Training method of object classification model and related device

Abstract

This application discloses a training method and related apparatus for an object classification model, which can be used in scenarios such as cloud technology, artificial intelligence, intelligent transportation, and assisted driving to obtain a sample object dataset including a first positive example sample set and an unlabeled sample set. Intermediate samples are removed from the first positive example sample set to obtain a second positive example sample set, and these intermediate samples are added to the unlabeled sample set to obtain the original unlabeled sample set. Based on the object features of the sample objects in the constructed sample object dataset, a neural network model is trained on the second positive example sample set and the original unlabeled sample set to obtain the object classification model. During the training of the object classification model, the trained neural network model is used to classify the sample objects in the original unlabeled sample set until each sample object in the original unlabeled sample set has a classification label, thereby balancing positive and negative samples, ensuring the classification accuracy of the sample objects in the original unlabeled sample set, and improving the business effect of information recommendation.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a training method and related apparatus for an object classification model. Background Technology

[0002] In the operation of various products, it is necessary to promote and advertise the products to continuously attract new users. Generally speaking, users who are likely to be interested in the product should be targeted with product-related information to improve the effectiveness of promotion.

[0003] Therefore, it is necessary to identify user intentions to determine potential customers who may be interested in the product. Current technologies can train object classification models to identify potentially interested users and then recommend information to them.

[0004] However, the object classification models provided by related technologies are not very accurate in identifying whether a user is interested in the product, which reduces the effectiveness of information recommendation services. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides a training method and related apparatus for an object classification model, which effectively balances the initial positive and negative sample samples, enabling it to have better generalization during the classification process, ensuring the classification accuracy of sample objects in the original unlabeled sample set, and further improving the business effect of subsequent information recommendation.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] On one hand, embodiments of this application provide a method for training an object classification model, the method comprising:

[0008] Obtain a sample object dataset for the target product. The sample object dataset includes a first positive example sample set and an unlabeled sample set. The sample objects in the first positive example sample set have positive example labels, which are classification labels used to mark that the sample objects have the intention to perform a conversion operation for the target product. The sample objects in the unlabeled sample set do not have classification labels, which are used to mark whether the sample objects have the intention to perform a conversion operation for the target product.

[0009] Construct the object features of the sample objects in the sample object dataset;

[0010] Multiple sample objects are selected from the first positive sample set as intermediate samples, and the intermediate samples are removed from the first positive sample set to obtain a second positive sample set. The intermediate samples are then added to the unlabeled sample set to obtain the original unlabeled sample set.

[0011] Based on the object characteristics, a neural network model is trained on the second positive example sample set and the original unlabeled sample set to obtain the object classification model;

[0012] In the process of training the neural network model to obtain the object classification model, the sample objects in the original unlabeled sample set are classified using the neural network model obtained during the training process until each sample object in the original unlabeled sample set has a classification label, thereby obtaining the target positive example set, the target negative example set, and the object classification model.

[0013] On one hand, embodiments of this application provide a method for training an object classification model, the apparatus comprising an acquisition unit, a construction unit, a determination unit, and a training unit:

[0014] The acquisition unit is used to acquire a sample object dataset for the target product. The sample object dataset includes a first positive example sample set and an unlabeled sample set. The sample objects in the first positive example sample set have positive example labels, which are classification labels used to mark that the sample objects have the intention to perform a conversion operation for the target product. The sample objects in the unlabeled sample set do not have classification labels, which are used to mark whether the sample objects have the intention to perform a conversion operation for the target product.

[0015] The construction unit is used to construct the object features of the sample objects in the sample object dataset;

[0016] The determining unit is configured to select multiple sample objects as intermediate samples from the first positive sample set, remove the intermediate samples from the first positive sample set to obtain a second positive sample set, and add the intermediate samples to the unlabeled sample set to obtain the original unlabeled sample set.

[0017] The training unit is used to train a neural network model on the second positive sample set and the original unlabeled sample set according to the object features, so as to obtain the object classification model.

[0018] During the process of training the neural network model to obtain the object classification model, the training unit is used to classify the sample objects in the original unlabeled sample set using the neural network model obtained during the training process, until each sample object in the original unlabeled sample set has a classification label, thereby obtaining the target positive example set, the target negative example set, and the object classification model.

[0019] On one hand, embodiments of this application provide a computer device, the computer device including a processor and a memory:

[0020] The memory is used to store program code and transmit the program code to the processor;

[0021] The processor is configured to execute the method described in any of the foregoing aspects according to the instructions in the program code.

[0022] In one aspect, embodiments of this application provide a computer-readable storage medium for storing program code for performing the methods described in any of the foregoing aspects.

[0023] On one hand, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the foregoing aspects.

[0024] As can be seen from the above technical solution, when training the object classification model, a sample object dataset targeting the target product can be obtained. This dataset includes a first positive example sample set and an unlabeled sample set. The sample objects in the first positive example sample set have positive labels, which indicate that the sample objects have the intention to perform a conversion operation on the target product. The sample objects in the unlabeled sample set do not have classification labels, which indicate whether the sample objects have the intention to perform a conversion operation on the target product. Multiple sample objects are selected from the first positive example sample set as intermediate samples, and these intermediate samples are removed to obtain a second positive example sample set. These intermediate samples are then added to the unlabeled sample set to obtain the original unlabeled sample set. Based on the object characteristics of the sample objects in the constructed sample object dataset, a neural network model is trained on the second positive example sample set and the original unlabeled sample set to obtain the object classification model. During the training of the neural network model to obtain the object classification model, the neural network model obtained during training is used to classify the sample objects in the original unlabeled sample set until each sample object in the original unlabeled sample set has a classification label, resulting in the target positive example set, the target negative example set, and the object classification model. Since intermediate samples are positive examples selected from the first set of positive examples, their behavior is consistent with that of unknown positive examples in the original set of unlabeled samples. Using intermediate samples allows for reliable evaluation of unknown positive examples, effectively balancing positive and negative samples and improving generalization during classification. This leads to better classification of the sample objects in the original set of unlabeled samples. Furthermore, by continuously updating the object classification model to divide the sample objects in the original set of unlabeled samples, the unlabeled sample objects in the original set of unlabeled samples are purified each time when the object classification model is at its best, ensuring the classification accuracy of the sample objects in the original set of unlabeled samples and further improving the business performance of subsequent information recommendation. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 An application scenario architecture diagram of a training method for an object classification model provided in an embodiment of this application;

[0027] Figure 2 A flowchart illustrating a training method for an object classification model provided in an embodiment of this application;

[0028] Figure 3An example diagram illustrating the technical architecture of a training method for an object classification model provided in this application embodiment;

[0029] Figure 4 An example diagram illustrating an aggregation method provided in an embodiment of this application;

[0030] Figure 5 A structural diagram of an ISSYST algorithm framework provided in an embodiment of this application;

[0031] Figure 6 A flowchart for parameter optimization of a neural network model based on QPSO is provided in this application embodiment;

[0032] Figure 7 A flowchart of a quantum measurement method provided in this application embodiment;

[0033] Figure 8 An example diagram of a CNN network structure provided in an embodiment of this application;

[0034] Figure 9 This is a schematic diagram of a convolution operation of a convolutional layer provided in an embodiment of this application;

[0035] Figure 10 This application provides a schematic diagram of a pooling operation for a pooling layer.

[0036] Figure 11 This is a schematic diagram of a fully connected layer provided in an embodiment of this application;

[0037] Figure 12 A detailed flowchart of a method for training an object classification model based on the ISPYST algorithm and BQPSO is provided for embodiments of this application;

[0038] Figure 13 This application provides a schematic diagram comparing the business performance of different methods in its embodiments.

[0039] Figure 14 A structural diagram of a training device for an object classification model provided in an embodiment of this application;

[0040] Figure 15 A structural diagram of a terminal provided in an embodiment of this application;

[0041] Figure 16 This is a structural diagram of a server provided in an embodiment of this application. Detailed Implementation

[0042] The embodiments of this application will now be described with reference to the accompanying drawings.

[0043] The object classification models provided in related technologies are based on non-deep learning methods and can classify objects. However, in scenarios involving "predicting whether an object has the intention to perform a conversion operation on the target product," the object characteristics are quite complex and difficult to express explicitly in data representation. These models are unsuitable for scenarios involving "predicting whether an object has the intention to perform a conversion operation on the target product," often leading to data silos and poor business results.

[0044] In addition, because labeling requires a lot of manpower and resources, there are few sample objects with classification labels, and there are even problems such as an abnormal ratio of positive and negative samples and a serious low number of positive samples.

[0045] To address the aforementioned technical problems, this application provides a training method and related apparatus for an object classification model. Positive samples selected from a first positive sample set are added as intermediate samples to an unlabeled sample set to obtain an original unlabeled sample set. The intermediate samples are then removed from the first positive sample set to obtain a second positive sample set. A neural network model is trained on the second positive sample set and the original unlabeled sample set to obtain the object classification model. Since the intermediate samples are positive samples selected from the first positive sample set, their behavior is consistent with that of unknown positive samples in the original unlabeled sample set. Using intermediate samples allows for reliable evaluation of unknown positive samples, effectively balancing positive and negative samples, resulting in better generalization during classification and thus better classification of the sample objects in the original unlabeled sample set. Furthermore, by continuously updating the object classification model to divide the sample objects in the original unlabeled sample set, the sample objects in the unlabeled original unlabeled sample set are purified each time when the object classification model is in its best state, ensuring the classification accuracy of the sample objects in the original unlabeled sample set and further improving the business effect of subsequent information recommendation.

[0046] This application's embodiments can be applied to various object classification scenarios. Object classification here is primarily used to distinguish whether an object has the intention to perform a conversion operation on the target product. This allows for the recommendation of relevant information about the target product to objects with the intention to perform a conversion operation during subsequent information recommendation processes, thereby improving business efficiency and directly contributing to refined operations. The conversion operation can be an operation that brings revenue or beneficial effects to operations, and the object or sample object can be, for example, a user.

[0047] In different scenarios, the willingness to perform a conversion action for a target product can vary. For example, in a video conferencing scenario, the target product could be paid video conferencing, and the willingness to perform a conversion action could be whether the user is willing to pay for video conferencing, i.e., the conversion action is payment. This video conferencing can utilize video conferencing software with features such as 300-person online meetings, one-click access across platforms, intelligent audio and video noise reduction, beauty filters, background blur, meeting locking, and screen watermarking. This software also provides real-time screen sharing and supports online document collaboration. Similarly, in a used car buying scenario, the target product could be a used car, and the willingness to perform a conversion action could be whether the user is willing to buy a used car, i.e., the conversion action is purchase. Likewise, in a video viewing scenario, the target product could be video, and the willingness to perform a conversion action could be whether the user is willing to watch the video, i.e., the conversion action is viewing, and so on.

[0048] It should be noted that the embodiments of this application can be implemented using cloud technology. Cloud technology refers to a hosting technology that unifies a series of resources such as hardware, software, and networks within a wide area network or local area network to realize data computing, storage, processing, and sharing. Cloud technology is a general term for network technology, information technology, integration technology, management platform technology, and application technology applied based on the cloud computing business model. It can form a resource pool, which can be used on demand, offering flexibility and convenience.

[0049] For example, cloud computing, a computing model, distributes computing tasks across a resource pool composed of a large number of computers, enabling various application systems to obtain computing power, storage space, and information services as needed. The network providing these resources is called the "cloud." From the user's perspective, the resources in the "cloud" are infinitely scalable, readily available, on-demand, expandable, and pay-as-you-go.

[0050] Another example involves big data in cloud technology. Big data refers to data sets that cannot be captured, managed, and processed within a certain timeframe using conventional software tools. It represents massive, rapidly growing, and diverse information assets that require new processing models to achieve stronger decision-making, insightful discovery, and process optimization capabilities. With the advent of the cloud era, big data has attracted increasing attention. Big data requires specialized technologies to effectively process large amounts of data within a tolerable timeframe. Technologies suitable for big data include massively parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the internet, and scalable storage systems.

[0051] This application's embodiments may also relate to artificial intelligence (AI). AI is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. AI software technology mainly includes several major areas such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0052] 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. For example, machine learning and deep learning can be used to train object classification models.

[0053] It should be noted that, in the embodiments of this application, the training method of the object classification model can be executed by a computer device, which can be a terminal, a server, or a combination of a terminal and a server. The embodiments of this application do not limit this.

[0054] like Figure 1 As shown, Figure 1 An application scenario architecture diagram of a training method for an object classification model is shown. Taking a computer device as a server as an example, this application scenario can include server 101 and terminal 102. Server 101 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Terminal 102 can be the terminal corresponding to the object to be identified. Terminal 102 can be a smartphone, tablet, laptop, desktop computer, intelligent voice interaction device, smart home appliance, vehicle terminal, aircraft, etc., but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, which is not limited herein.

[0055] The method provided in this application can be applied to various scenarios that require object classification, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving.

[0056] exist Figure 1In the application scenario architecture diagram shown, server 101 can be used to train an object classification model. After the object classification model is trained, server 101 can use the object classification model online to classify objects and then recommend relevant information about the target product based on the classification results.

[0057] When training the object classification model, server 101 can obtain a dataset of sample objects targeting the target product. The target product can be a product provided by the operation team, representing various conversion operations to be performed. The sample object dataset includes a first set of positive examples and an unlabeled set. The sample objects in the first set of positive examples have positive labels, which are used to indicate whether the sample object intends to perform a conversion operation on the target product. These positive labels can be pre-labeled manually. However, due to the significant manpower and resources required for labeling, many unlabeled sample objects exist; that is, the sample objects in the unlabeled set do not have classification labels. Classification labels are used to indicate whether the sample object intends to perform a conversion operation on the target product.

[0058] To obtain more labeled sample objects for training the neural network model and to maintain a balance between positive and negative samples, server 101 can select multiple sample objects as intermediate samples from the first positive sample set, remove the intermediate samples from the first positive sample set to obtain a second positive sample set, and add the intermediate samples to the unlabeled sample set to obtain the original unlabeled sample set. Then, based on the object features of the sample objects in the constructed sample object dataset, the neural network model is trained on the second positive sample set and the original unlabeled sample set to obtain the object classification model.

[0059] In the process of training a neural network model to obtain an object classification model, the neural network model obtained during training is used to classify the sample objects in the original unlabeled sample set until each sample object in the original unlabeled sample set has a classification label, resulting in a target positive example set, a target negative example set, and an object classification model. Since intermediate samples are positive examples selected from the first positive example set, their behavior is consistent with that of unknown positive examples in the original unlabeled sample set. Using intermediate samples can reliably evaluate unknown positive examples, effectively balancing the initial positive and negative examples, thus improving the generalization ability during the classification process and better classifying the sample objects in the original unlabeled sample set. Furthermore, by continuously updating the object classification model to divide the sample objects in the original unlabeled sample set, the unlabeled sample objects in the original unlabeled sample set can be purified each time when the object classification model is in its best state, ensuring the classification accuracy of the sample objects in the original unlabeled sample set and further improving the business effect of subsequent information recommendation.

[0060] After training the object classification model, server 101 can use the model to classify the objects to be identified, and then determine whether to recommend relevant information about the target product to the objects based on the classification results. If the objects to be identified intend to perform a conversion operation on the target product, relevant information about the target product can be recommended to them, for example, by sending the relevant information to the terminal 102 corresponding to the objects to be identified for display, thereby improving the conversion rate of the relevant information; if the objects to be identified do not intend to perform a conversion operation on the target product, then there is no need to recommend relevant information about the target product to them, thereby reducing invalid recommendations and improving business effectiveness.

[0061] It should be noted that the acquisition and processing of user-related data (such as the first positive example sample set, the unlabeled sample set, the object characteristics of sample objects in the sample object dataset, and related data required to construct object characteristics, etc.) involved in the embodiments of this application are authorized and permitted by the user.

[0062] Next, taking the training method of the object classification model executed by the server as an example, the training method of the object classification model provided in the embodiments of this application will be described in detail with reference to the accompanying drawings. See also Figure 2 , Figure 2 A flowchart illustrating a training method for an object classification model is shown, the method comprising:

[0063] S201. Obtain a dataset of sample objects for the target product. The dataset includes a first set of positive examples and an unlabeled set of samples. The sample objects in the first set of positive examples have positive labels, which are classification labels. The positive labels are used to mark whether the sample objects have the intention to perform a conversion operation for the target product. The sample objects in the unlabeled set do not have classification labels. The classification labels are used to mark whether the sample objects have the intention to perform a conversion operation for the target product.

[0064] The first set of positive examples can be represented by P, and the unlabeled set of examples can be represented by U. Classification labels can be represented by numbers, symbols, etc. Taking numbers as an example, when the classification label is 1, the classification label can be the positive example label, and when the classification label is 0, the classification label can be the negative example label.

[0065] The technical architecture of the object classification model training method provided in this application embodiment can be found in [reference needed]. Figure 3 As shown, this mainly includes sample creation (e.g.) Figure 3 As shown in Figure 301), feature construction (e.g.) Figure 3 As shown in Figure 302), a training framework combining concealment techniques and self-learning training (e.g.) Figure 3 The ISSYST algorithm framework shown in Figure 303 will first introduce sample creation.

[0066] When creating samples, manual annotation and business experience can be used to identify sample objects that are strongly relevant to the business, have normal data distribution, and reasonable object profiles. Strong relevance to the business can be based on experience to identify sample objects (e.g., users). For example, if the target product is paid video conferencing, then sample objects (e.g., users) strongly relevant to the business might be office workers, and their workplaces are often large listed companies. Normal data distribution means comparing various data sets of the sample objects, and the distribution should be relatively reasonable. For example, if the target product is paid video conferencing, and paying users are often users of large listed companies, then these users tend to have shorter gaming sessions. Based on this, positive sample objects with longer gaming sessions are removed based on the distribution of gaming time collected from the positive sample objects. Reasonable object profiles mean that the basic characteristics of the sample objects are reasonable. For example, if the target product is paid video conferencing, and paying users are often users of large listed companies, then the object profile should conform to this prior knowledge, such as an age range of 24-45 years old.

[0067] Typically, the ratio of positive to negative samples is 1:10,000, and the number of positive samples (samples with positive labels) is less than 50,000, which does not meet the minimum requirements for training an object classification model. Therefore, in this application, when training the object classification model, positive samples and unlabeled samples (i.e., samples without classification labels) are mainly used. This is combined with a training framework that utilizes hidden imagery techniques and self-learning to classify the unlabeled samples, thereby obtaining sufficient and reliable positive and negative samples (samples with negative labels, which indicate that the sample does not intend to perform a conversion operation on the target product). To this end, this application embodiment obtains a dataset of sample objects targeting the target product, which includes a first set of positive samples and a set of unlabeled samples.

[0068] The target product can be any product offered by the operations team and subject to various conversion operations, such as paid video conferencing services, video software memberships, a specific video, a product, or an advertisement. The conversion operation can be anything that brings revenue or beneficial results to the operations team. The willingness to perform a conversion operation on the target product can vary depending on the scenario. For example, in a video conferencing scenario, the target product could be paid video conferencing services, and the willingness to perform a conversion operation on the target product depends on whether the user is willing to pay for video conferencing services—the conversion operation is payment. Similarly, in a used car buying scenario, the target product could be a used car, and the willingness to perform a conversion operation on the target product depends on whether the user is willing to buy a used car—the conversion operation is purchase. Likewise, in a video watching scenario, the target product could be a video, and the willingness to perform a conversion operation on the target product depends on whether the user is willing to watch the video—the conversion operation is watching, and so on.

[0069] In one possible implementation, the sample object dataset for the target product can be obtained by acquiring an initial seed object for the target product, filtering out abnormal seed objects from the initial seed object to obtain the remaining seed objects, and then constructing the sample object dataset based on the remaining seed objects.

[0070] It is understood that in the embodiments of this application, filtering abnormal seed objects can be done in a variety of ways. One way is to first recall a batch of initial seed objects based on rules, then filter them by manual screening, and finally verify them based on business logic.

[0071] Another approach is to obtain a basic profile of the initial seed object. This basic profile includes non-privacy and interaction data about the sample object regarding the target product, such as whether it has installed the same mobile security app as the target product's manufacturer, and whether it uses the same spam blocking function as the target product's manufacturer's mobile security app.

[0072] Another possible implementation is to calculate anomaly detection metrics and then filter out anomalous seed objects based on these metrics. In real-world business scenarios, there may be instances of fake objects (users) or computer-controlled terminals. To eliminate the influence of non-real users on the training object classification model, anomaly detection metrics are set based on business experience, such as user traffic usage on other products from the same manufacturer as the target product, and the time distribution of traffic generation. These anomaly seed objects are then filtered out based on these anomaly detection metrics.

[0073] In another possible implementation, based on the distribution anomaly theorem, outlier seed objects are filtered. The "Laida criterion" is used as the outlier judgment standard. Specifically, it is assumed that a set of detection data contains only random errors. The standard deviation is calculated and processed, and an interval is determined according to a certain probability. Errors exceeding this interval are considered to be gross errors rather than random errors, and data containing such errors (such as the initial seed object) should be removed.

[0074] The above four methods for filtering abnormal seed objects can be used individually or in combination, and this application embodiment does not limit this.

[0075] After obtaining the sample object dataset, it can be stored offline, for example, in a distributed file system (HDFS).

[0076] S202. Construct the object features of sample objects in the sample object dataset.

[0077] Then, the server constructs object features for the sample objects in the sample object dataset. In one possible implementation, to construct high-dimensional feature vectors and improve the accuracy of training the object classification model, the server can construct basic features for the sample objects in the sample object dataset based on their interaction behavior, and construct vertical type features based on the business characteristics of the target product. Finally, the basic and vertical type features are merged to obtain the object features.

[0078] When constructing basic features, the server can build a rich object profile based on object interaction behavior as basic features. For example, this profile can include: object basic attributes, device basic attributes, network connection attributes, etc. For example: object basic attributes (age, gender), device basic attributes (terminal brand), network connection attributes (number of times connected to the network this week, such as 10 times).

[0079] When constructing vertical type features, vertical type features may include, for example, the click-through rate and conversion rate of sample objects for a specific type of advertisement.

[0080] When the server merges features based on basic features and vertical type features to obtain object features, it can first combine the time dimension to aggregate basic features and vertical type features across different time spans. Then, it can perform feature processing on the aggregated basic features and vertical type features, and finally merge the features based on the processed basic features and vertical type features to obtain the object features (e.g., Figure 3 (As shown in Figure 302).

[0081] Combining the time dimension, there are three ways to aggregate basic features and vertical type features across different time spans: summation, median, and standard deviation. For example, to aggregate the vertical type features of a sample object over a period of time, such as the last six months, the last three months, the last month, or the last week, the summation aggregation method can be used. (See [link to relevant documentation]). Figure 4 As shown, in Figure 4 In the diagram, (a) shows the vertical type features before aggregation, and (b) shows the vertical type features after aggregation. In both diagrams (a) and (b), the same fill pattern indicates the corresponding vertical type features before and after aggregation. For example, four squares filled with white (1, 2, 3, 0) result in a vertical type feature value of 1.5 after aggregation; four squares filled with a diagonal line from bottom left to top right (0, 3, 7, 2) result in a vertical type feature value of 3 after aggregation; four squares filled with a diagonal line from top left to bottom right (0, 3, 0, 4) result in a vertical type feature value of 1.75 after aggregation; and four squares filled with a curve (5, 1, 0, 4) result in a vertical type feature value of 2.5 after aggregation.

[0082] Feature processing methods for the aggregated basic features and vertical type features can include normalization to obtain normalized numerical features and discretization to obtain discretized non-numerical features. This application mainly describes discretization processing. The discretization processing method may differ for different features. Discretization processing mainly includes the following methods:

[0083] One-hot encoding, also known as one-bit valid encoding, uses an N-bit state register to encode N states. Each state has its own independent register bit, and at any given time, only one bit is valid. That is, only one bit is 1, and the rest are zero values. One-hot encoding is mainly used for basic features such as gender and region. For example, for the gender feature of a sample object (e.g., a user), one-hot encoding becomes: male: (1,0), female: (0,1).

[0084] Count encoding encodes data based on the number of times a category appears. For example, for the WiFi Point of Interest (POI) feature of a sample object (user), count encoding is used to identify the degree of interest of the sample object in that POI. For instance, a user visited the POI "Food - Chinese Food - Cantonese Food" a total of 3 times that week.

[0085] Consolidation encoding refers to the process of grouping multiple values ​​for a variable corresponding to a certain feature into a single piece of information. For example, the system version feature of a terminal for an operating system might have three values: "4.2", "4.4", and "5.0". Based on experience, these three values ​​can be grouped into "lower version of the operating system".

[0086] The processed basic features and vertical type features are merged to obtain object features, which are then stored offline in HDFS for easy access in subsequent processes. For each sample object in the sample object dataset, the object feature input to the neural network model is an N*1 numerical vector, such as (1,0,31,4,0.2,9.3,8.8,…,0,0,1,2,34).

[0087] The above method can produce high-dimensional feature vectors, thereby obtaining object features with richer information, which can improve the accuracy of training object classification models.

[0088] S203. Select multiple sample objects from the first positive sample set as intermediate samples, remove the intermediate samples from the first positive sample set to obtain the second positive sample set, and add the intermediate samples to the unlabeled sample set to obtain the original unlabeled sample set.

[0089] In this embodiment, collecting a large number of unlabeled sample objects is relatively easy, while obtaining a large number of labeled sample objects is much more difficult. Self-training algorithms are a widely studied approach in semi-supervised learning. This algorithm can train a classifier (also called an object classification model, denoted by C) using a small number of labeled sample objects. Then, it uses this object classification model to predict unlabeled sample objects and adds high-confidence results to the set of labeled sample objects. This process iterates until all unlabeled sample objects are classified. Essentially, the classifier uses its own predictions to improve itself.

[0090] In this embodiment, to address the issues of an abnormal positive-to-negative sample ratio and a severely low number of positive samples, a training framework combining hidden techniques and self-learning training can be used to train the neural network model. Self-learning training can be Positive and Unlabeled (PU) learning, where the training set consists only of positive samples and unlabeled samples. The training framework combining hidden techniques and self-learning training can be, for example, a PU learning method combining spy technology and semi-supervised self-training (SPYST), or an improved spy technique combined with semi-supervised self-training for PU learning (ISPYST). This embodiment will primarily use the ISSYST algorithm framework as an example. The classifier (neural network model) can be, for example, a Convolutional Neural Network (CNN), a Recurrent Neural Network (RNN), etc. This embodiment will primarily use a CNN as an example.

[0091] When training a neural network model to obtain an object classification model using a training framework that combines hidden vector techniques and self-learning training, the server first selects multiple sample objects from the positive sample set as intermediate samples. When selecting intermediate samples, the spatial distribution information of the first positive sample set can be mined. After understanding the spatial distribution information of the first positive sample set, its cluster centers are calculated, and intermediate samples are selected based on their distance from the cluster centers. Typically, sample objects that are closer to the cluster centers are selected as intermediate samples (also called hidden samples, denoted by S). Specifically, multiple sample objects can be selected from the positive sample set as intermediate samples by determining the centroid vector (i.e., the cluster center) of the first positive sample set, calculating the distance between each sample object in the first positive sample set and the centroid vector, and then determining the intermediate samples based on the distance. The distance between the intermediate samples and the centroid vector is generally small; that is, the intermediate samples are sample objects that are closer to the centroid vector.

[0092] In one possible implementation, the method for determining the centroid vector of the first positive sample set may include the following steps:

[0093] 1. Randomly select a sample object from the first positive sample set P as the initial centroid vector u;

[0094] 2. Randomly select t sample objects from P as intermediate samples;

[0095] 3. Initialize the cluster as an empty set;

[0096] 4. For j = 1, ..., n, calculate the distance Dj between the sample object Pj and the centroid vector u. The calculation formula is as follows:

[0097]

[0098] Where Pj represents the j-th sample object in the first positive sample set P, u represents the centroid vector, and Dj represents the distance between Pj and the centroid vector u;

[0099] 5. Update Cluster = Cluster∪{Pj};

[0100] 6. Recalculate the new centroid vector u for all sample objects in the Cluster. The calculation formula is as follows:

[0101]

[0102] Where P∈Cluster means that P is a sample object in the cluster, and Cluster, as the denominator, represents the number of sample objects in the cluster;

[0103] 7. Repeat steps 2-5 until the centroid vector u no longer changes, then output the centroid vector u.

[0104] The distance between each sample object and its centroid vector in the first positive sample set can be calculated using the following formula:

[0105]

[0106] Where d represents the distance between each sample object in the first positive sample set and the centroid vector, Pj represents the j-th sample object in the first positive sample set P, u represents the finally determined centroid vector, and n represents the number of sample objects in the first positive sample set P.

[0107] It should be noted that the purpose of determining intermediate samples based on distance in this embodiment is to identify sample objects that are closer to the centroid vector as intermediate objects. In this case, there are several ways to determine intermediate samples based on distance. One way is to select sample objects whose distance is less than a preset threshold as intermediate objects based on the distance corresponding to each sample object.

[0108] Another approach is to arrange the distances corresponding to each sample object in ascending order, and then select the first t sample objects as intermediate samples.

[0109] Another approach is to arrange the distances corresponding to each sample object in descending order, and then select the last t sample objects as intermediate samples.

[0110] Using the above method, the spatial distribution information of the first positive sample set can be extracted. This spatial distribution information reflects the spatial structure of the first positive sample set. After understanding the spatial structure of the first positive sample set, its centroid vector is calculated, and sample objects closer to the centroid vector are identified as intermediate samples. This makes the intermediate samples spatially closer to the centroid vector, containing more true information about the positive samples, and thus more effectively reflecting the distribution of unknown positive examples in the unlabeled sample set.

[0111] After selecting intermediate samples, they can be removed from the first positive sample set to obtain the second positive sample set, and then added to the unlabeled sample set to obtain the original unlabeled sample set. That is, the second positive sample set can be represented as PS, and the original unlabeled sample set can be represented as U+S. See also... Figure 5 As shown, Figure 5 This diagram illustrates the structure of the ISPYST algorithm framework. The second positive sample set P, the unlabeled sample set U, the second positive sample set PS, and the original unlabeled sample set U+S can be found in [references]. Figure 5 As shown.

[0112] S204. Based on the object characteristics, train the neural network model on the second positive sample set and the original unlabeled sample set to obtain the object classification model.

[0113] S205. In the process of training the neural network model to obtain the object classification model, the neural network model obtained during the training process is used to classify the sample objects in the original unlabeled sample set until each sample object in the original unlabeled sample set has a classification label, thereby obtaining the target positive example set, the target negative example set and the object classification model.

[0114] After obtaining the second set of positive examples and the original set of samples to be labeled, the server can train a neural network model on the second set of positive examples and the original set of samples to be labeled based on the object characteristics, and obtain an object classification model.

[0115] In the process of training a neural network model to obtain an object classification model, the neural network model obtained during training is used to classify the sample objects in the original unlabeled sample set until each sample object in the original unlabeled sample set has a classification label, resulting in a target positive example set, a target negative example set, and an object classification model. The target positive example set can be represented by RP, and the target negative example set can be represented by RN.

[0116] Since intermediate samples are positive examples selected from the first set of positive examples, their behavior is consistent with that of unknown positive examples in the original set of unlabeled samples. Using intermediate samples allows for reliable evaluation of unknown positive examples, effectively balancing positive and negative samples and improving generalization during classification. This results in better classification of the sample objects in the original set of unlabeled samples. Furthermore, by continuously updating the object classification model to divide the sample objects in the original set of unlabeled samples, the unlabeled sample objects in the original set of unlabeled samples are purified each time when the object classification model is at its best, ensuring the classification accuracy of the sample objects in the original set of unlabeled samples.

[0117] In one possible implementation, during the training of the neural network model to obtain the object classification model, the trained neural network model is used to classify the sample objects in the original unlabeled sample set until each sample object in the original unlabeled sample set has a classification label. The method to obtain the target positive example set, the target negative example set, and the object classification model can be as follows: the trained neural network model is used to classify the sample objects in the original unlabeled sample set to obtain an initial positive example set and an initial negative example set, and the initial positive example set is used as a new unlabeled sample set; the neural network model is retrained on the second positive example set and the initial negative example set; the trained neural network model is used to classify the new unlabeled sample set, and the second positive example set, the new unlabeled sample set, and the initial negative example set are updated based on the classification results. The step of retraining the neural network model on the second positive example set and the initial negative example set is repeated until each sample object in the new unlabeled sample set has a classification label, thus obtaining the target positive example set, the target negative example set, and the object classification model.

[0118] The initial positive example set can be represented by RN, and the initial negative example set can be represented by RP. The second positive example set is relabeled, and the sample objects in RN are relabeled as negative examples. RP is hidden, and RP is considered the new unlabeled sample set, which can be represented by U'. Then, the trained neural network model is used to classify U', and the second positive example set P, the new unlabeled sample set U', and the initial negative example set RN are updated based on the classification results. The step of retraining the neural network model on the second positive example set and the initial negative example set is repeated until each sample object in the new unlabeled sample set has a classification label, resulting in the target positive example set, the target negative example set, and the object classification model (see [link to documentation]). Figure 5 (As shown).

[0119] In one possible implementation, the initial positive and negative example sets are obtained by classifying the sample objects in the original unlabeled sample set using a trained neural network model. This classification yields the first posterior probability (Pru) of each sample object belonging to a positive sample. The second posterior probability (Prs) of each intermediate sample belonging to a positive sample is calculated. A probability threshold (theta) is then determined based on the second posterior probability Prs. Next, the initial negative example set is determined based on sample objects whose first posterior probability is less than the probability threshold, and the initial positive example set is determined based on the original unlabeled sample set and the initial negative example set. For the original unlabeled sample set U = {u1, u2, u3, ...}, where ui represents the i-th sample object, if the first posterior probability of ui is less than the probability threshold, it is added to the initial negative example set. The initial positive example set is the set obtained by removing sample objects belonging to the initial negative example set from the original unlabeled sample set, i.e., RP = U - RN.

[0120] In one possible implementation, the method of classifying the new unlabeled sample set using a trained neural network model and updating the second positive sample set, the new unlabeled sample set, and the initial negative sample set based on the classification results can be as follows: Classify the new unlabeled sample set using the trained neural network model to obtain the maximum posterior probability of the category to which the sample objects in the new unlabeled sample set belong, and use the maximum posterior probability as the classification result; arrange the maximum posterior probabilities of the categories to which the sample objects in the new unlabeled sample set belong in descending order; select the sample objects whose maximum posterior probabilities are in the top f positions; and update the second positive sample set, the new unlabeled sample set, and the initial negative sample set based on the top f sample objects and their corresponding classification labels.

[0121] The maximum posterior probability can be denoted as Pr. The maximum posterior probabilities of the categories to which the sample objects in the new unlabeled sample set belong are arranged in descending order. The larger the maximum posterior probability Pr, the higher the confidence level. Therefore, the sample objects with the top f maximum posterior probabilities are selected, and their corresponding classification labels are used to update the second positive sample set, the new unlabeled sample set, and the initial negative sample set. Typically, the selected f sample objects need to be subtracted from the new unlabeled sample set U' to update the new unlabeled sample set.

[0122] By selecting the top f samples, which are high-confidence samples carrying more effective distribution information, the neural network model can be retrained. This process is iterated until all samples in the new set of unlabeled samples are classified. This achieves the goal of refining the coarse classification results of the hiding technique, resulting in a more efficient PU learning framework.

[0123] The above describes the specific implementation method of the SPYST algorithm framework. The core function of the SPYST algorithm framework is to perform self-training purification on the unlabeled sample set, extract more reliable high-quality positive samples, reduce the skew ratio of positive and negative samples, and increase the absolute number of positive samples.

[0124] After training the object classification model, the server can use it to classify the objects to be identified, and then determine whether to recommend relevant information about the target product to the objects based on the classification results. Specifically, the server can obtain the objects to be identified, construct the target object features of the objects to be identified, and then determine the classification label of the objects to be identified through the object classification model based on the target object features. The classification label can be used as the classification result, and the classification label is used to indicate whether the objects to be identified have the intention to perform a conversion operation for the target product.

[0125] If the category label indicates that the target object is willing to perform a conversion operation for the target product, then relevant information about the target product can be recommended to the terminal corresponding to the target object. For example, the relevant information about the target product can be sent to the terminal corresponding to the target object for display, thereby improving the conversion rate of the relevant information. If the category label indicates that the target object is not willing to perform a conversion operation for the target product, then there is no need to recommend relevant information about the target product to the target object, thereby reducing invalid recommendations and improving business effectiveness.

[0126] As can be seen from the above technical solution, when training the object classification model, a sample object dataset targeting the target product can be obtained. This dataset includes a first positive example sample set and an unlabeled sample set. The sample objects in the first positive example sample set have positive labels, which indicate that the sample objects have the intention to perform a conversion operation on the target product. The sample objects in the unlabeled sample set do not have classification labels, which indicate whether the sample objects have the intention to perform a conversion operation on the target product. Multiple sample objects are selected from the first positive example sample set as intermediate samples, and these intermediate samples are removed to obtain a second positive example sample set. These intermediate samples are then added to the unlabeled sample set to obtain the original unlabeled sample set. Based on the object characteristics of the sample objects in the constructed sample object dataset, a neural network model is trained on the second positive example sample set and the original unlabeled sample set to obtain the object classification model. During the training of the neural network model to obtain the object classification model, the neural network model obtained during training is used to classify the sample objects in the original unlabeled sample set until each sample object in the original unlabeled sample set has a classification label, resulting in the target positive example set, the target negative example set, and the object classification model. Since intermediate samples are positive examples selected from the first set of positive examples, their behavior is consistent with that of unknown positive examples in the original set of unlabeled samples. Using intermediate samples allows for reliable evaluation of unknown positive examples, effectively balancing positive and negative samples and improving generalization during classification. This leads to better classification of the sample objects in the original set of unlabeled samples. Furthermore, by continuously updating the object classification model to divide the sample objects in the original set of unlabeled samples, the unlabeled sample objects in the original set of unlabeled samples are purified each time when the object classification model is at its best, ensuring the classification accuracy of the sample objects in the original set of unlabeled samples and further improving the business performance of subsequent information recommendation.

[0127] In this embodiment, not only is a training framework combining hiding techniques and self-learning training used to train the neural network model to improve the accuracy and recognition efficiency of the trained object classification model, but the parameter optimization process of the neural network model can also be improved during the training process to improve the training efficiency of the object classification model and further improve the accuracy and recognition efficiency of the trained object classification model.

[0128] In this case, the way to train the neural network model on the second positive sample set and the original unlabeled sample set according to the object characteristics is to update the parameters of the neural network model on the second positive sample set and the original unlabeled sample set according to the object characteristics to obtain the optimal network parameters; and determine the object classification model based on the optimal network parameters.

[0129] In one possible implementation, particle swarm optimization (PSO) algorithm can be used for parameter optimization. Based on this, improved PSO-based algorithms can also be used for parameter optimization. These improved PSO algorithms may include, for example, quantum-behaved particle swarm optimization (QPSO) and quantum-bahaved particle swarm optimization with binary encoding (BQPSO). This application primarily describes parameter optimization based on improved PSO algorithms; for more information on parameter optimization based on improved PSO algorithms, please refer to [link to relevant documentation]. Figure 3 As shown in Figure 304. If the neural network model includes multiple network layers, the parameters of the neural network model are updated on the second positive sample set and the original unlabeled sample set according to the object characteristics. The optimal network parameters can be obtained by encoding the parameters of multiple network layers separately to obtain multi-dimensional particle vectors; using an improved quantum particle swarm optimization algorithm, the global optimal position of the population corresponding to the multi-dimensional particle vector is determined; in the process of using the improved quantum particle swarm optimization algorithm to determine the global optimal position of the population corresponding to the multi-dimensional particle vector, if the termination condition of the improved quantum particle swarm optimization algorithm is reached, the global optimal position of the population at the time of reaching the termination condition is decoded and output to obtain the optimal network parameters.

[0130] Next, this application will mainly use the improved quantum particle swarm optimization algorithms QPSO and BQPSO as examples to describe in detail the method for determining the above-mentioned optimal network parameters. First, we will introduce the determination of optimal network parameters based on QPSO.

[0131] The flowchart for parameter optimization of neural network models based on QPSO can be found in [link / reference]. Figure 6 As shown:

[0132] S601. Initialize the quantum particle position information in QPSO and select the globally optimal position.

[0133] S602. Each quantum particle will be decoded into a neural network model structure according to a segmented compression coding strategy.

[0134] S603. Determine if the neural network model structure is valid. If yes, proceed to S604; otherwise, proceed to S607.

[0135] S604. Calculate the fitness value.

[0136] S605, will update the individual best position Pi and the global best position Pg.

[0137] S606. Determine if the maximum evolution cycle has been reached. If yes, end the process and output Pg; otherwise, return to S602.

[0138] S607, Perform quantum measurement.

[0139] When particle decoding produces invalid neural network structures, quantum particles can be repeatedly measured to obtain more new positional information, thereby decoding and generating more new CNN structures. Methods of quantum measurement can be found in [reference needed]. Figure 7 As shown, the method includes:

[0140] S701. Calculate the characteristic length and attractor of the current particle.

[0141] The formulas for calculating the characteristic length and the attractor can be as follows:

[0142]

[0143]

[0144] Where mbest is the average optimal position of the population, M is the population size, and Pid is the i-th particle in the d-th dimension of the population, where d takes values ​​between 1 and D, and D is the dimension of the particle vector. u and β are random decimals in the interval [0,1], β is a positive constant called the contraction-expansion coefficient, and a id Let represent a local attractor, and Pid and Pgd represent the position information of the local optimal solution and the global optimal solution of the i-th particle in the d-th dimension, respectively.

[0145] S702, Randomly generate a decimal number.

[0146] Normally, a decimal number can be randomly generated in the interval [0,1].

[0147] S703, generating new individuals.

[0148] In this embodiment of the application, a new individual can be generated according to the following formula:

[0149]

[0150] Where, x id (t+1) represents the new individual generated for the (t+1)th iteration, x id (t) represents the current particle, i.e., the individual in the t-th iteration, u represents a randomly generated decimal, and mbest d This represents the group's average optimal position in the d-th dimension.

[0151] S704. Decode the position information of the unit digit into a neural network structure according to the segmented compression coding strategy.

[0152] S705. Determine whether the neural network model structure is valid. If yes, proceed to S706; otherwise, return to S702.

[0153] S706. Treat this individual as the next generation individual.

[0154] If the maximum number of iterations is reached, the process ends.

[0155] The following steps describe how to determine the optimal network parameters based on BQPSO:

[0156] a) Initialize each particle xi in the swarm using a binary bit string, such that Pi = xi;

[0157] Where xi = {xi1, xi2, ..., xiD}, and xid represents the position information of the i-th particle in the d-th dimension.

[0158] b) Calculate the value of the population average optimal position (mbest). The calculation process is as follows:

[0159] Let P1, P2, ..., P M mbest is taken as input and output as output. The following code calculates mbest:

[0160] Set the initial value of sum to 0;

[0161] sum = sum + P i [j]

[0162] avg = sum / M;

[0163] When avg > 0.5, mbest[j] = 1; when avg < 0.5, mbest[j] = 0; when avg = 0.5, mbest[j] = 0 or 1.

[0164] Where j is a value between 1 and l, l is the length of each particle, Pi is the i-th particle in the population, i is a value between 1 and M, and M is the population size.

[0165] c) Calculate the fitness value of each particle in the swarm according to the fitness function, and compare it with the previous best value of the particle. If f(xi) > f(Pi), then Pi = xi; otherwise, do not update. The fitness function can include many methods. In this embodiment, the Schaffer function and the RA-Rastrigin function can be selected for calculating the fitness value.

[0166] d) Update the globally optimal particle pg in the swarm;

[0167] e) Calculate the local attractor a according to the algorithm. id The value;

[0168] a id It is calculated based on Pid and Pgd, where Pid and Pgd represent the position information of the local optimal solution and the global optimal solution of the i-th particle in the d-th dimension, respectively.

[0169] f) Calculate p according to the formula rd The value of p rd Let represent the probability in the d-th dimension, as shown in the formula below:

[0170]

[0171]

[0172] Where, x id mbest represents the position information of the i-th particle in the d-th dimension. d This represents the average optimal position of the population in the d-th dimension, u represents a randomly generated decimal between 0 and 1, β is a positive constant called the contraction-expansion coefficient, and ld represents x. id Length, d H () indicates a function to be computed.

[0173] g) Calculate x according to the algorithm id The value of is concatenated to generate xi;

[0174] x id It is based on a id and p rd Calculated.

[0175] h) Repeat the above steps until the termination condition of the improved quantum particle swarm optimization algorithm is met.

[0176] In the parameter optimization process based on the improved particle swarm optimization algorithm, the parameters of the neural network model first need to be encoded. Taking a CNN as an example, the network structure of a CNN can be found in [reference needed]. Figure 8 As shown, the network consists of multiple layers: one input layer, two convolutional layers, two pooling layers, and two fully connected layers (including hidden and output layers). Each layer has a different configuration: stride, kernel size, and feature maps are important parameters for convolutional layer configuration; stride, kernel size, and pooling type are the main attributes of pooling layers; and the number of neurons is the only attribute of fully connected layers. Below is a brief introduction to convolutional, pooling, and fully connected layers.

[0177] A CNN consists of one input layer, two convolutional layers, two pooling layers, and two fully connected layers. The number of neurons in the output layer is equal to the number of classes in a classification problem.

[0178] A schematic diagram of the convolution operation of a convolutional layer is shown below. Figure 9 As shown: Using a 3x3 convolution kernel, the convolution operation starts from the shaded position at the top left corner of the input image matrix. The result of this operation is 3. Since the stride is set to 1, the shaded position can also move left or right, ultimately resulting in a 3x3 feature map composed of 9 results.

[0179] A schematic diagram of the pooling operation of the pooling layer is shown below. Figure 10 As shown, to some extent, the more convolutional layers there are, the more comprehensive the extracted features. However, excessive convolutional kernel operations can lead to excessively large feature dimensions and require excessive computational resources. Therefore, pooling operations are needed to reduce the feature dimensionality. Pooling operations have two parameters: the pooling kernel and the stride. Common pooling operations include mean pooling and max pooling. The former takes the mean of the feature matrix for the corresponding region, while the latter takes the maximum value of the feature matrix for the corresponding region.

[0180] Fully connected layers such as Figure 11 As shown: the first layer is responsible for input data (e.g., x1, x2, x3), the last layer is responsible for output results (e.g., y), and the other layers are called hidden layers. Each neuron in a hidden layer is connected to a neuron in the layer above it, hence the name fully connected layer.

[0181] Taking the improved particle swarm optimization algorithm BQPSO as an example, the method for encoding the parameters of a CNN is as follows:

[0182] Binary encoding strategy:

[0183] The encoding process is implemented in the D-dimensional discrete search space of the BQPSO algorithm, where the D decision variables in each particle are fixed-length binary sequences, as shown in Table 1:

[0184] Table 1

[0185] ID Parameter 1 Parameter 2 ……

[0186] Each binary sequence carries the configuration of a network layer, including its ID and parameters.

[0187] The parameters of each CNN network layer type, after binary encoding, are shown below:

[0188] The encoding strategies for convolutional layers are shown in Table 2:

[0189] Table 2

[0190] ID nuclear size Feature map Step length total Range of values 0 or 7 [1,8] [1,128] [1,4] Binary bits 3 3 7 2 15 example 0(000) 2(001) 16(0001111) 2(01) 000001000111101

[0191] The pooling layer encoding strategy is shown in Table 3:

[0192] Table 3

[0193] ID nuclear size Step length type total Range of values 1 or 6 [1,4] [1,4] [1,2] Binary bits 3 2 2 1 8 example 0(001) 2(01) 2(01) 2(1) 00101011XXXXXXX

[0194] The encoding strategies for fully connected layers are shown in Table 4:

[0195] Table 4

[0196] ID Number of neurons total Range of values 2 or 5 [1,2048] Binary bits 3 11 14 example 2(010) 1024(01111111111) 010011111111111X

[0197] The blank layer encoding strategy is shown in Table 5:

[0198] Table 5

[0199] ID total Range of values 3 or 4 Binary bits 3 3 example 3(011) 011XXXXXXXXXXXX

[0200] Based on the encoding strategies of the different network layers and the network structure of CNN, the parameters of CNN can be encoded.

[0201] A binary-based quantum particle swarm optimization algorithm is used to optimize the parameters of a neural network model to achieve better performance. The optimized neural network model can then be used to predict the accuracy of all samples.

[0202] It should be noted that the detailed flowchart of the method for training an object classification model based on the ISPYST algorithm and BQPSO provided in this application embodiment can be found in [reference needed]. Figure 12 As shown, it includes:

[0203] S1201. Obtain the initial seed object for the target product.

[0204] S1202. Filter out abnormal seed objects from the initial seed objects to obtain the remaining seed objects.

[0205] S1203. Construct a sample object dataset based on the remaining seed objects.

[0206] S1204. Construct basic features.

[0207] S1205, Construct vertical type features.

[0208] S1206. Merge features based on basic features and vertical type features to obtain object features.

[0209] S1207. Select multiple sample objects from the first positive sample set as intermediate samples.

[0210] S1208. Remove intermediate samples from the first positive sample set to obtain the second positive sample set, and add the intermediate samples to the unlabeled sample set to obtain the original unlabeled sample set.

[0211] S1209. Use the trained neural network model to classify the sample objects in the original unlabeled sample set to obtain the initial positive example set and the initial negative example set.

[0212] S1210. Use the initial set of positive examples as the new set of samples to be labeled.

[0213] S1211. Retrain the neural network model on the second positive sample set and the initial negative sample set.

[0214] S1212. Use the trained neural network model to classify the new set of samples to be labeled, and update the second positive sample set, the new set of samples to be labeled, and the initial negative sample set based on the classification results.

[0215] S1213. Initialize each particle xi in the swarm using binary bit strings, such that Pi = xi.

[0216] S1214. Calculate the value of the population average optimal position mbest.

[0217] S1215. Calculate the fitness value of each particle in the population according to the fitness function and compare it with the previous best value of the particle. If f(xi)>f(Pi), then Pi=xi.

[0218] S1216, Update the globally optimal particle pg in the swarm.

[0219] S1217. Calculate the local attractor a according to the algorithm. id The value of .

[0220] S1218, Calculate p rd The value of .

[0221] S1219. Calculate x according to the algorithm. id The value of is obtained and concatenated to generate xi.

[0222] S1220. Repeat the above steps until the termination condition of the improved quantum particle swarm optimization algorithm is met.

[0223] The following is a comparative analysis of the business effectiveness of different methods. These methods include related technology 1, related technology 2, and the method provided in the embodiments of this application. Related technology 1 can be based on human experience to determine data rules; that is, product operations set rules for manual identification based on business experience. For example, objects "aged 35-40 with a high standard of living" are considered more likely to be "highly willing to pay for video conferencing." Related technology 2 can be based on non-deep learning methods. A comparative diagram of the business effectiveness of different methods is shown below. Figure 13 As shown:

[0224] exist Figure 13 As can be seen, in terms of ad click-through rate, the method provided in this application embodiment improves by an average of 182.7% compared to other technologies; in terms of ad conversion rate, the method provided in this application embodiment improves by an average of 178.41% compared to other technologies, that is, the business effect is significantly improved.

[0225] It should be noted that, based on the implementation methods provided in the above aspects, this application can be further combined to provide more implementation methods.

[0226] based on Figure 2 Corresponding to the object classification model training method provided in the embodiments, this application also provides an object classification model training apparatus 1400. See also Figure 14 The training device 1400 for the object classification model includes an acquisition unit 1401, a construction unit 1402, a determination unit 1403, and a training unit 1404.

[0227] The acquisition unit 1401 is used to acquire a sample object dataset for the target product. The sample object dataset includes a first positive example sample set and an unlabeled sample set. The sample objects in the first positive example sample set have positive example labels, which are classification labels used to mark that the sample objects have the intention to perform a conversion operation for the target product. The sample objects in the unlabeled sample set do not have classification labels, which are used to mark whether the sample objects have the intention to perform a conversion operation for the target product.

[0228] The construction unit 1402 is used to construct the object features of the sample objects in the sample object dataset;

[0229] The determining unit 1403 is configured to select multiple sample objects as intermediate samples from the first positive sample set, remove the intermediate samples from the first positive sample set to obtain a second positive sample set, and add the intermediate samples to the unlabeled sample set to obtain the original unlabeled sample set.

[0230] The training unit 1404 is used to train a neural network model on the second positive sample set and the original unlabeled sample set according to the object features, so as to obtain the object classification model.

[0231] During the process of training the neural network model to obtain the object classification model, the training unit 1404 is used to classify the sample objects in the original unlabeled sample set using the neural network model obtained during the training process, until each sample object in the original unlabeled sample set has a classification label, thereby obtaining the target positive example set, the target negative example set, and the object classification model.

[0232] In one possible implementation, the determining unit 1403 is specifically used for;

[0233] Determine the centroid vector of the first positive example sample set;

[0234] Calculate the distance between each sample object in the first positive sample set and the centroid vector;

[0235] The intermediate sample is determined based on the distance.

[0236] In one possible implementation, the determining unit 1403 is specifically used for:

[0237] Arrange the distances corresponding to each sample object in ascending order;

[0238] The first t sample objects are selected as the intermediate samples.

[0239] In one possible implementation, the training unit 1404 is specifically used for:

[0240] The trained neural network model is used to classify the sample objects in the original unlabeled sample set to obtain an initial positive example set and an initial negative example set, and the initial positive example set is used as a new unlabeled sample set.

[0241] The neural network model is retrained on the second positive sample set and the initial negative sample set;

[0242] The trained neural network model is used to classify the new set of samples to be labeled, and the second positive sample set, the new set of samples to be labeled, and the initial negative sample set are updated based on the classification results. The step of retraining the neural network model on the second positive sample set and the initial negative sample set is repeated until each sample object in the new set of samples to be labeled has a classification label, thus obtaining the target positive sample set, the target negative sample set, and the object classification model.

[0243] In one possible implementation, the training unit 1404 is specifically used for:

[0244] The trained neural network model is used to classify the sample objects in the original unlabeled sample set to obtain the first posterior probability that the sample objects in the original unlabeled sample set belong to positive samples.

[0245] Calculate the second posterior probability that the intermediate sample belongs to the positive sample;

[0246] The probability threshold is determined based on the second posterior probability;

[0247] The initial negative example set is determined based on sample objects whose first posterior probability is less than the probability threshold, and the initial positive example set is determined based on the original unlabeled sample set and the initial negative example set.

[0248] In one possible implementation, the training unit 1404 is specifically used for:

[0249] The trained neural network model is used to classify the new set of samples to be labeled, and the maximum posterior probability of the category to which the sample object in the new set of samples to be labeled is obtained. The maximum posterior probability is used as the classification result.

[0250] Arrange the maximum posterior probabilities of the categories to which the sample objects in the new unlabeled sample set belong in descending order;

[0251] Select the sample objects that are ranked in the top f positions by the maximum posterior probability;

[0252] The second positive sample set, the new unlabeled sample set, and the initial negative sample set are updated based on the first f sample objects and their corresponding classification labels.

[0253] In one possible implementation, the training unit 1404 is specifically used for;

[0254] Based on the object characteristics, the parameters of the neural network model are updated on the second positive sample set and the original unlabeled sample set to obtain the optimal network parameters;

[0255] The object classification model is determined based on the optimal network parameters.

[0256] In one possible implementation, the neural network model includes multiple network layers, and the training unit 1404 is specifically used for:

[0257] The parameters of the multiple network layers are encoded to obtain multidimensional particle vectors;

[0258] An improved particle swarm optimization algorithm is used to determine the global optimal position of the population corresponding to the multidimensional particle vector;

[0259] In the process of determining the global optimal position of the population corresponding to the multidimensional particle vector using the improved particle swarm optimization algorithm, if the termination condition of the improved quantum particle swarm optimization algorithm is reached, the global optimal position of the population at the time of reaching the termination condition is decoded and output to obtain the optimal network parameters.

[0260] In one possible implementation, the acquisition unit 1401 is specifically used for:

[0261] Obtain the initial seed object for the target product;

[0262] Filter out abnormal seed objects from the initial seed objects to obtain the remaining seed objects;

[0263] The sample object dataset is constructed based on the remaining seed objects.

[0264] In one possible implementation, the building unit 1402 is specifically used for:

[0265] Based on the object interaction behavior of the sample objects in the sample object dataset, construct the basic features of the sample objects in the sample object dataset;

[0266] Based on the business characteristics of the target product, construct vertical type features;

[0267] The object features are obtained by merging the basic features and the vertical type features.

[0268] In one possible implementation, the device further includes a classification unit:

[0269] The acquisition unit 1401 is also used to acquire the object to be identified;

[0270] The construction unit 1402 is also used to construct the target object features of the object to be identified;

[0271] The classification unit is used to determine the classification label of the object to be identified based on the characteristics of the target object and through the object classification model. The classification label is used to indicate whether the object to be identified has the intention to perform a conversion operation for the target product.

[0272] In one possible implementation, the device further includes a recommendation unit:

[0273] If the classification label indicates that the object to be identified has the intention to perform a conversion operation for the target product, the recommendation unit is used to recommend relevant information of the target product to the terminal corresponding to the object to be identified.

[0274] As can be seen from the above technical solution, when training the object classification model, a sample object dataset targeting the target product can be obtained. This dataset includes a first positive example sample set and an unlabeled sample set. The sample objects in the first positive example sample set have positive labels, which indicate that the sample objects have the intention to perform a conversion operation on the target product. The sample objects in the unlabeled sample set do not have classification labels, which indicate whether the sample objects have the intention to perform a conversion operation on the target product. Multiple sample objects are selected from the first positive example sample set as intermediate samples, and these intermediate samples are removed to obtain a second positive example sample set. These intermediate samples are then added to the unlabeled sample set to obtain the original unlabeled sample set. Based on the object characteristics of the sample objects in the constructed sample object dataset, a neural network model is trained on the second positive example sample set and the original unlabeled sample set to obtain the object classification model. During the training of the neural network model to obtain the object classification model, the neural network model obtained during training is used to classify the sample objects in the original unlabeled sample set until each sample object in the original unlabeled sample set has a classification label, resulting in the target positive example set, the target negative example set, and the object classification model. Since intermediate samples are positive examples selected from the first set of positive examples, their behavior is consistent with that of unknown positive examples in the original set of unlabeled samples. Using intermediate samples allows for reliable evaluation of unknown positive examples, effectively balancing positive and negative samples and improving generalization during classification. This leads to better classification of the sample objects in the original set of unlabeled samples. Furthermore, by continuously updating the object classification model to divide the sample objects in the original set of unlabeled samples, the unlabeled sample objects in the original set of unlabeled samples are purified each time when the object classification model is at its best, ensuring the classification accuracy of the sample objects in the original set of unlabeled samples and further improving the business performance of subsequent information recommendation.

[0275] This application also provides a computer device capable of executing a method for processing graph data. This computer device may be, for example, a terminal, such as a smartphone.

[0276] Figure 15 The diagram shown is a block diagram of a portion of the structure of a smartphone provided in an embodiment of this application. (Reference) Figure 15A smartphone includes components such as: a radio frequency (RF) circuit 1510, a memory 1520, an input unit 1530, a display unit 1540, a sensor 1550, an audio circuit 1560, a Wi-Fi module 1570, a processor 1580, and a power supply 1590. The input unit 1530 may include a touch panel 1531 and other input devices 1532, the display unit 1540 may include a display panel 1541, and the audio circuit 1560 may include a speaker 1561 and a microphone 1562. It is understood that... Figure 15 The smartphone structure shown does not constitute a limitation on smartphones and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0277] The memory 1520 can be used to store software programs and modules. The processor 1580 executes various functions and data processing of the smartphone by running the software programs and modules stored in the memory 1520. The memory 1520 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the smartphone (such as audio data, phonebook, etc.). In addition, the memory 1520 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0278] The processor 1580 is the control center of the smartphone, connecting various parts of the smartphone via various interfaces and lines. It performs various functions and processes data by running or executing software programs and / or modules stored in the memory 1520 and by accessing data stored in the memory 1520. Optionally, the processor 1580 may include one or more processing units; preferably, the processor 1580 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor 1580.

[0279] In this embodiment, the steps performed by the processor 1580 in the smartphone can be based on Figure 15 The structure shown is implemented.

[0280] The computer device in this application embodiment can also be a server; please refer to [link / reference]. Figure 16 As shown, Figure 16The diagram illustrates the structure of a server 1600 provided in this embodiment. The server 1600 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 1622 (e.g., one or more processors) and a memory 1632, and one or more storage media 1630 (e.g., one or more mass storage devices) for storing application programs 1642 or data 1644. The memory 1632 and storage media 1630 may be temporary or persistent storage. The program stored in the storage media 1630 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the server. Furthermore, the CPU 1622 may be configured to communicate with the storage media 1630 and execute the series of instruction operations stored in the storage media 1630 on the server 1600.

[0281] Server 1600 may also include one or more power supplies 1626, one or more wired or wireless network interfaces 1650, one or more input / output interfaces 1658, and / or one or more operating systems 1641, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0282] In this embodiment, the central processing unit 1622 in server 1600 can perform the following steps:

[0283] Obtain a sample object dataset for the target product. The sample object dataset includes a first positive example sample set and an unlabeled sample set. The sample objects in the first positive example sample set have positive example labels, which are classification labels used to mark that the sample objects have the intention to perform a conversion operation for the target product. The sample objects in the unlabeled sample set do not have classification labels, which are used to mark whether the sample objects have the intention to perform a conversion operation for the target product.

[0284] Construct the object features of the sample objects in the sample object dataset;

[0285] Multiple sample objects are selected from the first positive sample set as intermediate samples, and the intermediate samples are removed from the first positive sample set to obtain a second positive sample set. The intermediate samples are then added to the unlabeled sample set to obtain the original unlabeled sample set.

[0286] Based on the object characteristics, a neural network model is trained on the second positive example sample set and the original unlabeled sample set to obtain the object classification model;

[0287] In the process of training the neural network model to obtain the object classification model, the sample objects in the original unlabeled sample set are classified using the neural network model obtained during the training process until each sample object in the original unlabeled sample set has a classification label, thereby obtaining the target positive example set, the target negative example set, and the object classification model.

[0288] According to one aspect of this application, a computer-readable storage medium is provided for storing program code for executing the training method of the object classification model described in the foregoing embodiments.

[0289] According to one aspect of this application, a computer program product or computer program is provided, comprising 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 methods provided in various optional implementations of the above embodiments.

[0290] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0291] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0292] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0293] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0294] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0295] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0296] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A training method for an object classification model, characterized in that, The method includes: Obtain a sample object dataset for the target product. The sample object dataset includes a first positive example sample set and an unlabeled sample set. The sample objects in the first positive example sample set have positive example labels, which are classification labels used to mark that the sample objects have the intention to perform a conversion operation for the target product. The sample objects in the unlabeled sample set do not have classification labels, which are used to mark whether the sample objects have the intention to perform a conversion operation for the target product. Construct object features for the sample objects in the sample object dataset; the object features include basic features and vertical type features, the basic features include at least one of object basic attributes, device basic attributes and network connection attributes, and the vertical type features include the click-through rate or conversion rate of the sample object for the corresponding type of advertisement; Determine the centroid vector of the first positive example sample set; Calculate the distance between each sample object in the first positive sample set and the centroid vector; the amount of true information contained in the positive sample by the sample object is negatively correlated with the distance between the centroid vector and the centroid vector. Multiple sample objects are selected as intermediate samples based on the distance, and the intermediate samples are removed from the first positive sample set to obtain a second positive sample set. The intermediate samples are then added to the unlabeled sample set to obtain the original unlabeled sample set. Based on the object characteristics, the parameters of the neural network model are updated on the second positive sample set and the original unlabeled sample set to obtain the optimal network parameters; The object classification model is determined based on the optimal network parameters; In the process of training the neural network model to obtain the object classification model, the trained neural network model is used to classify the sample objects in the original unlabeled sample set to obtain an initial positive example set and an initial negative example set, and the initial positive example set is used as a new unlabeled sample set. The neural network model is retrained on the second positive sample set and the initial negative sample set; The trained neural network model is used to classify the new set of samples to be labeled, and the second positive sample set, the new set of samples to be labeled, and the initial negative sample set are updated based on the classification results. The step of retraining the neural network model on the second positive sample set and the initial negative sample set is repeated until each sample object in the new set of samples to be labeled has a classification label, thus obtaining the target positive sample set, the target negative sample set, and the object classification model. The neural network model includes multiple network layers. The step of updating the parameters of the neural network model based on the object features on the second positive sample set and the original unlabeled sample set to obtain optimal network parameters includes: The parameters of the multiple network layers are encoded to obtain multidimensional particle vectors; An improved particle swarm optimization algorithm is used to determine the global optimal position of the population corresponding to the multidimensional particle vector; In the process of determining the global optimal position of the population corresponding to the multidimensional particle vector using the improved particle swarm optimization algorithm, if the termination condition of the improved quantum particle swarm optimization algorithm is reached, the global optimal position of the population at the time of reaching the termination condition is decoded and output to obtain the optimal network parameters.

2. The method according to claim 1, characterized in that, Determining the intermediate sample based on the distance includes: Arrange the distances corresponding to each sample object in ascending order; The first t sample objects are selected as the intermediate samples.

3. The method according to claim 1, characterized in that, The process of classifying sample objects in the original unlabeled sample set using a trained neural network model to obtain an initial positive example set and an initial negative example set includes: The trained neural network model is used to classify the sample objects in the original unlabeled sample set to obtain the first posterior probability that the sample objects in the original unlabeled sample set belong to positive samples. Calculate the second posterior probability that the intermediate sample belongs to the positive sample; The probability threshold is determined based on the second posterior probability; The initial negative example set is determined based on sample objects whose first posterior probability is less than the probability threshold, and the initial positive example set is determined based on the original unlabeled sample set and the initial negative example set.

4. The method according to claim 1, characterized in that, The step of classifying the new set of samples to be labeled using a trained neural network model, and updating the second positive sample set, the new set of samples to be labeled, and the initial negative sample set based on the classification results, includes: The trained neural network model is used to classify the new set of samples to be labeled, and the maximum posterior probability of the category to which the sample object in the new set of samples to be labeled is obtained. The maximum posterior probability is used as the classification result. Arrange the maximum posterior probabilities of the categories to which the sample objects in the new unlabeled sample set belong in descending order; Select the sample objects that are ranked in the top f positions by the maximum posterior probability; The second positive sample set, the new unlabeled sample set, and the initial negative sample set are updated based on the first f sample objects and their corresponding classification labels.

5. The method according to any one of claims 1-4, characterized in that, The process of obtaining a sample object dataset for the target product includes: Obtain the initial seed object for the target product; Filter out abnormal seed objects from the initial seed objects to obtain the remaining seed objects; The sample object dataset is constructed based on the remaining seed objects.

6. The method according to any one of claims 1-4, characterized in that, The construction of object features for sample objects in the sample object dataset includes: Based on the object interaction behavior of the sample objects in the sample object dataset, construct the basic features of the sample objects in the sample object dataset; Based on the business characteristics of the target product, construct vertical type features; The object features are obtained by merging the basic features and the vertical type features.

7. The method according to any one of claims 1-4, characterized in that, The method further includes: Obtain the object to be identified; Construct the target object features of the object to be identified; Based on the characteristics of the target object, the classification label of the object to be identified is determined by the object classification model. The classification label is used to indicate whether the object to be identified has the intention to perform a conversion operation for the target product.

8. The method according to claim 7, characterized in that, The method further includes: If the classification label indicates that the object to be identified has the intention to perform a conversion operation for the target product, the relevant information of the target product is recommended to the terminal corresponding to the object to be identified.

9. A training device for an object classification model, characterized in that, The device includes an acquisition unit, a construction unit, a determination unit, and a training unit: The acquisition unit is used to acquire a sample object dataset for the target product. The sample object dataset includes a first positive example sample set and an unlabeled sample set. The sample objects in the first positive example sample set have positive example labels, which are classification labels used to mark that the sample objects have the intention to perform a conversion operation for the target product. The sample objects in the unlabeled sample set do not have classification labels, which are used to mark whether the sample objects have the intention to perform a conversion operation for the target product. The construction unit is used to construct object features of sample objects in the sample object dataset; the object features include basic features and vertical type features, the basic features include at least one of object basic attributes, device basic attributes and network connection attributes, and the vertical type features include the click-through rate or conversion rate of the sample object for the corresponding type of advertisement; The determining unit is configured to: determine the centroid vector of the first positive example sample set; calculate the distance between each sample object in the first positive example sample set and the centroid vector; determine intermediate samples based on the distance, remove the intermediate samples from the first positive example sample set to obtain a second positive example sample set, and add the intermediate samples to the unlabeled sample set to obtain the original unlabeled sample set; the amount of true information contained in the sample object as a positive example sample is negatively correlated with the distance between the centroid vector and the sample object. The training unit is used to update the parameters of the neural network model on the second positive sample set and the original unlabeled sample set according to the object features, so as to obtain the optimal network parameters; and to determine the object classification model according to the optimal network parameters. In the process of training the neural network model to obtain the object classification model, the training unit is used to classify the sample objects in the original unlabeled sample set using the neural network model obtained during the training process, until each sample object in the original unlabeled sample set has a classification label, thereby obtaining the target positive example set, the target negative example set, and the object classification model; The training unit is specifically used for: The trained neural network model is used to classify the sample objects in the original unlabeled sample set to obtain an initial positive example set and an initial negative example set, and the initial positive example set is used as a new unlabeled sample set. The neural network model is retrained on the second positive sample set and the initial negative sample set; The trained neural network model is used to classify the new set of samples to be labeled, and the second positive sample set, the new set of samples to be labeled, and the initial negative sample set are updated based on the classification results. The step of retraining the neural network model on the second positive sample set and the initial negative sample set is repeated until each sample object in the new set of samples to be labeled has a classification label, thus obtaining the target positive sample set, the target negative sample set, and the object classification model. The neural network model includes multiple network layers, and the training unit is specifically used for: The parameters of the multiple network layers are encoded to obtain multidimensional particle vectors; An improved particle swarm optimization algorithm is used to determine the global optimal position of the population corresponding to the multidimensional particle vector; In the process of determining the global optimal position of the population corresponding to the multidimensional particle vector using the improved particle swarm optimization algorithm, if the termination condition of the improved quantum particle swarm optimization algorithm is reached, the global optimal position of the population at the time of reaching the termination condition is decoded and output to obtain the optimal network parameters.

10. The apparatus according to claim 9, characterized in that, The determining unit is specifically used for: Arrange the distances corresponding to each sample object in ascending order; The first t sample objects are selected as the intermediate samples.

11. The apparatus according to claim 9, characterized in that, The training unit is specifically used for: The trained neural network model is used to classify the sample objects in the original unlabeled sample set to obtain the first posterior probability that the sample objects in the original unlabeled sample set belong to positive samples. Calculate the second posterior probability that the intermediate sample belongs to the positive sample; The probability threshold is determined based on the second posterior probability; The initial negative example set is determined based on sample objects whose first posterior probability is less than the probability threshold, and the initial positive example set is determined based on the original unlabeled sample set and the initial negative example set.

12. The apparatus according to claim 9, characterized in that, The training unit is specifically used for: The trained neural network model is used to classify the new set of samples to be labeled, and the maximum posterior probability of the category to which the sample object in the new set of samples to be labeled is obtained. The maximum posterior probability is used as the classification result. Arrange the maximum posterior probabilities of the categories to which the sample objects in the new unlabeled sample set belong in descending order; Select the sample objects that are ranked in the top f positions by the maximum posterior probability; The second positive sample set, the new unlabeled sample set, and the initial negative sample set are updated based on the first f sample objects and their corresponding classification labels.

13. The apparatus according to any one of claims 9-12, characterized in that, The acquisition unit is specifically used for: Obtain the initial seed object for the target product; Filter out abnormal seed objects from the initial seed objects to obtain the remaining seed objects; The sample object dataset is constructed based on the remaining seed objects.

14. The apparatus according to any one of claims 9-12, characterized in that, The building unit is specifically used for: Based on the object interaction behavior of the sample objects in the sample object dataset, construct the basic features of the sample objects in the sample object dataset; Based on the business characteristics of the target product, construct vertical type features; The object features are obtained by merging the basic features and the vertical type features.

15. The apparatus according to any one of claims 9-12, characterized in that, The device also includes a sorting unit: The acquisition unit is also used to acquire the object to be identified; The construction unit is also used to construct the target object features of the object to be identified; The classification unit is used to determine the classification label of the object to be identified based on the characteristics of the target object and through the object classification model. The classification label is used to indicate whether the object to be identified has the intention to perform a conversion operation for the target product.

16. The apparatus according to claim 15, characterized in that, The device also includes a recommendation unit: If the classification label indicates that the object to be identified has the intention to perform a conversion operation for the target product, the recommendation unit is used to recommend relevant information of the target product to the terminal corresponding to the object to be identified.

17. A computer device, characterized in that, The computer device includes a processor and memory: The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the method according to any one of claims 1-8 according to the instructions in the program code.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code that, when executed by a processor, causes the processor to perform the method according to any one of claims 1-8.

19. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-8.

Images