

[0160] In the above object recognition method, a candidate object set corresponding to the target push content is obtained. The candidate object set includes multiple candidate objects. The object samples of the candidate objects are input into the shared feature extraction network of the trained object recognition model for feature extraction to obtain the object extraction features corresponding to the candidate objects. The object extraction features are input into the push object recognition network in the object recognition model for recognition. Based on the recognition conversion probability, the first content conversion probability corresponding to the candidate object is obtained. The object extraction features are input into the reference object recognition network in the object recognition model for recognition. Based on the recognition conversion probability, the second content conversion probability corresponding to the candidate object is obtained. The probability difference between the first content conversion probability and the second content conversion probability corresponding to the candidate object is obtained. Based on the probability difference, the candidate objects that meet the difference conditions are selected from the candidate object set as the push objects corresponding to the target push content. The difference conditions include at least one of the following: the probability difference ranking is before the preset ranking or the probability difference is greater than the difference threshold. Thus, the push objects corresponding to the target push content are quickly and conveniently identified using the trained object recognition model, improving the accuracy of object recognition.

[0161] In some embodiments, the step of obtaining a trained object recognition model includes: obtaining a training sample set corresponding to the object recognition model to be trained, the training sample set including training samples corresponding to content push objects and training samples corresponding to reference objects; inputting the training samples in the training sample set into the shared feature extraction network in the object recognition model for feature extraction to obtain sample extraction features corresponding to the training samples; determining the target sample type corresponding to the sample extraction features; inputting the sample extraction features of the target sample type being push sample type into the push object recognition network in the object recognition model for recognition to obtain the push object content conversion degree corresponding to the content push object; inputting the sample extraction features of the target sample type being reference sample type into the reference object recognition network in the object recognition model for recognition to obtain the reference object content conversion degree corresponding to the reference object; obtaining a push recognition loss value based on the push object content conversion degree corresponding to the content push object, and obtaining a reference recognition loss value based on the reference object content conversion degree corresponding to the reference object; adjusting the parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining the trained object recognition model based on the adjusted shared feature extraction network.

[0162] In this embodiment, since the sample extraction features corresponding to the training samples of the content push object are extracted by a shared feature extraction network, and the sample extraction features corresponding to the training samples of the reference object are also extracted by a shared feature extraction network, the training samples of the content push object and the training samples of the reference object share a shared feature extraction network. This allows the shared feature extraction network to accurately identify samples even when the number of samples is unbalanced, improving the sample identification accuracy and thus improving the accuracy of the object recognition model. Since the object recognition model is used for object recognition, the accuracy of object recognition is improved.

[0163] This application also provides an application scenario in which the above-described object recognition method is applied. Specifically, the object recognition method is applied in this scenario as follows:

[0164] 1. Identify the target ads to be pushed.

[0165] 2. Obtain the candidate object set; the candidate object set includes multiple candidate objects.

[0166] 3. Obtain the trained object recognition model, which includes a shared feature extraction network, a push object recognition network, and a reference object recognition network. The push object recognition network includes a first transformation probability recognition network and a first probability decay factor recognition network, and the reference object recognition network includes a second transformation probability recognition network and a second probability decay factor recognition network.

[0167] 4. Input the object samples corresponding to the candidate objects into the shared feature extraction network for feature extraction to obtain the object extraction features corresponding to the candidate objects.

[0168] 5. Input the extracted features of the object into the first conversion probability recognition network to obtain the first total conversion probability corresponding to the candidate object. Input the extracted features of the object into the first probability decay factor recognition network to obtain the first probability decay factor corresponding to the candidate object.

[0169] 6. Input the extracted features of the object into the second conversion probability recognition network to obtain the second total conversion probability corresponding to the candidate object. Input the extracted features of the object into the second probability decay factor recognition network to obtain the second probability decay factor corresponding to the candidate object.

[0170] 7. Obtain the target probability decay time. Based on the first total conversion probability, the first probability decay factor, and the target probability decay time, obtain the first content conversion probability corresponding to the candidate object.

[0171] 8. Based on the second total conversion probability, the second probability decay factor, and the target probability decay time, the second content conversion probability corresponding to the candidate object is obtained.

[0172] 9. Obtain the difference in probability between the first content conversion probability and the second content conversion probability corresponding to the candidate object.

[0173] 10. Based on the probability difference of the candidate objects, select the candidate objects that meet the difference conditions from the candidate object set and use them as the push objects corresponding to the target advertisement.

[0174] The difference conditions include at least one of the following: the energy difference is ranked before the preset ranking or the probability difference is greater than the difference threshold.

[0175] 11. Push the target advertisement to the terminal corresponding to the target audience.

[0176] Traditionally, uplift models can be used to screen for marketing-sensitive users. The modeling method for uplift models differs from that of traditional response models. Uplift models predict the probability of a user converting under a certain intervention (treatment). Uplift models can screen for marketing-sensitive users by calculating a gain value, which is the difference between the probability of a user converting under intervention and the probability of a user converting without intervention.

[0177] Modeling methods for gain models, such as τ i =Y i (1)-Yi (0)(9), where i represents a user in a user group consisting of N users, Y i (1) Y represents the probability that user i will convert after intervention, such as the probability that user i will purchase a product after a coupon is issued to user i. i (0) represents the probability of user i converting without intervention, such as the probability of user i purchasing a product without receiving a coupon. τ i τ represents the difference between the conversion probabilities before and after intervention, specifically the difference between the probability of a user converting with intervention and the probability of a user converting without intervention. i This represents the gain value, also known as the increment or causal effect of user i, which reflects the increase in conversion probability when there is intervention compared to when there is no intervention.

[0178] Since user i cannot simultaneously be in an interventionist and a controlled state—for example, a treatment agent can either send coupons to user i or not send them—it cannot both send coupons to and not send them. Therefore, the conditional average treatment effect (CATE) can be used to characterize group variability. This means using the average causal effect of a group to represent the causal effect of an individual. The formula for the conditional average treatment effect is as follows: CATE:τ(X i )=E[Y i (1)|X i ]-E[Y i (0)|X i (10).

[0179] Since whether or not an action is taken against the same user in the same event is a mutually exclusive event, the observed Y can be... i Rewritten as formula Y i obs =W i Y i (1)+(1-W i )Y i (0)(11), when user feature X i With intervention condition W i When the causal effects are independent of each other (Conditional Independence Assumption, CIA), the observed differences in causal effects can be expressed as Equation (12).

[0180] τ(X i )=E[Y iobs |X i =x,W=1]-E[Y i obs |X i =x,W=0] (12)

[0181] In some embodiments, the modeling method for the gain model may include at least one of the following: a two-model uplift approach, a one-model uplift approach, or a transformation approach. The two-model approach includes an intervention group model and a control group model. For example... Figure 6 The diagram shows the flowchart for training the two models. During training, the intervention group model and the control group model use independent samples. The intervention group model is trained using samples from the intervention group, and the control group model is trained using samples from the control group. The intervention group samples are those corresponding to users who received the intervention, and the control group samples are those corresponding to users who did not receive the intervention. Figure 7 The diagram illustrates the flowchart for gain value prediction using two trained models. During prediction, the intervention group model is used to predict the conversion score of the prediction group samples, yielding the intervention conversion score. The control group model is used to predict the conversion score of the prediction group samples, yielding the control conversion score. The difference between the intervention conversion score and the control conversion score is calculated, and this difference is used as the conversion gain score corresponding to the prediction group sample. The intervention group model can be represented by G... T This indicates that the control group model can use G... C This indicates that the results obtained from the intervention group model can be represented by G. T (Y i |X i The result obtained from the control group model can be represented by G. C (Y i |X i The conversion gain fraction (i.e., the gain value) can be represented as in the two-model approach. τ(X i ) = G T (Y i |X i )-G C (Y i |X i (13). The control group model and the intervention group model can be collectively referred to as the base model G. The base model G can be implemented using a classification or regression model, such as a multilayer neural network based on cross-entropy loss, to predict user conversion rates and other predicted values ​​as a ranking criterion. The intervention group can also be referred to as the experimental group.

[0182] In the two-model approach, the intervention group model and the control group model are trained using independent samples, which may lead to the accumulation of scoring errors. Therefore, a single-model approach is proposed, in which the intervention status is incorporated as a feature into the training process. Specifically, in the single-model approach, the original user feature X is modified along the sample dimension. i To expand the model, intervention-related features T are introduced, such as T=1 representing the intervention group feature and T=0 representing the control group feature. The single-model structure is as follows: Figure 8 As shown, the output of a single model can be represented by G(Y). i |X i ,T) indicates that, in a single model, the conversion gain fraction can be expressed as τ(X i )=G(Y i |X i ,T=1)-G(Y i |X i ,T=0)(14.

[0183] The category transformation method is a single-model method, which can only be applied to binary intervention scenarios (such as whether or not intervention is applied). The category transformation method introduces an intermediate variable Z to optimize τ(X). i The model structure of the category conversion method is as follows: Figure 9 As shown. The intermediate variable Z can be represented by formula (15). That is, if the user is intervened and converted, Z = 1; if the user is not intervened and not converted, Z = 1; otherwise, Z = 0. The output of the single model in the category conversion method can be represented by G(Z). i |X i )express.

[0184]

[0185] When the sample sizes of the intervention group and the control group are the same, that is, when the probability of a user coming from the intervention group is equal to the probability of coming from the control group, the conversion gain score can be expressed by formula (16), that is, learning τ(X i This is equivalent to learning P[Z] i =1|X i Equation (15) can be used to convert the original label Y (Y in the formula) to Z, and model G is used to optimize Z. When making predictions, the predicted P[Z] is used. i =1|X i Simply sort them.

[0186]

[0187] In category transformation methods, the sample sizes of the intervention and control groups need to be exactly equal, which is difficult to achieve in reality. Generally, resampling is used to make the sample sizes of the two groups identical, or user preference scores can be introduced to make the sample sizes identical. Category transformation methods are only applicable to binary cases where there is no intervention; they cannot be used if there are multiple intervention options (such as different amounts of red envelopes).

[0188] Therefore, in traditional gain modeling methods, the samples of the intervention and control groups need to meet the conditional independence assumption. To satisfy this condition, in practice, online randomized A / B experiments are needed to set up intervention and control groups, ensuring that the feature distributions of the two groups are consistent. For example, in an A / B experiment, participants are randomly selected and divided into two groups of 25 each. Group A uses keyboard layout A, and Group B uses keyboard layout B. They are asked to type a standard 20-word text message within 30 seconds, and the number of typos is recorded. However, in financial scenarios, due to low user conversion rates and infrequent user behavior, it is difficult to reserve a large unbiased control group. That is, unbiased control samples are scarce in financial scenarios, user decision-making costs are high, conversion is difficult, and the ratio of positive to negative samples is skewed. This results in the sample size of the control group being much smaller than that of the intervention group. Furthermore, since the control group sample has not undergone intervention, the positive samples rely entirely on natural user conversion, leading to an even more skewed ratio of positive to negative samples in the control group. Therefore, traditional gain modeling methods are not suitable for financial scenarios. The reasons are as follows: For the two-model method, insufficient samples in the control group will lead to insufficient training of the control group model and increased error; for the single-model method, insufficient samples in the control group will lead to a small impact of the intervention feature T on the model's loss value, and the intervention feature T in the model will be difficult to play its expected role, resulting in a small or even no difference in the model output score when T=1 and T=0 during prediction; for the class conversion method, since the class conversion method requires the number of samples in the intervention group to be equal to the number of samples in the control group, it is usually used to make the sample size of the two the same by resampling, or by introducing user preference scores to make the sample size of the two the same. However, the resampling method will make the sample size of the intervention group as scarce as that of the control group, and the introduction of user preference scores and other methods will also lead to the prediction score being mainly contributed by the intervention group, and the class conversion method can only be used in scenarios where the intervention factor is binary (such as whether to send a red envelope).

[0189] In the object recognition method provided in this application, since the push training samples (i.e., intervention group samples) and the reference training samples (i.e., control group samples) share a shared feature extraction network, that is, the same shared feature extraction network is trained through the intervention group samples and the control group samples. Therefore, even if the number of intervention group samples and the number of control group samples are not the same, for example, the number of intervention group samples is greater than the number of control group samples, a network that can distinguish between intervention group samples and control group samples can still be trained. Thus, the object recognition method provided in this application can be applied to scenarios with imbalanced samples, such as in financial scenarios, such as advertising push or coupon distribution in financial scenarios.

[0190] Traditional gain models typically use a base model G based on cross-entropy loss (classification) or square root loss (regression), neglecting the latency of user conversion. However, in financial scenarios, user decisions are heavily influenced by external factors like market conditions, leading to high costs for user subscription decisions. After receiving marketing information or rewards, conversion is often delayed by hours or even days. For example, a user might receive a red envelope on day N, but might not convert until day N+D. Therefore, before conversion, these unconverted samples are not negative samples but rather positive samples (called gray samples) that haven't yet reached their conversion time. Because user conversion in financial scenarios is often delayed, and the collected samples contain noise, traditional gain models, by not considering this latency, lead to inaccurate predictions of user conversion, resulting in inaccurate gain calculations.

[0191] The object recognition method provided in this application considers the impact of time delay on the user's conversion probability, which solves the problem of inaccurate model prediction caused by user delayed conversion, improves the accuracy of conversion probability prediction, and thus improves the accuracy of gain calculation.

[0192] Experiments demonstrated the effectiveness of the object recognition method provided in this application. This method significantly improves performance compared to traditional methods, reduces harassment of non-converting users, and enhances user experience. Specific results are shown in Table 1, where the number of converted users represents the total number of users who converted through the marketing campaign within one week in the intervention group. Overall, the Uplift modeling method, compared to the response model, achieves a higher ROI (return on investment) when the number of converted users remains constant or even slightly increases. Among the Uplift modeling methods, TransUplift and TransDelayUplift outperform other methods.

[0193] Table 1 Experimental Results

[0194] Model Name Number of converted relative improvement ROI relative improvement Response Model 10389 - 1.151 - Two Model Approach 10442 0.51％ 1.185 2.95％ One Model Approach 10428 0.38％ 1.173 1.91％ TransUplift 10505 1.12％ 1.192 3.56％ TransDelayUplift 10536 1.41％ 1.211 5.21％

[0195] It should be understood that, although Figures 2-9 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figures 2-9 At least some of the steps in the process may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0196] In some embodiments, such as Figure 10 As shown, an object recognition device is provided. This device can be a software module, a hardware module, or a combination of both integrated into a computer device. Specifically, the device includes: a training sample set acquisition module 1002, a sample feature extraction module 1004, a target sample type determination module 1006, a push object content conversion degree acquisition module 1008, a reference object content conversion degree acquisition module 1010, a recognition loss value acquisition module 1012, and a trained object recognition model acquisition module 1014, wherein:

[0197] The training sample set acquisition module 1002 is used to acquire the training sample set corresponding to the object recognition model to be trained. The training sample set includes the training samples corresponding to the content push object and the training samples corresponding to the reference object.

[0198] The sample extraction feature acquisition module 1004 is used to input the training samples in the training sample set into the shared feature extraction network in the object recognition model to extract features and obtain the sample extraction features corresponding to the training samples.

[0199] The target sample type determination module 1006 is used to determine the target sample type corresponding to the extracted sample features.

[0200] The module 1008 for obtaining the conversion degree of the push object content is used to input the extracted features of the target sample type (push sample type) into the push object recognition network in the object recognition model for recognition, and to obtain the conversion degree of the push object content corresponding to the content push object.

[0201] The reference object content conversion degree module 1010 is used to input the sample extracted features of the target sample type as the reference sample type into the reference object recognition network in the object recognition model for recognition, and to obtain the reference object content conversion degree corresponding to the reference object.

[0202] The identification loss value acquisition module 1012 is used to obtain the push identification loss value based on the content conversion degree of the push object corresponding to the content push object, and to obtain the reference identification loss value based on the content conversion degree of the reference object corresponding to the reference object.

[0203] The trained object recognition model is obtained by module 1014, which is used to adjust the parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtain the trained object recognition model based on the adjusted shared feature extraction network.

[0204] In some embodiments, the target sample type determination module includes: a sample label value acquisition unit, used to acquire the sample label value of the training sample corresponding to the sample extraction feature; and a target sample type determination unit, used to determine the target sample type corresponding to the sample extraction feature based on the sample label value.

[0205] In some embodiments, the conversion degree of the push object content is used to represent the conversion probability of the content push object for the push content. The identification loss value acquisition module includes: an object identification loss value acquisition unit, which is used to multiply the sample label value and the conversion degree of the push object content to obtain the object identification loss value corresponding to the content push object; and a push identification loss value acquisition unit, which is used to statistically analyze the object identification loss values ​​of multiple content push objects to obtain the push identification loss value.

[0206] In some embodiments, the trained object recognition model acquisition module includes: a first parameter adjustment unit, configured to adjust the parameters of the shared feature extraction network based on the push recognition loss value, to obtain the shared feature extraction network after adjusting the push recognition loss value; a second parameter adjustment unit, configured to adjust the parameters of the shared feature extraction network after adjusting the push recognition loss value based on the reference recognition loss value, to obtain the reference recognition loss value and the shared feature extraction network after adjusting the push recognition loss value; and a trained object recognition model acquisition unit, configured to obtain the trained object recognition model based on the reference recognition loss value and the shared feature extraction network after adjusting the push recognition loss value.

[0207] In some embodiments, the first parameter adjustment unit is further configured to: determine the push parameter adjustment related value corresponding to the push object recognition network based on the push recognition loss value; adjust the parameters of the push object recognition network based on the push parameter adjustment related value to obtain the parameter-adjusted push object recognition network; determine the first feature parameter adjustment related value corresponding to the shared feature extraction network based on the push recognition loss value and the first loss value adjustment direction; and adjust the parameters of the shared feature extraction network based on the first feature parameter adjustment related value to obtain the shared feature extraction network after the push recognition loss value is adjusted.

[0208] In some embodiments, the second parameter adjustment unit is further configured to: determine reference parameter adjustment related values ​​for the reference object recognition network based on the reference recognition loss value; adjust the parameters of the reference object recognition network based on the reference parameter adjustment related values ​​to obtain the parameter-adjusted reference object recognition network; determine the second feature parameter adjustment related values ​​for the shared feature extraction network after the push recognition loss value is adjusted based on the reference recognition loss value and the second loss value adjustment direction; and adjust the parameters of the shared feature extraction network after the push recognition loss value is adjusted based on the second feature parameter adjustment related values ​​to obtain the reference recognition loss value and the shared feature extraction network after the push recognition loss value is adjusted.

[0209] In some embodiments, the push object identification network includes a conversion probability identification network and a probability decay factor identification network; the push object content conversion degree obtaining module includes: a total conversion probability obtaining unit, used to extract features from samples whose target sample type is push sample type, input them into the conversion probability identification network in the push object identification network for conversion probability identification, and obtain the total conversion probability; a probability decay factor obtaining unit, used to extract features from samples whose target sample type is push sample type, input them into the probability decay factor identification network in the push object identification network for decay factor identification, and obtain the probability decay factor; and a push object content conversion degree determining unit, used to determine the push object content conversion degree of the content push object based on the total conversion probability and the probability decay factor.

[0210] In some embodiments, the push object content conversion degree determination unit is further configured to obtain the probability decay duration; determine the sub-conversion probability degree of the content push object based on the probability decay duration and the probability decay factor, wherein the sub-conversion probability degree is used to reflect the probability degree of conversion of the object corresponding to the time interval between the content push time and the probability decay duration; and determine the push object content conversion degree of the content push object based on the sub-conversion probability degree and the total conversion probability degree.

[0211] In some embodiments, the training sample set acquisition module includes: a training push content acquisition unit for acquiring training push content; an object partitioning unit for partitioning the training object set to obtain a content push object set and a reference object set; a content push unit for pushing the training push content to each content push object in the content push object set and acquiring the content push time of the training push content, wherein the reference objects in the reference object set are shielded from automatic push of training push content; a backward operation behavior record acquisition unit for acquiring the backward operation behavior record of the content push object in response to the training push content; a content conversion result determination unit for determining the content conversion result of the content push object in the content push object based on the backward operation behavior record; a content conversion result acquisition unit for acquiring the content conversion result of the reference object in the reference object set after the content push time; and a positive and negative sample acquisition unit for taking the training samples corresponding to the objects whose content conversion result is converted as positive samples in the training sample set and taking the training samples corresponding to the objects whose content conversion result is not converted as negative samples in the training sample set.

[0212] In some embodiments, such as Figure 11 As shown, an object recognition device is provided. This device can be a software module, a hardware module, or a combination of both integrated into a computer device. Specifically, the device includes: a candidate object set acquisition module 1102, an object feature extraction module 1104, a first content conversion probability acquisition module 1106, a second content conversion probability acquisition module 1108, a probability difference acquisition module 1110, and a push object determination module 1112, wherein:

[0213] The candidate object set acquisition module 1102 is used to acquire the candidate object set corresponding to the target push content; the candidate object set includes multiple candidate objects;

[0214] The object extraction feature acquisition module 1104 is used to input the object sample of the candidate object into the shared feature extraction network of the trained object recognition model for feature extraction, and obtain the object extraction features corresponding to the candidate object;

[0215] The first content conversion probability module 1106 is used to input the extracted features of the object into the push object recognition network in the object recognition model for recognition, and obtain the first content conversion probability corresponding to the candidate object based on the conversion probability obtained from the recognition.

[0216] The second content conversion probability module 1108 is used to input the extracted features of the object into the reference object recognition network in the object recognition model for recognition, and to obtain the second content conversion probability corresponding to the candidate object based on the conversion probability obtained from the recognition.

[0217] The probability difference acquisition module 1110 is used to obtain the probability difference between the first content conversion probability and the second content conversion probability corresponding to the candidate object;

[0218] The push target determination module 1112 is used to filter out candidate objects that meet the difference conditions from the candidate object set based on the probability difference of the candidate objects, and use them as the push objects corresponding to the target push content; the difference conditions include at least one of the following: the probability difference is ranked before the preset ranking or the probability difference is greater than the difference threshold.

[0219] In some embodiments, the apparatus further includes an object recognition model training module, which includes: a training sample set acquisition module, configured to acquire a training sample set corresponding to the object recognition model to be trained, the training sample set including training samples corresponding to content push objects and training samples corresponding to reference objects; a sample feature extraction unit, configured to input the training samples in the training sample set into the shared feature extraction network in the object recognition model for feature extraction to obtain sample extraction features corresponding to the training samples; a target sample type determination unit, configured to determine the target sample type corresponding to the sample extraction features; and a push object content conversion degree acquisition unit, configured to input the sample extraction features whose target sample type is the push sample type into the push object recognition network in the object recognition model. The object recognition model is divided into several units: a reference object content conversion degree unit, a reference object content conversion degree unit, and a training object recognition model unit. The former is used to identify the content push object in the network and obtain the content conversion degree of the corresponding push object. The latter is used to extract features from samples of the target sample type (reference sample type) and input them into the reference object recognition network in the object recognition model to obtain the reference object content conversion degree. The latter is used to obtain the push recognition loss value based on the content push object content conversion degree and the reference recognition loss value. The former is used to adjust the parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtain the trained object recognition model based on the adjusted shared feature extraction network.

[0220] For specific limitations regarding the object recognition device, please refer to the limitations of the object recognition method above, which will not be repeated here. Each module in the aforementioned object recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0221] In some embodiments, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 12 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements an object recognition method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0222] In some embodiments, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 13 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data involved in the object recognition method. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements an object recognition method.

[0223] Those skilled in the art will understand that Figure 12 and Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0224] In some embodiments, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0225] In some embodiments, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0226] In some embodiments, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above-described method embodiments.

[0227] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0228] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0229] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method of object recognition, characterized by, The method comprises: obtaining a training sample set corresponding to an object recognition model to be trained, the training sample set comprising training samples corresponding to content push objects and training samples corresponding to reference objects; inputting the training samples in the training sample set into a shared feature extraction network in the object recognition model respectively for feature extraction, to obtain sample extraction features corresponding to the training samples; determining target sample types corresponding to the sample extraction features; inputting sample extraction features of a target sample type being a push sample type into a push object recognition network in the object recognition model for recognition, to obtain push object content conversion degrees corresponding to the content push objects; inputting sample extraction features of a target sample type being a reference sample type into a reference object recognition network in the object recognition model for recognition, to obtain reference object content conversion degrees corresponding to the reference objects; obtaining a push recognition loss value based on the push object content conversion degrees corresponding to the content push objects, and obtaining a reference recognition loss value based on the reference object content conversion degrees corresponding to the reference objects; adjusting parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining a trained object recognition model based on the adjusted shared feature extraction network, the trained object recognition model being used to recognize marketing sensitive objects, to push content to the marketing sensitive objects, and to encourage the marketing sensitive objects to convert.

2. The method of claim 1, wherein, The determination of the target sample types corresponding to the sample extraction features comprises: obtaining sample label values of the training samples corresponding to the sample extraction features; determining the target sample types corresponding to the sample extraction features based on the sample label values.

3. The method of claim 2, wherein, The push object content conversion degrees are used to represent conversion probabilities of the content push objects for push content, and the obtaining of the push recognition loss value based on the push object content conversion degrees corresponding to the content push objects comprises: multiplying the sample label values and the push object content conversion degrees to obtain object recognition loss values corresponding to the content push objects; statistically obtaining push recognition loss values based on the object recognition loss values of the content push objects.

4. The method of claim 1, wherein, The adjustment of the parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and the obtaining of the trained object recognition model based on the adjusted shared feature extraction network comprise: adjusting the parameters of the shared feature extraction network based on the push recognition loss value, to obtain a shared feature extraction network adjusted based on a push recognition loss value; adjusting the parameters of the shared feature extraction network adjusted based on the push recognition loss value based on the reference recognition loss value, to obtain a shared feature extraction network adjusted based on a reference recognition loss value and a push recognition loss value; obtaining the trained object recognition model based on the shared feature extraction network adjusted based on the reference recognition loss value and the push recognition loss value.

5. The method of claim 4, wherein, The adjustment of the parameters of the shared feature extraction network based on the push recognition loss value, to obtain a shared feature extraction network adjusted based on a push recognition loss value comprises: determine a push parameter adjustment related value corresponding to the push object recognition network based on the push recognition loss value; perform parameter adjustment on the push object recognition network based on the push parameter adjustment related value, to obtain a push object recognition network after parameter adjustment; determine a first feature parameter adjustment related value corresponding to the shared feature extraction network based on the push recognition loss value and the first loss value adjustment direction; perform parameter adjustment on the shared feature extraction network based on the first feature parameter adjustment related value, to obtain a shared feature extraction network after push recognition loss value adjustment.

6. The method of claim 4, wherein, The parameter adjustment on the shared feature extraction network after push recognition loss value adjustment based on the reference recognition loss value comprises: determine a reference parameter adjustment related value corresponding to the reference object recognition network based on the reference recognition loss value; perform parameter adjustment on the reference object recognition network based on the reference parameter adjustment related value, to obtain a reference object recognition network after parameter adjustment; determine a second feature parameter adjustment related value corresponding to the shared feature extraction network after push recognition loss value adjustment based on the reference recognition loss value and the second loss value adjustment direction; perform parameter adjustment on the shared feature extraction network after push recognition loss value adjustment based on the second feature parameter adjustment related value, to obtain a shared feature extraction network after reference recognition loss value and push recognition loss value adjustment.

7. The method of claim 1, wherein, The push object recognition network comprises a conversion possibility recognition network and a possibility decay factor recognition network; the identification of the sample feature of the sample of the target sample type being the push sample type in the push object recognition network in the object recognition model comprises: input the sample feature of the sample of the target sample type being the push sample type into the conversion possibility recognition network in the push object recognition network for conversion possibility recognition, to obtain a total conversion possibility; input the sample feature of the sample of the target sample type being the push sample type into the possibility decay factor recognition network in the push object recognition network for decay factor recognition, to obtain a possibility decay factor; determine the push object content conversion degree of the content push object based on the total conversion possibility and the possibility decay factor.

8. The method of claim 7, wherein, The determination of the push object content conversion degree of the content push object based on the total conversion possibility and the possibility decay factor comprises: obtain a possibility decay duration; determine a sub-conversion possibility of the content push object based on the possibility decay duration and the possibility decay factor; the sub-conversion possibility is used to reflect the possibility of the object conversion corresponding to the time interval between the content push time and the possibility decay duration; determine the push object content conversion degree of the content push object based on the sub-conversion possibility and the total conversion possibility.

9. The method of claim 1, wherein, The obtaining of the training sample set corresponding to the object recognition model to be trained comprises: obtain training push content; Divide the training object set to obtain a content push object set and a reference object set; Push the training push content to each content push object in the content push object set, and obtain a content push time of the training push content, wherein the reference object in the reference object set shields automatic push of the training push content; Obtain a backward operation behavior record of the content push object for the training push content; Determine a content conversion result of the content push object in the content push object set based on the backward operation behavior record; For the reference object in the reference object set, obtain a content conversion result of the reference object after the content push time; The content conversion result is converted into a training sample corresponding to the object that has been converted as a positive sample in the training sample set, and the content conversion result is converted into a training sample corresponding to the object that has not been converted as a negative sample in the training sample set.

10. An object recognition method characterized by, The method comprises: Obtain a candidate object set corresponding to target push content; the candidate object set includes a plurality of candidate objects; Input the object sample of the candidate object into the shared feature extraction network of the trained object recognition model for feature extraction, to obtain the object extraction feature corresponding to the candidate object; Input the object extraction feature into the push object recognition network in the object recognition model for recognition, and obtain the first content conversion possibility corresponding to the candidate object based on the conversion possibility obtained by recognition; Input the object extraction feature into the reference object recognition network in the object recognition model for recognition, and obtain the second content conversion possibility corresponding to the candidate object based on the conversion possibility obtained by recognition; Obtain the possibility difference between the first content conversion possibility and the second content conversion possibility corresponding to the candidate object; Based on the possibility difference corresponding to the candidate object, the candidate object satisfying the difference condition is screened from the candidate object set as the push object corresponding to the target push content, and the target push content is pushed to the push object; the difference condition includes at least one of the following: the sorting of the possibility difference is before the preset sorting, or the possibility difference is greater than the difference threshold.

11. The method of claim 10, wherein, The step of obtaining the trained object recognition model comprises: Obtain a training sample set corresponding to the object recognition model to be trained, the training sample set including a training sample corresponding to a content push object and a training sample corresponding to a reference object; Input the training sample in the training sample set into the shared feature extraction network in the object recognition model for feature extraction, to obtain the sample extraction feature corresponding to the training sample; Determine the target sample type corresponding to the sample extraction feature; Input the sample extraction feature with the target sample type as the push sample type into the push object recognition network in the object recognition model for recognition, to obtain the push object content conversion degree corresponding to the content push object by recognition; inputting the sample extraction features of the samples of the target sample type into a reference object recognition network in the object recognition model for recognition, to obtain a reference object content conversion degree corresponding to the reference object; obtaining a push recognition loss value based on the push object content conversion degree corresponding to the content push object, and obtaining a reference recognition loss value based on the reference object content conversion degree corresponding to the reference object; adjusting parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtaining the trained object recognition model based on the adjusted shared feature extraction network.

12. An object recognition method apparatus characterized by, The device comprises: a training sample set acquisition module configured to acquire a training sample set corresponding to an object recognition model to be trained, the training sample set comprising training samples corresponding to content push objects and training samples corresponding to reference objects; a sample extraction feature obtaining module configured to input the training samples in the training sample set into a shared feature extraction network in the object recognition model for feature extraction, to obtain sample extraction features corresponding to the training samples; a target sample type determination module configured to determine a target sample type corresponding to the sample extraction features; a push object content conversion degree obtaining module configured to input the sample extraction features of the samples of the target sample type into a push object recognition network in the object recognition model for recognition, to obtain a push object content conversion degree corresponding to the content push object; a reference object content conversion degree obtaining module configured to input the sample extraction features of the samples of the target sample type into a reference object recognition network in the object recognition model for recognition, to obtain a reference object content conversion degree corresponding to the reference object; an identification loss value obtaining module configured to obtain a push recognition loss value based on the push object content conversion degree corresponding to the content push object, and obtain a reference recognition loss value based on the reference object content conversion degree corresponding to the reference object; a trained object recognition model obtaining module configured to adjust parameters of the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtain a trained object recognition model based on the adjusted shared feature extraction network, the trained object recognition model being used to identify marketing sensitive objects, to push content to the marketing sensitive objects, and to encourage the marketing sensitive objects to convert.

13. The apparatus of claim 12, wherein, The target sample type determination module is further configured to: acquire a sample label value of the training sample corresponding to the sample extraction features; determine the target sample type corresponding to the sample extraction features based on the sample label value.

14. The apparatus of claim 13, wherein, The push object content conversion degree is used to represent a conversion possibility of the content push object for push content, and the identification loss value obtaining module is further configured to: multiply the sample label value and the push object content conversion degree to obtain an object recognition loss value corresponding to the content push object; statistically obtain a push recognition loss value based on object recognition loss values of multiple content push objects.

15. The apparatus of claim 12, wherein, The trained object recognition model obtaining module is further configured to: adjust parameters of the shared feature extraction network based on the push recognition loss value, to obtain the shared feature extraction network after adjustment of the push recognition loss value; adjust parameters of the shared feature extraction network after adjustment of the push recognition loss value based on the reference recognition loss value, to obtain the shared feature extraction network after adjustment of the reference recognition loss value and the push recognition loss value; obtain the trained object recognition model based on the shared feature extraction network after adjustment of the reference recognition loss value and the push recognition loss value.

16. The apparatus of claim 15, wherein, The trained object recognition model obtaining module is further configured to: determine a push parameter adjustment related value corresponding to the push object recognition network based on the push recognition loss value; adjust parameters of the push object recognition network based on the push parameter adjustment related value, to obtain the push object recognition network after parameter adjustment; determine a first feature parameter adjustment related value corresponding to the shared feature extraction network based on the push recognition loss value and the first loss value adjustment direction; adjust parameters of the shared feature extraction network based on the first feature parameter adjustment related value, to obtain the shared feature extraction network after adjustment of the push recognition loss value.

17. The apparatus of claim 15, wherein, The trained object recognition model obtaining module is further configured to: determine a reference parameter adjustment related value corresponding to the reference object recognition network based on the reference recognition loss value; adjust parameters of the reference object recognition network based on the reference parameter adjustment related value, to obtain the reference object recognition network after parameter adjustment; determine a second feature parameter adjustment related value corresponding to the shared feature extraction network after adjustment of the push recognition loss value based on the reference recognition loss value and the second loss value adjustment direction; adjust parameters of the shared feature extraction network after adjustment of the push recognition loss value based on the second feature parameter adjustment related value, to obtain the shared feature extraction network after adjustment of the reference recognition loss value and the push recognition loss value.

18. The apparatus of claim 12, wherein, The push object recognition network includes a conversion possibility recognition network and a possibility attenuation factor recognition network; the push object content conversion degree obtaining module is further configured to: extract features of a sample of a target sample type being a push sample type, and input the features into the conversion possibility recognition network in the push object recognition network to perform conversion possibility recognition, to obtain a total conversion possibility; extract features of a sample of a target sample type being a push sample type, and input the features into the possibility attenuation factor recognition network in the push object recognition network to perform attenuation factor recognition, to obtain a possibility attenuation factor; determine the push object content conversion degree of the content push object based on the total conversion possibility and the possibility attenuation factor.

19. The apparatus of claim 18, wherein, The push object content conversion degree obtaining module is further configured to: obtain a possibility attenuation duration; determine a sub-conversion possibility of the content push object based on the possibility attenuation duration and the possibility attenuation factor; the sub-conversion possibility is used to reflect a possibility of conversion of an object corresponding to a time interval between the content push time and the time of the possibility attenuation duration; Based on the sub-conversion possibility and the total conversion possibility, a conversion degree of a push object content of the content push object is determined.

20. The apparatus of claim 12, wherein, The training sample set obtaining module is further configured to: obtain training push content; divide the training object set to obtain a content push object set and a reference object set; push the training push content to each content push object in the content push object set to obtain a content push time of the training push content, wherein a reference object in the reference object set shields automatic pushing of the training push content; obtain a backward operation behavior record of the content push object for the training push content; determine a content conversion result of the content push object in the content push object set based on the backward operation behavior record; for a reference object in the reference object set, obtain a content conversion result of the reference object after the content push time; take a training sample corresponding to a content-converted object as a positive sample in the training sample set, and take a training sample corresponding to a non-converted object as a negative sample in the training sample set.

21. An object recognition apparatus characterized by comprising: The apparatus comprises: a candidate object set obtaining module configured to obtain a candidate object set corresponding to target push content; the candidate object set comprises a plurality of candidate objects; an object extraction feature obtaining module configured to input an object sample of the candidate object into a shared feature extraction network of a trained object recognition model to perform feature extraction, and obtain an object extraction feature corresponding to the candidate object; a first content conversion possibility obtaining module configured to input the object extraction feature into a push object recognition network in the object recognition model to perform recognition, and obtain a first content conversion possibility corresponding to the candidate object based on a conversion possibility obtained by the recognition; a second content conversion possibility obtaining module configured to input the object extraction feature into a reference object recognition network in the object recognition model to perform recognition, and obtain a second content conversion possibility corresponding to the candidate object based on a conversion possibility obtained by the recognition; a possibility difference obtaining module configured to obtain a possibility difference between the first content conversion possibility and the second content conversion possibility corresponding to the candidate object; a push object determining module configured to select a candidate object satisfying a difference condition from the candidate object set based on the possibility difference corresponding to the candidate object, as a push object corresponding to the target push content, and push the target push content to the push object; the difference condition comprises at least one of the following: the sorting of the possibility difference is before a preset sorting, or the possibility difference is greater than a difference threshold.

22. The apparatus of claim 21, wherein, The module for obtaining the trained object recognition model is configured to: obtain a training sample set corresponding to a to-be-trained object recognition model; the training sample set comprises a training sample corresponding to a content push object and a training sample corresponding to a reference object; input the training sample in the training sample set into a shared feature extraction network in the object recognition model to perform feature extraction, and obtain a sample extraction feature corresponding to the training sample. determine a target sample type corresponding to the sample extraction feature; input the sample extraction feature of which the target sample type is the push sample type into a push object recognition network in the object recognition model for recognition, and obtain a push object content conversion degree corresponding to the content push object through the recognition; input the sample extraction feature of which the target sample type is the reference sample type into a reference object recognition network in the object recognition model for recognition, and obtain a reference object content conversion degree corresponding to the reference object through the recognition; obtain a push recognition loss value based on the push object content conversion degree corresponding to the content push object, and obtain a reference recognition loss value based on the reference object content conversion degree corresponding to the reference object; adjust the shared feature extraction network based on the push recognition loss value and the reference recognition loss value, and obtain the trained object recognition model based on the adjusted shared feature extraction network. 23.A computer device, comprising a memory and a processor, wherein the memory stores a computer program. The processor executes the computer program to implement the steps of the method of any one of claims 1 to 11.

24. A computer readable storage medium storing a computer program, wherein the computer program comprises instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1-23. The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 11.

25. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 11. The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 11.

Images