Object classification method, classification model training method, related equipment and program product

Abstract

The invention discloses an object classification method, a classification model training method, related equipment and a program product, and relates to the technical field of machine learning. The method comprises the steps of obtaining parameter data of a to-be-classified object; performing feature extraction on the parameter data to obtain parameter features; the parameter features are sent to a classification model to obtain a classification result, the classification model is obtained by training a random forest model through a reinforcement learning strategy, and the weight parameter of each decision tree in the random forest model is updated and fixed in the reinforcement learning training process. Through the reinforcement learning strategy, the weight of the decision tree can be automatically adjusted, so that the classification accuracy and robustness of the classification model on complex and unbalanced data are improved. In addition, compared with a deep learning model, the parameter quantity is smaller, and the required computing resources are lower.

Description

Technical Field

[0001] This application relates to the field of machine learning technology. More specifically, it relates to an object classification method, a classification model training method, related devices, and program products. Background Art

[0002] The classification task refers to dividing the object to be classified into the category it belongs to according to the set category set. There are various classification tasks in practical applications. Taking the classification of mobile applications as an example, mobile applications can be divided into various types such as shopping applications, game applications, news and information applications, etc. according to their different functions.

[0003] Currently, the solutions for classification tasks mainly focus on the application of traditional machine learning models and deep learning models. Traditional machine learning models such as Support Vector Machine (SVM), decision trees, and K-Nearest Neighbor (KNN), etc. These models are prone to overfitting when dealing with high-dimensional sparse data, and lack effective coping strategies for the problem of class imbalance, resulting in the impact on classification accuracy. Although the random forest model based on ensemble learning can alleviate the overfitting problem to a certain extent, there are still certain limitations when dealing with the problem of class imbalance. Deep learning models have a large number of parameters and require high computing resources, and their processing effect on imbalanced data sets is poor. Summary of the Invention

[0004] In view of the above problems, this application is proposed to provide an object classification method, a classification model training method, related devices, and program products, which can improve the classification effect in the scenario of data class imbalance with a relatively small parameter scale. The specific solutions are as follows:

[0005] In the first aspect of this application, an object classification method is provided, including:

[0006] Obtain the parameter data of the object to be classified;

[0007] Extract features from the parameter data to obtain parameter features;

[0008] Send the parameter features into the classification model to obtain a classification result, where the classification model is obtained by training a random forest model using a reinforcement learning strategy, and the weight parameters of each decision tree in the random forest model are updated and fixed during the reinforcement learning training process.

[0009] In a possible design, in another implementation manner of the first aspect of the embodiments of this application, the training process of the classification model includes:

[0010] Obtain training samples labeled with classification result labels, and extract features from the training samples to obtain a set of feature samples;

[0011] Obtain an initial random forest model as the classification model to be trained, where the random forest model includes a number of decision trees;

[0012] Send the feature samples in the feature sample set into the random forest model for classification processing to obtain the classification result output by the random forest model, where the classification result is obtained by weighted fusion of the classification results of the decision trees according to the weight parameters of each decision tree;

[0013] According to the classification result output by the random forest model and the classification result label corresponding to the feature sample, adopt a reinforcement learning strategy to adjust the structural parameters and weight parameters of each decision tree until the set training end condition is reached, and obtain the trained classification model.

[0014] In a possible design, in another implementation manner of the first aspect of the embodiments of the present application, the parameter data includes multi-modal parameter data;

[0015] Then the process of extracting features from the parameter data to obtain parameter features includes:

[0016] Extract features from each type of modal parameter data respectively to obtain each type of modal parameter features, and fuse various modal parameter features to obtain the fused parameter features.

[0017] In a possible design, in another implementation manner of the first aspect of the embodiments of the present application, it further includes:

[0018] Resample the feature sample set based on the proportion of feature samples of each category in the feature sample set to balance the proportion of feature samples of each category.

[0019] In a possible design, in another implementation manner of the first aspect of the embodiments of the present application, the process of adopting a reinforcement learning strategy to adjust the structural parameters and weight parameters of each decision tree according to the classification result output by the random forest model and the classification result label corresponding to the feature sample includes:

[0020] According to the classification result output by the random forest model and the classification result label corresponding to the feature sample, calculate the classification accuracy, model confidence of the random forest model, and the contribution degree of each decision tree in the classification task;

[0021] Based on the classification accuracy, model confidence, and the contribution degree of each decision tree in the classification task, calculate the value of the reward function;

[0022] Update the structural parameters and weight parameters of each decision tree in the random forest model according to the value of the reward function.

[0023] In a possible design, in another implementation manner of the first aspect of the embodiments of the present application, the object to be classified is an application to be classified, and the parameter data includes at least one of the following: the description text of the application, the display image of the application, and the user behavior data related to the application.

[0024] In a possible design, in another implementation manner of the first aspect of the embodiments of the present application, it further includes:

[0025] Sending the parameter features into at least one other type of base learner to obtain the classification results output by the base learner, where the base learner is configured to have the ability of data classification;

[0026] Fusing the classification results of the classification model and each of the base learners through weighted average or voting mechanism to obtain the final classification result.

[0027] In a possible design, in another implementation manner of the first aspect of the embodiments of the present application, it further includes:

[0028] Sending the parameter features into at least one other type of base learner to obtain the classification results output by the base learner, where the base learner is configured to have the ability of data classification;

[0029] Inputting the classification results output by the classification model and the classification results output by each of the base learners into a meta-learner to obtain the final classification result output by the meta-learner, where the meta-learner is trained with the classification results of the classification model and each of the base learners for the same training sample as the input sample and the classification result label of the same training sample as the sample label.

[0030] In the second aspect of the present application, a classification model training method is provided, including:

[0031] Obtaining training samples labeled with classification result labels, and performing feature extraction on the training samples to obtain a set of feature samples;

[0032] Obtaining an initial random forest model as the classification model to be trained, where the random forest model includes a plurality of decision trees;

[0033] Sending the feature samples in the set of feature samples into the random forest model for classification processing to obtain the classification results output by the random forest model, where the classification results are obtained by weighted fusion of the classification results of the decision trees according to the weight parameters of each decision tree;

[0034] According to the classification result output by the random forest model and the classification result label corresponding to the feature sample, an enhanced learning strategy is adopted to adjust the structural parameters and weight parameters of each decision tree until the set training end condition is reached, and a trained classification model is obtained.

[0035] In the third aspect of the present application, an object classification device is provided, including:

[0036] A parameter acquisition unit, configured to acquire parameter data of an object to be classified;

[0037] A feature extraction unit, configured to extract features from the parameter data to obtain parameter features;

[0038] A classification prediction unit, configured to send the parameter features into a classification model to obtain a classification result, where the classification model is obtained by training a random forest model using an enhanced learning strategy, and the weight parameters of each decision tree in the random forest model are updated and fixed during the enhanced learning training process.

[0039] In the fourth aspect of the present application, an electronic device is provided, including: a memory and a processor;

[0040] The memory is configured to store a program;

[0041] The processor is configured to execute the program to implement each step of the object classification method described in any one of the first aspects of the present application, or to implement each step of the classification model training described in the second aspect of the present application.

[0042] In the fifth aspect of the present application, a readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, each step of the object classification method described in any one of the first aspects of the present application is implemented, or each step of the classification model training described in the second aspect of the present application is implemented.

[0043] In the sixth aspect of the present application, a computer program product is provided, including a computer program. When the computer program is executed by a processor, each step of the object classification method described in any one of the first aspects of the present application is implemented, or each step of the classification model training described in the second aspect of the present application is implemented.

[0044] With the above technical solution, the present application combines the integrated advantages of traditional machine learning and the adaptive adjustment ability of reinforcement learning, optimizes and trains the decision trees in the random forest model using integrated learning with a reinforcement learning strategy, and obtains a trained classification model, making up for the deficiencies of the traditional random forest model in aspects such as feature selection, class imbalance, and handling of difficult-to-classify samples, and improving the classification effect. Specifically, the reinforcement learning strategy enables the classification model to adaptively adjust its weights according to the performance of each decision tree, that is, dynamically adjusts the weights of each decision tree in the random forest model through the reinforcement learning strategy. In the traditional random forest model, since each decision tree assigns the same weight to all samples, minority class samples and difficult-to-classify samples are easily ignored, thus affecting the classification accuracy of the model. However, through the reinforcement learning strategy, the present application can automatically adjust the training strategy, thereby obtaining more learning opportunities on these samples and improving the robustness and classification accuracy of the classification model. In addition, the classification model of the present application uses a random forest model architecture, which still belongs to the traditional machine learning model, with fewer parameters and lower computational resources required compared to deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] By reading the detailed description of the preferred embodiments below, various other advantages and benefits will become clear to those of ordinary skill in the art. The drawings are only for the purpose of illustrating the preferred embodiments and are not considered to be a limitation of the present application. Moreover, throughout the drawings, the same reference numerals are used to represent the same components. In the drawings:

[0046] Figure 1 is a schematic flow chart of an object classification method provided by an embodiment of the present application;

[0047] Figure 2 is a schematic flow chart of a classification model training method provided by an embodiment of the present application;

[0048] Figure 3 is a schematic structural diagram of an object classification device provided by an embodiment of the present application;

[0049] Figure 4 is a schematic structural diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] Before introducing the solution of the present application, some term concepts involved in the present application are first introduced.

[0051] Random forest model:

[0052] Random Forest is an Ensemble Learning method and belongs to the Bagging (Bootstrap Aggregating) class of algorithms. It improves the accuracy and robustness of the model by constructing multiple decision trees and combining their prediction results.

[0053] Random Forest is an ensemble model composed of multiple Decision Trees and operates in the following ways:

[0054] Randomly sample the training data: Each decision tree is trained on a different subset generated by Bootstrap Sampling of the original data (i.e., some samples may be drawn repeatedly while some may be ignored).

[0055] Randomly select features: When splitting nodes in each tree, only the optimal splitting feature is selected from a randomly chosen subset of features (rather than all features), increasing the diversity of the trees.

[0056] Voting or averaging: In classification tasks, the final result is determined by the voting (majority rule) of all trees; in regression tasks, the result is the average of the predictions of all trees.

[0057] Reinforcement Learning:

[0058] Reinforcement learning (RL), also known as re-inforcement learning, evaluation learning, or enhanced learning, etc., is used to describe and solve the problem that an agent learns a strategy to maximize rewards or achieve specific goals during the interaction with the environment.

[0059] Reinforcement learning is a way for an agent to learn by "trial and error". The agent interacts with the environment through actions and is guided by the rewards obtained from the interaction. The goal is to maximize the rewards obtained by the agent. Usually, reinforcement learning can operate without a training dataset or generate training data by interacting with the environment. In reinforcement learning, the reinforcement signal provided by the environment (i.e., the reward) evaluates the quality of the actions taken, rather than telling the reinforcement learning system how to generate correct actions. Since the information provided by the external environment is scarce, the agent must learn from its own experiences. In this way, the agent acquires knowledge in an action-evaluation (i.e., reward) environment and improves its action plan to adapt to the environment.

[0060] It can be understood that reinforcement learning is a machine learning paradigm used to solve the problem of maximizing rewards or completing specific machine learning tasks by an agent learning a strategy during the interaction with the environment. The agent learns in a trial-and-error manner, obtains rewards through interaction with the environment, and uses them to guide the selection of actions. It finds the optimal strategy in the current state and selects appropriate actions according to this strategy to maximize the benefits.

[0061] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0062] It can be understood that before using the technical solutions disclosed in the embodiments of the present application, the types, usage scopes, usage scenarios, etc. of the personal information involved in the present application should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0063] Currently, the widely adopted classification solutions in the market (such as mobile application classification solutions) mainly focus on the application of traditional machine learning models and deep learning models.

[0064] Some typical traditional machine learning models such as: Support Vector Machine (SVM), K-Nearest Neighbor (KNN), decision tree, etc. These models are trained by manually extracting features (such as TF-IDF, bag-of-words model, etc.), but they face the problem of high-dimensional sparsity of text data. In addition, these models usually cannot fully mine the deep semantic information in text data and perform poorly in the case of class imbalance. Although the random forest based on ensemble learning can alleviate the overfitting problem to a certain extent, there are still certain limitations when dealing with complex text features and data imbalance.

[0065] Deep learning models: With the rise of deep learning, deep learning methods such as Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM) have been widely applied to text classification tasks. In the field of mobile application classification, deep learning methods can effectively extract high-level semantic features in text through automated feature learning. These methods usually have strong generalization ability and perform well especially on large-scale datasets. However, deep learning models require a large amount of computing resources, and may encounter difficulties in training or a decrease in classification accuracy when facing small samples and imbalanced datasets.

[0066] Compared with the above solutions, the method of optimizing a random forest model based on reinforcement learning in this application combines the integration advantages of traditional machine learning and the adaptive adjustment ability of reinforcement learning, and can achieve better classification results in scenarios with data imbalance and complex features. At the same time, by dynamically adjusting the weights of different decision trees through the reinforcement learning mechanism, it can effectively solve the deficiencies of traditional random forests on complex data sets and improve the robustness and accuracy of the model.

[0067] The solution of this application involves the training stage of the classification model and the inference stage of the classification model (i.e., the process of applying the trained classification model to execute the object classification method).

[0068] The solution of this application involves a method for training a classification model and an object classification method based on the trained classification model. The object classification method of this application can be applied to various scenarios that require classification, such as classifying application programs, classifying emails, classifying users, etc. In the following embodiments of this application, the application program classification scenario is used as an example for illustration.

[0069] The method provided in this application can be divided into a training stage and an inference stage. The training stage is the stage of training the classification model, and the inference stage is the process of using the trained classification model to execute the object classification method. Among them, the training stage and the inference stage can be deployed on the same device or on different devices. For example, the training stage can be deployed in the cloud or on a server, and the inference stage can be deployed on a smart terminal, such as a mobile phone, a tablet, a smart car, a robot, or a wearable device, etc.

[0070] For ease of understanding, this application introduces the processes of the training stage and the inference stage respectively.

[0071] I. Training Stage

[0072] Referring to Figure 1 , the embodiment of this application provides a schematic flow chart of a method for training a classification model. The method for training a classification model may specifically include the following steps:

[0073] Step S100: Obtain training samples labeled with classification result labels, and perform feature extraction on the training samples to obtain a set of feature samples.

[0074] Specifically, the original data can be collected from the target domain to form a data set. Generally, the data set can contain sufficient sample diversity, especially being representative in aspects such as class distribution, feature dimension, and data noise. Among them, the target domain is the domain to which the classification model is to be applied. Taking the application classification task as an example, the target domain can be the domain containing various types of applications, so as to collect the parameter data of various types of applications, such as the description text of the application, the display image of the application, the user behavior data related to the application, etc.

[0075] After the data set is collected, preprocessing can be performed on the data set, such as data cleaning, denoising, etc. Through data cleaning, missing values, duplicate values, and outliers can be cleaned up, and the data format and type can be unified. Through denoising processing, low-quality samples can be removed to improve the quality of the data set.

[0076] The training samples in the data set are labeled with classification result labels. Further, feature extraction can be performed on the training samples in the data set to obtain a set of feature samples. The classification result labels of the training samples can be used as the labels of the feature samples extracted from the training samples, so as to establish the corresponding relationship between the feature samples and the classification result labels for the subsequent training of the model.

[0077] In some embodiments, the training samples can include single-modal or multi-modal data. For example, the training samples can include multi-modal data such as text modality, image modality, audio-video modality, etc. When performing feature extraction on the training samples, appropriate feature extraction methods can be adopted for different modality data.

[0078] Taking the application classification task as an example, the text data included in the training samples, such as application names, function descriptions, user comments, etc., usually contains a large amount of semantic information, and this information is crucial for accurately identifying the category, function, and potential user needs of the application. Due to the complexity and high dimensionality of application data, how to effectively extract and preprocess text features has become a key link in improving the model performance.

[0079] In some embodiments, for the text data in the training samples, word embedding technology can be used to extract text features. Examples of word embedding technology include pre-trained word vector models such as Word2Vec and GloVe. Further optionally, after obtaining the text features, dimensionality reduction processing can be performed on the text features to reduce the computational complexity.

[0080] Traditional text feature extraction methods (such as TF-IDF, bag-of-words model) can capture the surface information of text to a certain extent, but they ignore the semantic relationships between words, resulting in unsatisfactory results when dealing with complex texts. To better capture the semantic information in text, word embedding technology can be adopted in this embodiment. Through pre-trained word vector models (such as Word2Vec, GloVe, etc.), we map each word to a high-dimensional vector space, thereby being able to better represent the semantic similarity between words. This method can not only capture the surface word frequency information of text, but also deeply explore the semantic connections between words, improving the expressive ability of text features.

[0081] In some embodiments, in addition to using word embedding technology to extract text features in this embodiment, other methods can be further adopted to extract text features from text data. For example, the term frequency-inverse document frequency TF-IDF features of the text data are extracted, and the TF-IDF features and the text features extracted by word embedding technology are fused to obtain the fused text features.

[0082] TF-IDF features can measure the importance of words in a document, while word embedding technology can further map words to a high-dimensional space to express the semantic relationships between words. By combining the advantages of both, the fused text features can not only express word frequency information, but also fully reflect the semantic connections of words.

[0083] In some embodiments, the parameter data of training samples may be diverse and complex. In this embodiment, deep learning models such as convolutional neural network (CNN) and long short-term memory network (LSTM) can be further combined to extract features from the text data contained in the training samples, so as to fully capture the context information in the text. The text features extracted by the deep learning model can be fused with the text features extracted in the foregoing embodiments to obtain the fused text features.

[0084] In some embodiments, when the training samples contain image data (such as in the application classification task, the icons, screenshots or video preview images of the application, etc., and the image data can convey important information about the visual style, function layout, etc. of the application), a convolutional neural network CNN or other neural network models with image processing capabilities can be used to extract image features from the image data.

[0085] In some embodiments, when the training samples contain user behavior data (such as in the application classification task, user behaviors such as clicks, downloads, ratings of the application, etc.), time series modeling methods (such as using models such as long short-term memory network LSTM, gated recurrent unit GRU, etc.) can be used to extract behavior features from the user behavior data.

[0086] For the case where the training samples contain multi-modal data, feature extraction can be performed separately for each type of modal data to obtain the feature of each modal parameter. On this basis, the feature of each modal parameter can be fused to obtain the fused parameter feature, which is used as the feature sample corresponding to the training sample.

[0087] In some embodiments, when fusing the feature of each modal parameter, a suitable target feature fusion strategy can be selected according to factors such as the importance of each modal parameter feature in the current classification task, the dimension matching of each modal parameter feature, and the requirements of the classification model for the input data, so as to merge the multi-modal parameter features into a unified representation. Some possible feature fusion strategies are as follows:

[0088] Concatenation fusion: directly combine the feature of each modal parameter side by side according to the dimension. This feature fusion strategy is suitable for scenarios where there is no obvious redundancy between the feature of each modal parameter and the classification model can handle high-dimensional input.

[0089] Weighted average fusion: assign weights to different modal parameter features according to their importance in the classification task and calculate the weighted sum. This feature fusion strategy is suitable for cases where dimension reduction is required or there is a certain complementary relationship between modalities.

[0090] In this embodiment, by fusing the multi-modal parameter features of the training samples and using the fused feature samples as the input of the classification model, the following effects can be achieved:

[0091] Information supplement and enhancement: Different modalities of data provide complementary information. Image data can convey visual features, text data can reveal semantic information, and user behavior data can capture user preferences. By effectively fusing these data, the classification model can obtain a more comprehensive feature representation.

[0092] Enhance the robustness of the classification model: Due to the differences between multi-modal data, the model can still rely on the information of other modalities for effective classification when some modal data fails or has a large amount of noise. This enhances the adaptability of the model to different scenarios and reduces the dependence on a single modality of data.

[0093] Handle the problem of data imbalance: The introduction of multi-modal data makes the sample information in the classification task more abundant. Especially in the case of class imbalance, it can effectively utilize the additional information of minority class samples, thereby improving the classification effect. The integrated classification strategy that fuses multi-modal data can effectively improve the model performance in the classification task through reasonable preprocessing and feature fusion. By combining text, image, and user behavior data, the classification model can comprehensively understand the features of the training samples, and thus show stronger classification ability in complex and diverse data environments.

[0094] In some embodiments, after obtaining the feature sample set, the feature samples in the feature sample set can also be standardized to make the scales of different feature dimensions consistent, avoiding unnecessary impacts on model training caused by excessive scale differences among certain feature samples. Standardization processing can adopt standardization algorithms such as zero-mean unit-variance Z-score standardization and Min-Max normalization (mapping the data to the range [0, 1]).

[0095] In some embodiments, the feature samples in the feature sample set can also be dimensionally reduced. For example, dimensional reduction methods such as principal component analysis PCA or t-SNE (t-Distributed Stochastic Neighbor Embedding) can be used to reduce the dimensionality of the feature space and relieve the computational burden.

[0096] In some embodiments, resampling processing of the feature sample set can also be added. Specifically, in the face of data imbalance and high-dimensional sparsity problems, resampling strategies can effectively improve the training effect of the model and enhance the classification accuracy.

[0097] Taking the application classification task as an example, in practical applications, there is usually a large imbalance problem in the class distribution of applications. For example, some application categories may contain a large number of applications, while other categories may have only a few applications. Traditional classification methods tend to be overly biased towards majority-class samples, resulting in a decrease in the prediction accuracy of minority-class samples. In this embodiment, resampling technology can be introduced to reduce the impact brought by data imbalance by balancing the proportions of samples in each category. In this embodiment, resampling can be performed on the feature sample set based on the proportions of feature samples in each category in the feature sample set to balance the proportions of feature samples in each category.

[0098] Resampling strategies can include two methods: over-sampling and under-sampling. In over-sampling, the number of minority-class samples is increased by replicating minority-class samples or generating new samples (such as using the SMOTE technique) to balance the class distribution in the dataset. In under-sampling, the majority-class samples are randomly deleted to reduce their proportion in the dataset so that each category has a relatively fair position in the training process.

[0099] In some embodiments, in order to improve the diversity and representativeness of samples during the resampling process, some resampling strategies can be adopted to avoid model bias caused by excessive replication of certain samples. This embodiment provides some optional resampling strategies, including but not limited to:

[0100] 1. Prioritize the use of synthetic sample methods over simple replication: Generate new synthetic samples by interpolating between minority-class samples to avoid repeated stacking of original samples, thereby reducing the risk of overfitting.

[0101] 2. Limit the number of replications or generations of each original sample: Set a threshold (e.g., at most k new samples can be generated for each sample) to prevent a sample from dominating the distribution of the minority class due to frequent selection.

[0102] 3. Combine undersampling strategies: Appropriately undersample the majority class and, in conjunction with oversampling, balance the data as a whole and mitigate oversampling bias.

[0103] 4. Use clustering-assisted oversampling: First, cluster the minority-class samples, and then perform oversampling from each cluster center to ensure a more balanced and diverse distribution of the generated samples.

[0104] Through the resampling strategies in the above examples, the generalization ability of the classification model can be improved, and overfitting to specific patterns of minority-class samples can be avoided.

[0105] Step S110: Obtain an initial random forest model as the classification model to be trained. The random forest model includes a number of decision trees.

[0106] The random forest model obtains the final classification result by constructing multiple decision trees and voting or weighted averaging their prediction results. In this embodiment, the initial random forest model is used as the classification model to be trained. By integrating the results of multiple decision trees, the random forest model can reduce the bias of a single decision tree model and improve the classification accuracy.

[0107] When training, the random forest model generates multiple different training subsets through sampling with replacement (Bootstrap Sampling), and each decision tree is trained on an independent subset. When splitting the nodes of each decision tree, only the optimal splitting feature is selected from a randomly selected subset of features, reducing the correlation between the trees. Each decision tree only predicts one target class. Decision tree initialization settings: When initializing each decision tree, set the maximum depth (to control overfitting), the minimum number of samples for splitting (to avoid excessive branching), and the maximum number of leaf nodes.

[0108] In this embodiment, the process of obtaining the initial random forest model may include constructing the random forest model. In this embodiment, features related to the classification task are selected from the pre-collected original features to form a feature set. The feature set serves as the basis for splitting the nodes of the decision tree, that is, when splitting the nodes, a subset of features is randomly selected from the feature set, and the optimal splitting feature is selected from the subset of features.

[0109] Among them, the original features can be obtained from the data collection stage. For example, after preprocessing the collected data, a set of original features that can be used for modeling is extracted. The original feature set often contains redundant, noisy, and task-irrelevant features. Therefore, features relevant to the classification task can be selected from the original features to form a feature set.

[0110] The process of feature selection can be completed manually or through machine automatic selection algorithms. In some possible implementations, manual selection and algorithm automatic selection can also be combined. The machine automatic selection algorithm first needs to determine the relevance of the features to the classification task, and then select features whose relevance meets the requirements according to the magnitude of the relevance. The process of determining the relevance of features to the classification task can adopt statistical methods such as mutual information, chi-square test, Pearson correlation coefficient, etc.; it can also adopt model-driven methods such as random forest, L1-regularized logistic regression, embedded methods (such as recursive feature elimination), etc.

[0111] The selected feature set is used as the basis for node splitting in the decision tree. During the training process, the decision tree will preferentially select features that can improve the data purity (such as the largest information gain, Gini index change) for node division, so as to improve the discrimination ability of the model.

[0112] In the initial random forest model, the weights of each decision tree are equal. In the subsequent training process using the reinforcement learning strategy, the weights of each decision tree will be adaptively adjusted.

[0113] Step S120: Send the feature samples in the feature sample set into the random forest model for classification processing, and obtain the classification result output by the random forest model. The classification result is obtained by weighted fusion of the classification results of the decision trees according to the weight parameters of each decision tree.

[0114] Step S130: According to the classification result output by the random forest model and the classification result label corresponding to the feature sample, adopt the reinforcement learning strategy to adjust the structure parameters and weight parameters of each decision tree until the set training end condition is reached, and obtain the trained classification model.

[0115] During the process of using the feature sample set to perform reinforcement learning training on the random forest model, in each round of training, all decision trees perform adaptive adjustment according to their feedback during the training process.

[0116] Multiple reinforcement learning algorithms can be selected for the reinforcement learning strategy, and examples include the Q-learning algorithm, etc. A reward (punishment) mechanism is introduced: when the classification result output by the random forest model is correct, a positive reward is given, and when the classification result output by the random forest model is incorrect, a negative reward is given.

[0117] Definition of state space in the reinforcement learning algorithm: Define the "state" of the decision tree as the classification accuracy of the current decision tree, the depth of the tree, and its performance in the classification task.

[0118] Definition of action space: Define the actions taken by the decision trees in the model, such as reweighting specific classification samples, adjusting the weights of training samples, etc.

[0119] Calculation of reward function: Evaluate the performance of each decision tree in the current state through the reward function, and calculate the weight adjustment strategy of the decision tree accordingly.

[0120] In one possible implementation, the reinforcement learning training process in step S130 can be implemented as follows:

[0121] According to the classification results output by the random forest model and the classification result labels corresponding to the feature samples, calculate the classification accuracy, model confidence, and the contribution of each decision tree in the classification task of the random forest model.

[0122] Based on the classification accuracy, model confidence, and the contribution of each decision tree in the classification task, calculate the value of the reward function.

[0123] Update the structure parameters and weight parameters of each decision tree in the random forest model according to the value of the reward function.

[0124] Where:

[0125] Classification accuracy: It is a basic criterion to measure whether the model prediction is correct. If the prediction result of the model is consistent with the actual label, it is regarded as a correct prediction, otherwise it is an incorrect prediction. The higher the accuracy, the greater the reward.

[0126] Model confidence: It refers to the confidence level of the model in its prediction results. Usually, when the model performs classification, it will output a probability value indicating the possibility of predicting a certain category. A prediction with high confidence means that the model has a high confidence in this prediction. Correspondingly, the reward will also be higher when the confidence is relatively high.

[0127] Contribution of decision tree: Each decision tree plays a different role in the random forest model. Some decision trees may be more effective in certain specific tasks and have a greater contribution. The contribution of a decision tree can be determined by evaluating the performance of the decision tree or the information gain it generates in the classification task. Trees with greater contribution will receive higher rewards.

[0128] When calculating the value of the reward function, it can be calculated by combining the classification accuracy, model confidence, and the contribution of the decision tree. Among them, the classification accuracy and model confidence will affect the direct reward of each prediction, while the contribution of the decision tree can be used to measure the impact of this decision tree on the final result. Trees with greater contribution will receive higher rewards.

[0129] In this embodiment, through the reward and punishment mechanism of the reinforcement learning strategy, it is possible to promote the trained classification model to more accurately identify and adjust its own learning process. Different from the fixed weights of each decision tree in the traditional random forest model, the weights of each decision tree in the trained classification model of this application are dynamically adjustable, so as to better adapt to different characteristics of the data.

[0130] The classification model training method provided by the embodiments of this application uses a reinforcement learning strategy to train a random forest model. The reinforcement learning strategy drives the learning process of the model through rewards and punishments. When the model successfully classifies a sample into the correct category, the reinforcement learning gives a positive reward, otherwise a negative reward. This process adjusts the parameters and weights of each decision tree in the model, enabling the model to gradually improve its prediction ability. Specifically, in each training iteration, the model calculates the reward based on the prediction results and dynamically updates the model's parameters based on this feedback information. In this way, the decision tree can be continuously optimized, the occurrence of misclassification can be reduced, and the adaptability to complex data sets can be enhanced.

[0131] Data imbalance is a common problem in the field of machine learning. Especially in classification tasks, the classification of minority class samples is often less accurate than that of majority class samples. When facing data imbalance, the traditional random forest model may tend to the majority class, resulting in a decrease in classification accuracy. However, the trained classification model of this application enables the classification model to identify the misclassification of minority class samples during the training process through the reinforcement learning strategy, and then strengthen the learning of these samples. The reinforcement learning strategy can adjust the weights of the decision trees, so that the minority class samples receive more attention during the model training process. This strategy helps to balance the learning of the classification model, avoid the bias of the majority class samples, and thus improve the overall classification accuracy.

[0132] In some embodiments, for the training process of the classification model, cross-validation and model adjustment processing can also be added.

[0133] Cross-validation: K-fold cross-validation (such as 5-fold or 10-fold cross-validation) can be used to evaluate the training effect of each decision tree. Cross-validation helps to avoid overfitting of the decision tree on a specific data partition and evaluate its generalization ability on different data subsets.

[0134] The model adjustment processing can include:

[0135] Hyperparameter optimization: Methods such as Grid Search or Random Search can be used to adjust the hyperparameters of the enhanced random forest model, such as the number of trees, maximum depth, learning rate, minimum sample split number of trees, etc., to achieve the best performance of the model.

[0136] Number of trees: Increasing the number of trees generally improves model stability but also increases computational overhead. Select an optimal value through cross - validation.

[0137] Learning rate adjustment: During the training of the classification model, control the update step size by adjusting the learning rate to avoid unstable training caused by too large a step size or slow convergence caused by too small a step size.

[0138] Use the optimized hyperparameters and reinforcement learning strategies to train the classification model. Conduct multiple iterations of optimization on the training set, update the weights of each decision tree, and finally obtain an enhanced random forest model with better performance as the trained classification model.

[0139] In some embodiments, for the obtained trained classification model, the performance of the splitting model can be evaluated. For example, use standard classification evaluation metrics (such as accuracy, recall, F1 - score, AUC, etc.) to evaluate the classification model and check the performance of the classification model on the validation set.

[0140] In some embodiments, after the classification model training is completed, the final classification model can be persisted to save the structure and weights of each decision tree for subsequent prediction and online learning.

[0141] In some embodiments, to enhance the interpretability of the classification model, model interpretation techniques (such as SHAP values, LIME, etc.) can be used in this embodiment to perform interpretability analysis on the trained classification model to help users understand the decision - making process of the model and identify important features.

[0142] Display the training process and results through visualization tools (such as decision tree visualization, feature importance graph, etc.) to enhance the transparency and credibility of the classification model.

[0143] In some embodiments of the present application, a model integration scheme is further provided.

[0144] After obtaining the trained classification model in the foregoing embodiments, the classification model can be further integrated with at least one other type of base learner configured. The other type of base learner is configured to have the ability of data classification. For example, a support vector machine (SVM), k - nearest neighbor (KNN) neural network, etc. can be used as the other type of base learner.

[0145] For the classification model and each base learner in the integration framework, different integration methods (such as weighted voting, Stacking, etc.) can be used for integration. The specific integration method can be determined according to the nature and performance requirements of the classification task.

[0146] Among them, weighted voting: for classification problems, a weighted voting strategy is adopted. The prediction results of each base learner and classification model are assigned weights according to their performance, and finally the prediction result of the ensemble model is obtained through weighted averaging.

[0147] Stacking: By training a new "meta-learner" to integrate the prediction results of each base learner and classification model, more accurate results can be obtained.

[0148] Among them, the training process of the meta-learner includes:

[0149] Collect the classification results of the classification models and each base learner after training in the above embodiments for the same training sample to form an input sample. Using the classification result label of the same training sample as the sample label corresponding to the input sample, the meta-learner is trained with the input sample and the sample label. So that the meta-learner learns how to give the final classification decision result based on the prediction results of the base learner and the classification model.

[0150] The meta-learner can adopt model architectures such as logistic regression, decision tree, neural network, etc. Optimization algorithms (such as gradient descent, Adam, etc.) can be used during training to adjust the weights of the meta-learner to ensure that it can make full use of the advantages of the base learner and the classification model and improve the accuracy of the overall ensemble model.

[0151] Ensemble model evaluation and tuning:

[0152] Evaluation criteria: This application can use metrics such as accuracy, precision, recall, F1 score, AUC, etc. to evaluate the ensemble model to ensure that the ensemble model has excellent performance in multiple aspects.

[0153] Hyperparameter optimization: Use grid search or random search methods to optimize the hyperparameters in the ensemble model. Mainly optimize the parameters in the ensemble model, such as the number of base learners, the number of training rounds for each base learner, the hyperparameters of the meta-learner, etc.

[0154] Ensemble model validation: Perform multiple cross-validations on the ensemble model to ensure that it can achieve stable performance under different data partitions.

[0155] Model performance comparison:

[0156] Comparison between single model and ensemble model: Compare the performance differences between the classification model and the ensemble model obtained by training the random forest model using the reinforcement learning strategy in the foregoing embodiments, and evaluate whether the ensemble model has obvious advantages in terms of classification accuracy, computational efficiency, generalization ability, etc.

[0157] Multi-task verification: Verify on multiple datasets to check the adaptability of the integrated model to different tasks (such as image classification, text classification, etc.), and verify whether the integrated model has good cross-task generalization ability.

[0158] Model optimization and iterative update:

[0159] Online learning mechanism: Design an online learning mechanism for the integrated model so that the integrated model can dynamically adjust its structure and weights to adapt to new data samples and environmental changes. This is particularly important for processing real-time data streams or making predictions in a production environment.

[0160] Regular update and maintenance: Regularly update the integrated model according to the feedback and performance changes after the deployment of the integrated model. Continuously improve the prediction ability of the integrated model by adding new base learners, updating the meta-learner, etc.

[0161] Deployment of the integrated model: Deploy the final integrated model to the actual application scenario to ensure that the model can perform efficient real-time prediction in the production environment. Optimization in aspects such as model inference time, memory occupancy, and computing resources needs to be considered during the deployment process.

[0162] Performance monitoring of the integrated model: After deployment, establish a real-time monitoring system to track the prediction effect and performance of the integrated model in actual applications. If a performance decline or a change in data distribution is found, retrain or update the integrated model in a timely manner. Through the above steps, the integrated model is completely constructed, and combined with the random forest model (classification model) trained by reinforcement learning and other base learners, it provides a powerful and flexible model optimization and verification solution. This framework can not only improve the accuracy of the model, but also achieve excellent performance in a variety of actual scenarios through flexible tuning of various integration methods and meta-learners.

[0163] II. Inference stage

[0164] Based on the model training stage introduced in the foregoing embodiments, the embodiments of the present application further provide an object classification method, which can be applicable to a variety of classification task scenarios.

[0165] Refer to Figure 2 , Figure 2 to provide a schematic flow diagram of an object classification method.

[0166] As Figure 2 shown, the object classification method may include the following steps:

[0167] Step S200, obtain parameter data of the object to be classified.

[0168] The object classification method of the present application can be applied to a variety of different classification task scenarios. Taking the classification task of application programs as an example, the object to be classified is the application program to be classified. The parameter data may include at least one of the following:

[0169] The description text of the application program, the display image of the application program, the user behavior data related to the application program, etc.

[0170] Step S210: Extract features from the parameter data to obtain parameter features.

[0171] Specifically, the parameter data of the object to be classified may include parameter data of one or more modalities. For different modality parameter data, corresponding feature extraction methods can be used. Among them, the specific implementation of feature extraction for different modality parameter data can refer to the relevant description in the previous classification model training stage, which will not be elaborated here. In this step, through feature extraction, the parameter features of the object to be classified are obtained.

[0172] Step S220: Feed the parameter features into the classification model to obtain a classification result. Among them, the classification model is obtained by training a random forest model using a reinforcement learning strategy, and the weight parameters of each decision tree in the random forest model are updated and fixed during the reinforcement learning training process.

[0173] In the traditional random forest model, since each decision tree assigns the same weight to all samples, minority class samples and difficult-to-classify samples are easily ignored, thus affecting the classification accuracy of the model. In the present application, by using a reinforcement learning strategy to dynamically adjust the weights of each decision tree in the random forest model, more learning opportunities can be obtained on these samples, improving the robustness and classification accuracy of the classification model. In addition, the classification model of the present application adopts a random forest model architecture, which still belongs to a traditional machine learning model and has fewer parameters and lower computational resources required compared to deep learning models.

[0174] The training process of the classification model adopted in step S220 of this embodiment can refer to the relevant introduction in the previous model training stage, which will not be elaborated here.

[0175] In some embodiments, taking the application classification task as an example. The parameter data of an application usually comes from multiple different sources, involving different types and formats. Traditional single data source or unimodal classification methods are difficult to comprehensively mine the rich information in the data, especially when multiple modal information such as the function description of the application, user comments, image data, and user behavior coexist. To better utilize these diverse data sources, this application proposes an integrated classification method for fusing multi-modal data based on an enhanced random forest framework, aiming to improve the classification accuracy and the generalization ability of the model through the fusion of multi-modal data. Feature extraction is performed on the multi-modal parameter data of the application respectively, and the multi-modal features are fused. The fused features are sent into the classification model to obtain the classification result. By fully integrating different types of data (such as text, images, user behavior data, etc.), an ensemble learning model is constructed, and the input information of different modalities can be fused at the feature level or the decision level, thereby improving the classification accuracy.

[0176] In application classification, common modal data includes: Text data: Such as the function description of the application, application name, and user comments, which contain rich semantic information. Image data: Icons, screenshots, or video preview images of the application, etc., which can convey important information about the visual style, function layout, etc. of the application. User behavior data: Including user click behavior, download volume, user rating, etc., which reflect the interaction relationship between the user and the application and the market performance of the application.

[0177] In some embodiments of this application, based on the introduction of the foregoing model training stage, this application also provides a model integration framework, which integrates a trained classification model and at least one other type of base learner.

[0178] On this basis, this embodiment provides another object classification method. On the basis of the classification result output by the object classification method introduced in the foregoing embodiment, the following steps may further be included:

[0179] Send the parameter features obtained in step S210 into at least one other type of base learner to obtain the classification result output by the base learner, where the base learner is configured to have the ability to classify data.

[0180] After obtaining the classification results output by each base learner, the classification results output by the base learner and the classification results output by the classification model can be integrated to determine the final classification result.

[0181] An optional integration method is, for example, to fuse the classification results of the classification model and each base learner through a weighted average or voting mechanism to obtain the final classification result. Among them, the weights of each base learner and the classification model can be assigned corresponding weights according to the performance of the base learner and the classification model, and finally the final classification result of the integrated model is obtained through weighted average.

[0182] Another optional integration method is, for example, to input the classification result output by the classification model and the classification results output by each base learner into a meta-learner to obtain the final classification result output by the meta-learner. The meta-learner is trained with the classification results of the classification model and each base learner for the same training sample as the input sample and the classification result label of the same training sample as the sample label.

[0183] By training the meta-learner, the meta-learner can make full use of the advantages of the base learner and the classification model to improve the accuracy of the overall integrated model.

[0184] Next, the object classification device provided in the embodiments of the present application will be described. The object classification device described below can be correspondingly referred to the object classification method described above.

[0185] See Figure 3 , Figure 3 which is a schematic structural diagram of an object classification device disclosed in the embodiments of the present application.

[0186] As Figure 3 shown, the device may include:

[0187] A parameter acquisition unit 11, configured to acquire parameter data of an object to be classified;

[0188] A feature extraction unit 12, configured to extract features from the parameter data to obtain parameter features;

[0189] A classification prediction unit 13, configured to send the parameter features into a classification model to obtain a classification result, where the classification model is obtained by training a random forest model using a reinforcement learning strategy, and the weight parameters of each decision tree in the random forest model are updated and fixed during the reinforcement learning training process.

[0190] In a possible implementation, the object to be classified is an application program to be classified, and the parameter data includes at least one of the following: the description text of the application program, the display image of the application program, and the user behavior data related to the application program.

[0191] In a possible implementation, the device of the present application further includes:

[0192] A base learner classification unit for sending the parameter features into at least one other type of base learner to obtain the classification results output by the base learner, where the base learner is configured to have the ability to classify data;

[0193] A first classification result fusion unit for fusing the classification results of the classification model and each of the base learners through a weighted average or voting mechanism to obtain a final classification result.

[0194] In a possible implementation, the device of the present application further includes:

[0195] A base learner classification unit for sending the parameter features into at least one other type of base learner to obtain the classification results output by the base learner, where the base learner is configured to have the ability to classify data;

[0196] A second classification result fusion unit for inputting the classification results output by the classification model and the classification results output by each of the base learners into a meta-learner to obtain the final classification result output by the meta-learner. The meta-learner is trained with the classification results of the classification model and each of the base learners for the same training sample as the input sample and the classification result label of the same training sample as the sample label.

[0197] In a possible implementation, the device of the present application further includes:

[0198] A classification model training unit for performing the training process of the classification model, and this training process may include:

[0199] Obtain training samples labeled with classification result labels, and perform feature extraction on the training samples to obtain a set of feature samples;

[0200] Obtain an initial random forest model as the classification model to be trained, and the random forest model includes several decision trees;

[0201] Send the feature samples in the set of feature samples into the random forest model for classification processing to obtain the classification results output by the random forest model, and the classification results are obtained by weighted fusion of the classification results of the decision trees according to the weight parameters of each decision tree;

[0202] According to the classification results output by the random forest model and the classification result labels corresponding to the feature samples, adopt a reinforcement learning strategy to adjust the structural parameters and weight parameters of each decision tree until the set training end condition is reached to obtain the trained classification model.

[0203] In a possible implementation, the parameter data includes multi-modal parameter data; the process of the classification model training unit performing feature extraction on the parameter data to obtain parameter features includes:

[0204] Feature extraction is performed on each type of modal parameter data to obtain the features of each type of modal parameter, and the features of various modal parameters are fused to obtain the fused parameter features.

[0205] In a possible implementation, the process of the classification model training unit performing the training process of the classification model further includes:

[0206] Based on the proportions of the feature samples of each category in the feature sample set, resample the feature sample set to balance the proportions of the feature samples of each category.

[0207] In a possible implementation, the process of the classification model training unit adjusting the structural parameters and weight parameters of each decision tree by using a reinforcement learning strategy according to the classification result output by the random forest model and the classification result label corresponding to the feature sample includes:

[0208] According to the classification result output by the random forest model and the classification result label corresponding to the feature sample, calculate the classification accuracy, model confidence, and contribution degree of each decision tree in the classification task of the random forest model;

[0209] Based on the classification accuracy, model confidence, and contribution degree of each decision tree in the classification task, calculate the value of the reward function;

[0210] Update the structural parameters and weight parameters of each decision tree in the random forest model according to the value of the reward function.

[0211] An electronic device is also provided in an embodiment of the present application. Refer to Figure 4 As shown, it shows a schematic structural diagram of an electronic device suitable for implementing the electronic device in an embodiment of the present application. The electronic device in an embodiment of the present application may include, but is not limited to, terminals such as mobile phones, tablet computers, teaching large screens, wearable devices, servers, and the like. Figure 4 The electronic device shown is only an example and should not bring any limitation to the functions and usage scopes of the embodiments of the present application.

[0212] As Figure 4As shown, the electronic device may include a processing device (such as a central processing unit, a graphics processing unit, etc.) 601, which may perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 602 or the program loaded from the storage device 608 into the random access memory (RAM) 603, so as to implement the object classification method or classification model training method of the foregoing embodiments of the present application. When the electronic device is powered on, various programs and data required for the operation of the electronic device are also stored in the RAM 603. The processing device 601, the ROM 602, and the RAM 603 are connected to each other through a bus 604. The input / output (I / O) interface 605 is also connected to the bus 604.

[0213] Generally, the following devices may be connected to the I / O interface 605: an input device 606 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 607 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 608 including, for example, a memory card, a hard disk, etc.; and a communication device 609. The communication device 609 may allow the electronic device to communicate with other devices wirelessly or wiredly to exchange data. Although Figure 4 an electronic device with various devices is shown, it should be understood that it is not required to implement or have all the shown devices. Instead, more or fewer devices may be implemented or had.

[0214] An embodiment of the present application also provides a computer program product including computer-readable instructions. When the computer-readable instructions run on an electronic device, the electronic device is enabled to implement any object classification method or classification model training method provided by the embodiments of the present application.

[0215] An embodiment of the present application also provides a computer-readable storage medium. The storage medium carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can be enabled to implement any object classification method or classification model training method provided by the embodiments of the present application.

[0216] In addition, it should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the drawings of the device embodiments provided in the present application, the connection relationships between the modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines.

[0217] Through the description of the above embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software plus necessary general hardware. Of course, it can also be implemented by dedicated hardware including application-specific integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. Generally, functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structures for implementing the same function can also be various, such as analog circuits, digital circuits, or dedicated circuits, etc. However, for the present application, in more cases, software program implementation is a better implementation method. Based on such an understanding, the technical solution of the present application, in essence, or the part that makes a contribution to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disc of a computer, etc., and includes several instructions to enable a computer device (which can be a personal computer, training device, or network device, etc.) to execute the methods described in various embodiments of the present application.

[0218] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product.

[0219] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general computer, a dedicated computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from a website, computer, training device, or data center to another website, computer, training device, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or a wireless manner (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can store, or a data storage device such as a training device or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0220] The various embodiments in this specification are described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

Claims

1. A method for object classification, characterized in that, Including: Obtain parameter data of the object to be classified; Extract features from the parameter data to obtain parameter features; Send the parameter features into a classification model to obtain a classification result, where the classification model is obtained by training a random forest model using a reinforcement learning strategy, and the weight parameters of each decision tree in the random forest model are updated and fixed during the reinforcement learning training process.

2. The method according to claim 1, wherein The training process of the classification model includes: Obtain training samples labeled with classification result labels, and extract features from the training samples to obtain a set of feature samples; Obtain an initial random forest model as the classification model to be trained, and the random forest model includes several decision trees; Send the feature samples in the set of feature samples into the random forest model for classification processing to obtain a classification result output by the random forest model, and the classification result is obtained by weighted fusion of the classification results of the decision trees according to the weight parameters of each decision tree; According to the classification result output by the random forest model and the classification result label corresponding to the feature sample, adopt a reinforcement learning strategy to adjust the structure parameters and weight parameters of each decision tree until a set training end condition is reached to obtain a trained classification model.

3. The method according to claim 2, wherein The parameter data includes multi-modal parameter data; Then the process of extracting features from the parameter data to obtain parameter features includes: Extract features from each type of modal parameter data respectively to obtain each type of modal parameter feature, and fuse various modal parameter features to obtain a fused parameter feature.

4. The method according to claim 2, wherein Also included: Based on the proportion of feature samples of each category in the set of feature samples, resample the set of feature samples to balance the proportion of feature samples of each category.

5. The method according to claim 2, characterized in that, The process of adopting a reinforcement learning strategy to adjust the structure parameters and weight parameters of each decision tree according to the classification result output by the random forest model and the classification result label corresponding to the feature sample includes: According to the classification result output by the random forest model and the classification result label corresponding to the feature sample, calculate the classification accuracy, model confidence level of the random forest model, and the contribution degree of each decision tree in the classification task; Based on the classification accuracy, model confidence level, and the contribution degree of each decision tree in the classification task, calculate the value of the reward function; Update the structure parameters and weight parameters of each decision tree in the random forest model according to the value of the reward function.

6. The method according to claim 1, characterized in that, The object to be classified is an application program to be classified, and the parameter data includes at least one of the following: the description text of the application program, the display image of the application program, and the user behavior data related to the application program.

7. The method according to any one of claims 1-6, characterized in that, Also included: Send the parameter features into at least one other type of base learner to obtain a classification result output by the base learner, where the base learner is configured to have the ability to classify data; Fuse the classification results of the classification model and each base learner through a weighted average or voting mechanism to obtain a final classification result.

8. The method according to any one of claims 1 to 6, characterized in that, Also included: Send the parameter features into at least one other type of base learner to obtain a classification result output by the base learner, where the base learner is configured to have the ability to classify data; The classification results output by the classification model and the classification results output by each of the base learners are input into a meta-learner to obtain the final classification result output by the meta-learner. The meta-learner is trained with the classification results of the classification model and each of the base learners for the same training sample as the input samples and the classification result label of the same training sample as the sample label.

9. A method for training a classification model, characterized in that, It includes: Obtain training samples labeled with classification result labels, and perform feature extraction on the training samples to obtain a set of feature samples; Obtain an initial random forest model as the classification model to be trained, and the random forest model includes a number of decision trees; Send the feature samples in the set of feature samples into the random forest model for classification processing to obtain the classification results output by the random forest model, and the classification results are obtained by weighted fusion of the classification results of the decision trees according to the weight parameters of each decision tree; According to the classification results output by the random forest model and the classification result labels corresponding to the feature samples, adopt a reinforcement learning strategy to adjust the structural parameters and weight parameters of each decision tree until the set training end condition is reached to obtain the trained classification model.

10. An object classification device, characterized in that, It includes: A parameter acquisition unit for acquiring parameter data of an object to be classified; A feature extraction unit for performing feature extraction on the parameter data to obtain parameter features; A classification prediction unit for sending the parameter features into a classification model to obtain a classification result, where the classification model is obtained by training a random forest model using a reinforcement learning strategy, and the weight parameters of each decision tree in the random forest model are updated and fixed during the reinforcement learning training process.

11. An electronic device, characterized in that, It includes: A memory and a processor; The memory is used to store programs; The processor is used to execute the program to implement each step of the object classification method according to any one of claims 1 to 8, or to implement each step of the classification model training method according to claim 9.

12. A readable storage medium, on which a computer program is stored, characterized in that, When the computer program is executed by the processor, it implements each step of the object classification method according to any one of claims 1 to 8, or implements each step of the classification model training method according to claim 9.

13. A computer program product, comprising a computer program, characterized in that, When this computer program is executed by the processor, it implements each step of the object classification method according to any one of claims 1 to 8, or implements each step of the classification model training method according to claim 9.

Images