Object attribute identification method, apparatus, device and storage medium

Abstract

This application discloses an object attribute recognition method, apparatus, device, and storage medium, which can be applied to various scenarios such as cloud technology, artificial intelligence, intelligent transportation, and the Internet of Vehicles. The method includes: acquiring features of an object to be identified; performing object attribute recognition processing on the features of the object to be identified based on an object attribute recognition model to obtain target attribute labels corresponding to the features of the object to be identified; wherein, the training method of the object attribute recognition model includes: acquiring sample object features; pre-training the sample object features based on a preset network to obtain a first attribute recognition model; obtaining an initial attribute recognition model based on the first attribute recognition model and a first output result; and training the object attribute recognition model by updating the loss function of the initial attribute recognition model based on an objective function. The object attribute recognition model of this application improves the accuracy of object attribute recognition.

Description

Technical Field

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

[0002] In related technologies, data rules are typically determined by human experience to identify object attribute labels. Product operations, based on business experience, set rules for identifying user lifestyles; for example, users "aged 25-35, residing in first-tier cities, and in specific technical fields" are considered more likely to be "homebodies." This method of determining data rules based on human experience not only uses a very limited number of rules but also fails to capture the high-dimensional feature information of interactions between rules. Most importantly, it cannot determine the optimal parameters for each rule, resulting in low accuracy in identifying object attributes.

[0003] Therefore, it is necessary to provide an object attribute identification method, apparatus, device, and storage medium to improve the accuracy of object attribute identification. Summary of the Invention

[0004] This application provides a method, apparatus, device, and storage medium for object attribute identification, which can improve the accuracy of object attribute identification.

[0005] On the one hand, this application provides a method for identifying object attributes, the method comprising:

[0006] Obtain the features of the object to be identified;

[0007] Based on the object attribute recognition model, the features of the object to be identified are processed to obtain the target attribute label corresponding to the features of the object to be identified.

[0008] The training method for the object attribute recognition model includes:

[0009] Obtain sample object features; the sample object features are labeled with sample attribute tags, and the sample attribute tags characterize the sample object attributes corresponding to the sample object features;

[0010] The first attribute recognition model is obtained by pre-training the features of the sample object based on a preset network.

[0011] Based on the first attribute recognition model and the first output result, an initial attribute recognition model is obtained; the first output result is the output result of the target pooling layer corresponding to the first attribute recognition model when training converges;

[0012] An updated model is obtained by updating the loss function of the initial attribute recognition model based on the objective function, and the updated model is used to train the sample object features to obtain the object attribute recognition model.

[0013] On the other hand, an object attribute recognition device is provided, the device comprising:

[0014] The object feature acquisition module is used to acquire the features of the object to be identified;

[0015] The attribute tag recognition module is used to perform object attribute recognition processing on the features of the object to be identified based on the object attribute recognition model, so as to obtain the target attribute tag corresponding to the features of the object to be identified.

[0016] The model training module is used to train the object attribute recognition model;

[0017] The model training module includes:

[0018] The sample object feature acquisition submodule is used to acquire sample object features; the sample object features are labeled with sample attribute tags, and the sample attribute tags represent the sample object attributes corresponding to the sample object features.

[0019] The first attribute recognition model determination submodule is used to pre-train the features of the sample object based on a preset network to obtain the first attribute recognition model;

[0020] The initial attribute recognition model determination submodule is used to obtain an initial attribute recognition model based on the first attribute recognition model and the first output result; the first output result is the output result of the target pooling layer corresponding to the first attribute recognition model when training converges;

[0021] The object attribute recognition model determination submodule is used to update the loss function of the initial attribute recognition model based on the objective function to obtain an updated model, and to train the sample object features according to the updated model to obtain the object attribute recognition model.

[0022] On the other hand, an object attribute recognition device is provided, the device including a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the object attribute recognition method as described above.

[0023] On the other hand, a computer storage medium is provided, which stores at least one instruction or at least one program, which is loaded and executed by a processor to implement the object attribute identification method as described above.

[0024] On the other hand, a computer program product or computer program is provided, which includes 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 object attribute identification method as described above.

[0025] The object attribute identification method, apparatus, device, and storage medium provided in this application have the following technical advantages:

[0026] This application obtains the features of an object to be identified; based on an object attribute recognition model, it performs object attribute recognition processing on the features of the object to be identified to obtain the target attribute label corresponding to the features of the object to be identified; wherein, the training method of the object attribute recognition model includes: obtaining sample object features; the sample object features are labeled with sample attribute labels, the sample attribute labels representing the sample object attributes corresponding to the sample object features; pre-training the sample object features based on a preset network to obtain a first attribute recognition model; obtaining an initial attribute recognition model based on the first attribute recognition model and a first output result; the first output result is the output result of the target pooling layer corresponding to the first attribute recognition model when training converges; updating the loss function of the initial attribute recognition model based on an objective function to obtain an updated model, and training the sample object features according to the updated model to obtain the object attribute recognition model. This application obtains an initial attribute recognition model with high accuracy based on the first attribute recognition model and the first output result; then updates the loss function of the initial attribute recognition model based on an objective function to obtain an updated model, and continues training to obtain an object attribute recognition model that can accurately identify object attributes; thereby improving the accuracy of object attribute recognition. Attached Figure Description

[0027] To more clearly illustrate the technical solutions and advantages 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.

[0028] Figure 1 This is a schematic diagram of an object attribute recognition system provided in an embodiment of this application;

[0029] Figure 2 This is a flowchart illustrating an object attribute identification method provided in an embodiment of this application;

[0030] Figure 3This is a flowchart illustrating a method for obtaining features of a sample object according to an embodiment of this application;

[0031] Figure 4 This is an example of a method for determining an initial attribute recognition model provided in an embodiment of this application;

[0032] Figure 5 This is a second example of a method for determining an initial attribute recognition model provided in the embodiments of this application;

[0033] Figure 6 This is a flowchart illustrating a method for training an initial attribute recognition model according to an embodiment of this application;

[0034] Figure 7 This is a flowchart illustrating a verification method for an object attribute recognition model provided in an embodiment of this application;

[0035] Figure 8 This is a flowchart illustrating a method for determining the features of a target sample object based on sorting results, provided in an embodiment of this application.

[0036] Figure 9 This is a schematic diagram of the structure of the DenseNet model provided in the embodiments of this application;

[0037] Figure 10 This is a schematic diagram illustrating the label recognition principle of the DenseNet model provided in the embodiments of this application;

[0038] Figure 11 This is a schematic diagram comparing the AUC of different models provided in the embodiments of this application for predicting user lifestyles;

[0039] Figure 12 This is a schematic diagram comparing the effects of different models provided in the embodiments of this application on predicting user lifestyles;

[0040] Figure 13 This is a schematic diagram of the structure of an object attribute recognition device provided in an embodiment of this application.

[0041] Figure 14 This is a schematic diagram of the structure of a server provided in an embodiment of this application. Detailed Implementation

[0042] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0043] This application provides an object attribute identification method, apparatus, device, and storage medium. Specifically, the object attribute identification method of this application can be executed by a computer device, which can be a terminal or server, etc. This application can be applied to various scenarios such as data security, cloud technology, artificial intelligence, and smart transportation.

[0044] First, some of the nouns or terms that appear in the description of the embodiments of this application are explained as follows:

[0045] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0046] Intelligent transportation fully utilizes next-generation information technologies such as the Internet of Things, spatial sensing, cloud computing, and mobile internet across the entire transportation sector. It comprehensively applies theories and tools from transportation science, systems methods, artificial intelligence, and knowledge mining. With the goals of comprehensive perception, deep integration, proactive service, and scientific decision-making, it builds a real-time dynamic information service system, deeply mines transportation-related data, forms problem analysis models, and enhances the industry's ability to optimize resource allocation, improve public decision-making capabilities, enhance industry management capabilities, and improve public service capabilities. This promotes safer, more efficient, more convenient, more economical, more environmentally friendly, and more comfortable operation and development of transportation, and drives the transformation and upgrading of transportation-related industries.

[0047] 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.

[0048] Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0049] Deep Learning (DL) is a branch of machine learning that attempts to perform high-level abstractions of data using multiple processing layers with complex structures or multiple nonlinear transformations. Deep learning learns the inherent patterns and hierarchical representations of the features of sample objects. The information gained during this learning process greatly aids in interpreting data such as text, images, and sound. The ultimate goal of deep learning is to enable machines to possess analytical and learning capabilities similar to humans, allowing them to recognize data such as text, images, and sound. Deep learning is a complex machine learning algorithm, and its performance in speech and image recognition far surpasses previous related technologies.

[0050] Neural Network (NN): A deep learning model in the fields of machine learning and cognitive science that mimics the structure and function of biological neural networks.

[0051] Representation learning: In the field of machine learning, representation learning (or feature learning) is a set of techniques that transform raw data into something that can be effectively utilized by machine learning. Before the advent of feature learning algorithms, machine learning researchers needed to use techniques such as manual feature engineering to build features from the domain knowledge of the raw data before deploying relevant machine learning algorithms. While manual feature engineering is effective for applying machine learning, it is also difficult, expensive, time-consuming, and relies on strong expertise. Feature learning fills this gap, enabling machines not only to learn the features of data but also to use those features to complete a specific task.

[0052] Metric learning: The objects are typically the distances between sample feature vectors. The goal of metric learning is to reduce or limit the distances between samples of the same class, while increasing the distances between samples of different classes, through training and learning. Metric learning studies how to learn a distance function for a specific task, such that this distance function helps nearest neighbor-based algorithms (kNN, k-means, etc.) achieve better performance.

[0053] User lifestyle: refers to the user's current long-term lifestyle. Values ​​include: workaholic, super-lifestyle enthusiast, and homebody.

[0054] The k-nearest neighbors algorithm is one of the most fundamental methods used for classification and regression. When the distribution of the training data is unknown, the k-nearest neighbors algorithm is often the first approach that comes to mind. The k-nearest neighbors algorithm takes the feature vectors of instances as input and outputs the class of the instance. The algorithm's core idea is that, given a training dataset and corresponding data points in the input space, to determine the class of a new data point, it takes the k nearest neighbors to the target data point, counts the percentage of each class among these k neighbors, and selects the class with the highest percentage as the target data point's class. The k-nearest neighbors algorithm typically uses Euclidean distance to represent the distance between two data points; the smaller the distance, the closer the two data points are.

[0055] Company A's WiFi Manager: Company A's WiFi Manager is a free WiFi management software under Company A. It supports hundreds of millions of public WiFi hotspots, allowing one-click connection without entering a password. It also conducts a comprehensive evaluation of these hotspots based on Company A's five-star WiFi standard, assessing their security, connection speed, and network quality to ensure that there are no zombie viruses, no risks, and no fake WiFi networks.

[0056] AUC (Area Under the ROC Curve) is an important and commonly used metric for measuring the classification performance of machine learning models, but it can only be used for binary classification. AUC considers the quality of the model's predictions, reflecting the proportion of positive examples that the model ranks before negative examples (if AUC = 1, it means the model ranks all positive examples before negative examples 100%). ROC stands for Receiver Operating Characteristic curve, which is plotted with the false positive rate (FPR) on the horizontal axis and the true positive rate (TPR) on the vertical axis. The closer the ROC curve is to the upper left corner, the better the model's performance. If the ROC curve of model A completely "encloses" the ROC curve of model B, it can be asserted that model A performs better than model B. However, the ROC curves of two models often intersect, and in this case, AUC is needed to compare performance.

[0057] It should be noted that the terms "first," "second," etc., in the specification, claims, 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 in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server 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 devices.

[0058] This application's embodiments establish an automated recognition system for "user lifestyle prediction" based on deep learning technology, which is accurate, effective, and highly reusable. User lifestyles can include categories such as "homebody," "workaholic," and "superhuman." Theoretically, by simply adjusting the characteristics of sample objects from other scenarios, the recognition probability of the corresponding scenario can be determined, such as "user's technical field" or "user's highest education level." Identifying user lifestyles directly helps improve the refined operation of app traffic.

[0059] It should be noted that the training and prediction processes of the object attribute recognition model can be completed either on the server or on the terminal. When both training and prediction are completed on the server, the set of objects to be recognized can be input into the server. After the server completes the prediction, it sends the predicted label information of the objects to be recognized to the terminal for display. Alternatively, the trained model file (model file) can be ported to the terminal. If prediction is needed on the input set of objects to be recognized, the set can be input into the trained model file (model file), and the predicted label information of the objects can be obtained through calculation.

[0060] Please see Figure 1 , Figure 1 This is a schematic diagram of an object attribute recognition system provided in an embodiment of this application, such as... Figure 1As shown, the object attribute recognition system may include at least server 01 and client 02.

[0061] Specifically, in this embodiment, server 01 may include a standalone server, a distributed server, or a server cluster composed of multiple servers. It may also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Server 01 may include network communication units, processors, and memory, etc. Specifically, server 01 can be used to train an object attribute recognition model and perform object attribute recognition processing on the features of the object to be recognized based on the model to obtain the target attribute label corresponding to the features of the object to be recognized.

[0062] Specifically, in this embodiment, the client 02 may include physical devices such as smartphones, desktop computers, tablets, laptops, digital assistants, smart wearable devices, smart speakers, in-vehicle terminals, and smart TVs. It may also include software running on the physical device, such as web pages provided to users by service providers, or applications provided by those service providers. Specifically, the client 02 can be used to query online the target attribute tags corresponding to the features of the object to be identified.

[0063] The following describes an object attribute recognition method according to this application. Figure 2 This is a flowchart illustrating an object attribute identification method provided in an embodiment of this application. This specification provides the operational steps of the method described in the embodiments or flowcharts, but based on conventional or non-inventive labor, more or fewer operational steps may be included. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual system or server product execution, the method can be executed sequentially or in parallel (e.g., in a parallel processor or multi-threaded processing environment) as shown in the embodiments or accompanying drawings. Specifically, as shown... Figure 2 As shown, the method may include:

[0064] S201: Obtain the features of the object to be identified.

[0065] In this embodiment of the application, the features of the object to be identified can be the features corresponding to the target object, and the features of the object to be identified and the features of the sample object are of the same category; the features of the object to be identified can be one or more, and the feature type can be determined according to the attribute label to be predicted; for example, if the attribute label to be predicted is the user's lifestyle, then the features of the object to be identified are features associated with the user's lifestyle; if the attribute label to be predicted is the user's highest education level, then the features of the object to be identified are features associated with education level.

[0066] S203: Based on the object attribute recognition model, perform object attribute recognition processing on the features of the object to be identified to obtain the target attribute label corresponding to the features of the object to be identified.

[0067] In the embodiments of this application, the object attribute recognition model can be pre-trained, or the object attribute recognition model can be generated in real time after the features of the object to be recognized are obtained, and the target attribute label can be predicted.

[0068] The training method for the object attribute recognition model includes:

[0069] S301: Obtain sample object features; the sample object features are labeled with sample attribute tags, and the sample attribute tags represent the sample object attributes corresponding to the sample object features;

[0070] In this embodiment, the sample attribute tags can include various types, specifically determined based on the attributes of the sample object. Object attributes can include a user's lifestyle, highest educational level, or technical field. Specifically, a user's lifestyle can include three types: workaholic, high-achieving individual, and homebody. Identification information can be used as attribute tags to represent these three lifestyles. For example, the sample attribute tags can be the numbers 1, 2, and 3, where 1 represents a workaholic, 2 represents a high-achieving individual, and 3 represents a homebody.

[0071] In this embodiment, based on manual annotation and business experience, sample objects that are strongly relevant to the business, have normal data distribution, and reasonable object association data can be identified from the original sample objects. Seed objects are then selected from the sample objects, and abnormal seed objects are filtered out, thereby obtaining the sample objects. For example, the object can be a user, and the object attribute can be the user's lifestyle.

[0072] In the embodiments of this application, such as Figure 3 As shown, obtaining sample object features as described above may include:

[0073] S30101: Based on recall rules or business logic, obtain a seed object with label information from the original sample object;

[0074] In this embodiment, the recall rule includes the fact that the terminal device used by the object has applications installed in the application whitelist, and the business logic includes the object's activity trajectory. For example, a batch of candidate seed objects is roughly recalled based on the recall rule, then the candidate seed objects are filtered by manual screening, and finally the candidate seed objects are verified based on the business logic to obtain seed objects with tag information. For example, the tag information can be the object's lifestyle tag, which can include 1, 2, 3; where 1 represents a workaholic, 2 represents a superhuman, and 3 represents a homebody; the tag information can also be the object's highest education level tag, which can include undergraduate, master's, and doctoral degrees; the tag information can also be the object's technical field tag, which can include teachers, doctors, civil servants, etc. When the tag information is the object's lifestyle tag (superhuman), the recall rule can be that the terminal device used by the object has applications installed in the application whitelist, such as food shopping apps, clothing shopping apps, utility payment apps, parenting apps, financial management apps, etc. For example, a batch of candidate seed objects can be roughly recalled based on the terminal device used by the object to install a financial management app. This business logic can include the object's activity trajectory in the target area, such as the object must frequently appear in supermarkets or shopping malls near its residence. If a candidate seed object has never been to supermarkets or similar places, it needs to be removed to obtain seed objects with tag information, thereby obtaining the sample object characteristics.

[0075] S30103: Obtain the basic association data of the seed object;

[0076] In this embodiment, the basic association data includes non-privacy behavior data of the seed object within the target application set. For example, the target application set can be a series of apps from Company A. The basic association data includes some non-privacy behavior data of the seed object within these apps, such as whether Company A's WiFi Manager is installed, whether Company A's Mobile Manager is installed, whether Company A's Mobile Manager's harassment blocking function is used, and whether the time the object spends using a specific app reaches a certain time threshold. This basic association data of the seed object is not all of its basic association data features, but only includes some of them. It is used to evaluate the quality of the seed object. Generally, the basic association data of the seed object does not undergo cross-processing in the time dimension. After obtaining the basic association data of the seed object, it is generally determined whether the basic association data meets specific anomaly indicator conditions to verify whether the values ​​of the seed object in certain association data dimensions are outliers. For example, if candidate seed object 'a' with the tag 'homebody' spends more than 8 hours a day browsing a certain official website, then they may not actually be a homebody, but rather working from home. Therefore, candidate seed object 'a' can be classified as an abnormal seed object. All the aforementioned basic association data was obtained with the user's consent, and the corresponding data use is within the scope of laws and regulations.

[0077] S30105: Calculate the abnormal object type evaluation index based on the basic association data of the seed object;

[0078] In this embodiment, the abnormal object type evaluation index is used to evaluate whether there are abnormal seed objects among the seed objects. In real business scenarios, there may be fake objects or situations where a computer controls a mobile phone. In order to eliminate the impact of non-real objects on modeling and analysis, an abnormal object type evaluation index will be set based on business experience. For example, the abnormal object type evaluation index includes, but is not limited to: the traffic usage of the object in the target application set (A company's series of products), the time distribution of traffic generation, etc.

[0079] S30107: Based on the abnormal object type evaluation index and the distribution anomaly theorem, abnormal seed objects are filtered out from the seed objects to obtain updated seed objects with label information, thus obtaining sample objects.

[0080] In this embodiment, the distribution anomaly theorem may include the Raida criterion, which can be used as a criterion for judging outliers. The Raida criterion, also known as the 3σ criterion, assumes that a set of detection data contains only random errors. The standard deviation is calculated, and an interval is determined with a certain probability. Errors exceeding this interval are considered gross errors, not random errors, and data containing such errors should be discarded. In a normal distribution, σ represents the standard deviation, and μ represents the mean. x = μ is the axis of symmetry of the image. The 3σ principle states that the probability of a value falling within (μ-σ, μ+σ) is 0.6827; the probability of a value falling within (μ-2σ, μ+2σ) is 0.9544; and the probability of a value falling within (μ-3σ, μ+3σ) is 0.9974. It can be considered that the values ​​of Y are almost entirely concentrated within the (μ-3σ, μ+3σ) interval, with the probability of exceeding this range being less than 0.3%. When filtering outlier seed objects, these outlier seed objects can be directly deleted from the candidate sample objects.

[0081] S30109: Based on the sample object, obtain the sample object features.

[0082] In this embodiment of the application, the sample object features corresponding to the sample object can be obtained, thereby enabling the next step of model training.

[0083] In this embodiment, basic associated data features of the sample object can be constructed based on the sample object's characteristics, and a high-dimensional feature vector can be generated based on the sample object's business vertical type features, combined with the time dimension and different feature processing methods. For example, this feature processing can be offline feature processing or online feature processing.

[0084] Optionally, before model training, multiple sample object features can be combined into a sample object feature set; the sample object feature set can be processed into a feature vector to obtain the feature vector of each sample object feature in the sample object feature set, and the model can be trained based on the feature vector.

[0085] The step of performing feature vector processing on the sample object feature set to obtain the feature vector of each sample object feature in the sample object feature set includes:

[0086] Construct the basic associated data features for each feature of the sample object in the sample object feature set.

[0087] In this embodiment of the application, the construction of the basic associated data features for each sample object feature in the sample object feature set includes:

[0088] S1: Based on the historical behavior data of each sample object feature in the sample object feature set, construct the basic association data features of each sample object feature in the sample object feature set, wherein the basic association data features include basic object attributes, basic device attributes, and network connection attributes.

[0089] Specifically, based on the historical behavior data of each sample object feature in the sample object feature set, basic associated data features are constructed for each sample object feature in the sample object feature set. These basic associated data features may include: basic object attributes, basic device attributes, and network connection attributes. For example, basic object attributes may include gender, age group (1-10 years old, 11-20 years old, etc.). Basic device attributes may include the brand and model of the terminal device used by the object. Network connection attributes may include the number of times the object connects to Wi-Fi per day and the Wi-Fi coverage area.

[0090] S2: Based on the business characteristics of each sample object feature in the sample object feature set, construct the business vertical type feature of each sample object feature in the sample object feature set;

[0091] In this embodiment, business characteristics may include the presence of apps on the whitelist installed on the terminal device used by the object. When the tag information is the object's lifestyle tag (Life Superman), the whitelist includes food shopping apps, clothing shopping apps, utility payment apps, parenting apps, financial management apps, etc. The business vertical type characteristics may include the object's click-through rate and conversion rate for specific types of advertisements. If the specific type of advertisement is an e-commerce advertisement, the business vertical type characteristics may also include product purchase rate, etc. All acquisition of the aforementioned apps has been with the user's consent, and the corresponding data usage is within the scope of laws and regulations.

[0092] S3: Combine the basic association data features and business vertical type features of each sample object feature in the sample object feature set with a preset time dimension to aggregate the aggregated basic association data features and aggregated business vertical type features with different time dimensions.

[0093] In this embodiment, the aggregated basic related data features of the object over the past six months / three months / one month / one week are calculated. For example, the aggregation method can be summation, median, standard deviation, etc. For example, if the business vertical type feature of the object using a specific App is feature A, the preset time dimensions include four time dimensions: the past day, the past week, the past month, and the past three months. After cross-processing feature A with the four time dimensions, it becomes four features: A1, A2, A3, and A4. Then, the features corresponding to the above four different time dimensions (A1, A2, A3, and A4) are aggregated to obtain the aggregated business vertical type feature A'. Here, the basic related data features and business vertical type features are the basic features of the sample object features. In order to enrich some features in the basic related data features and business vertical type features that can be combined with the time dimensions, it is necessary to combine the time dimensions to aggregate the aggregated basic related data features and aggregated business vertical type features with different time dimensions.

[0094] S4: Perform feature vector processing on the basic associated data features and business vertical type features before and / or after aggregation to obtain the first feature information of each sample object feature in the sample object feature set;

[0095] In this embodiment of the application, the feature vector processing of the basic associated data features and business vertical type features before and / or after aggregation includes:

[0096] The numerical features in the basic associated data features and business vertical type features before and / or after aggregation are normalized; the non-numerical features in the basic associated data features and business vertical type features before and / or after aggregation are discretized.

[0097] Specifically, the numerical features in the basic associated data features and business vertical type features before and / or after aggregation are normalized to map the original feature values ​​of the relevant features to a certain range (e.g., [0, 1] or [-1, 1]), thereby eliminating the adverse effects caused by singular sample object feature data. For example, the non-numerical features in the basic associated data features and business vertical type features before and / or after aggregation are discretized, which can transform the relevant features into features that are easier for the model to recognize. For example, age group is a numerical feature, and gender is a non-numerical feature.

[0098] For example, Gaussian normalization can be chosen as the normalization method.

[0099] For example, discretization can include the following methods, and the appropriate method can be selected based on the characteristics of the feature itself:

[0100] (1) One-Hot Encoding. For example, for features such as the gender of an object, One-Hot Encoding is used for discretization.

[0101] (2) Count Encoding. For example, for the WiFi map Point of Interest (POI) feature of an object, Count Encoding is used to identify the object's degree of interest in this POI. For example, the object visited the POI "Food - Chinese Food - Cantonese Food" a total of 3 times in the current week.

[0102] (3) Category Embedding. Data analysis revealed that many category features exhibit strong sparsity. To avoid model overfitting and improve model stability, neural networks were introduced for Category Embedding to transform high-dimensional sparse categorical variables into low-dimensional dense embedding variables.

[0103] (4) Missing value embedding (NaN Embedding). For handling missing values ​​of features, we have experimented with methods such as "removal", "mean imputation" and "missing label". The experimental results show that using NaN Embedding to convert missing values ​​into embedding representations has the greatest positive benefit to the model performance.

[0104] (5) Consolidation Encoding. Multiple values ​​under certain category variables can be summarized into the same information. For example, the system version feature of an Android phone includes three values: "4.2", "4.4", and "5.0". Based on experience, these three values ​​can be summarized as "lower version Android system". Experiments have shown that the Consolidation Encoding method brings greater positive benefits than directly one-hot encoding the "Android system version" feature.

[0105] S5: Process the category features in the first feature information of each sample object feature in the sample object feature set to obtain the second feature information of each sample object feature in the sample object feature set;

[0106] In this embodiment of the application, the second feature information includes embedded features, which include WiFi trajectory embedded features and App traffic embedded features.

[0107] Specifically, the category features from the first feature information are input into a DNN model for training to obtain the second feature information, which includes embedding features. The task of deep learning is to map high-dimensional raw data (such as images or sentences) to a low-dimensional manifold, making the high-dimensional raw data separable after mapping to the low-dimensional manifold. This mapping is called embedding. Embedding features can be understood as extracting the low-dimensional, dense weight features from the fully connected layers of the neural network to replace the original feature representation. The dimension is typically a power of 2, such as 128. The steps for training embedding features may include:

[0108] (1) Obtaining WiFi Trajectory Embedding Features. Based on the MST-CNN deep learning network, embedding is performed on the WiFi connection trajectory data of the object to capture the pattern information of the object's Wi-Fi behavior. MST (multiple spanning tree) is an IEEE (Institute of Electrical and Electronics Engineers) standard implemented from Cisco's proprietary Multiple Instance Spanning Tree Protocol (MISTP). Convolutional Neural Networks (CNNs) are a class of feedforward neural networks with deep structures that include convolutional computations, and are one of the representative algorithms of deep learning. By performing embedding, discrete variables can be transformed into continuous vectors, which can also be used to visualize the relationships between different discrete variables. For example, pattern information can be used to represent the different expressions of different types of objects in Wi-Fi connections. This expression is often implicitly reflected in embedding. By embedding the Wi-Fi connection trajectory data of an object, the pattern information of the object's Wi-Fi behavior can be captured, such as pattern information like "homebody type" or "social type".

[0109] (2) Obtain App Traffic Embedding Features. Based on the List-Embedding method, for an object, traffic from different categories of apps within the target application set (e.g., Company A's series) is used to extract embeddings using behavioral sequences. For example, using the traffic embeddings of Company A's series of social apps, low-dimensional and dense object behavioral features are obtained. List-Embedding is an embedding method that, unlike other embedding methods, concatenates multiple input values ​​into a list.

[0110] S6: Merge the first feature information and the second feature information of each sample object feature in the sample object feature set to obtain the feature vector of each sample object feature in the sample object feature set.

[0111] In this embodiment of the application, merging the first feature information and the second feature information of each sample object feature in the sample object feature set to obtain the feature vector of each sample object feature in the sample object feature set includes:

[0112] The low-dimensional and high-dimensional features in the first and second feature information are concatenated column by column to obtain the feature vector of each sample object feature in the sample object feature set.

[0113] After merging the first and second feature information of each sample object in the sample object feature set to obtain the feature vector of each sample object feature in the sample object feature set, the feature vector of each sample object feature in the obtained sample object feature set can be stored offline in the Hadoop Distributed File System (HDFS) for easy access in subsequent processes. HDFS is used for file storage and can achieve large-capacity, high-speed, and reliable data storage and access through a large-scale distributed server cluster. During the merging process, the first feature information of each sample object feature in the processed sample object feature set can be concatenated column-wise with the low-dimensional and high-dimensional features in the second feature information to obtain the feature vector of each sample object feature in the sample object feature set.

[0114] In some embodiments, after step S6, the method may further include:

[0115] S7: Solidify the feature processing logic, perform scheduled offline automated calculations, and push the offline calculation results to the online storage engine. Specifically, solidify the feature processing logic code and run the code on a scheduled basis.

[0116] S303: Pre-train the features of the sample object based on a preset network to obtain a first attribute recognition model;

[0117] In this embodiment, the preset network can be DenseNet; DenseNet (Dense Convolutional Network) is a convolutional neural network with dense connections. In this network, there is a direct connection between any two layers. That is, the input of each layer is the union of the outputs of all preceding layers, and the feature map learned by that layer is directly passed to all subsequent layers as input. Its characteristic is the dense connection between all preceding layers and subsequent layers, and feature reuse is achieved through connections in the channel dimension of features. These characteristics allow DenseNet to achieve better performance than other network structures with fewer parameters and computational costs. In a traditional convolutional neural network, if there are L layers, there will be L connections; however, in DenseNet, there will be L(L+1) / 2 connections. Simply put, the input of each layer comes from the outputs of all preceding layers. The structure of the DenseNet model is as follows: Figure 9 As shown, the input of H1 is x0, the input of H2 is x0 and x1 (x1 is the output of H1), the input of H3 is x0, x1 and x2 (x2 is the output of H2), and the input of H4 is x0, x1, x2 and x3 (x3 is the output of H3).

[0118] Compared to other network architectures, DenseNet proposes a more aggressive dense connection mechanism: all layers are interconnected, meaning each layer receives all preceding layers as additional input. Therefore, in DenseNet, each layer is concatenated with all preceding layers along the channel dimension, serving as input to the next layer.

[0119] like Figure 10As shown, DenseNet consists of three Dense Blocks, each with the same number of layers. Before entering the first Dense Block, the feature map passes through a Conv (16 output channels) to maintain its size. The Dense Block contains dense connections. A transition layer separates two Dense Blocks. After the last Dense Block, an average pooling layer is followed by a Softmax classifier. The feature map sizes of the three Dense Blocks are 32×32, 16×16, and 8×8, respectively.

[0120] In this embodiment, the DenseNet network includes the following key features:

[0121] a) Pooling Layer. DenseConnectivity cannot be used when the feature map size is different. DenseNet divides the network into multiple densely connected Dense Blocks. Downsampling operations (BN + 1×1 Conv + 2×2 average pooling) are performed between Blocks (called Transition Layers). Here, BN (Batch Normalization) is the batch normalization layer, Conv (Convolutional layer) is the convolutional layer, and average pooling is the average pooling layer.

[0122] b) Dense Connectivity Layers. Dense connectivity is used within the block. Connections are made between any two layers to all subsequent layers. Inter-layer connections are BN + ReLU + 3×3 Conv. Here, ReLU (Rectified Liner Uints) are sparse activation function layers.

[0123] c) Growth Rate. If each layer produces k feature maps, then the number of feature maps input to the Lth layer is k0 + k(L-1), where k0 is the input channels. DenseNet differs from other network architectures in that its layers can be very narrow, for example, k = 12. Here, k is called the network's growth rate.

[0124] d) Bottleneck Layers. A 1×1 convolutional layer can serve as a bottleneck layer before a 3×3 convolutional layer, reducing the dimensionality of the input feature map and thus improving computational efficiency. This structure was introduced in DenseNet-B, where the 1×1 convolutional layer outputs a 4k feature map.

[0125] In this embodiment, DenseNet has two outstanding contributions compared to other models:

[0126] a) Optimized Natural integrates the features of identity mapping, deep supervision, and diverse depths. It alleviates the vanishing gradient problem, enhances feature propagation, encourages feature reuse, and significantly reduces the number of parameters.

[0127] b) Each layer has a gradient directly from the original input to the loss function, which realizes implicit deep supervision and helps train deep networks; dense connections have a regularization effect, which reduces overfitting on small training sets.

[0128] In this embodiment, the entire training sample can be divided into several batches. The batch size (size of each batch of samples) for training DenseNet can be set to 32. The learning rate is initially 0.1, and then divided by 5 at 50% and 75% of the epochs. An epoch is defined as the process of training all training samples once, after a complete dataset has passed through the neural network once and returned once. (That is, all training samples undergo one forward propagation and one backward propagation in the neural network). However, when the number of samples in an epoch (i.e., all training samples) may be too large (for a computer), it needs to be divided into multiple smaller blocks, i.e., multiple batches, for training.

[0129] After the model completes its predictions, information can be recommended to users with different tags based on the prediction results. For example, targeted ads can be pushed to users with specific tags, and the success rate of ad impressions and the click-through rate of ads can be counted to evaluate the accuracy of the model's predictions.

[0130] Evaluation of offline experiments:

[0131] 1) Evaluation of mathematical indicators:

[0132] a) AUC (Area under Curve): The larger the AUC value, the more likely the current classification algorithm is to rank positive samples ahead of negative samples and obtain better classification results.

[0133] 2) Online experiment evaluation:

[0134] a) Evaluate the model's performance based on online traffic from A / B testing.

[0135] b) Evaluation metrics include: ad impression success rate and ad click-through rate.

[0136] S305: Based on the first attribute recognition model and the first output result, an initial attribute recognition model is obtained; the first output result is the output result of the target pooling layer corresponding to the first attribute recognition model when training converges;

[0137] In the embodiments of this application, such as Figure 4 As shown, the process of obtaining the initial attribute recognition model based on the first attribute recognition model and the first output result includes:

[0138] S3051: Obtain the feature matrix corresponding to the features of the sample object when the training of the first attribute recognition model converges;

[0139] In this embodiment of the application, when the training of the first attribute recognition model converges, the feature matrix corresponding to the features of the sample object is obtained. The feature matrix can be the weight matrix of the feature layer.

[0140] S3053: Perform singular value decomposition on the feature matrix to obtain feature layer parameters;

[0141] In the embodiments of this application, singular value decomposition can be performed on the feature matrix, and the feature layer parameters can be determined by the decomposed matrix. Singular value decomposition (SVD) is an algorithm widely used in the field of machine learning. It can be used not only for feature decomposition in dimensionality reduction algorithms, but also for recommendation systems and natural language processing.

[0142] In the embodiments of this application, the feature matrix can be subjected to singular value decomposition after each convergence of subsequent model training.

[0143] S3055: Based on the first attribute recognition model, the feature layer parameters, and the first output result, an initial attribute recognition model is obtained.

[0144] In the embodiments of this application, the singular value decomposition algorithm can remove the correlation of the fully connected layers of the network, improve the discriminative ability of the DenseNet feature layers, and maintain the orthogonality of the feature layers.

[0145] In this embodiment of the application, obtaining the initial attribute recognition model based on the first attribute recognition model, the feature layer parameters, and the first output result includes:

[0146] Using the feature layer parameters as constraints, the first output result is input into the first attribute recognition model for constrained training to obtain the second attribute recognition model;

[0147] In this embodiment, the feature layer parameters after singular value decomposition can be fixed, and the parameters of other layers can be updated until the first attribute recognition model converges again; this training process is defined as constrained training. Through constrained training, the second attribute recognition model is obtained.

[0148] Remove the constraints on the feature layer parameters, input the second output result into the second attribute recognition model for relaxation training, and obtain the initial attribute recognition model; the second output result is the output result of the target pooling layer corresponding to the second attribute recognition model when the training converges.

[0149] In this embodiment, the model training process that removes constraints on the feature layer parameters is called relaxation training. The target pooling layer can be the fifth pooling layer in the network corresponding to the model. When the second attribute recognition model training converges, the output of the target pooling layer corresponding to the model is used again as the input feature for the next round of model training, thereby obtaining the initial attribute recognition model.

[0150] In this embodiment, after constrained training, the constraints on the feature layer parameters can be removed, and the second attribute recognition model can continue to be trained until the model converges, thus obtaining the initial attribute recognition model. Through constrained training and relaxation training, the accuracy of the initial attribute recognition model can be improved.

[0151] In some embodiments, such as Figure 5 As shown, the step of inputting the second output result into the second attribute recognition model for relaxation training to obtain the initial attribute recognition model includes:

[0152] S501: Input the second output result into the second attribute recognition model for relaxation training to obtain the third attribute recognition model;

[0153] S503: Obtain the updated feature matrix corresponding to the features of the sample object when the training of the third attribute recognition model converges;

[0154] In this embodiment, after the relaxation training ends, the updated feature matrix corresponding to the model convergence can be obtained again, and subsequent repeated training steps can be performed.

[0155] S505: Perform singular value decomposition on the updated feature matrix to obtain the updated feature layer parameters;

[0156] In this embodiment, singular value decomposition is performed on the updated feature matrix to obtain the decomposed updated matrix, thereby obtaining the corresponding updated feature layer parameters.

[0157] S507: Based on the third attribute recognition model, the updated feature layer parameters, and the third output result, repeat the constraint training and the relaxation training to obtain the initial attribute recognition model; the third output result is the output result of the target pooling layer corresponding to the third attribute recognition model when the training converges.

[0158] In this embodiment, after one round of constraint training and relaxation training, the trained model can be used to continue the next round of constraint training and relaxation training. Through repeated training, the stability of the initial attribute recognition model can be improved.

[0159] In the embodiments of this application, such as Figure 6 As shown, the initial attribute recognition model is obtained by repeating the constraint training and relaxation training based on the third attribute recognition model, the updated feature layer parameters, and the third output result, including:

[0160] S5071: Based on the third attribute recognition model, the updated feature layer parameters, and the third output result, perform constraint training and relaxation training to obtain an updated recognition model; and use the third attribute recognition model as the current recognition model;

[0161] In the embodiments of this application, both constraint training and relaxation training are the same as the steps described above, and each subsequent training round uses the model trained in the previous round and the updated feature parameters.

[0162] S5073: Obtain the first loss value when the current recognition model converges and the second loss value when the updated recognition model converges;

[0163] In this embodiment of the application, each time the model converges during training, there is a corresponding loss value. This loss value is usually less than a preset value, which can be set according to the actual situation.

[0164] S5075: Calculate the target value based on the first loss value and the second loss value; the target value is the absolute value or ratio of the difference between the first loss value and the second loss value;

[0165] S5077: If the target value is less than or equal to a preset threshold, the updated recognition model is determined as the initial attribute recognition model.

[0166] In this embodiment of the application, the training endpoint of the model can be determined based on the magnitude of the target value.

[0167] In this embodiment of the application, the method further includes:

[0168] S5079: If the target value is greater than the preset threshold, the updated recognition model will be used as the current recognition model again;

[0169] S50711: Perform constraint training and relaxation training based on the current recognition model to obtain the target model, and use the target model as the updated recognition model;

[0170] S50713: Repeat the steps of obtaining the first loss value when the current recognition model converges and the second loss value when the updated recognition model converges to perform constraint training and relaxation training based on the current recognition model to obtain the target model, and use the target model as the updated recognition model again.

[0171] In this embodiment of the application, the repeated steps include: obtaining a first loss value when the current recognition model converges and a second loss value when the updated recognition model converges; calculating a target value based on the first loss value and the second loss value; if the target value is greater than the preset threshold, using the updated recognition model as the current recognition model again; and performing constraint training and relaxation training based on the current recognition model to obtain a target model, and using the target model as the updated recognition model again.

[0172] In this embodiment, the stability of the model can be determined based on the target value corresponding to the loss value. If the target value is small, it indicates that the model has good stability after two consecutive training iterations, and training can be terminated. If the target value is large, it indicates that the model has poor stability, and training needs to be repeated.

[0173] In some embodiments, during model training, repeated constraint training and relaxation training can be performed to obtain the loss value at each convergence. The stability of the model can be determined using multiple loss values, thereby determining whether to terminate the training process. For example, multiple loss values ​​can be grouped, with each group containing two values. The target value corresponding to the loss value of each group can be calculated, and model training terminates when the target values ​​corresponding to the loss values ​​of each group are all less than or equal to a preset threshold. The last model is then used as the initial attribute recognition model.

[0174] S307: Based on the objective function, update the loss function of the initial attribute recognition model to obtain an updated model, and train the sample object features according to the updated model to obtain the object attribute recognition model.

[0175] In this embodiment, the original loss function of the initial attribute recognition model can be replaced with an objective function to continue model training. The objective function can be AM-Softmax (additive margin softmax), which normalizes the features and weights of the regression objective of the Softmax loss. Softmax, used in multi-class classification, maps the outputs of multiple neurons to the (0, 1) interval, which can be understood as probabilities for multi-class classification. Unlike Softmax, AM-Softmax is a strategy that reduces intra-class distance and increases inter-class distance. Softmax can only define the boundaries between classes, while AM-Softmax can reduce intra-class distance and increase inter-class distance, narrowing the class range to the target region while simultaneously generating inter-class distances with margin sizes.

[0176] In this embodiment, the AM-SoftMax loss function is used for metric learning, which can improve the model's recognition ability and recognition accuracy.

[0177] In the embodiments of this application, such as Figure 7 As shown, the updated model is obtained by updating the loss function of the initial attribute recognition model based on the objective function, and the updated model is used to train the sample object features to obtain the object attribute recognition model, including:

[0178] S3071: The updated model is obtained by updating the loss function of the initial attribute recognition model based on the objective function;

[0179] In the embodiments of this application, the original loss function of the initial attribute recognition model can be replaced by the objective function to obtain the updated model.

[0180] In this embodiment of the application, after the model training is completed, the model can be verified, thereby improving the recognition accuracy of the model.

[0181] S3073: Train the sample object feature set according to the updated model to obtain the attribute recognition model to be verified; the sample object feature set includes at least two sample object features;

[0182] In this embodiment of the application, the feature set of the sample object is trained according to the updated model to obtain the attribute recognition model to be verified.

[0183] S3075: Obtain the test object feature set; the test object features in the test object feature set are labeled with test attribute tags;

[0184] In this embodiment of the application, a test object feature set may include multiple test object features.

[0185] S3077: Based on the attribute recognition model to be verified, determine the target sample object features from the sample object feature set;

[0186] In the embodiments of this application, such as Figure 8 As shown, determining the target sample object features from the sample object feature set based on the attribute recognition model to be verified includes:

[0187] S30771: Based on the attribute recognition model to be verified, determine the comprehensive distance between the feature of each sample object and the feature of the target test object; the feature of the target test object is any feature in the set of test object features;

[0188] In the embodiments of this application, the comprehensive distance can characterize the similarity between the features of the sample object and the features of the target test object.

[0189] In this embodiment of the application, determining the comprehensive distance between the features of each sample object and the features of the target test object based on the attribute recognition model to be verified includes:

[0190] Based on the attribute recognition model to be verified, the original distance between the features of the target test object and the features of each sample object is calculated;

[0191] In this embodiment, the original distance can be Euclidean distance, which can be used to calculate the Euclidean distance between the features of the target test object and the features of each sample object.

[0192] Based on the original distance corresponding to each sample object feature, a preset number of candidate sample object features are determined;

[0193] In this embodiment of the application, the features of the candidate sample object can be determined based on the original distance; then, the attribute labels corresponding to the features of the target test object can be determined based on the K-nearest neighbor method.

[0194] In this embodiment, the preset number can be K, and the features of the sample objects can be reordered according to the K-nearest neighbor method, thereby improving the recognition ability of the model.

[0195] K-Nearest Neighbors (KNN) is one of the most fundamental methods used for classification and regression. When the distribution of the training data is unknown, KNN is often the first approach that comes to mind. KNN takes the feature vectors of instances as input and outputs the class of the instance. The algorithm works by taking the K nearest neighbors of a target data point from the training dataset and calculating the percentage of each of these K nearest neighbors in the input space. The class with the highest percentage of nearest neighbors is then chosen as the target data point's class. KNN typically uses Euclidean distance to represent the distance between two data points; the smaller the distance, the closer the two data points are.

[0196] In some embodiments, the K-nearest neighbor set can be expanded into a more robust set before calculating the distance. Specifically, features in the set can be added based on the current set. For example, the radius of the circle can be expanded outward from the center of the current set to identify more candidate sample object features.

[0197] In some embodiments, after determining the features of the candidate sample objects, the test labels of the test object features can be directly determined according to the K nearest neighbor method; based on the candidate sample labels corresponding to each feature in the K candidate sample object features, the proportion of each candidate sample label in the K labels is determined, the candidate sample label with the highest proportion is determined as the test label, and the obtained test label is matched with the labeled label, so as to verify the model based on the matching result.

[0198] Construct a target sample feature set based on the preset number of candidate sample object features;

[0199] Calculate the Jaccard distance between the feature set of the test object and the feature set of the target sample;

[0200] Based on the original distance and Jaccard distance corresponding to each candidate sample object feature, the comprehensive distance between each candidate sample object feature and the target test object feature is determined.

[0201] In this embodiment, Jaccard distance is a metric used to measure the difference between two sets. The combined distance for each candidate sample object feature can be determined by combining the original distance and the Jaccard distance. Before determining the combined distance, weights can be set for the original distance and the Jaccard distance respectively, and the sum of the weighted distances is calculated based on these weights to obtain the combined distance.

[0202] S30773: Sort the sample object features in the sample object feature set based on the comprehensive distance corresponding to each sample object feature;

[0203] In this embodiment of the application, the data can be sorted according to the comprehensive distance from largest to smallest or from smallest to largest.

[0204] S30775: Determine the characteristics of the target sample object based on the sorting results.

[0205] In some embodiments, the feature of the sample object with the smallest comprehensive distance can be determined as the target sample object feature based on the ranking results.

[0206] In this embodiment of the application, the probability value of the sample attribute label corresponding to each sample object feature being the target sample attribute label can also be determined based on the comprehensive distance; the probability values ​​corresponding to each sample object feature are sorted, and the sample attribute label with the highest probability value is determined as the target sample attribute label.

[0207] S3079: Obtain the target sample attribute label corresponding to the feature of the target sample object;

[0208] S30711: If the target sample attribute label matches the test attribute label, the attribute recognition model to be verified is determined to be the object attribute recognition model.

[0209] In this embodiment, matching the target sample attribute label with the test attribute label means that they are the same label. If they match, the attribute recognition model to be verified is determined to have passed verification, and the attribute recognition model to be verified is an object attribute recognition model. If they do not match, the above training steps are repeated for the attribute recognition model to be verified.

[0210] In some embodiments, after determining a preset number of candidate sample object features based on the original distance corresponding to each sample object feature, the method may further include:

[0211] Determine the candidate attribute labels corresponding to the features of each candidate sample object;

[0212] Based on the candidate attribute labels, the features of a preset number of candidate sample objects are divided into a target number of candidate feature sets; each candidate sample object in the candidate feature set corresponds to the same candidate attribute label.

[0213] In this embodiment of the application, the features of candidate sample objects can be classified according to attribute labels, and features with the same attribute label can be formed into a set.

[0214] Calculate the first distance between the candidate features in each candidate feature set and the test sample;

[0215] In this embodiment of the application, the first distance can be a Euclidean distance.

[0216] Calculate the second distance between each candidate feature set and the test sample set, where the second distance is the Jaccard distance;

[0217] In this test sample set, the test samples all have the same test attribute label;

[0218] Based on the first distance and the second distance, determine the comprehensive distance between the test sample and each candidate feature set;

[0219] The target candidate feature set is determined based on the comprehensive distance corresponding to each candidate feature set;

[0220] The candidate attribute labels corresponding to the target candidate feature set are determined as the output labels of the test samples;

[0221] If the output label matches the test attribute label, the model verification is confirmed to be successful, and the attribute recognition model to be verified is determined as the object attribute recognition model.

[0222] If the output label does not match the test attribute label, the attribute recognition model to be verified is trained again.

[0223] In this embodiment of the application, after the model training is completed, the recognition accuracy of the model can be improved by validating the model.

[0224] In this embodiment, the attribute type corresponding to the sample feature can be changed and applied to different scenarios; for example, it can be used to predict the "user's highest education level tag". The server collects the corresponding log data and uses the same feature concatenation, feature processing and model training methods to obtain the corresponding prediction model, and then applies it.

[0225] In one specific embodiment, please refer to Figure 11 and Figure 12 , Figure 11 This diagram illustrates a comparison of the AUC (Average Values) of different models in predicting user lifestyles. Figure 12 This diagram illustrates the comparison of the effectiveness of different models in predicting user lifestyles.

[0226] Among them, from Figure 11 The comparative analysis of the model performance shows that: a) in terms of offline AUC performance, the solution in this embodiment improves by an average of 33.54% compared with other technical solutions; b) in terms of online AUC performance, the solution in this embodiment improves by an average of 31.42% compared with other technical solutions.

[0227] Among them, from Figure 12The comparative analysis of business performance shows that: a) in terms of ad click-through rate, the solution in this embodiment is on average 387.48% higher than other technical solutions; b) in terms of ad conversion rate, the solution in this embodiment is on average 387.72% higher than other technical solutions.

[0228] Analysis of model performance and business performance shows that the object attribute recognition model provided in this application is significantly better than other models. The object attribute recognition model constructed using the training method provided in this application effectively improves the accuracy of object attribute recognition.

[0229] As can be seen from the technical solutions provided in the embodiments of this application above, the embodiments of this application obtain the features of the object to be identified; based on the object attribute recognition model, the features of the object to be identified are processed for object attribute recognition to obtain the target attribute label corresponding to the features of the object to be identified; wherein, the training method of the object attribute recognition model includes: obtaining sample object features; the sample object features are labeled with sample attribute labels, the sample attribute labels characterizing the sample object attributes corresponding to the sample object features; pre-training the sample object features based on a preset network to obtain a first attribute recognition model; obtaining an initial attribute recognition model based on the first attribute recognition model and a first output result; the first output result is the output result of the target pooling layer corresponding to the first attribute recognition model when training converges; updating the loss function of the initial attribute recognition model based on the objective function to obtain an updated model, and training the sample object features according to the updated model to obtain the object attribute recognition model. This application obtains an initial attribute recognition model with high accuracy based on the first attribute recognition model and the first output result; then updates the loss function of the initial attribute recognition model based on the objective function to obtain an updated model, and continues training to obtain the object attribute recognition model; thereby improving the recognition accuracy of the object attribute recognition model.

[0230] This application's embodiments can be implemented using cloud technology or blockchain network technology. Cloud technology refers to a hosting technology that unifies hardware, software, network, and other resources within a wide area network (WAN) or local area network (LAN) to achieve data computation, 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 resource pools, which can be used on demand, offering flexibility and convenience. The backend services of technical network systems require a large amount of computing and storage resources, such as video websites, image websites, and many portal websites; therefore, cloud technology needs to be supported by cloud computing.

[0231] It's important to note that cloud computing is a computing model that distributes computing tasks across a resource pool comprised of numerous computers, enabling various application systems to access computing power, storage space, and information services as needed. The network providing these resources is called the "cloud." From the user's perspective, resources in the "cloud" are infinitely scalable, readily available, and can be used on demand, expanded at any time, and paid for based on usage. As the foundational providers of cloud computing capabilities, they establish cloud resource pool platforms, often referred to as cloud platforms or Infrastructure as a Service (IaaS). These platforms deploy various types of virtual resources within the resource pool for external customers to choose from. The cloud resource pool primarily includes: computing devices (which can be virtualized machines containing operating systems), storage devices, and network devices.

[0232] This application also provides an object attribute recognition device, such as... Figure 13 As shown, the device includes:

[0233] The object feature acquisition module 1310 is used to acquire the features of the object to be identified.

[0234] The attribute tag recognition module 1320 is used to perform object attribute recognition processing on the features of the object to be recognized based on the object attribute recognition model, so as to obtain the target attribute tag corresponding to the features of the object to be recognized.

[0235] Model training module 1330 is used to train the object attribute recognition model;

[0236] The model training module 1330 includes:

[0237] The sample object feature acquisition submodule 13310 is used to acquire sample object features; the sample object features are labeled with sample attribute tags, and the sample attribute tags represent the sample object attributes corresponding to the sample object features.

[0238] The first attribute recognition model determination submodule 13320 is used to pre-train the features of the sample object based on a preset network to obtain the first attribute recognition model;

[0239] The initial attribute recognition model determination submodule 13330 is used to obtain an initial attribute recognition model based on the first attribute recognition model and the first output result; the first output result is the output result of the target pooling layer corresponding to the first attribute recognition model when training converges;

[0240] The object attribute recognition model determination submodule 13340 is used to update the loss function of the initial attribute recognition model based on the objective function to obtain an updated model, and to train the sample object features according to the updated model to obtain the object attribute recognition model.

[0241] In some embodiments, the initial attribute recognition model determination submodule may include:

[0242] The feature matrix acquisition unit is used to acquire the feature matrix corresponding to the features of the sample object when the training of the first attribute recognition model converges.

[0243] The feature layer parameter determination unit is used to perform singular value decomposition on the feature matrix to obtain the feature layer parameters;

[0244] The initial attribute recognition model determination unit is used to obtain an initial attribute recognition model based on the first attribute recognition model, the feature layer parameters, and the first output result.

[0245] In some embodiments, the initial attribute recognition model determination unit may include:

[0246] The first input subunit is used to input the first output result into the first attribute recognition model for constrained training with the feature layer parameters as constraints, so as to obtain the second attribute recognition model.

[0247] The second input subunit is used to remove the constraints on the feature layer parameters and input the second output result into the second attribute recognition model for relaxation training to obtain the initial attribute recognition model; the second output result is the output result of the target pooling layer corresponding to the second attribute recognition model when the training converges.

[0248] In some embodiments, the second input subunit may include:

[0249] The third attribute recognition model determines the sub-unit, which is used to input the second output result into the second attribute recognition model for relaxation training, so as to obtain the third attribute recognition model;

[0250] The feature matrix update acquisition subunit is used to acquire the updated feature matrix corresponding to the features of the sample object when the training of the third attribute recognition model converges.

[0251] The updated feature layer parameter determination subunit is used to perform singular value decomposition on the updated feature matrix to obtain the updated feature layer parameters;

[0252] The repeating subunit is used to repeat the constraint training and the relaxation training based on the third attribute recognition model, the updated feature layer parameters, and the third output result to obtain the initial attribute recognition model; the third output result is the output result of the target pooling layer corresponding to the third attribute recognition model when the training converges.

[0253] In some embodiments, the object attribute recognition model determination submodule may include:

[0254] The model update unit is used to update the loss function of the initial attribute recognition model based on the objective function to obtain the updated model;

[0255] The attribute recognition model training unit is used to train the sample object feature set according to the updated model to obtain the attribute recognition model to be verified; the sample object feature set includes at least two sample object features;

[0256] The test object feature set acquisition unit is used to acquire the test object feature set; the test object features in the test object feature set are labeled with test attribute tags;

[0257] The target sample object feature determination unit is used to determine the target sample object features from the sample object feature set based on the attribute recognition model to be verified.

[0258] A target sample attribute label acquisition unit is used to acquire the target sample attribute labels corresponding to the features of the target sample object.

[0259] The model determination unit is used to determine the attribute recognition model to be verified as the object attribute recognition model if the target sample attribute label matches the test attribute label.

[0260] In some embodiments, the target sample object feature determination unit may include:

[0261] The comprehensive distance determination subunit is used to determine the comprehensive distance between each sample object feature and the target test object feature based on the attribute recognition model to be verified; the target test object feature is any feature in the set of test object features;

[0262] The sorting subunit is used to sort the sample object features in the sample object feature set based on the comprehensive distance corresponding to each sample object feature;

[0263] The target sample object feature determination subunit is used to determine the features of the target sample object based on the sorting results.

[0264] In some embodiments, the integrated distance determination subunit may include:

[0265] The original distance calculation subunit is used to calculate the original distance between the features of the target test object and the features of each sample object based on the attribute recognition model to be verified.

[0266] The candidate sample object feature determination subunit is used to determine a preset number of candidate sample object features based on the original distance corresponding to each sample object feature;

[0267] The target sample feature set construction subunit is used to construct a target sample feature set based on the preset number of candidate sample object features;

[0268] The Jaccard distance calculation subunit is used to calculate the Jaccard distance between the test object feature set and the target sample feature set.

[0269] The distance determination subunit is used to determine the comprehensive distance between each candidate sample object feature and the target test object feature based on the original distance and Jaccard distance corresponding to each candidate sample object feature.

[0270] The apparatus and method embodiments described herein are based on the same inventive concept.

[0271] This application provides an object attribute recognition device, which includes a processor and a memory. The memory stores at least one instruction or at least one program, which is loaded and executed by the processor to implement the object attribute recognition method provided in the above method embodiments.

[0272] The embodiments of this application also provide a computer storage medium, which can be disposed in a terminal to store at least one instruction or at least one program related to implementing an object attribute recognition method in the method embodiments. The at least one instruction or at least one program is loaded and executed by the processor to implement the object attribute recognition method provided in the above method embodiments.

[0273] Embodiments of this application also provide a computer program product or computer program, which includes 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 object attribute identification method provided in the above-described method embodiments.

[0274] Optionally, in this embodiment, the storage medium may be located at at least one of the multiple network servers in a computer network. Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0275] The memory described in this application embodiment can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for the functions, etc.; the data storage area may store data created according to the use of the device, etc. In addition, the memory 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. Accordingly, the memory may also include a memory controller to provide the processor with access to the memory.

[0276] The object attribute recognition method embodiments provided in this application can be executed on mobile terminals, computer terminals, servers, or similar computing devices. Taking running on a server as an example, Figure 14 This is a hardware structure block diagram of a server for an object attribute identification method provided in an embodiment of this application. For example... Figure 14As shown, the server 1400 can vary significantly due to different configurations or performance. It may include one or more Central Processing Units (CPUs) 1410 (CPUs 1410 may include, but are not limited to, microprocessors (MCUs) or programmable logic devices (FPGAs), a memory 1430 for storing data, and one or more storage media 1420 (e.g., one or more mass storage devices) for storing application programs 1423 or data 1422. The memory 1430 and storage media 1420 may be temporary or persistent storage. The program stored in the storage media 1420 may include one or more modules, each module may include a series of instruction operations on the server. Furthermore, the CPU 1410 may be configured to communicate with the storage media 1420 and execute the series of instruction operations stored in the storage media 1420 on the server 1400. Server 1400 may also include one or more power supplies 1460, one or more wired or wireless network interfaces 1450, one or more input / output interfaces 1440, and / or one or more operating systems 1421, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0277] The input / output interface 1440 can be used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of server 1400. In one example, the input / output interface 1440 includes a network interface controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In one example, the input / output interface 1440 can be a radio frequency (RF) module for wireless communication with the Internet.

[0278] Those skilled in the art will understand that Figure 14 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, server 1400 may also include... Figure 14 The more or fewer components shown, or having the same Figure 14 The different configurations shown.

[0279] As can be seen from the embodiments of the object attribute recognition method, apparatus, device, or storage medium provided in this application above, this application acquires the features of an object to be identified; based on an object attribute recognition model, it performs object attribute recognition processing on the features of the object to be identified to obtain the target attribute label corresponding to the features of the object to be identified; wherein, the training method of the object attribute recognition model includes: acquiring sample object features; the sample object features are labeled with sample attribute labels, and the sample attribute labels characterize the sample object attributes corresponding to the sample object features; pre-training the sample object features based on a preset network to obtain a first attribute recognition model; obtaining an initial attribute recognition model based on the first attribute recognition model and a first output result; the first output result is the output result of the target pooling layer corresponding to the first attribute recognition model when training converges; updating the loss function of the initial attribute recognition model based on the objective function to obtain an updated model, and training the sample object features according to the updated model to obtain the object attribute recognition model. This application obtains an initial attribute recognition model with high accuracy based on the first attribute recognition model and the first output result; then, it updates the loss function of the initial attribute recognition model based on the objective function to obtain an updated model, and continues to train to obtain an object attribute recognition model that can accurately identify object attributes; thereby improving the accuracy of object attribute recognition.

[0280] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments have been described above. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0281] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0282] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer storage medium, such as a read-only memory, a disk, or an optical disk.

[0283] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for identifying object attributes, characterized in that, The method includes: Obtain the features of the object to be identified; Based on the object attribute recognition model, the features of the object to be identified are processed to obtain the target attribute label corresponding to the features of the object to be identified. The training method for the object attribute recognition model includes: Obtain sample object features; the sample object features are labeled with sample attribute tags, and the sample attribute tags characterize the sample object attributes corresponding to the sample object features; The first attribute recognition model is obtained by pre-training the features of the sample object based on a preset network. Obtain the feature matrix corresponding to the features of the sample object when the training of the first attribute recognition model converges; and perform singular value decomposition on the feature matrix to obtain the feature layer parameters; Using the feature layer parameters as constraints, the first output result is input into the first attribute recognition model for constrained training to obtain the second attribute recognition model; the first output result is the output result of the target pooling layer corresponding to the first attribute recognition model when training converges; Remove the constraints on the feature layer parameters, input the second output result into the second attribute recognition model for relaxation training, and obtain the initial attribute recognition model; the second output result is the output result of the target pooling layer corresponding to the second attribute recognition model when the training converges; An updated model is obtained by updating the loss function of the initial attribute recognition model based on the objective function, and the updated model is used to train the sample object features to obtain the object attribute recognition model.

2. The method according to claim 1, characterized in that, The step of inputting the second output result into the second attribute recognition model for relaxation training to obtain the initial attribute recognition model includes: The second output result is input into the second attribute recognition model for relaxation training to obtain the third attribute recognition model; Obtain the updated feature matrix corresponding to the features of the sample object when the training of the third attribute recognition model converges; The updated feature matrix is ​​subjected to singular value decomposition to obtain the updated feature layer parameters; Based on the third attribute recognition model, the updated feature layer parameters, and the third output result, the constraint training and the relaxation training are repeated to obtain the initial attribute recognition model; the third output result is the output result of the target pooling layer corresponding to the third attribute recognition model when the training converges.

3. The method according to any one of claims 1-2, characterized in that, The process of updating the loss function of the initial attribute recognition model based on the objective function to obtain an updated model, and then training the sample object features based on the updated model to obtain the object attribute recognition model, includes: The updated model is obtained by updating the loss function of the initial attribute recognition model based on the objective function; The sample object feature set is trained according to the updated model to obtain the attribute recognition model to be verified; the sample object feature set includes at least two sample object features; Obtain the test object feature set; the test object features in the test object feature set are labeled with test attribute tags; Based on the attribute recognition model to be verified, the features of the target sample object are determined from the feature set of the sample object. Obtain the target sample attribute labels corresponding to the features of the target sample object; If the target sample attribute label matches the test attribute label, the attribute recognition model to be verified is determined to be the object attribute recognition model.

4. The method according to claim 3, characterized in that, The step of determining the target sample object features from the sample object feature set based on the attribute recognition model to be verified includes: Based on the attribute recognition model to be verified, the comprehensive distance between each sample object feature and the target test object feature is determined; the target test object feature is any feature in the test object feature set. Based on the comprehensive distance corresponding to each sample object feature, the sample object features in the sample object feature set are sorted. Based on the sorting results, the characteristics of the target sample objects are determined.

5. The method according to claim 4, characterized in that, The determination of the comprehensive distance between the features of each sample object and the features of the target test object based on the attribute recognition model to be verified includes: Based on the attribute recognition model to be verified, the original distance between the features of the target test object and the features of each sample object is calculated; Based on the original distance corresponding to each sample object feature, a preset number of candidate sample object features are determined; Construct a target sample feature set based on the preset number of candidate sample object features; Calculate the Jaccard distance between the feature set of the test object and the feature set of the target sample; Based on the original distance and Jaccard distance corresponding to each candidate sample object feature, the comprehensive distance between each candidate sample object feature and the target test object feature is determined.

6. An object attribute recognition device, characterized in that, The device includes: The object feature acquisition module is used to acquire the features of the object to be identified; The attribute tag recognition module is used to perform object attribute recognition processing on the features of the object to be identified based on the object attribute recognition model, so as to obtain the target attribute tag corresponding to the features of the object to be identified. The model training module is used to train the object attribute recognition model; The model training module includes: The sample object feature acquisition submodule is used to acquire sample object features; the sample object features are labeled with sample attribute tags, and the sample attribute tags represent the sample object attributes corresponding to the sample object features. The first attribute recognition model determination submodule is used to pre-train the features of the sample object based on a preset network to obtain the first attribute recognition model; The initial attribute recognition model determination submodule is used to obtain an initial attribute recognition model based on the first attribute recognition model and the first output result; the first output result is the output result of the target pooling layer corresponding to the first attribute recognition model when training converges; The object attribute recognition model determination submodule is used to update the loss function of the initial attribute recognition model based on the objective function to obtain an updated model, and to train the sample object features based on the updated model to obtain the object attribute recognition model; The initial attribute recognition model determination submodule includes: The feature matrix acquisition unit is used to acquire the feature matrix corresponding to the features of the sample object when the training of the first attribute recognition model converges; the feature layer parameter determination unit is used to perform singular value decomposition on the feature matrix to obtain the feature layer parameters; the initial attribute recognition model determination unit is used to obtain the initial attribute recognition model based on the first attribute recognition model, the feature layer parameters and the first output result. The initial attribute recognition model determination unit includes: The first input subunit is used to input the first output result into the first attribute recognition model for constrained training with the feature layer parameters as constraints, so as to obtain the second attribute recognition model. The second input subunit is used to remove the constraints on the feature layer parameters and input the second output result into the second attribute recognition model for relaxation training to obtain the initial attribute recognition model; the second output result is the output result of the target pooling layer corresponding to the second attribute recognition model when the training converges.

7. The apparatus according to claim 6, characterized in that, The second input subunit includes: The third attribute recognition model determines the sub-unit, which is used to input the second output result into the second attribute recognition model for relaxation training, so as to obtain the third attribute recognition model; The feature matrix update acquisition subunit is used to acquire the updated feature matrix corresponding to the features of the sample object when the training of the third attribute recognition model converges. The updated feature layer parameter determination subunit is used to perform singular value decomposition on the updated feature matrix to obtain the updated feature layer parameters; The repeating subunit is used to repeat the constraint training and the relaxation training based on the third attribute recognition model, the updated feature layer parameters, and the third output result to obtain the initial attribute recognition model; the third output result is the output result of the target pooling layer corresponding to the third attribute recognition model when the training converges.

8. The apparatus according to any one of claims 6-7, characterized in that, The object attribute recognition model determination submodule includes: The model update unit is used to update the loss function of the initial attribute recognition model based on the objective function to obtain the updated model; The attribute recognition model training unit is used to train the sample object feature set according to the updated model to obtain the attribute recognition model to be verified; the sample object feature set includes at least two sample object features; The test object feature set acquisition unit is used to acquire the test object feature set; the test object features in the test object feature set are labeled with test attribute tags; The target sample object feature determination unit is used to determine the target sample object features from the sample object feature set based on the attribute recognition model to be verified. A target sample attribute label acquisition unit is used to acquire the target sample attribute labels corresponding to the features of the target sample object. The model determination unit is used to determine the attribute recognition model to be verified as the object attribute recognition model if the target sample attribute label matches the test attribute label.

9. The apparatus according to claim 8, characterized in that, The target sample object feature determination unit includes: The comprehensive distance determination subunit is used to determine the comprehensive distance between each sample object feature and the target test object feature based on the attribute recognition model to be verified; the target test object feature is any feature in the set of test object features; The sorting subunit is used to sort the sample object features in the sample object feature set based on the comprehensive distance corresponding to each sample object feature; The target sample object feature determination subunit is used to determine the features of the target sample object based on the sorting results.

10. The apparatus according to claim 9, characterized in that, The integrated distance determination subunit includes: The original distance calculation subunit is used to calculate the original distance between the features of the target test object and the features of each sample object based on the attribute recognition model to be verified. The candidate sample object feature determination subunit is used to determine a preset number of candidate sample object features based on the original distance corresponding to each sample object feature; The target sample feature set construction subunit is used to construct a target sample feature set based on the preset number of candidate sample object features; The Jaccard distance calculation subunit is used to calculate the Jaccard distance between the test object feature set and the target sample feature set. The distance determination subunit is used to determine the comprehensive distance between each candidate sample object feature and the target test object feature based on the original distance and Jaccard distance corresponding to each candidate sample object feature.

11. An object attribute recognition device, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction or at least one program, the at least one instruction or the at least one program being loaded and executed by the processor to implement the object attribute recognition method as described in any one of claims 1-5.

12. A computer storage medium, characterized in that, The computer storage medium stores at least one instruction or at least one program, which is loaded and executed by a processor to implement the object attribute recognition method as described in any one of claims 1-5.

13. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the object attribute recognition method as described in any one of claims 1-5.

Images