Object type identification method, device, electronic device and storage medium

Abstract

The present application discloses an object type recognition method, device, electronic device and storage medium, the method comprising: obtaining feature information of an object to be recognized; inputting the feature information of the object to be recognized into a first target type recognition model to obtain a first probability value that the object to be recognized belongs to the target type; the first target type recognition model is obtained by training a first neural network model based on meta-learning; inputting the feature information of the object to be recognized into a second target type recognition model to obtain a second probability value that the object to be recognized belongs to the target type; the second target type recognition model is obtained by training the second neural network model based on a genetic algorithm; and determining the type of the object to be recognized based on the first probability value and the second probability value. In this way, the accuracy of object type recognition can be greatly improved, and the problems of low efficiency and poor effect of object type recognition existing in traditional manual or traditional non-deep learning models can be solved.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to an object type recognition method, device, electronic device, and storage medium. Background Art

[0002] During the application promotion process, different objects have different requirements for applications. By predicting the type of objects, we can better understand the needs of objects and help provide them with applications that they actually need.

[0003] In related technologies, there are two main methods for identifying object types: the first is to determine classification rules based on manual experience; the second is to achieve it through traditional non-deep learning models.

[0004] However, the first method has the problems of using a very limited number of rules, being unable to capture high-dimensional feature information of interactions between rules, being unable to determine the optimal parameters of each rule, and having high costs and low accuracy. In the second method, due to the limited learning ability of traditional non-deep learning models, the model performance is prone to bottlenecks for objects with more complex features, resulting in poor recognition results. Summary of the Invention

[0005] In order to solve the problems of low efficiency and poor effect in object type identification through manual or traditional non-deep learning models in the existing technology, this application provides an object type identification method, device, electronic device and storage medium:

[0006] According to a first aspect of the present application, a method for identifying an object type is provided, comprising:

[0007] Acquire characteristic information of the object to be identified; the characteristic information includes at least one characteristic information of the object attribute characteristics, device attribute characteristics, network connection attribute characteristics and behavior characteristics in a preset application;

[0008] Inputting feature information of the object to be identified into a first target type recognition model to obtain a first probability value of the object to be identified belonging to the target type; the first target type recognition model is obtained by training a first neural network model based on a meta-learning method; the target type is a type of person who is willing to use the target application function;

[0009] Inputting feature information of the object to be identified into a second target type recognition model to obtain a second probability value of the object to be identified belonging to the target type; the second target type recognition model is obtained by training the second neural network model based on a genetic algorithm;

[0010] The type of the object to be identified is determined according to the first probability value and the second probability value.

[0011] According to a second aspect of the present application, there is provided an object type identification device, comprising:

[0012] An acquisition module is configured to acquire feature information of an object to be identified;

[0013] The first recognition module is configured to input feature information of the object to be recognized into a first target type recognition model to obtain a first probability value of the object to be recognized belonging to the target type; the first target type recognition model is obtained by training the first neural network model based on meta-learning;

[0014] The second recognition module is configured to input feature information of the object to be identified into a second target type recognition model to obtain a second probability value of the object to be identified belonging to the target type; the second target type recognition model is obtained by training the second neural network model based on a genetic algorithm;

[0015] The determination module is configured to determine the type of the object to be identified based on the first probability value and the second probability value.

[0016] According to a third aspect of the present application, an electronic device is provided, which includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by the processor to implement the object type identification method of the first aspect of the present application.

[0017] According to the fourth aspect of the present application, a computer storage medium is provided, in which at least one instruction or at least one program is stored, and the at least one instruction or at least one program is loaded and executed by a processor to implement the object type identification method of the first aspect of the present application.

[0018] According to the fifth aspect of the present application, a computer program product is provided, which includes at least one instruction or at least one program segment, and the at least one instruction or at least one program segment is loaded and executed by a processor to implement the object type identification method of the first aspect of the present application.

[0019] The object type identification method, device, electronic device, and storage medium provided in this application have the following technical effects:

[0020] The embodiment of the present application obtains a first target type recognition model through training based on meta-learning, and obtains a second target type recognition model through training based on a genetic algorithm. The first target type recognition model and the second target type recognition model are used to respectively identify the feature information of the object to be identified, and the predicted probability values of the two models are combined to determine whether the object to be identified belongs to the target type. This can greatly improve the recognition accuracy of the object type. In this way, it can solve the problems of low efficiency and poor effect in the prior art of identifying object types through manual or traditional non-deep learning models. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present application or the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0022] Figure 1 is a schematic diagram of an application environment provided by an embodiment of the present application;

[0023] Figure 2 This is a flow chart of an object type identification method provided in an embodiment of the present application;

[0024] Figure 3 is a schematic diagram of an average pooling process provided in an embodiment of the present application;

[0025] Figure 4 This is a schematic diagram of a meta-learning process provided by an embodiment of the present application;

[0026] Figure 5 This is a flowchart of a method for generating a first target type recognition model provided in an embodiment of the present application;

[0027] Figure 6 This is a flow chart of a method for generating a second target type recognition model provided in an embodiment of the present application;

[0028] Figure 7 This is a flow chart of a method for training a second neural network model using a genetic algorithm, provided in an embodiment of the present application;

[0029] Figure 8 This is a schematic diagram of a genetic algorithm operation provided by an embodiment of the present application;

[0030] Figure 9 This is a schematic diagram of a crossover process provided by an embodiment of the present application;

[0031] Figure 10is a schematic diagram of the main stages of a training process provided by an embodiment of the present application;

[0032] Figure 11 This is a block diagram of a device for identifying an object type provided in an embodiment of the present application;

[0033] Figure 12 This is a structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0034] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0035] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or devices.

[0036] See also Figure 1 , Figure 1 : is a schematic diagram of an application environment provided by an embodiment of the present application, which may include a client 10 and a server 20. The client 10 and the server 20 may be directly or indirectly connected via wired or wireless communication. The server 20 may provide better services to the client 10 user by identifying the type of the user. Specifically, the server 20 obtains characteristic information of the client 10 user; then, the characteristic information is input into a first target type recognition model to obtain a first probability value that the client 10 user belongs to the target type, and at the same time, the characteristic information is input into a second target type recognition model to obtain a second probability value that the client 10 user belongs to the target type; finally, the server 20 may determine whether the client 10 user belongs to the target type based on the first probability value and the second probability value.

[0037] The client 10 may be a physical device such as a smartphone, a computer (e.g., a desktop computer, tablet computer, laptop computer), an augmented reality (AR) / virtual reality (VR) device, a digital assistant, an intelligent voice interaction device (e.g., a smart speaker), a smart wearable device, a smart home appliance, or an in-vehicle terminal. It may also be software running on a physical device, such as a computer program. The operating system supported by the client may be Android, iOS (a mobile operating system developed by Apple), Linux, or Microsoft Windows.

[0038] The server 20 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. The server can include a network communication unit, a processor, a memory, etc. The server can provide background services for the corresponding client.

[0039] It should be noted that Figure 1 This is merely an example. In some application scenarios, the server 20 may also identify the types of other entities or non-entity objects. In addition, for data such as user information involved in the embodiments of this application, when the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0040] The following describes a specific embodiment of an object type identification method of the present application. Figure 2 It is a flowchart of an object type identification method provided by an embodiment of the present application. The present application provides method operation steps as described in the embodiment or flowchart, but may include more or fewer operation steps based on conventional or non-creative labor. The order of steps listed in the embodiment is only one way of executing the steps among many steps, and does not represent the only execution order. When the actual system or product is executed, it can be executed in sequence or in parallel (for example, in a parallel processor or multi-threaded processing environment) according to the method shown in the embodiment or the accompanying drawings. Specifically, Figure 2 As shown, the method may include:

[0041] S201: Acquire feature information of an object to be identified.

[0042] In the embodiments of the present application, the object to be identified may have different representations depending on the application scenario. For example, in the application scenario of user-type identification, the object to be identified is the user; in the application scenario of image-type identification, the object to be identified may be an image; and in the application scenario of video-type identification, the object to be identified may be a video. Accordingly, the content of the feature information of the object to be identified depends on the actual representation of the object.

[0043] The object type identification described in the embodiments of the present application mainly identifies whether the object belongs to a specified target type. The final result may include that the object to be identified belongs to the target type, or that the object to be identified does not belong to the target type; wherein, the target type is also determined based on the representation of the object in actual application.

[0044] In a specific application scenario, when users use a client to open overseas websites to search for required information or take online courses, they often encounter problems such as the website cannot be opened or the online course playback is stuck, resulting in the user being unable to obtain relevant information and unable to hear the teacher's content clearly. In this case, the user can solve this problem by downloading network acceleration software. Therefore, in the process of promoting network acceleration software, if it is possible to know in advance which users have such needs, it can be recommended to such users to help them effectively solve the problems they encounter.

[0045] Therefore, in some possible embodiments, the target type is a type that is willing to use the target application function; wherein the target application function may be the network acceleration function provided by the above-mentioned network acceleration software;

[0046] Accordingly, the above-mentioned acquisition of feature information of the object to be identified may include:

[0047] At least one feature information of an object attribute feature, a device attribute feature, a network connection attribute feature, and a behavior feature in a preset application of the object to be identified is obtained.

[0048] Among them, the object attribute characteristics may include at least one of the user's age, gender, province, and city where the user is located; the device attribute characteristics may include at least one of the brand, model, and version number of the mobile device used; the network connection attribute characteristics may include at least one of the number of times the wireless network is connected within a preset time period, the location information corresponding to the wireless network, and the traffic usage corresponding to the wireless network; the behavioral characteristics in the preset application may include the traffic usage of the preset application and / or the time information of the traffic generation. In this case, the preset application and the target application may be two applications of the same category or that implement similar functions; or, the behavioral characteristics in the preset application may include the usage of a specific category in the preset application, and clicking on a specific category indicates that the user has a certain degree of willingness to use the target application function; for example, the number of times English songs are listened to in music software, the number of times translation software is downloaded in download software, etc.

[0049] In some possible embodiments, feature information includes multiple features corresponding to multiple dimensions; multiple features may include a combination of any multiple features in the above embodiments, and may also include aggregated features of the same dimension under different time dimensions. Among them, different time dimensions may include the past six months, the past three months, the past month, the past week, and every day; the aggregation method may be summation, median, standard deviation, average pooling, etc. For example, the number of times a user connects to a wireless network every day can be summed to obtain the number of times the user connects to the wireless network in the past week, the past month, or other time dimensions. In this way, the richness of the features can be increased. In addition, the average pooling method is as follows: Figure 3 As shown, Figure 3 The left side shows the network connection attribute features, where the number in each small square represents the number of times a user connects to the wireless network each day. Setting the pooling size to 2*2 and the step size to 2 significantly reduces the dimensionality of the resulting aggregated features after average pooling. This allows aggregating features with a large amount of raw data to reduce the amount of model input data without losing effective information, helping to speed up subsequent model recognition.

[0050] In some possible embodiments, after obtaining multiple features of the object to be identified and before inputting the multiple features of the object to be identified into the first object type recognition model, the method of the embodiment of the present application may further include the following steps:

[0051] Normalization is performed on the numerical features among the multiple features, and discretization is performed on the non-numerical features among the multiple features to obtain processed feature information.

[0052] In a specific embodiment, the discretization processing includes the following methods: One-Hot Encoding, for example, for user gender, male can be represented as "1" and female can be represented as "0"; Frequency Encoding, for example, the click situation of a specific category in a preset application, such as the user clicked the category of overseas study counseling a total of 3 times; Consolidation Encoding, multiple values under certain category variables can be summarized into the same information. For example, the multiple values of the version number feature of the mobile device used include "4.2", "4.4" and "5.0". Based on experience, these three values can be summarized as "low version Android system" and represented by "0", and the other version numbers can be summarized as "high version Android system" and represented by "1".

[0053] Ultimately, the processed feature information can be converted into an N*1 numerical vector, such as (1, 0, 31, 4, 0.2, 9.3, 8.8, …, 0, 0, 1, 2, 34). This N*1 numerical vector corresponding to each object to be identified is then fed into the subsequent model for machine recognition. Furthermore, the processed feature information can be stored offline in the Hadoop Distributed File System (HDFS), facilitating quick access in subsequent processes.

[0054] S203: Inputting feature information of the object to be identified into a first target type identification model to obtain a first probability value that the object to be identified belongs to the target type.

[0055] Among them, the first target type recognition model is obtained by training the first neural network model based on meta-learning.

[0056] In an embodiment of the present application, a pre-built first neural network model is trained based on a meta-learning approach to obtain a trained first target type recognition model; meta-learning (Meta Learning): also known as learning to learn (Learn to learn), that is, to enable the intelligent agent to acquire the ability to "learn to learn" so that it can quickly learn new tasks based on past experience.

[0057] like Figure 4As shown, the meta-learning method includes a meta-training stage and a meta-testing stage. In the meta-training stage, other task data sets with high relevance to the target task (i.e., identifying the target type) are first introduced, and the gradient descent method is used for training to obtain a trained meta-model; then the meta-testing stage is entered. In the meta-testing stage, the trained meta-model is migrated to the dataset for object type recognition in this application for model fine-tuning, and finally a first target type recognition model for realizing object type recognition is obtained; the meta-training stage and the meta-testing stage will be described in detail below and will not be repeated here. The first target type recognition model is used to determine whether the object to be identified is the target type based on the feature information of the object.

[0058] In some possible embodiments, the generation method of the first target type recognition model may include: Figure 5 The following steps are shown:

[0059] S501: Construct a first neural network model.

[0060] Among them, the structure of the first neural network model can use the deep neural network model structure commonly used in the type recognition field or its improved model structure.

[0061] S503: Obtain a first training data set.

[0062] The first training dataset includes multiple sub-datasets corresponding to multiple training tasks; each of the multiple sub-datasets includes multiple training sample inputs and type label information for each of the multiple training sample inputs. In the sub-datasets corresponding to each training task, the type label information of the training sample inputs indicates whether the training sample inputs are of the corresponding type, which may be different from the target type.

[0063] For example, in the specific application scenario mentioned above, the target type is the type of user who is willing to use the network acceleration function. Considering that overseas students generally have the willingness to use the network acceleration function when visiting overseas websites, and overseas students are usually masters or doctors, it is believed that identifying whether the user is a user type willing to use the network acceleration function has certain similarities with identifying whether the user is a master or doctor. Therefore, one of the above-mentioned multiple training tasks can be to identify whether the user is a master or doctor. Correspondingly, in the corresponding sub-dataset, the type label information of each training sample input can be used to indicate whether each training sample input is a master or doctor.

[0064] S505: Perform meta-training on the first neural network model according to the first training data set to obtain a trained meta-model.

[0065] Among them, combined Figure 4The meta-learning process shown in the figure, the meta-training phase can include the following steps:

[0066] First, the parameters of the first neural network model are randomly initialized;

[0067] Secondly, the loop starts, randomly sampling several training tasks from the first training dataset to form a batch;

[0068] Next, the first gradient update is performed, which uses each training task in the batch to update the model parameters.

[0069] After the first gradient update is complete, a second gradient update is calculated based on the parameters obtained from the first gradient update. During this second gradient update, the sum of the losses corresponding to all training tasks in a batch is calculated, and stochastic gradient descent is performed on the gradients. The gradients calculated during the second gradient update are applied to the original model to update the model parameters. After this, the model ends training for that batch and returns to the second step to continue sampling the next batch.

[0070] Finally, the cycle ends until a preset training end condition is met, and a trained meta-model is obtained. The preset training end condition may include ending the training when a preset number of iterations is reached or when a target optimization parameter is reached.

[0071] S507: Obtain a second training data set.

[0072] The second training data set includes a plurality of training sample objects and target type label information of each of the plurality of training sample objects.

[0073] This step begins the meta-testing phase, that is, using the data set (second training data set) of the target task to train the meta-model to obtain a first target type recognition model suitable for the target task.

[0074] In a specific application scenario, S507 may include: roughly recalling a group of seed users based on preset rules, where seed users refer to users who are willing to use the network acceleration function; the preset rules may include screening rules for whether they have overseas study experience, screening rules for whether they are students, etc.; then, filtering is performed based on manual screening, and the filtered seed users can be used as training sample objects, and the target type label information of each training sample object indicates the willingness to use the network acceleration function.

[0075] S509: Train the meta-model according to the second training data set to obtain a first target type recognition model.

[0076] In a specific embodiment, S509 may include the following steps:

[0077] S5091: Obtain feature information of each training sample object.

[0078] The feature information of each training sample object corresponds to the feature information of the object to be identified in actual application.

[0079] In a specific embodiment, the feature information of each training sample object may include at least one of object attribute features, device attribute features, network connection attribute features, and behavior features in a preset application. The specific content of each feature can be referred to in the above embodiment and will not be repeated here.

[0080] S5093: Input the feature information of each training sample object into the meta-model to obtain a prediction value of each training sample object belonging to the target type.

[0081] S5095: Train the meta-model according to the predicted value of each training sample object belonging to the target type and the corresponding target type label information to obtain a first target type recognition model.

[0082] Specific, combined Figure 4 As shown, when training the meta-model, parameters are no longer randomly initialized, but instead are initialized using a trained meta-model. The feature information of each training sample object is input into the meta-model, and a prediction value for each training sample object belonging to the target type is output. Then, based on the corresponding target type label information, a loss value is calculated, and the meta-model parameters are updated using gradient descent. Multiple iterations are performed until a preset training end condition is met, resulting in a trained first target type recognition model. The preset training end condition can include terminating training when a preset number of iterations is reached or when the target optimization parameters are reached.

[0083] In the above embodiment, training is performed based on a meta-learning approach, and the network initialization parameters obtained during the meta-training process are used. In the recognition task of this application, only a few steps of fine-tuning are required to obtain network parameters suitable for the recognition task of this application. This not only speeds up the generation of the first target type recognition model, but also solves the problem of small amounts of training sample object data in actual application scenarios. In addition, the first target type recognition model built based on the deep neural network model can learn higher-dimensional features of the object to be recognized, increase the richness of the feature dimensions, and thus help accurately identify the type of the object to be recognized.

[0084] In some possible embodiments, before the above step S509, a plurality of training sample objects may be screened for abnormal sample objects, which may specifically include the following steps:

[0085] S5081: Obtain characteristic anomaly determination indicators corresponding to characteristic information.

[0086] S5083: According to the feature abnormality determination index, determine abnormal training sample objects from multiple training sample objects, and delete the abnormal training samples.

[0087] This application takes into account that in actual business scenarios, there may be fake users and computers controlling mobile phones. Therefore, in order to eliminate the influence of non-real users on model training, characteristic anomaly judgment indicators can be set based on business experience; for example, in a specific application scenario, the characteristic information of the training sample object may include the traffic usage of the seed user in the preset application and the time information of the traffic generation; accordingly, the characteristic anomaly judgment indicator corresponding to the traffic usage in the preset application may be that the traffic used in one day exceeds 20G; the characteristic anomaly judgment indicator corresponding to the time information of the traffic generation may be that the traffic is generated for 24 consecutive hours. Thus, seed users who use more than 20G of traffic in one day or generate traffic for 24 consecutive hours can be identified as abnormal users and then deleted.

[0088] In the embodiment of the present application, after acquiring the feature information of the object to be identified, the feature information of the object to be identified is input into a trained first target type recognition model, which outputs a first probability value indicating that the object to be identified belongs to the target type. In some application scenarios, this first probability value can be used directly to determine whether the object to be identified belongs to the target type. However, in addition to using the first target type recognition model to identify the object to be identified, the embodiment of the present application also uses a second target type recognition model to identify the object to be identified. Based on the recognition results of the two models, the determination of whether the object to be identified belongs to the target type is made, thereby ensuring recognition accuracy.

[0089] S205: Inputting the feature information of the object to be identified into a second target type identification model to obtain a second probability value that the object to be identified belongs to the target type.

[0090] The second target type recognition model is obtained by training the second neural network model based on a genetic algorithm.

[0091] In an embodiment of the present application, the hyperparameters of the pre-built second neural network model are globally tuned based on a genetic algorithm to determine the optimal hyperparameters of the model and obtain a trained pre-trained model; then, based on the pre-trained model, the data set of the object type recognition task in the present application is migrated to perform model fine-tuning, and finally a second target type recognition model for realizing object type recognition is obtained.

[0092] In some possible embodiments, the generation method of the second target type recognition model may include: Figure 6 The following steps are shown:

[0093] S601: Construct a second neural network model.

[0094] The structure of the second neural network model can use a deep neural network model structure commonly used in the field of type recognition or an improved model structure thereof. The structure of the second neural network model can be the same as or different from the structure of the first neural network model.

[0095] S603: Obtain a third training data set.

[0096] The third training data set includes multiple training sample inputs and type label information of each training sample input in the multiple training sample inputs.

[0097] Here, the third training data set may include one or more sub-data sets mentioned in step S503 of the above embodiment.

[0098] S605: Based on the third training data set, the second neural network model is trained using a genetic algorithm to obtain a pre-trained model.

[0099] In a specific embodiment, combining Figure 7 and Figure 8 As shown, S605 may include the following steps:

[0100] S701: Generate an initialization population.

[0101] The initialized population includes multiple second neural network models; the multiple initial hyperparameters corresponding to the multiple second neural network models are different.

[0102] S703: Obtain feature information of each training sample input.

[0103] This step may refer to the embodiment of obtaining feature information of the object to be identified in actual application or the embodiment of S5091, and will not be described in detail here.

[0104] S705: According to each second neural network model in the initialized population, feature information of each training sample input is identified to obtain a corresponding prediction value.

[0105] S707: Determine the fitness value of each second neural network model based on the predicted value and reward function corresponding to each second neural network model.

[0106] The fitness value represents the quality of each second neural network model.

[0107] In a specific application scenario, the fitness value can be calculated based on the reward function in reinforcement learning. The fitness value can be determined according to the following formula (1):

[0108] fitness=min(ep_r) / ep_step……(1)

[0109] Where fitness represents the fitness value, which is determined by the minimum round reward in all rounds; ep_r represents the reward value in a round, which is calculated based on the corresponding predicted value of each round; ep_step represents the total number of running steps in a round.

[0110] S709: Based on the fitness value of each second neural network model, perform crossover recombination operation, mutation operation and replacement operation to iterate the population.

[0111] Among them, during crossover recombination, the selection operation is repeated twice, two networks are selected as parents, and a new offspring is generated through crossover. The offspring will share some properties of the parents. The crossover process is as follows Figure 9 shown.

[0112] Mutation includes connection mutation and node mutation. For connection mutation, the currently active connection of the best-performing individual in the current population is first obtained based on the fitness value, and all nodes on the currently active connection are derived. Then, two nodes are selected from all nodes, and a connection is formed between the two nodes. For node mutation, the currently active connection of the best-performing individual in the population is obtained based on the fitness value, a random connection is selected from all active connections, and a node is added to this connection.

[0113] S711: When the preset number of iterations is reached, the second neural network model in the current population whose fitness value is greater than or equal to the preset value is used as a pre-training model.

[0114] The preset value may be the maximum fitness value in the current population. That is, when the preset number of iterations is reached, the network architecture of the best performing individual in the current population is retained and used as a pre-training model for subsequent training.

[0115] S607: Obtain a fourth training data set.

[0116] The fourth training data set includes a plurality of training sample objects and target type label information of each of the plurality of training sample objects.

[0117] This step can refer to the above embodiment S507 and will not be repeated here.

[0118] S609: Based on the fourth training data set, optimize the model parameters of the pre-training model to obtain a second target type recognition model.

[0119] In a specific embodiment, S609 may include the following steps:

[0120] S6091: Obtain feature information of each training sample object.

[0121] S6093: Input the feature information of each training sample object into the pre-training model to obtain a prediction value of each training sample object belonging to the target type.

[0122] S6095: Train the pre-training model according to the predicted value of each training sample object belonging to the target type and the corresponding target type label information to obtain a second target type recognition model.

[0123] In the above embodiment, training is performed based on a genetic algorithm, and the network architecture of the trained pre-trained model is used. In the recognition task of this application, only a few steps of fine-tuning are required to obtain a network model suitable for the recognition task of this application. This not only speeds up the generation of the second target type recognition model, but also solves the problem of small amount of training sample object data in actual application scenarios. In addition, the second target type recognition model built based on the deep neural network model can learn higher-dimensional features of the object to be recognized, increase the richness of the feature dimension, and thus help to accurately identify the type of the object to be recognized.

[0124] In a specific embodiment, Figure 10 As shown, the present application mainly includes four stages in the training process, namely the sample preparation stage, the feature processing stage, the first target type recognition model training stage and the second target type recognition model stage; first, the sample preparation stage can mainly include obtaining training sample objects, obtaining feature information of training sample objects, and filtering abnormal training sample objects; second, the feature processing stage can mainly include constructing aggregate features of training sample objects, normalizing numerical features and discretizing non-numerical features, and storing the processed feature information offline in the Hadoop distributed file system; second, the first target type recognition model training stage mainly includes training through the first training data set to obtain a meta-model, and training the meta-model through the second training data set to obtain a first target type recognition model; second, the second target type recognition model stage can include training through the third training data set to obtain a pre-trained model, and training the pre-trained model through the fourth training data set to obtain a second target type recognition model.

[0125] S207: Determine the type of the object to be identified according to the first probability value and the second probability value.

[0126] In some possible embodiments, determining the type of the object to be identified based on the first probability value and the second probability value may include the following steps:

[0127] S2071: Determine an average probability value based on the first probability value and the second probability value.

[0128] S2073: Determine whether the average probability value is greater than or equal to a preset probability value; if the average probability value is greater than or equal to the preset probability value, execute step S2075; otherwise, execute step S2077.

[0129] S2075: Determine that the type of the object to be identified is the target type.

[0130] S2077: Determine that the type of the object to be identified is not the target type.

[0131] In the above embodiment, the feature information of the object to be identified is respectively identified by the first target type identification model and the second target type identification model to determine the first probability value and the second probability value that the object to be identified belongs to the target type. Thus, based on the average of the first probability value and the second probability value, it is determined whether the object to be identified belongs to the target type.

[0132] In a specific application scenario, this application compares the manual method, traditional recognition model method and the recognition method of this application in the existing technology, and delivers advertisements of network acceleration software to the client of the user identified as the target type, and obtains the click-through rate and conversion rate of the user on the advertisement under different methods; according to the advertisement click-through rate in the figure, the present application solution is improved by an average of 182.7% compared with other technical solutions, and from the perspective of advertisement conversion rate, the present application solution is improved by an average of 178.41% compared with other technical solutions.

[0133] To sum up, the embodiment of the present application obtains a first target type recognition model through training based on meta-learning, and obtains a second target type recognition model through training based on a genetic algorithm. The first target type recognition model and the second target type recognition model are used to respectively identify the feature information of the object to be identified, and the predicted probability values of the two models are combined to determine whether the object to be identified belongs to the target type. This can greatly improve the recognition accuracy of the object type. In this way, it can solve the problems of low efficiency and poor effect in the existing technology of identifying object types through manual or traditional non-deep learning models.

[0134] The present application also provides an object type identification device, such as Figure 11 As shown, the object type identification device 110 includes:

[0135] An acquisition module 1101 is configured to acquire feature information of an object to be identified;

[0136] The first recognition module 1102 is configured to input feature information of the object to be recognized into a first target type recognition model to obtain a first probability value of the object to be recognized belonging to the target type; the first target type recognition model is obtained by training the first neural network model based on meta-learning;

[0137] The second recognition module 1103 is configured to input feature information of the object to be recognized into a second target type recognition model to obtain a second probability value of the object to be recognized belonging to the target type; the second target type recognition model is obtained by training the second neural network model based on a genetic algorithm;

[0138] The determination module 1104 is configured to determine the type of the object to be identified based on the first probability value and the second probability value.

[0139] In some possible embodiments, a generation module of a first target type recognition model is further included, which is configured to execute the construction of a first neural network model; obtain a first training data set; the first training data set includes multiple sub-data sets corresponding to multiple training tasks; each sub-data set in the multiple sub-data sets includes multiple training sample inputs and type label information of each training sample input in the multiple training sample inputs; meta-train the first neural network model according to the first training data set to obtain a trained meta-model; obtain a second training data set; the second training data set includes multiple training sample objects and target type label information of each training sample object in the multiple training sample objects; train the meta-model according to the second training data set to obtain a first target type recognition model.

[0140] In some possible embodiments, the generation module of the first target type recognition model is further configured to execute the acquisition of feature information of each training sample object; input the feature information of each training sample object into the meta-model to obtain the prediction value of each training sample object belonging to the target type; train the meta-model according to the prediction value of each training sample object belonging to the target type and the corresponding target type label information to obtain the first target type recognition model.

[0141] In some possible embodiments, the generation module of the first target type recognition model is further configured to execute a feature anomaly determination indicator corresponding to the acquired feature information; based on the feature anomaly determination indicator, determine abnormal training sample objects from multiple training sample objects and delete the abnormal training samples.

[0142] In some possible embodiments, a generation module of a second target type recognition model is further included, which is configured to execute the construction of a second neural network model; obtain a third training data set; the third training data set includes multiple training sample inputs and type label information of each training sample input in the multiple training sample inputs; based on the third training data set, the second neural network model is trained using a genetic algorithm to obtain a pre-trained model; obtain a fourth training data set; the fourth training data set includes multiple training sample objects and target type label information of each training sample object in the multiple training sample objects; based on the fourth training data set, the model parameters of the pre-trained model are optimized to obtain a second target type recognition model.

[0143] In some possible embodiments, the generation module of the second target type recognition model is further configured to generate an initialization population; the initialization population includes multiple second neural network models; the multiple initial hyperparameters corresponding to the multiple second neural network models are different; the feature information of each training sample input is obtained; according to each second neural network model in the initialization population, the feature information of each training sample input is identified to obtain the corresponding prediction value; based on the prediction value and reward function corresponding to each second neural network model, the fitness value of each second neural network model is determined; based on the fitness value of each second neural network model, a cross-recombination operation, a mutation operation and a replacement operation are performed to iterate the population; when the preset number of iterations is reached, the second neural network model in the current population whose fitness value is greater than or equal to the preset value is used as a pre-training model.

[0144] In some possible embodiments, the feature information includes multiple features corresponding to multiple dimensions; the device also includes a feature processing module, which is configured to perform normalization processing on numerical features among the multiple features and discretization processing on non-numerical features among the multiple features to obtain processed feature information.

[0145] In some possible embodiments, the target type is a type that is willing to use the target application function; the acquisition module 1101 is also configured to execute the acquisition of at least one feature information of the object attribute characteristics, device attribute characteristics, network connection attribute characteristics and behavioral characteristics in the preset application of the object to be identified.

[0146] In some possible embodiments, the determination module 1104 is further configured to determine an average probability value based on the first probability value and the second probability value; if the average probability value is greater than or equal to a preset probability value, determine that the type of the object to be identified is the target type.

[0147] It should be noted that the device and method embodiments in the device embodiment are based on the same inventive concept.

[0148] An embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the object type identification method provided in the above method embodiment.

[0149] Furthermore, Figure 12 The hardware structure diagram of an electronic device for implementing the object type identification method provided in the embodiment of the present application is shown. The electronic device may participate in or include the object type identification device provided in the embodiment of the present application. Figure 12As shown, the electronic device 100 may include one or more (illustrated as 1002a, 1002b, ..., 1002n in the figure) processors 1002 (the processor 1002 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 1004 for storing data, and a transmission device 1006 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply and / or a camera. It will be understood by those skilled in the art that Figure 12 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 12 More or fewer components than shown, or with Figure 12 Different configurations shown.

[0150] It should be noted that the one or more processors 1002 and / or other data processing circuits described above may generally be referred to herein as "data processing circuitry". The data processing circuitry may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuitry may be a single independent processing module, or may be incorporated in whole or in part into any of the other components of the electronic device 100 (or mobile device). As described in the embodiments of the present application, the data processing circuitry serves as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0151] The memory 1004 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the object type identification method described in the embodiments of the present application. The processor 1002 executes various functional applications and data processing by running the software programs and modules stored in the memory 1004, that is, implementing the above-mentioned object type identification method. The memory 1004 may include a high-speed random access memory and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1004 may further include a memory remotely located relative to the processor 1002, and these remote memories may be connected to the electronic device 100 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0152] The transmission device 1006 is used to receive or send data via a network. Specific examples of the aforementioned network may include a wireless network provided by the communication provider of the electronic device 100. In one embodiment, the transmission device 1006 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In one embodiment, the transmission device 1006 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0153] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the electronic device 100 (or mobile device).

[0154] An embodiment of the present application also provides a computer-readable storage medium, which can be set in an electronic device to store at least one instruction or at least one program related to implementing an object type identification method in a method embodiment. The at least one instruction or the at least one program is loaded and executed by the processor to implement the object type identification method provided by the above method embodiment.

[0155] Optionally, in this embodiment, the storage medium may be located in at least one of a plurality of 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 a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk, or an optical disk.

[0156] It should be noted that the order of the embodiments of the present application described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. The above description is of specific embodiments of the present application. 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 an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0157] The various embodiments in this application are described in a progressive manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the device and electronic device embodiments are generally similar to the method embodiments, so the description is relatively simple. For relevant parts, refer to the partial description of the method embodiments.

[0158] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0159] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A method for identifying object types, characterized in that: include: Obtaining feature information of the object to be identified; The feature information includes at least one feature information of an object attribute feature, a device attribute feature, a network connection attribute feature, and a behavior feature in a preset application; Inputting the feature information of the object to be identified into a first target type recognition model to obtain a first probability value of the object to be identified belonging to the target type; the first target type recognition model is obtained by training a meta-model based on a second training data set; The meta-model is obtained by meta-training the first neural network model based on the first training data set; the target type is the type of person who is willing to use the target application function; the first training data set includes multiple training sample inputs corresponding to multiple training tasks and type label information of each training sample input in the multiple training sample inputs; the second training data set includes multiple training sample objects and target type label information of each training sample object in the multiple training sample objects; Inputting the feature information of the object to be identified into a second target type recognition model to obtain a second probability value of the object to be identified belonging to the target type; the second target type recognition model is obtained by optimizing model parameters of a pre-trained model based on a fourth training data set; The pre-trained model is obtained by training the second neural network model using a genetic algorithm based on a third training data set; the third training data set includes multiple training sample inputs and type label information of each training sample input in the multiple training sample inputs; the fourth training data set includes multiple training sample objects and target type label information of each training sample object in the multiple training sample objects; The type of the object to be identified is determined according to the first probability value and the second probability value.

2. The object type identification method according to claim 1, characterized in that: The method for generating the first target type recognition model includes: Constructing the first neural network model; Obtain a first training data set; the first training data set includes multiple sub-data sets corresponding to multiple training tasks; Each sub-dataset in the plurality of sub-datasets includes a plurality of training sample inputs and type label information of each training sample input in the plurality of training sample inputs; Performing meta-training on the first neural network model according to the first training data set to obtain a trained meta-model; Obtain a second training data set; The meta-model is trained according to the second training data set to obtain the first target type recognition model.

3. The object type identification method according to claim 2, characterized in that: The step of training the meta-model according to the second training data set to obtain the first target type recognition model includes: Acquiring feature information of each training sample object; Inputting the characteristic information of each training sample object into the meta-model to obtain a prediction value of each training sample object belonging to the target type; The meta-model is trained according to the predicted value of each training sample object belonging to the target type and the corresponding target type label information to obtain the first target type recognition model.

4. The object type identification method according to claim 3, characterized in that: Before the meta-model is trained according to the second training data set to obtain the first target type recognition model, the method further includes: Obtaining a characteristic abnormality determination indicator corresponding to the characteristic information; According to the characteristic abnormality determination index, abnormal training sample objects are determined from the multiple training sample objects, and the abnormal training samples are deleted.

5. The object type identification method according to claim 1, characterized in that: The method for generating the second target type recognition model includes: Constructing the second neural network model; Obtain a third training data set; Based on the third training data set, the second neural network model is trained using a genetic algorithm to obtain a pre-trained model; Obtain a fourth training data set; Based on the fourth training data set, the model parameters of the pre-training model are optimized to obtain the second target type recognition model.

6. The object type identification method according to claim 5, characterized in that: The method of training the second neural network model based on the third training data set using a genetic algorithm to obtain a pre-trained model includes: generating an initialization population; wherein the initialization population includes a plurality of second neural network models; and the plurality of second neural network models correspond to different initial hyperparameters; Obtaining feature information of each training sample input; According to each second neural network model in the initialized population, identifying the feature information of each training sample input to obtain a corresponding prediction value; Determining the fitness value of each of the second neural network models based on the predicted value and the reward function corresponding to each of the second neural network models; Based on the fitness value of each of the second neural network models, performing a crossover recombination operation, a mutation operation, and a replacement operation to iterate the population; When the preset number of iterations is reached, the second neural network model in the current population whose fitness value is greater than or equal to the preset value is used as the pre-training model.

7. The object type identification method according to claim 1, characterized in that: The feature information includes multiple features corresponding to multiple dimensions; before inputting the feature information of the object to be identified into the first target type recognition model, the method further includes: Normalization processing is performed on numerical features among the multiple features, and discretization processing is performed on non-numerical features among the multiple features to obtain processed feature information.

8. The object type identification method according to any one of claims 1 to 7, characterized in that: The determining the type of the object to be identified according to the first probability value and the second probability value includes: determining an average probability value based on the first probability value and the second probability value; If the average probability value is greater than or equal to a preset probability value, it is determined that the type of the object to be identified is the target type.

9. An object type identification device, characterized in that: include: An acquisition module is configured to acquire feature information of an object to be identified; The feature information includes at least one feature information of an object attribute feature, a device attribute feature, a network connection attribute feature, and a behavior feature in a preset application; a first recognition module configured to input feature information of the object to be recognized into a first target type recognition model to obtain a first probability value of the object to be recognized belonging to the target type; the first target type recognition model is obtained by training a meta-model based on a second training data set; The meta-model is obtained by meta-training the first neural network model based on the first training data set; the target type is the type of person who is willing to use the target application function; the first training data set includes multiple training sample inputs corresponding to multiple training tasks and type label information of each training sample input in the multiple training sample inputs; the second training data set includes multiple training sample objects and target type label information of each training sample object in the multiple training sample objects; a second recognition module configured to input the feature information of the object to be recognized into a second target type recognition model to obtain a second probability value of the object to be recognized belonging to the target type; the second target type recognition model is obtained by optimizing model parameters of a pre-trained model based on a fourth training data set; The pre-trained model is obtained by training the second neural network model using a genetic algorithm based on a third training data set; the third training data set includes multiple training sample inputs and type label information of each training sample input in the multiple training sample inputs; the fourth training data set includes multiple training sample objects and target type label information of each training sample object in the multiple training sample objects; The determination module is configured to determine the type of the object to be identified based on the first probability value and the second probability value.

10. The object type identification device according to claim 9, characterized in that: Also includes: A generation module of a first target type recognition model is configured to execute construction of a first neural network model; Obtain a first training data set; The first training data set includes multiple sub-data sets corresponding to multiple training tasks; Each of the multiple sub-datasets includes multiple training sample inputs and type label information of each of the multiple training sample inputs; Performing meta-training on the first neural network model according to the first training data set to obtain a trained meta-model; Obtain a second training data set; The second training data set includes multiple training sample objects and target type label information of each training sample object in the multiple training sample objects; the meta-model is trained according to the second training data set to obtain a first target type recognition model.

11. The object type identification device according to claim 10, characterized in that: The generation module of the first target type recognition model is further configured to execute: Acquire feature information of each training sample object; input the feature information of each training sample object into the meta-model to obtain a prediction value of each training sample object belonging to the target type; train the meta-model according to the prediction value of each training sample object belonging to the target type and the corresponding target type label information to obtain the first target type recognition model.

12. The object type identification device according to claim 11, characterized in that: The generation module of the first target type recognition model is further configured to execute: Acquire a feature anomaly determination index corresponding to the feature information; determine an abnormal training sample object from the multiple training sample objects based on the feature anomaly determination index, and delete the abnormal training sample.

13. The object type identification device according to claim 9, characterized in that: Also includes: A second target type recognition model generation module is configured to execute construction of the second neural network model; Obtain a third training data set; Based on the third training data set, the second neural network model is trained using a genetic algorithm to obtain a pre-trained model; and a fourth training data set is obtained; Based on the fourth training data set, the model parameters of the pre-training model are optimized to obtain the second target type recognition model.

14. The object type identification device according to claim 13, characterized in that: The generation module of the second target type recognition model is further configured to execute: generating an initialization population; wherein the initialization population includes a plurality of second neural network models; and the plurality of second neural network models correspond to different initial hyperparameters; Obtaining feature information of each training sample input; According to each second neural network model in the initialized population, identifying the feature information of each training sample input to obtain a corresponding prediction value; Determining the fitness value of each of the second neural network models based on the predicted value and the reward function corresponding to each of the second neural network models; Based on the fitness value of each of the second neural network models, performing a crossover recombination operation, a mutation operation, and a replacement operation to iterate the population; When the preset number of iterations is reached, the second neural network model in the current population whose fitness value is greater than or equal to the preset value is used as the pre-training model.

15. The object type identification device according to claim 9, characterized in that: The feature information includes multiple features corresponding to multiple dimensions; the device further includes: The feature processing module is configured to perform normalization processing on the numerical features among the multiple features and discretization processing on the non-numerical features among the multiple features to obtain processed feature information.

16. The object type identification device according to any one of claims 9 to 15, characterized in that: The determining module is further configured to execute: An average probability value is determined based on the first probability value and the second probability value; if the average probability value is greater than or equal to the preset probability value, the type of the object to be identified is determined to be the target type.

17. An electronic device, characterized in that: The electronic device includes a processor and a memory, wherein the memory stores at least one instruction or at least one program, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the object type identification method according to any one of claims 1 to 8.

18. A computer storage medium, characterized in that The storage medium stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by the processor to implement the object type identification method according to any one of claims 1 to 8.

19. A computer program product, characterized in that The computer program product includes at least one instruction or at least one program segment, and the at least one instruction or the at least one program segment is loaded and executed by a processor to implement the object type identification method according to any one of claims 1 to 8.

Images