A state recognition method, device, storage medium and equipment

Abstract

Embodiments of the present application disclose a state recognition method and device, a storage medium and equipment, which can be applied to the fields of artificial intelligence and Internet of Vehicles. The method comprises: performing linear state recognition on a target object according to object characteristic information of the target object about a target service, to obtain first processing state information for reflecting the target object about the target service; performing cross state recognition on the target object according to an information amount contained in the object characteristic information and the object characteristic information, to obtain second processing state information for reflecting the target object about the target service; fusing the object characteristic information, the first processing state information and the second processing state information, to obtain target fusion characteristic information; and performing combined state recognition on the target object according to the target fusion characteristic information, to obtain target processing state information for reflecting the target object about the target service. Through the present application, the recognition accuracy of the object about the service processing state can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a state recognition method, a storage medium and equipment. BACKGROUND

[0002] With the acceleration of people's life pace, the business that an object needs to handle every day also increases, and a mobile business platform for recommending business to the object is used, and is widely used in the fields of finance, medical treatment, education institutions and the like, and can effectively improve the business handling efficiency of the user. The mobile business platform identifies the handling state information of the object about the business, and recommends the business to be handled to the object according to the handling state information.

[0003] At present, the mobile business platform mainly sets the identification rule through artificial experience, and identifies the handling state information of the object about the business based on the identification rule, but the artificial experience is easily affected by subjective factors and the like, and the artificial experience is limited, so that the accuracy of the handling state information obtained through identification is relatively low. SUMMARY

[0004] The technical problem to be solved by the embodiments of the present application is to provide a state recognition method, device, storage medium and equipment, so as to improve the recognition accuracy of the handling state of the object about the business.

[0005] The embodiments of the present application provide a state recognition method, comprising:

[0006] obtaining object feature information of a target object about a target business, and an information amount contained in the object feature information;

[0007] performing linear state recognition on the target object according to the object feature information, to obtain first handling state information for reflecting the target object about the target business;

[0008] performing cross state recognition on the target object according to the information amount contained in the object feature information and the object feature information, to obtain second handling state information for reflecting the target object about the target business;

[0009] fusing the object feature information, the first handling state information and the second handling state information, to obtain target fusion feature information;

[0010] performing combined state recognition on the target object according to the target fusion feature information, to obtain target handling state information for reflecting the target object about the target business.

[0011] The embodiments of the present application provide a state recognition device, comprising:

[0012] The first obtaining module is configured to obtain object characteristic information of a target object about a target service and an information amount contained in the object characteristic information.

[0013] The linear state recognition module is configured to perform linear state recognition on the target object according to the object characteristic information, to obtain first processing state information reflecting the target object about the target service.

[0014] The cross state recognition module is configured to perform cross state recognition on the target object according to the information amount contained in the object characteristic information and the object characteristic information, to obtain second processing state information reflecting the target object about the target service.

[0015] The fusion module is configured to fuse the object characteristic information, the first processing state information and the second processing state information, to obtain target fusion characteristic information.

[0016] The combined state recognition module is configured to perform combined state recognition on the target object according to the target fusion characteristic information, to obtain target processing state information reflecting the target object about the target service.

[0017] The linear state recognition module includes:

[0018] The deep feature extraction unit is configured to call a feature extraction layer of the linear state recognition model, to perform deep feature extraction on the object characteristic information, to obtain deep characteristic information of the target object about the target service.

[0019] The linear state recognition unit is configured to call a state recognition layer of the linear state recognition model, to perform linear state recognition on the target object according to the deep characteristic information, to obtain the first processing state information reflecting the target object about the target service.

[0020] The deep feature extraction unit is specifically configured to:

[0021] Call the feature extraction layer of the linear state recognition model, to perform correlation relationship recognition on the object characteristic information in each dimension of the object characteristic information, to obtain object correlation characteristic information.

[0022] Perform linear processing on the object correlation characteristic information, to obtain linear characteristic information.

[0023] According to the linear characteristic information, determine the deep characteristic information of the target object about the target service.

[0024] The cross state recognition module includes:

[0025] The first determination unit is configured to call an influence degree determination layer of the cross state recognition model, to determine an influence degree of the object characteristic information according to the information amount contained in the object characteristic information.

[0026] The cross-state recognition unit is used to call the state recognition layer of the cross-state recognition model, and perform cross-state recognition on the target object based on the influence degree and object feature information to obtain the second processing state information of the target object regarding the target business.

[0027] Specifically, the cross-state recognition unit is used for:

[0028] The state recognition layer of the cross-state recognition model is invoked to determine the weighting parameters of the object feature information based on the influence degree.

[0029] The object feature information is weighted according to the weighting parameters to obtain the weighted object feature information;

[0030] The weighted object feature information and the object feature information are combined by feature cross-combination to obtain cross-combined feature information;

[0031] State recognition is performed on the cross-combined feature information to obtain second processing state information that reflects the target object's relationship with the target business.

[0032] The combined state recognition module includes:

[0033] The acquisition unit is used to call the weight prediction layer of the combined state recognition model to obtain the combined weight between any two dimensions of feature information in the target fusion feature information;

[0034] The feature combination unit is used to call the feature combination layer of the combined state recognition model. Based on the combination weight, it performs feature combination on any two dimensions of the target fusion feature information to obtain the target fusion feature information after feature combination.

[0035] The combined state recognition unit is used to call the state recognition layer of the combined state recognition model to perform state recognition on the target fusion feature information after feature combination, and obtain target processing state information that reflects the target object with respect to the target business.

[0036] The status recognition device also includes:

[0037] The second acquisition module is used to acquire sample object feature information of the sample object regarding the target business and the target processing status information label of the sample object regarding the target business;

[0038] The first calling module is used to call the linear state recognition model, perform linear state recognition on the sample object based on the sample object feature information, and obtain the first recognition processing state information reflecting the sample object with respect to the target business.

[0039] The second calling module is configured to call the cross-state identification model, perform cross-state identification on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, and obtain second identification processing state information reflecting the sample object with respect to the target business;

[0040] The third calling module is configured to call the initial combined-state identification model, fuse the sample object feature information, the first identification processing state information, and the second identification processing state information, obtain target sample fusion feature information, perform combined-state prediction on the sample object according to the target sample fusion feature information, and obtain predicted target processing state information reflecting the sample object with respect to the target business.

[0041] The first adjusting module is configured to adjust the initial combined-state identification model according to the target processing state information label and the predicted target processing state information, and obtain a combined-state identification model. The combined-state identification model is configured to determine target processing state information reflecting a target object with respect to a target business according to object feature information of the target object.

[0042] The first adjusting module is configured to determine a combined-state prediction error of the initial combined-state identification model according to the target processing state information label and the predicted target processing state information. If the combined-state prediction error is not in a convergence state, the initial combined-state identification model is adjusted according to the combined-state prediction error. The adjusted initial combined-state identification model is determined as the combined-state identification model.

[0043] The state identification device further includes:

[0044] The fourth calling module is configured to call the initial linear-state identification model, perform linear-state prediction on the sample object according to the sample object feature information, and obtain first predicted processing state information reflecting the sample object with respect to the target business.

[0045] The second adjusting module is configured to adjust the initial linear-state identification model according to the target processing state information label and the first predicted processing state information, and obtain a linear-state identification model. The linear-state identification model is configured to perform linear-state prediction on an object and obtain first processing state information of the object with respect to a business.

[0046] The second adjusting module includes: determining a linear-state prediction error of the initial linear-state identification model according to the target processing state information label and the first predicted processing state information. If the linear-state prediction error is not in a convergence state, the initial linear-state identification model is adjusted according to the linear-state prediction error. The adjusted initial linear-state identification model is determined as the linear-state identification model.

[0047] The state recognition device further includes:

[0048] The fifth calling module is configured to call the initial cross-state recognition model, perform cross-state prediction on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, and obtain second prediction processing state information reflecting the sample object about the target service.

[0049] The third adjusting module is configured to determine a second loss function according to the target processing state information label and the second prediction processing state information, adjust the initial cross-state recognition model according to the second loss function, and obtain the cross-state recognition model; and the cross-state recognition model is configured to perform cross-state prediction on the object to obtain second processing state information of the object about the service.

[0050] The third adjusting module is configured to determine a cross-state prediction error of the initial cross-state recognition model according to the target processing state information label and the second prediction processing state information; if the cross-state prediction error is not in a convergence state, adjust the initial cross-state recognition model according to the cross-state prediction error; and determine the adjusted initial cross-state recognition model as the cross-state recognition model.

[0051] The computer device provided in the embodiment of the present application includes a processor and a memory.

[0052] The processor is connected to the memory, and the memory is configured to store a computer program, which, when executed by the processor, causes the computer device to perform the method provided in the embodiment of the present application.

[0053] The computer readable storage medium provided in the embodiment of the present application stores a computer program, and the computer program is adapted to be loaded and executed by a processor, so that a computer device having the processor performs the method provided in the embodiment of the present application.

[0054] The computer program product or the computer program provided in the embodiment of the present application 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 the processor executes the computer instructions, so that the computer device performs the method provided in the embodiment of the present application.

[0055] In the embodiments of the present application, the first processing state information reflecting the target object about the target service is obtained by performing linear state recognition on the target object according to the object characteristic information of the target object. Meanwhile, the second processing state information reflecting the target object about the target service is obtained by performing cross state recognition on the target object according to the information amount contained in the object characteristic information and the object characteristic information. Further, the target fusion characteristic information is obtained by fusing the object characteristic information of the target object and the state information (i.e., the first processing state information and the second processing state information) obtained by using different state recognition modes. The target processing state information reflecting the target object about the target service is obtained by performing combined state recognition on the target object according to the target fusion characteristic information. That is, the target processing state information reflecting the target object about the target service is obtained by performing optimized recognition on the processing state of the target object about the target service according to the target fusion characteristic information, which can improve the recognition accuracy of the target processing state information reflecting the target object about the target service. BRIEF DESCRIPTION OF DRAWINGS

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0057] Figure 1 is a schematic diagram of the architecture of a state recognition system provided by the embodiments of the present application;

[0058] Figure 2 is a schematic diagram of an application scenario of state recognition provided by the embodiments of the present application;

[0059] Figure 3 is a schematic diagram of the flow of a state recognition method provided by the embodiments of the present application;

[0060] Figure 4 is a schematic diagram of the network structure of a linear state recognition model provided by the embodiments of the present application;

[0061] Figure 5 is a schematic diagram of the structure of a cross state recognition model provided by the embodiments of the present application;

[0062] Figure 6 is a schematic diagram of feature cross combination provided by the embodiments of the present application;

[0063] Figure 7 is a schematic diagram of the structure of a combined state recognition model provided by the embodiments of the present application;

[0064] Figure 8 is a flowchart of a process for obtaining target processing state information provided by an embodiment of the present application;

[0065] Figure 9 is a flowchart of a state recognition method provided by an embodiment of the present application;

[0066] Figure 10 is a schematic diagram of model training provided by an embodiment of the present application;

[0067] Figure 11 is a schematic diagram of combined state recognition model training provided by an embodiment of the present application;

[0068] Figure 12 is a schematic diagram of model prediction effect provided by an embodiment of the present application;

[0069] Figure 13 is a schematic diagram of model business effect provided by an embodiment of the present application;

[0070] Figure 14 is a structural schematic diagram of a state recognition apparatus provided by an embodiment of the present application;

[0071] Figure 15 is a structural schematic diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0072] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0073] Artificial Intelligence (AI) is the theory, method, technology and application system of using digital computer or digital computer controlled machine to simulate, extend and expand human intelligence, perceive environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that machines have the functions of perception, reasoning and decision making. Artificial intelligence technology is a comprehensive discipline, which involves a wide range of fields, including hardware and software technologies. Artificial intelligence basic technologies generally include technologies such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction system, mechatronics, etc. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, automatic driving, intelligent transportation and several other major directions.

[0074] Machine Learning (ML) is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a subject that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rule-based learning. In this application, the object feature information of the target object can be linearly identified by a linear state identification model to obtain first processing state information reflecting the target object about the target business, and the object feature information of the target object can be cross-identified by a cross-state identification model to obtain second processing state information reflecting the target object about the target business. Then, according to the first processing state information, the second processing state information and the object feature information, the target object is combined and identified by a combined state identification model to obtain target processing state information reflecting the target object about the target business. Through this application, the linear state identification model, the cross-state identification model and the combined state identification model can be used to simulate or implement human learning behavior to identify the state of the target object, obtain the target processing state information reflecting the target object about the target business, and further improve the accuracy of processing state information identification.

[0075] See Figure 1 , Figure 1 is a structural schematic diagram of a state identification system provided by an embodiment of the present application. As shown inFigure 1 As shown, the status identification system may include a server 10 and a user terminal cluster. The user terminal cluster may include one or more user terminals; the number of user terminals is not limited here. Figure 1 As shown, it can specifically include user terminal 100a, user terminal 100b, user terminal 100c, ..., user terminal 100n. Figure 1 As shown, user terminals 100a, 100b, 100c, ..., 100n can each connect to the server 10 via a network, so that each user terminal can interact with the server 10 through the network connection.

[0076] Each user terminal in this user terminal cluster can include: smartphones, tablets, laptops, desktop computers, wearable devices, smart home devices, head-mounted devices, in-vehicle terminals, and other intelligent terminals with status recognition capabilities. It should be understood that, for example... Figure 1 Each user terminal in the user terminal cluster shown can have the target application (i.e., the application client) installed. When the application client runs on each user terminal, it can interact with the aforementioned... Figure 1 Data interaction occurs between the servers 10 shown.

[0077] Among them, such as Figure 1 As shown, the server 10 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides 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, and big data and artificial intelligence platforms.

[0078] For ease of understanding, the embodiments of this application may be described in detail below. Figure 1 From the plurality of user terminals shown, one user terminal is selected as the target user terminal. The target user terminal may include: a smartphone, tablet computer, laptop computer, desktop computer, smart TV, or other smart terminal with status recognition capabilities. For example, for ease of understanding, embodiments of this application may... Figure 1The user terminal 100a shown as a target user terminal can obtain object characteristic information of a target object about a target service, the target object can refer to a service user, and the object characteristic information can include object attribute information possessed by the service user and service operation characteristic information when the service user handles the target service, etc. The user terminal 100a can send the object characteristic information of the target object about the target service to the server 10, which includes a linear state recognition model, a cross state recognition model, and a combined state recognition model. The server 10 can call the linear state recognition model, perform linear state recognition on the target object based on the object characteristic information of the target object, and obtain first handling state information reflecting the target object about the target service. For example, the target service can refer to a provident fund payment service, and the first handling state information can be a probability that the target object is in a normal provident fund payment state or a probability that the target object is in an abnormal provident fund payment state.

[0079] Further, the server 10 can call the cross state recognition model, obtain an information amount contained in the object characteristic information of the target object, perform cross state recognition on the target object based on the object characteristic information of the target object and the information amount contained in the object characteristic information, and obtain second handling state information reflecting the target object about the target service. For example, the target service can also refer to a provident fund payment service, and the second handling state information can be a probability that the target object is in a normal provident fund payment state or a probability that the target object is in an abnormal provident fund payment state.

[0080] Further, the server 10 can call a combined state recognition model, fuse the first processing state information, the second processing state information and the object feature information to obtain target fusion feature information, and perform combined state recognition on the target object according to the target fusion feature information to obtain target processing state information reflecting the target object about the target service. Specifically, since the linear state recognition model and the cross state recognition model have different processing modes for the object feature information, the obtained processing state information is also different. Therefore, the first processing state information output by the linear state recognition model and the second processing state information output by the cross state recognition model are taken as new feature information of the target object by the combined state recognition model. The new feature information of the target object and the object feature information are combined and taken as input of the combined state recognition model to perform combined state recognition on the target object to obtain target processing state information reflecting the target object about the target service. In this way, compared with manually identifying the target processing state information of the target object about the target service, the present application can perform supervised weight optimization on the outputs of the linear state recognition model and the cross state recognition model through the combined state recognition model, and can improve the accuracy and efficiency of the target processing state recognition of the target object about the target service. The embodiments of the present application can be applied to various scenes, including but not limited to cloud technology, artificial intelligence, intelligent transportation, assisted driving, etc.

[0081] For ease of understanding, further, Figure 2 , Figure 2 is a state recognition application scenario provided by the embodiments of the present application. As shown in Figure 2 , the server 20b can be the server 10 described above, as shown in Figure 2 , the target user terminal 20a can be any one of the user terminal cluster shown in the above Figure 1 , for example, the target user terminal 20a can be the user terminal 100b described above. As shown in Figure 2As shown, the target user terminal 20a obtains the object characteristic information of the target object about the housing fund payment service, the target object can refer to a service user, and the object characteristic information can refer to user basic attribute information (such as age, position, gender, etc.) of the service user, user equipment attribute information (such as the type of equipment owned by the service user, online time, user Wi-Fi connection time, etc.) and service operation characteristic information (such as click rate and conversion rate for a specific advertisement, etc.) of the service user when handling the service. The target user terminal 20a can send the object characteristic information 20c of the target object to the server 20b, perform linear state recognition on the object characteristic information 20c of the target object through the linear state recognition model 20d in the server 20b, and output first handling state information for the target object about the target service. The object characteristic information 20c of the target object is recognized through the cross-state recognition model 20e in the server 20b, and second handling state information for reflecting the target object about the target service is output. Further, the object characteristic information 20c of the target object, the first handling state information output by the linear state recognition model 20d, and the second handling state information output by the cross-state recognition model 20e can be fused through the combined state recognition model 20f to obtain target fusion characteristic information, and the target fusion characteristic information is recognized through the combined state recognition model 20f to output target handling state information 20g for reflecting the target object about the target service.

[0082] Further, the server 20b can return the target handling state information 20g for reflecting the target object about the target service to the target user terminal 20a. The target user terminal 20a can perform service recommendation on the target object according to the target handling state information 20g for reflecting the target object about the target service, and input recommendation information 20i in the service recommendation interface 20h. For example, when the target service is a marriage recommendation service, if the target handling state information 20g indicates that the target object is not married, the recommendation information 20i of a social application can be input in the recommendation service interface 20h. Through the present application, the target handling state information of the target object about the target service can be accurately recognized, and the target handling state information can provide a reference basis for the service handling associated with the target object.

[0083] Please refer to Figure 3 , Figure 3 is a flow diagram of a state recognition method provided by an embodiment of the present application. The state recognition method can be executed by a computer device, which can be a server (such as the server 10 in the above Figure 1 ) or a user terminal (such as any user terminal in the user terminal cluster in the above Figure 1 ). The present application does not limit this. As Figure 3As shown, the state recognition method can include, but is not limited to, the following steps:

[0084] In S101, object characteristic information of a target object about a target service is obtained, and an information amount contained in the object characteristic information is obtained.

[0085] Specifically, the computer device can perform state recognition on the target object according to the object characteristic information of the target object about the target service, and obtain target processing state information reflecting the target object about the target service. The target processing state information can provide a reference basis for business processing associated with the target service. Specifically, the computer device can obtain object characteristic information of a target object about a target service. The target object can be a business user, a vehicle, an article, an animal, etc. When the target object is a business user, the object characteristic information of the target object is obtained through authorization of the user. Specifically, the following embodiments are described by taking the target object as a business user as an example. The object characteristic information of the target object about the target service can be determined according to the business field and the business type of the target service.

[0086] For example, if the target service is a public accumulation fund payment service, the object characteristic information of the target object about the public accumulation fund payment service can include object attribute information and business processing characteristic information when the target object processes related services (including public accumulation fund payment services and other services). The object attribute information can include object basic information, device basic attribute information, network connection attribute information, etc. The business processing characteristic information can include the click rate and conversion rate of the target object for a specific service (such as a feature advertisement). The object basic information can include attribute information possessed by the user, such as object gender, object age, object place of origin, object working place, object residence city, etc. The device basic attribute information can include device resolution, device CPU core number, device model, device price, etc. The network connection attribute can include online duration, number of Wi-Fi connections, and earliest time of daily Wi-Fi connection, etc. Further, the computer device can obtain an information amount contained in each dimension of the object characteristic information of the target object. The information amount can be used to reflect the importance of each object characteristic information. If the information amount contained in the object characteristic information of a certain dimension is more, it means that the importance of the object characteristic information of the certain dimension is higher. If the information amount contained in the object characteristic information of a certain dimension is less, it means that the importance of the object characteristic information of the certain dimension is lower.

[0087] Optionally, the computer device can acquire online feature information of the target object in real time or periodically, the online feature information can be real-time generated feature information of the target object, i.e., current feature information currently generated by the target object, and the computer device can also acquire offline feature information stored offline in advance, the offline feature information can be historical generated feature information of the target object, i.e., the offline feature information can be historical feature information obtained by aggregating historical generated feature information of the target object. The computer device can respectively perform feature preprocessing on the online feature information and the offline feature information of the target object to obtain preprocessed online feature information and preprocessed offline feature information, and splice the preprocessed online feature information and the preprocessed offline feature information to obtain object feature information of the target object. The computer device can splice the preprocessed online feature information and the preprocessed offline feature information in a direct connection manner, i.e., splice the preprocessed offline feature information behind the preprocessed online feature information, or splice the preprocessed online feature information behind the preprocessed offline feature information, to obtain a long vector feature information. For example, the preprocessed online feature information is (1, 0, 2, 1.1), the preprocessed offline feature information is (41, …, 3, 1, 14), and the preprocessed offline feature information (41, …, 3, 1, 14) is directly spliced behind the preprocessed online feature information (1, 0, 2, 1.1) to obtain a long vector (1, 0, 2, 1.1, 41, …, 3, 1, 14).

[0088] Specifically, the online feature information of the target object can include general feature information and service feature information of the target object about multiple services, etc. The computer device can acquire the general feature information of the target object from the online real-time log data through a routine timing feature acquisition component. The routine timing feature acquisition component is based on fixed calculation logic and a calculation timer, acquires the general feature information of the target object from the online real-time log data at a fixed time, and performs feature calculation (i.e., feature encoding) on the general feature information of the target object. The computer device can also acquire the service operation feature information of the target object about multiple services from the online real-time log data through a service feature acquisition component set by a service side. The service feature acquisition component is also based on fixed calculation logic and a calculation timer, acquires the service operation feature information of the target object about multiple services from the online real-time log data at a fixed time, and performs feature calculation (i.e., feature conversion) on the service operation feature information of the target object. Specifically, the general feature information of the target object can include object basic information (such as gender, age, position, and belonging place), device basic attribute information (such as device type and device brand), and network connection attribute information (such as the number and position of Wi-Fi connections on the same day), etc. Specifically, the service operation feature information of the target object about multiple services can include the total time length of Wi-Fi use of the target object per day and other use information, and can also include the click behavior and conversion rate of the target object for multiple services, etc. The feature acquisition granularity of the computer device acquiring the online feature information of the target object can be half an hour (i.e., acquiring the online feature information of the target object every half an hour), one hour (i.e., acquiring the online feature information of the target object every hour), etc. Further, the computer device can splice the general feature information of the target object and the service operation feature information of the target object about multiple services as the online feature information of the target object. The spliced online feature information can be a high-dimensional vector.

[0089] Optionally, the computer device can perform abnormal feature information filtering processing and feature coding on the online feature information of the target object to obtain preprocessed online feature information. Specifically, the computer device can filter the online feature information of the target object based on an abnormal feature filtering rule to obtain online feature information after abnormal feature filtering. For example, the abnormal feature information is filtered according to the flow usage of the target object when using the product, and the time distribution of the flow, etc. For example, the feature information of each online time exceeding 24 hours is filtered. Specifically, when the computer device codes the online feature information after abnormal feature filtering, the information type of the online feature information of each dimension in the online feature information after abnormal feature filtering can be obtained, and according to the information type of the online feature information of each dimension, a target coding mode for coding the corresponding online feature information is determined to obtain the online feature information after feature coding.

[0090] Among them, the computer device can perform feature cleaning, filtering and verification on the object feature information of the target object. Specifically, the specific way in which the computer device performs feature cleaning, filtering and verification on the object feature information of the target object can include: since the object feature information is obtained through multiple strategy logic paths, the data quality of the object feature information is difficult to guarantee, so feature data quality detection is needed. The specific method is as follows: first, the computer device can formulate a feature quality standard based on business experience to clean, filter and verify the calculated features, and the quality standard includes but is not limited to: the length of time using the application program per day is less than 16 hours, etc. Second, the computer device can verify the existing features according to the preset strategy, and eliminate invalid and abnormal features, such as the length of time that the target object uses a certain application program cannot be greater than 24 hours, etc. Third, the computer device can splice the features that meet the business requirements, and mark the features that do not meet the requirements as missing.

[0091] Specifically, the computer device can encode the online feature information in One-Hot encoding, Count encoding, Category encoding, NaN encoding, Consolidation encoding, Trajectory encoding, Traffic encoding, or the like. One-Hot encoding, also known as one-hot encoding, mainly uses a bit state register to encode a state, each state has its own independent register bit, and only one bit is valid at any time. For example, the One-Hot encoding result of the object gender feature is: male: (1, 0), female: (0, 1). Count encoding is mainly used for binary classification problems, that is, each category is replaced by the number of samples with the corresponding label "1", that is, the number of occurrences of a certain category is calculated. For example, the Wi-Fi POI feature of the object (i.e., the point of interest feature) can be encoded using Count encoding to identify the interest degree of the object and the POI (i.e., the point of interest). For example, the object visited the POI (i.e., the point of interest) "food-Chinese food-Cantonese food" 3 times. Category encoding is a coding method for converting high-dimensional sparse classification variables into low-dimensional dense classification variables. According to data analysis, many features have strong sparsity. In order to improve the stability of feature extraction of the model, Category encoding is introduced to convert high-dimensional sparse classification variables into low-dimensional dense classification variables using a neural network. NaN encoding is a method for identifying missing data using "non-number" markers. For missing value processing of features, methods such as "elimination", "average value filling", and "missing marker" are used. Experimental results show that the non-number expression method for missing values can improve the accuracy of feature extraction. Consolidation encoding is a method for abstracting feature information using a coarse-grained label, that is, it is a method for abstracting feature information using a coarse-grained label. For example, multiple values under a certain classification variable can be abstracted into the same information, such as "4.2", "4.4", and "5.0" in the system version feature of an Android phone. Based on experience, these three values can be abstracted into "low version Android system". Experiments show that the Consolidation encoding processing method, i.e., replacing the traditional "Android system" with "low version Android system", can improve the accuracy of feature extraction. Trajectory encoding is a method for connecting the object behavior trajectory of a target object to obtain the behavior trajectory feature information of the target object. For example, based on Trajectory encoding, the Wi-Fi connection trajectory data of the target object is encoded to capture the user Wi-Fi behavior connection trajectory information.The traffic encoding can be used to list feature encode the use behavior sequence of the target object, to obtain low-dimensional dense behavior features. For example, based on the traffic encoding mode, the traffic use behavior sequence of the target object about different kinds of Apps is encoded and extracted, to obtain low-dimensional dense behavior features.

[0092] Optionally, the computer device can also perform abnormal feature information filtering processing and feature encoding on offline feature information of the target object, to obtain preprocessed offline feature information. The specific mode can refer to the above-mentioned mode of preprocessing online feature information of the target object, and details are not described herein again.

[0093] In S102, linear state recognition is performed on the target object according to the object feature information, to obtain first processing state information for reflecting the target object about the target business.

[0094] Specifically, the computer device can perform linear state recognition on the target object according to the object feature information of the target object, to obtain first processing state information for reflecting the target object about the target business. The first processing state can be used to reflect the business processing state of the target object for the target business. For example, when the target business is a public accumulation payment business, the first processing state information can be a probability that the target object belongs to a normal public accumulation payment state, or the first processing state information can be a probability that the target object belongs to an abnormal public accumulation payment state.

[0095] Optionally, the specific mode of performing linear state recognition on the target object according to the object feature information to obtain the first processing state information for reflecting the target object about the target business can include: calling a feature extraction layer of a linear state recognition model, performing deep feature extraction on the object feature information, to obtain deep feature information of the target object about the target business. Calling a state recognition layer of the linear state recognition model, performing linear state recognition on the target object according to the deep feature information, to obtain the first processing state information for reflecting the target object about the target business.

[0096] Specifically, the computer device can invoke a linear state recognition model, which can be used to perform linear state recognition on the target object according to the object feature information of the target object, to obtain the processing state information of the target object about the target business. The computer device can perform deep feature extraction on the object feature information of the target object through a feature extraction layer in the linear state recognition model, to obtain the deep feature information of the target object about the target business. The deep feature extraction can refer to deep learning-based feature extraction, and the deep feature information can refer to the internal rules and representation levels of the object feature information. The computer device can input the deep feature information of the target object about the target business into a state recognition layer in the linear state recognition model, to perform linear state recognition on the target object, to obtain the first processing state information of the target object about the target business. It can be understood that the state recognition layer in the linear state recognition model can classify the deep feature information of the target object about the target business, to obtain the first processing state information of the target object about the target business.

[0097] Optionally, the specific manner in which the computer device performs deep feature extraction on the object feature information to obtain the deep feature information of the target object about the target business can include: invoking a feature extraction layer of a linear state recognition model, performing association relationship recognition on the object feature information of each dimension in the object feature information, to obtain object association feature information. Performing linear processing on the object association feature information to obtain linear feature information, and determining the deep feature information of the target object about the target business according to the linear feature information.

[0098] Specifically, the feature extraction layer in the linear state recognition model can include an association feature extraction sub-layer and a linear feature extraction sub-layer. The computer device can invoke the association feature extraction sub-layer in the feature extraction layer to perform association relationship recognition on the object feature information in each dimension of the object feature information of the target object, and obtain object association feature information. It can be understood that in the association feature extraction sub-layer, the object feature information in any two dimensions can be combined to extract the object association feature information between the object feature information in the any two dimensions, so that more feature information can be extracted to accurately identify the first processing state information of the target object with respect to the target business. Further, the computer device can perform linear processing on the object association feature information of the target object through the linear feature extraction sub-layer in the feature extraction layer to obtain linear feature information, and determine the deep feature information of the target object with respect to the target business according to the linear feature information. It can be understood that in the linear feature extraction sub-layer, the object association feature information of the target object is linearly combined or nonlinearly combined to extract the linear relationship between the object association feature information of the target object, and further obtain more feature information to accurately identify the first processing state information of the target object with respect to the target business.

[0099] Optionally, the linear state recognition model can be an FNN model (Factorisation Machine supported Neural Network, i.e., a neural network based on a factorization machine). The FNN model uses the hidden layer and the weight thereof obtained by pre-training of an FM network structure (i.e., a factorization machine model) as the initial value of the first layer of the neural network, and then continuously stacks fully connected layers, and finally outputs the predicted click rate. The FNN model can be understood as a special embedding (i.e., embedding encoding layer) + MLP (i.e., multi-layer perceptron). The FNN model requires that the feature domain feature dimensions after embedding of each feature are consistent, and the initialization of the embedding weight is pre-trained by FM. The linear state recognition model can obtain an embedding feature vector of the object feature information through the FM network structure, and then input the embedding feature vector into a DNN network structure (i.e., a deep neural network structure), which is equivalent to introducing prior expert experience obtained from FM, can accelerate the extraction of feature information, and can also reduce the feature extraction overhead and improve the efficiency of feature extraction.

[0100] When the object feature information of the target object is input into the linear state recognition model, this object feature information is a large-scale discrete and sparse vector. The linear state recognition model can divide this large-scale discrete and sparse object feature information into object feature information in N dimensions. In each dimension, only one value is 1, and the rest are 0 (i.e., one-hot feature encoding and partitioning of the object feature information). The batch size (i.e., the number of samples in one training session) of the linear state recognition model can be 1024, and the embedding dimension can be 16. The Adam optimizer (Adaptive Gradient) can be used to adjust the model parameters in the linear state recognition model. The Adam optimizer can adjust different learning rates for different parameters, updating frequently changing parameters with smaller step sizes and sparse parameters with larger step sizes. At the same time, the Dropout parameter in the linear state recognition model can be set to 0.5. The Dropout parameter can effectively alleviate the overfitting problem of the linear state recognition model, thereby enabling the training of a deeper and wider network.

[0101] Specifically, the correlation feature extraction sublayer in a linear state recognition model can refer to an FM network structure, which can be used to combine pairwise features to extract the correlation between them. The linear feature extraction sublayer in a linear state recognition model can refer to a DNN network structure, which can extract the linear relationship between input and output features, i.e., perform multiple linear or non-linear combinations of input features to extract higher-order feature information.

[0102] like Figure 4 As shown, Figure 4 This is a schematic diagram of the network structure of a linear state recognition model provided in an embodiment of this application, such as... Figure 4 As shown, the computer device can input the object feature information of each dimension in the object feature information into the association feature extraction sublayer 40a in the linear state recognition model. Through the association feature extraction sublayer 40a in the linear state recognition model, the association relationship of the object feature information of each dimension in the target object's object feature information can be identified to obtain the object association feature information. The computer device can then input the object association feature information into the first linear feature extraction sublayer 40b in the linear state recognition model to perform linear processing on the object association feature information to obtain the first linear feature information.

[0103] The calculation formula for the first linear feature extraction sublayer 40b can be shown in the following formula (1):

[0104] L1=tanh(w1z)+b1 (1)

[0105] Wherein, L1 in formula (1) refers to the first linear feature extraction sub-layer, W1 refers to the weight matrix in the first linear feature extraction sub-layer, z refers to the object-related feature information input into the first linear feature extraction sub-layer, and b1 refers to the bias constant in the first linear feature extraction sub-layer.

[0106] Further, as shown in FIG. 4, the computer device can input the first linear feature information output by the first linear feature extraction sub-layer into the second linear feature extraction sub-layer 40c in the linear state recognition model, perform linear processing on the first linear feature information output by the first linear feature extraction sub-layer, and obtain second linear feature information. Figure 4

[0107] Wherein, the calculation formula of the second linear feature extraction sub-layer 40b can be as shown in the following formula (2):

[0108] L2=tanh(w2L1)+b2 (2)

[0109] Wherein, L2 in formula (1) refers to the second linear feature extraction sub-layer, W2 refers to the weight matrix in the second linear feature extraction sub-layer, L1 refers to the first linear feature information output by the first linear feature extraction sub-layer, and b2 refers to the bias constant in the second linear feature extraction sub-layer.

[0110] Further, the computer device can perform linear processing on the object-related feature information output by the related feature extraction sub-layer 40a through multiple perception layers to obtain deep feature information. The deep feature information is classified through the state recognition layer 40d in the linear state recognition model to obtain the first processing state information reflecting the target object with respect to the target service. The specific content of the linear state recognition model can refer to the content of step S102, and the embodiments of the present application will not be repeated here.

[0111] S103, cross-state recognition of the target object is performed according to the information amount contained in the object feature information and the object feature information, and second processing state information reflecting the target object with respect to the target service is obtained.

[0112] ​Specifically, the computer device can obtain an information amount contained in the object characteristic information, perform cross-state recognition on the target object according to the information amount contained in the object characteristic information and the object characteristic information, and obtain second processing state information reflecting the target object with respect to the target service. It can be understood that, compared with the linear state recognition of the target object in step S102, the computer device performs cross-state recognition on the target object according to the information amount contained in the object characteristic information, extracts the object characteristic information of the target object from a different angle, and further identifies the processing state information of the target object with respect to the target service under different recognition angles. For example, when the target service is a public accumulation payment service, the computer device can perform cross-state recognition on the target object according to the object characteristic information of the target object, and obtain second processing state information of the target object under the cross-state recognition angle, that is, under the cross-state recognition mode, determine the probability that the target object is in the normal payment state of the public accumulation, or the probability that the target object is in the non-normal payment state of the public accumulation.

[0113] Optionally, the specific manner in which the computer device obtains the second processing state information reflecting the target object with respect to the target service can include: calling an influence degree determination layer of the cross-state recognition model, determining the influence degree of the object characteristic information according to the information amount contained in the object characteristic information. Calling a state recognition layer of the cross-state recognition model, performing cross-state recognition on the target object according to the influence degree and the object characteristic information, and obtaining the second processing state information reflecting the target object with respect to the target service.

[0114] Specifically, the computer device can call the cross-state recognition model, which can perform cross-state recognition on the target object according to the object feature information of the target object to obtain the second processing state information of the target object about the target business. Through the influence degree determination layer of the cross-state recognition model, the amount of information contained in the object feature information of the target object is obtained, and the influence degree of the object feature information is determined according to the amount of information contained in the object feature information. Wherein, the more information the object feature information contains, the higher the corresponding influence degree is; the less information the object feature information contains, the lower the corresponding influence degree is. The influence degree determination layer is used to learn the importance (i.e. influence degree) of different features, which can be used for feature weighting of the features, and then the features with more information are enhanced and the features with less information are weakened. Wherein, the computer device can perform object feature information summary statistics on the object feature information of the target object in each dimension, i.e. compress the representation of the object feature information in each dimension to obtain the amount of information of the object feature information in each dimension. Specifically, the computer device can determine the influence degree (i.e. importance) of the object feature information in each dimension according to the amount of information of the object feature information in each dimension, and further, the computer device can call the state recognition layer in the cross-state recognition model to perform weighting processing on the object feature information according to the influence degree to obtain the weighted object feature information. According to the weighted object feature information, cross-state recognition is performed to obtain the second processing state information of the target object about the target business.

[0115] Optionally, the computer device performs cross-state recognition on the target object according to the influence degree and the object feature information, and the specific way of reflecting the second processing state information of the target object about the target business can include: calling the state recognition layer of the cross-state recognition model, determining the weighting parameter of the object feature information according to the influence degree. The object feature information is weighted processed according to the weighting parameter to obtain the weighted object feature information. The weighted object feature information and the object feature information are combined to obtain the cross-combined feature information. The cross-combined feature information is recognized to obtain the second processing state information of the target object about the target business.

[0116] Specifically, the computer device can call a state recognition layer of the cross-state recognition model, which can include a weighted parameter determination sub-layer, a weighted processing sub-layer, a cross combination sub-layer, and a state recognition sub-layer. The computer device can determine the weighted parameters of the object feature information according to the influence degree of the object feature information through the weighted parameter determination sub-layer. Specifically, the greater the influence degree of the object feature information, the greater the weighted parameters of the object feature information; the smaller the influence degree of the object feature information, the smaller the weighted parameters of the object feature information. After the computer device obtains the weighted parameters of the object feature information in each dimension of the object feature information, the computer device can perform weighted processing on the corresponding object feature information according to the weighted parameters in each dimension through the weighted processing sub-layer to obtain the object feature information after weighted processing. In this way, the object feature information containing more information can be enhanced, and the object feature information containing less information can be weakened, which can improve the accuracy of feature extraction.

[0117] Further, the computer device can cross-combine the object feature information after weighted processing and the object feature information (i.e., the object feature information without weighted processing) in a cross-combination manner through the cross-combination sub-layer to obtain cross-combination feature information. Specifically, the cross-combination manner can include an inner product manner, a Hadamard product manner, and a combination manner of the inner product and the Hadamard product, etc., wherein the inner product manner is also called a scalar product, which is a binary operation that accepts two vectors on real numbers R and returns a real number value scalar. The Hadamard product manner is that the Hadamard product is a kind of operation of matrix. If matrix A and matrix B are two matrices of the same order, then the Hadamard product of matrix A and matrix B is equal to the dot product between matrix A and matrix B, that is, the Hadamard product is the dot product of two matrices of the same order. The combination manner of the inner product and the Hadamard product is a cross-combination manner that combines the inner product and the Hadamard product and introduces an additional parameter matrix W. Further, the computer device can perform state recognition on the cross-combination feature information through the state recognition sub-layer to obtain the second processing state information reflecting the target object about the target business. In this way, the computer device can cross-combine the object feature information after weighted processing and the object feature information to obtain cross-combination feature information, and recognize the second processing state information reflecting the target object about the target business according to the cross-combination feature information. Compared with linear state recognition, cross-state recognition is to recognize the state of the object feature information of the target object from different angles, that is, to recognize the state of the object feature information of the target object from different angles to obtain the recognition result of the target object under different recognition angles.

[0118] Optionally, the cross-state recognition model can be a FiBiNet model (i.e., Combining Feature Importance and Bilinear feature Interaction for Click-Through Rate Prediction), the essence of which is to dynamically learn the importance of features by using a Squeeze-Excitation Network (i.e., compression + weight extraction network) structure and to better model cross features by using a bilinear function. The main role of the SENET Layer (i.e., impact degree determination layer) in the FiBiNet model is to learn the importance of different features and to weight the corresponding features according to the importance. In this way, features with little information can be weakened, and features with more information can be enhanced. The target object feature information can be input into the SENET layer (i.e., impact degree determination layer) in the FiBiNet model, and the object feature information after weighting processing is output. Specifically, the SENET layer (i.e., impact degree determination layer) in the FiBiNet model can include the following specific implementation process: vector encoding the object feature information on each dimension of the object feature information of the target object to obtain the object feature vector of the object feature information on each dimension, combining the object feature vector of the object feature information on each dimension to obtain the object feature group E of the object feature information of the target object as input, generating the weight vector group A = [a1,..., ai,... af] corresponding to the object feature group E, and the object feature vector of the object feature information on one dimension corresponds to one weight vector. Multiply the object feature group E of the object feature information on multiple dimensions by the weight vector group A to obtain a new group of encoding vectors V = [v1,..., vi,... vf], i.e., the object feature information after weighting processing.

[0119] Specifically, the SENET layer in the FiBiNet model includes a Squeeze structure layer (i.e., an information amount extraction sublayer, also referred to as a compression layer), an Excitation structure layer (i.e., a weighting parameter determination sublayer), and a Re-Weight structure layer (i.e., a weighting processing sublayer). The Squeeze structure layer mainly performs an operation of summarizing statistics on the object feature vectors of the object feature information in each dimension of the object feature information, uses a pooling operation to compress the object feature groups E = [e1,..., ei,..., ef] of the object feature information in multiple dimensions to obtain a statistical vector Z = [z1,..., zi,..., zf], where zi represents the global information (i.e., the information amount) of the object feature information in the i-th dimension. The Excitation structure layer is used to learn the importance weight (i.e., the weight parameter) of the object feature information in each dimension based on the statistical vector Z = [z1,..., zi,..., zf], and a two-layer neural network can be used to learn the importance weight (i.e., the weight parameter) of the object feature information in each dimension, the first layer being a dimension reduction layer and the second layer being a dimension promotion layer. The Re-Weight structure layer can use the weight parameter obtained by the Excitation structure layer to perform weighting processing on the object feature groups E = [e1,..., ei,..., ef] of the object feature information to obtain the weighted object feature information. In this way, the object feature information with less information amount can be weakened, and the object feature information with more information amount can be enhanced.

[0120] As shown in Figure 5 , Figure 5 is a schematic diagram of a structure of a cross-state recognition model provided by an embodiment of the present application, as shown in Figure 5As shown, the computer device can input the object feature information 50a of the target object into the embedding vector encoding layer 50b in the cross-state recognition model, perform embedding vector encoding on the object feature information of the target object, and obtain the object embedding vector of the object feature information. The computer device can input the object embedding feature into the SENET layer (i.e., the influence degree determination layer) 50c of the cross-state recognition model, obtain the influence degree of the object embedding vector of the object feature information, and perform weighted processing on the object embedding vector of the object feature information according to the influence degree, to obtain the weighted processed object feature information. Further, the computer device can input the weighted processed object feature information into the embedding vector encoding layer 50d to perform feature encoding on the weighted processed object feature information again, and then input the weighted processed object feature information into the first feature cross layer 50e in the cross-state recognition model to perform feature cross on the weighted processed object feature information, to obtain the first feature cross information. The computer device can input the object embedding vector of the object feature information into the second feature cross layer 50f in the cross-state recognition model to perform feature cross on the object embedding vector of the object feature information, to obtain the second feature cross information. The computer device can combine the first feature cross information and the weighted processed second feature cross information of the object feature information according to the feature combination layer 50g in the cross-state recognition model, to obtain the cross combination feature information. Further, the computer device performs convolutional fully connected processing on the cross combination feature information through the fully connected layer 50h in the cross-state recognition model, and performs classification processing on the fully connected processed cross feature information through the state recognition layer 50i in the cross-state recognition model, to input the second processing state information reflecting the target object with respect to the target business. The specific content of the cross-state recognition model can refer to the content of step S103, which will not be described herein again.

[0121] Further, after obtaining the weighted object feature information, the computer device can use the Bilinear-Interaction structure layer (i.e., the cross combination sublayer) in the FiBiNet model to perform cross combination of the weighted object feature information and the object feature information to obtain cross combination feature information. The Bilinear-Interaction structure layer in the FiBiNet model uses the combination of inner product and Hadamard product to perform cross combination of the weighted object feature information and the object feature information to obtain the cross combination feature information. Specifically, in the cross combination sublayer in the FiBiNet model, the computer device can first use the inner product to perform point multiplication of the object feature information and the parameter matrix W to obtain a first cross combination result. The computer can then use the Hadamard product to perform point multiplication of the first cross combination result and a same-order matrix to obtain first cross information. Similarly, the computer device can first use the inner product to perform point multiplication of the weighted object feature information and the parameter matrix W to obtain a second cross combination result. The computer can then use the Hadamard product to perform point multiplication of the second cross combination result and a same-order matrix to obtain second cross information. The first cross information and the second cross information are combined to obtain the cross combination feature information.

[0122] As shown in Figure 6 , Figure 6 is a schematic diagram of feature cross combination provided by an embodiment of the present application, as shown in Figure 6 , the computer device can use any one of the inner product method 60a, the Hadamard method 60b, or the combination of the inner product and the Hadamard method 60c to perform cross combination of the weighted object feature information and the object feature information to obtain cross combination feature information. Optionally, as shown in Figure 6 , the computer device can use the inner product method 60a to obtain a same-order matrix of the same order as the object feature information, invert the same-order matrix to obtain an inverted matrix, and perform point multiplication of the object feature information and the inverted matrix to obtain first cross information of the object feature information. Similarly, the inner product method 60a is used to obtain second cross information of the weighted object feature information, and the first cross information and the second cross information are combined to obtain the cross combination feature information.

[0123] Optionally, as shown in Figure 6As shown, the computer device can adopt Hadamard mode 60b to obtain a same order matrix with the object feature information, and multiply the same order matrix with the object feature information to obtain first cross feature information. Similarly, the second cross feature information of the weighted object feature information is obtained by adopting Hadamard mode 60b, and the first cross feature information and the second cross feature information are combined to obtain cross combined feature information.

[0124] Optionally, as shown in FIG. 6, the computer device can adopt the combination of inner product and Hadamard mode 60c to multiply the object feature information with a weight parameter W, and then multiply the object feature information with a same order matrix to obtain first cross feature information. Similarly, the second cross feature information of the weighted object feature information is obtained by adopting the combination of inner product and Hadamard mode 60c, and the first cross feature information and the second cross feature information are combined to obtain cross combined feature information. Figure 6

[0125] When the cross state combination model is a FiBiNet model, compared with a traditional embedding layer, a SENET layer is added in the FiBiNet model to encode the object feature information of the target object to obtain an influence degree of the object feature information, and the object feature information of the target object is weighted processed according to the influence degree. In this way, the object feature information with a larger influence degree (i.e., more information) can be enhanced, and the object feature information with a smaller influence degree (i.e., less information) can be weakened, thereby improving the accuracy of feature extraction. Meanwhile, the computer device can determine a weight parameter for weighting processing the object feature information of the target object according to the influence degree of the object feature information, and weight process the object feature information of the target object according to the weight parameter to obtain weighted object feature information. In addition, the FiBiNet model does not use a traditional inner product (i.e., Inner Product) and Hadamard mode (i.e., Hadamard Product) to cross combine the weighted object feature information and the object feature information. Instead, the inner product and Hadamard mode are combined, and a parameter matrix is introduced to cross combine the weighted object feature information and the object feature information to obtain cross combined feature information, so as to obtain the correlation between the object feature information, thereby improving the accuracy of feature extraction.

[0126] ​The FiBiNet model can have good feature extraction functions for features containing a large amount of classification information (i.e., Categorical Feature), features with high feature cardinality (i.e., quantity or dimension) (i.e., Feature Cardinality), and relatively sparse features. The text description type feature can be processed into a sentence sequence feature by using a tokenizer (i.e., Tokenizer), and the sentence sequence feature can be used as the input of the FiBiNet model, so as to improve the accuracy of feature extraction, and then accurately identify the second processing state information of the target object about the target business. When training the cross-state recognition model (i.e., FiBiNet model), the ReduceRatio (i.e., model gradient descent rate) in the FiBiNet model can be set to 6-8, so as to accelerate the convergence effect of the FiBiNet model, and also to make the performance of the FiBiNet model better.

[0127] S104, fuse the object feature information, the first processing state information and the second processing state information to obtain target fusion feature information.

[0128] Specifically, the computer device can fuse the object feature information, the first processing state information and the second processing state information of the target object to obtain the target fusion feature information. Specifically, the computer device can concatenate the first processing state information and the second processing state information after the object feature information of the target object to obtain the target fusion feature information. In this way, the linear state recognition model and the cross-state recognition model are used to output the first processing state information and the second processing state information of the object feature information of the target object about the target business, and the first processing state information and the second processing state information are used as new feature information of the target object.

[0129] S105, according to the target fusion feature information, the target object is combined to recognize the state to obtain the target processing state information reflecting the target object about the target business.

[0130] Specifically, the computer device can perform combined state recognition on the target object according to the target fusion feature information, and obtain target processing state information reflecting the target object about the target business. The target processing state information is associated with the target business. For example, the target business can be a provident fund payment business, and the target processing state information reflecting the target object about the provident fund payment business can be a probability that the target object is in a normal provident fund payment state, or a probability that the target object is in an abnormal provident fund payment state. The computer device fuses the first processing state information and the second processing state information with the object feature information of the target object to obtain object feature information in multiple dimensions. According to the object feature information in multiple dimensions, the target processing state information reflecting the target object about the target business is obtained by performing combined state recognition on the target object, which can improve the recognition accuracy of the target processing state information reflecting the target object about the target business.

[0131] Optionally, the computer device can perform combined state recognition on the target object to obtain the target processing state information reflecting the target object about the target business in the following manner. The computer device can call a weight prediction layer of the combined state recognition model to obtain a combination weight between feature information in any two dimensions of the target fusion feature information. The computer device can call a feature combination layer of the combined state recognition model to perform feature combination on the feature information in each dimension of the target fusion feature information based on the combination weight, and obtain feature-combined target fusion feature information. The computer device can call a state recognition layer of the combined state recognition model to perform state recognition on the feature-combined target fusion feature information, and obtain the target processing state information reflecting the target object about the target business.

[0132] Specifically, the computer device can call a combined state recognition model that can perform combined state recognition on the target object according to the target fusion feature information, and obtain the target processing state information reflecting the target object about the target business. The computer device can call a weight prediction layer in the combined state recognition model to obtain a combination weight between feature information in any two dimensions of the target fusion feature information, that is, the weight prediction layer can identify the representation ability of the combined feature information in any two dimensions. If the representation ability of the combined feature information is stronger, the corresponding combination weight is higher; if the representation ability of the combined feature information is weaker, the corresponding combination weight is lower. In this way, the feature information with strong representation ability combined together can be combined together, and the accuracy of feature extraction can be improved.

[0133] Furthermore, the computer device can invoke the feature combination layer in the combined state recognition model, and combine the feature information in the corresponding two dimensions based on the combination weights between feature information in any two dimensions, to obtain the target fusion feature information after feature combination. Specifically, the computer device can obtain feature information i in the target fusion feature information, where feature information i is feature information in any dimension of the target fusion feature information, and obtain the combination weights between feature information i and feature information in other dimensions of the target fusion feature information. The computer device can determine the importance of combining feature information i with other feature information based on the combination weights between feature information i and feature information in other dimensions of the target fusion feature information; that is, the more important the combination, the higher the combination weight, and the less important the combination, the lower the combination weight. Furthermore, the computer device can perform weighted processing on the combined feature information between feature information i and other feature information based on the combination weights between feature information i and other feature information, to obtain the feature combination information between feature information i and other feature information. The feature combination information between feature information in each dimension of the target fusion feature information and other feature information is summarized to obtain the target fusion feature information after feature combination.

[0134] Furthermore, the computer device can invoke the state recognition layer in the combined recognition model to perform state recognition on the target fusion feature information after feature combination. That is, it can classify and process the target fusion feature information after feature combination to obtain target processing state information reflecting the target object's status regarding the target business. This target processing state information can provide a reference for the business processing of other businesses related to the target object. For example, if the target business is housing provident fund payment, business recommendations can be made to the target object based on its housing provident fund payment status (i.e., normal payment (i.e., already paid) or abnormal payment (i.e., not paid)). For instance, advertising can be placed on the target object based on its housing provident fund payment status. For example, if the target business is dating recommendation, the target processing state information is the target object's dating status. If the target object is currently single, dating and social networking software can be recommended to it.

[0135] Optionally, the combined state recognition model can be an AutoInt model (i.e., Automatic Feature Interaction Learning via Self-Attentive Neural Networks). The AutoInt model is essentially a method for automatically finding high-order cross-features, which can compensate for the weakness of the multilayer perceptron in the linear state recognition model in capturing the ability of multiplicative feature combinations, and can better explain which feature combinations have strong representation ability. The bottom layer of the AutoInt model is an encoding layer (i.e., an embedding layer), which is similar to the design of DeepFM (a network structure for extracting low-order features and high-order features, including a neural network part and a factorization machine part). All discrete and continuous features are mapped into an equal-length encoding vector. The discrete features can be directly viewed by a feature encoding table, and the multi-value discrete features are mapped into an encoding vector using average pooling. The continuous features are equivalent to multiplying a Dense layer without a bias. The purpose of the Dense layer is to extract the correlation between the features through the Dense layer for nonlinear transformation, and finally map them to the output space.

[0136] As shown in Figure 7 , Figure 7 is a schematic diagram of a structure of a combined state recognition model provided by an embodiment of the present application, as shown in Figure 7 , the computer device can input the target fusion feature information 70a into the encoding layer 70b in the combined state recognition model, and output the fusion encoding vector of the target fusion feature information 70a. The computer device can input the fusion encoding vector into the weight prediction layer in the feature cross layer 70c in the combined state recognition model, and obtain the combination weight between the feature information of any two dimensions in the target fusion feature information. Further, the computer device can call the feature combination layer in the feature cross layer 70c, and perform feature combination on the feature information of each dimension in the target fusion feature information based on the combination weight, to obtain the target fusion feature information after feature combination. Further, the computer device can call the output layer (i.e., the state recognition layer) in the combined state recognition model, and perform state recognition on the target fusion feature information after feature combination, to obtain the target processing state information reflecting the target object about the target business. The specific content of the combined state recognition model can be referred to the content of step S105, which will not be described here.

[0137] When the combination state recognition model is an AutoInt model, high-order feature cross can be automatically learned, and the explainability is improved. Meanwhile, a new method based on a self-attentive neural network can automatically learn high-dimensional feature cross, and effectively improves the accuracy of the combination state as the target processing state information of the target object about the target business predicted by the model. The batch size of the combination state recognition model can be 512, and the embedding dimension can be 10. Meanwhile, the model parameters in the combination state recognition model can be adjusted using an Adam optimizer (Adaptive Gradient). The Adam optimizer can adjust different learning rates for different parameters, and update the parameters that change frequently with a smaller step size, and update the sparse parameters with a larger step size. Meanwhile, the Dropout parameter in the combination state recognition model can be set to 0.5. The Dropout parameter can effectively alleviate the overfitting problem of the linear state recognition model, so that a deeper and wider network is trained.

[0138] The first processing state information output by the linear state recognition model and the second processing state information output by the cross state recognition model can be used as new feature information of the target sample by the combination state recognition model. Meanwhile, the first processing state information output by the linear state recognition model and the second processing state information output by the cross state recognition model can be supervised weight optimized to improve the prediction accuracy of the target processing state information of the target sample about the target business.

[0139] When the above embodiments of the present application are applied to specific products or technologies, the permission or consent of the user (i.e., the target object is a business user) is required, and the collection, use and processing of the object feature information of the target object and the log data of the target object about the multimedia data need to comply with relevant laws, regulations and standards of the country and region. That is, the computer device can obtain the object feature information of the target object and the log data of the target object about the multimedia data only when the computer device obtains the authorization information of the user for the above information. That is, the object feature information of the target object and the log data of the target object about the multimedia data are obtained only after the user authorizes.

[0140] For example, the computer device displays a permission prompt interface in the multimedia interface of the multimedia platform, which is used to prompt the user that the object feature information of the target object and the log data of the target object about the multimedia data are currently being collected. After obtaining the confirmation operation of the user on the permission prompt interface, the step of obtaining the object feature information of the target object and the log data of the target object about the multimedia data is started, otherwise it is ended.

[0141] As Figure 8 shown, Figure 8 is a flowchart of obtaining target processing state information provided by an embodiment of the present application, as Figure 8 shown, when triggering the computer device to identify the target processing state information of the target object, the computer device can perform a start operation 80a. The computer device can perform reading a trained solidification model offline 80b, which can be a trained linear state identification model, a cross state identification model, and a combined state identification model. The computer device can pull the online real-time features of the target object 80c, and read the offline features 80d, which can be referred to the description in step S101, and will not be repeated here. The computer device can splice the online real-time features and offline features of the target object 80e to obtain the object feature information of the target object. Further, the computer device can input the spliced features into the model 80f, it can be understood that the computer device can input the spliced object feature information into the linear state identification model to perform linear state identification on the object feature information of the target object, and output the first processing state information reflecting the target object about the target business. The computer device can input the spliced object feature information into the cross state identification model to perform cross state identification on the object feature information of the target object, and output the second processing state information reflecting the target object about the target business. The computer device can input the spliced object feature information, the first processing state information, and the second processing state information into the combined state identification model to perform cross state identification on the target object, and return the calculation probability of the combined state identification model 80g, which refers to the probability information of the processing state of the target object about the target business identified by the combined state identification model. Further, the computer device can compare the calculation probability returned by the combined model with the preset positive example probability of product operation (i.e., the preset positive example probability) 80h to obtain the target processing state information, and perform an end operation 80i. For example, the target business is a public accumulation payment business operation, and the combined state identification model can output the probability that the target object is in a normal public accumulation payment state, and compare the probability that the target object is in a normal public accumulation payment state with the preset positive example probability of normal public accumulation payment. If the probability that the target object is in a normal public accumulation payment state is greater than or equal to the preset positive example probability of normal public accumulation payment, it is determined that the target processing state information of the target object is normal public accumulation payment. If the probability that the target object is in a normal public accumulation payment state is less than the preset positive example probability of normal public accumulation payment, it is determined that the target processing state information of the target object is abnormal public accumulation payment.

[0142] In the embodiment of the present application, the first processing state information reflecting the target object with respect to the target service is obtained by linear state recognition of the target object according to the object characteristic information of the target object. Meanwhile, the second processing state information reflecting the target object with respect to the target service is obtained by cross state recognition of the target object according to the information amount contained in the object characteristic information and the object characteristic information. Further, the target fusion characteristic information is obtained by fusing the object characteristic information of the target object and the state information (i.e., the first processing state information and the second processing state information) obtained by using different state recognition methods. The target processing state information reflecting the target object with respect to the target service is obtained by combined state recognition of the target object according to the target fusion characteristic information. That is, the target processing state information reflecting the target object with respect to the target service is obtained by optimized recognition of the processing state of the target object with respect to the target service according to the target fusion characteristic information, which can improve the recognition accuracy of the target processing state information reflecting the target object with respect to the target service. In addition, the object characteristic information of the target object is extracted according to two different feature extraction methods, i.e., linear state recognition and linear state recognition, to further obtain the processing state information of the target object at different angles. The first processing state information and the second processing state information of the target object at different angles are fused with the object characteristic information to further obtain the object characteristic information of the target object in multiple dimensions. The recognition accuracy of the target processing state information reflecting the target object with respect to the target service can be improved by combined state recognition of the object characteristic information of the target object in multiple dimensions. In addition, the first processing state information of the target object with respect to the target service is output by the linear state recognition model, and the second processing state information of the target object with respect to the target service is output by the cross state recognition model. The first processing state information input by the linear state recognition model, the second processing state information output by the cross state recognition model, and the object characteristic information are fused by the combined state recognition model to obtain the target fusion characteristic information of the target object in multiple dimensions. The target processing state information reflecting the target object with respect to the target service is recognized according to the target fusion characteristic information, which can improve the recognition accuracy of the target processing state information reflecting the target object with respect to the target service.

[0143] Please refer to Figure 9 , Figure 9 is a flowchart of a state recognition method provided by the embodiment of the present application. The state recognition method can be executed by a computer device, which can be a server (such as the server 10 in the above Figure 1 ) or a user terminal (such as any user terminal in the user terminal cluster in the above Figure 1 ). The present application does not limit this. As shown in Figure 9 , the state recognition method can include but is not limited to the following steps:

[0144] S201, obtain sample object feature information about a target business of a sample object and target processing state information labels about the target business of the sample object.

[0145] Specifically, the computer device can obtain sample object feature information about a target business of a sample object, and obtain target processing state information labels about the target business of the sample object, which can be manually labeled.

[0146] Optionally, the specific manner in which the computer device obtains sample object feature information about a target business of a sample object can include that the computer device can find positive and negative training samples that are strongly related to the target business, have normal data distribution, and have reasonable feature portraits. For example, when the target business is a public accumulation payment business, the positive and negative training sample data can be sample objects that normally pay public accumulation and sample objects that do not normally pay public accumulation.

[0147] Optionally, the computer device can obtain initial sample object feature information about a target business of a sample object, filter the initial sample object feature information based on an anomaly detection criterion to obtain first candidate object feature information, encode the first candidate object feature information to obtain second candidate object feature information, call a feature vector conversion model to convert the second candidate object feature information to obtain vector-converted second candidate object feature information, and determine the vector-converted second candidate object feature information as the sample object feature information about the target business of the sample object.

[0148] Specifically, the computer device can prepare sample training data, and the computer device can obtain seed objects (i.e., sample objects) with label information based on manual labeling and sample verification business logic. The computer device can recall a plurality of sample objects based on sample acquisition rules, and then filter the initial sample object feature information based on an anomaly detection criterion to obtain first candidate object feature information. The anomaly detection criterion refers to a rule for filtering sample object feature information used to train the initial combined state recognition model. The anomaly detection criterion is used to filter out false objects and computer-controlled mobile phones. The anomaly detection criterion can include an anomaly detection index for excluding the influence of non-real objects on modeling analysis, such as setting an index for anomaly detection of data such as traffic usage and time distribution of traffic generation of the sample object when using product software. The computer device can filter abnormal sample objects based on the anomaly detection index in the anomaly detection criterion, such as using the "Pareto criterion" for outlier judgment. The "Pareto criterion" is specifically as follows: assuming that a set of detection data only contains random errors, the standard deviation is obtained by calculating and processing the detection data, an interval is determined according to a certain probability, and it is considered that any error exceeding the interval is not a random error but a gross error, and the data containing the error should be excluded. Further, the computer device can call a feature vector conversion model to convert the second candidate object feature information to obtain vector-converted second candidate object feature information, and determine the vector-converted second candidate object feature information as sample object feature information of the sample object about the target business.

[0149] Further, the computer device can store the filtered normal sample objects and the object feature information of the sample objects offline in the HDFS (The Hadoop Distributed File System, i.e., a distributed file system running on commodity hardware), to facilitate quick access of subsequent processes. By constructing the portrait features of the training sample objects and based on the vertical characteristics of the features, combined with the time dimension and different feature processing methods, a high-dimensional feature vector is output. The specific content of the sample object feature information of the sample object can refer to the content of step S101 in the above embodiment, which will not be described here again. Figure 3

[0150] Optionally, a feature type of the first candidate object feature information is obtained, a target encoding mode for feature encoding of the first candidate object feature information is determined from a set of encoding modes according to the feature type, and the first candidate object feature information is encoded by using the target encoding mode to obtain second candidate object feature information.

[0151] ​Optionally, the feature type includes basic portrait features and business vertical types, and the computer device can perform offline feature processing on the object feature information of the filtered sample object, that is, perform feature coding on the sample object feature information of the sample object. Specifically, the computer device can construct basic portrait features of the sample object. For example, based on historical behavior data of the object, rich basic portrait features of the sample object can be constructed, and the basic portrait features include user basic attributes, device basic attributes, network connection attributes, and the like. The computer device can construct business vertical type features based on business characteristics. The vertical type features include click rates and conversion rates of the object on specific types of advertisements. Further, the computer device can combine time dimensions to aggregate portrait features and business features of different time spans from the basic portrait features and the business vertical type features, respectively, and determine the aggregated portrait features and business features of different time spans as second candidate object feature information. For example, the computer device aggregates portraits of the object in the last half year, the last three months, the last month, and the last week, and the aggregation methods include summation, median, and standard deviation. Further, the computer device can perform feature processing on the sample object feature information of the sample object, such as normalizing numerical features and discretizing non-numerical features, and the normalization method can be Gaussian normalization. The specific content of the feature preprocessing performed by the computer device on the sample object feature information of the sample object can be referred to the content of step S101 in the foregoing Figure 3

[0152] Further, the computer device can merge the processed object feature information and store it offline in the HDFS for quick access by subsequent processes. The computer device can solidify the feature processing logic for feature preprocessing of the sample feature information, perform automatic offline calculation of the object feature information of the sample object at regular intervals, and push the offline calculation result to the online storage engine.

[0153] S202, calling a linear state recognition model, performing linear state recognition on the sample object according to the sample object feature information, and obtaining first recognition processing state information reflecting the sample object about the target business.

[0154] Specifically, the computer device can call a linear state recognition model, perform linear state recognition on the sample object according to the sample object feature information of the sample object, and obtain first recognition processing state information reflecting the sample object about the target business. The content of the linear state recognition on the sample object can be referred to the content of step S102 in the foregoing Figure 3

[0155] ​​S203, calling a cross-state recognition model, performing cross-state recognition on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, and obtaining second recognition processing state information reflecting the sample object about the target business.

[0156] Specifically, the computer device can call the cross-state recognition model, perform cross-state recognition on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, and obtain second recognition processing state information reflecting the sample object about the target business. The content of cross-state recognition on the sample object can refer to the content of step S103 in the foregoing embodiment, and will not be described herein again. Figure 3

[0157] S204, calling an initial combination state recognition model, fusing the sample object feature information, the first recognition processing state information and the second recognition processing state information to obtain target sample fusion feature information, performing combination state prediction on the sample object according to the target sample fusion feature information, and obtaining predicted target processing state information reflecting the sample object about the target business.

[0158] Specifically, the computer device can call the initial combination state recognition model, fuse the sample object feature information, the first recognition processing state information and the second recognition processing state information to obtain target sample fusion feature information. According to the target sample fusion feature information, the combination state of the sample object is predicted, and the predicted target processing state information reflecting the sample object about the target business is obtained. The specific content of combination state recognition on the sample can refer to the content of step S105 in the foregoing embodiment, and will not be described herein again. Figure 3

[0159] S205, adjusting the initial combination state recognition model according to the target processing state information label and the predicted target processing state information to obtain a combination state recognition model; the combination state recognition model is used to predict target processing state information reflecting a target object about a target business according to object feature information of the target object.

[0160] ​​Specifically, the computer device can adjust the initial combined state recognition model according to the target processing state information label and the predicted target processing state information, and when the adjusted initial combined state recognition model meets the convergence condition, the initial combined state recognition model after the parameter adjustment is determined as the combined state recognition model. The combined state recognition model is used to predict target processing state information reflecting a target object about a target business according to object feature information of the target object. If the initial combined state recognition model after the parameter adjustment does not meet the convergence condition, the initial combined state recognition model that does not meet the convergence condition is iteratively trained until the initial combined state recognition model meets the convergence condition. The convergence condition can be that the number of training times of the initial combined state recognition model reaches a target number, or the model loss of the initial combined state recognition model is less than or equal to a target loss value.

[0161] Optionally, the computer device can adjust the initial combined state recognition model according to the target processing state information label and the predicted target processing state information to obtain the specific manner of the combined state recognition model, which can include: determining the combined state prediction error of the initial combined state recognition model according to the target processing state information label and the predicted target processing state information. If the combined state prediction error is not in a convergence state, the initial combined state recognition model is adjusted according to the combined state prediction error. The adjusted initial combined state recognition model is determined as the combined state recognition model.

[0162] Specifically, the convergence condition includes a convergence state. The computer device can determine a combined state prediction error of the initial combined state recognition model according to a difference between the target processing state information label and the predicted target processing state information. Further, the computer device can detect whether the combined state prediction error is in the convergence state. The convergence state can be whether the combined state prediction error is less than or equal to a target threshold. If the combined state prediction error is less than or equal to the target threshold, it is determined that the combined state prediction error is in the convergence condition, and the initial combined state recognition model is determined as the combined state recognition model. If the combined state prediction error is greater than the target threshold, it is determined that the combined state prediction error is not in the convergence condition, and the first model adjustment parameter for adjusting the initial combined state recognition model is determined according to the combined state prediction error. Further, the computer device can replace the initial model parameter in the initial combined state recognition model with the first model adjustment parameter to obtain the initial combined state recognition model after parameter adjustment. The computer device can again obtain the combined state prediction error corresponding to the initial combined state recognition model after parameter adjustment, detect whether the combined state recognition error obtained again satisfies the convergence state, and if so, determine the initial combined state recognition model after parameter adjustment as the combined state recognition model. If not, the model training of the initial combined state recognition model after parameter adjustment is continued until the combined state recognition error corresponding to the initial combined state recognition model after parameter adjustment satisfies the convergence state.

[0163] The initial combined state recognition model can output the first recognition processing state information output by the linear state recognition model and the second recognition processing state information output by the cross state recognition model as new sample feature information of the sample object. At the same time, the first recognition processing state information output by the linear state recognition model and the second recognition processing state information output by the cross state recognition model can also be supervised weight optimized to improve the model performance of the combined state recognition model obtained by training, and to improve the prediction accuracy of the combined state recognition model on the target processing state information of the target sample about the target business.

[0164] Optionally, the computer device can pre-train the linear state recognition model and directly call it when used. Specifically, the specific manner in which the computer device trains the linear state recognition model can include: calling an initial linear state recognition model, performing linear state prediction on the sample object according to the sample object feature information to obtain first predicted processing state information of the sample object about the target business. Adjusting the initial linear state recognition model according to the target processing state information label and the first predicted processing state information to obtain the linear state recognition model; the linear state recognition model is used for linear state prediction on the target object to obtain first processing state information of the target object.

[0165] Specifically, the computer device can divide the sample object feature information of the sample object to obtain a training set and a verification set. For example, the sample object feature information can be divided according to time windows to which the sample object belongs. The sample object feature information of an earlier time window is used as the training set, and the sample object feature information of a later time window is used as the verification set. The ratio of the training set to the verification set is ensured to be 5:1. An initial linear state recognition model is called, and the sample object is linearly state predicted according to the training set in the sample object feature information to obtain first predicted processing state information for reflecting the sample object about the target business. Further, the initial linear state recognition model is adjusted according to the target processing state information label and the first predicted processing state information to obtain a linear state recognition model. When the linear state recognition model is trained, the linear state recognition model can be called, and the sample object is linearly state predicted according to the verification set in the sample object feature information to obtain first processing state information for reflecting the sample object about the target business. The content of cross-state recognition of the sample object can refer to the content of step S103 in the method, and details are not described herein. Figure 3 The content of step S103 in the method can be referred to, and details are not described herein.

[0166] Optionally, the computer device adjusts the initial linear state recognition model according to the target processing state information label and the first predicted processing state information to obtain a linear state recognition model. The specific manner can include: determining a linear state prediction error of the initial linear state recognition model according to the target processing state information label and the first predicted processing state information. If the linear state prediction error is not in a convergence state, the initial linear state recognition model is adjusted according to the linear state prediction error, and the adjusted initial linear state recognition model is determined as the linear state recognition model.

[0167] Specifically, the computer device can obtain a state information difference between the target processing state information and the first predicted processing state information, and determine a linear state prediction error of an initial linear state recognition model according to the state information difference. The computer device can detect whether the linear state prediction error of the initial linear state recognition model is in a convergence state, and perform parameter adjustment on the initial linear state recognition model. When the initial linear state recognition model after parameter adjustment meets a convergence condition, the initial linear state recognition model after parameter adjustment is determined as the linear state recognition model. If the initial linear state recognition model after parameter adjustment does not meet the convergence condition, the initial linear state recognition model that does not meet the convergence condition is continuously iteratively trained until the initial linear state recognition model meets the convergence condition. The convergence condition can be that the number of training times of the initial linear state recognition model reaches a target number, or the model loss of the initial linear state recognition model is less than or equal to a target loss value. After the computer device completes training of the linear state recognition model, the linear state recognition model can be used to perform linear state recognition on a verification set in the sample object feature information, to obtain first recognition processing state information of the sample object about the target business. The first recognition processing state information can be a probability of the sample object about an identification processing state in the target business. For example, when the target business is a provident fund payment business, the first recognition processing state information can be a probability of the sample object about normal provident fund payment.

[0168] Optionally, the computer device can pre-train the cross-state recognition model and directly call the cross-state recognition model when used. Specifically, the computer device can train the cross-state recognition model in the following manner: calling an initial cross-state recognition model, performing cross-state prediction on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, to obtain second predicted processing state information of the sample object about the target business. Adjusting the initial cross-state recognition model according to the target processing state information label and the second predicted processing state information, to obtain the cross-state recognition model. The cross-state recognition model is used to perform cross-state prediction on the object, to obtain second processing state information of the object about the business.

[0169] Optionally, the above-mentioned adjusting the initial cross-state recognition model according to the target processing state information label and the second predicted processing state information, to obtain the cross-state recognition model, includes: determining a cross-state prediction error of the initial cross-state recognition model according to the target processing state information label and the second predicted processing state information; if the cross-state prediction error is not in a convergence state, adjusting the initial cross-state recognition model according to the cross-state prediction error; and determining the adjusted initial cross-state recognition model as the cross-state recognition model.

[0170] Specifically, the computer device can invoke the initial cross-state recognition model and, based on the amount of information contained in the training set of the sample object's feature information, perform cross-state recognition on the sample object to obtain second predictive processing state information reflecting the sample object's relationship to the target business. The specific details of performing cross-state recognition on the sample object can be found above. Figure 3 The content of step S103 will not be repeated here in the embodiments of this application.

[0171] Furthermore, the computer device can determine the cross-state prediction error of the initial cross-state recognition model based on the target processing state information label, the second predicted processing state information, and the loss function of the initial cross-state recognition model. This cross-state prediction error reflects the state prediction accuracy of the initial cross-state recognition model. Therefore, if the cross-state prediction error is in a convergent state, it indicates that the state prediction accuracy of the initial cross-state recognition model is relatively high, and the initial cross-state recognition model can be identified as a cross-state recognition model. If the cross-state prediction error is not in a convergent state, it indicates that the state prediction accuracy of the initial cross-state recognition model is relatively low, and the initial cross-state recognition model can be adjusted based on the cross-state prediction error. If the initial cross-state recognition model after parameter adjustment does not meet the convergence condition, iterative training of the initial cross-state recognition model that does not meet the convergence condition continues until the cross-state prediction error of the initial cross-state recognition model meets the convergence condition. The convergence condition can refer to the model loss of the initial cross-state recognition model being less than or equal to the target loss value. After the computer equipment has trained the cross-state recognition model, it can use the cross-state recognition model to perform cross-state recognition on the sample object's feature information based on the validation set. This results in the sample object's second recognition processing state information regarding the target business. This second recognition processing state information can refer to the probability that the sample object is in the second recognition processing state. For example, if the target business is housing provident fund payment, this second recognition processing state information can refer to the probability that the sample object is making a normal housing provident fund payment.

[0172] Wherein, the computer device can use the AUC (Area under Curve) evaluation index to evaluate the model parameters when performing model tuning on the model parameters in the initial combination state recognition model, the initial linear state recognition model and the initial cross state recognition model. Parameter tuning refers to grid optimization of the hyperparameters of the selected model, so as to expect that the evaluation index AUC can be improved. Wherein, the AUC index itself is irrelevant to the absolute value of the model prediction result, and only focuses on the ranking effect, which is more close to the needs of actual business. In addition, the calculation method of AUC considers the classification ability of the learner for positive examples and negative examples, and can still make a reasonable evaluation of the classifier in the case of sample imbalance. Wherein, the larger the AUC value is, the more likely the current classification algorithm is to arrange the positive samples in front of the negative samples, and obtain a better classification result. Wherein, the computer device can also perform online experiment evaluation on the combination state recognition model, the linear state recognition model and the cross state recognition model after training, such as online traffic based on A / B (i.e. control group and experimental group) experiment, to evaluate the effect of the model. The evaluation indexes include: advertisement click rate, advertisement conversion rate, etc.

[0173] Wherein, after the computer device trains the combination state recognition model, the linear state recognition model and the cross state recognition model, the combination state recognition model, the linear state recognition model and the cross state recognition model can be solidified. Wherein, the trained model can be solidified based on the Saver method of TensorFlow, to generate checkpoint (i.e. text file, recording the path information list of the model file), model.ckpt.data (i.e. model parameter file, recording the model network weight information), model.ckpt.index.data (i.e. binary file, saving the variable weight information in the model), index (i.e. binary file, saving the variable weight information in the model). In this way, the computer device can directly call when it is necessary to recognize the target processing state information of the target object about the target business.

[0174] As shown in Figure 10 , Figure 10 is a schematic diagram of model training provided by an embodiment of the present application, as shown in Figure 10As shown, the computer device can prepare samples by acquiring labeled sample objects based on manual annotation or preset business logic. The computer device can then create a basic profile (i.e., basic object feature information) of these labeled sample objects, calculate evaluation metrics for abnormal object types, filter out abnormal sample objects, and store the filtered normal sample objects offline. Furthermore, the computer device can perform feature processing on the offline stored sample object feature information. First, the computer device can construct the basic profile features of the sample objects (i.e., basic object feature information)... Figure 3 The general feature information in step S101, and the construction of business vertical features ( Figure 3 In step S101, the target object's business characteristic information pertains to multiple business operations. Secondly, the computer equipment can aggregate the sample object's feature information by incorporating the time dimension, obtaining basic profile features and business-specific vertical features across different practice spans. Then, the computer equipment can perform feature preprocessing and feature encoding on the basic profile features and business-specific vertical features across different practice spans to obtain processed sample feature information. For details, please refer to the above. Figure 3 The content of step S101 will not be repeated here in this embodiment. Then, the computer device can merge the processed sample object feature information and store it offline to HDFS for easy subsequent retrieval. The computer device can also solidify the feature processing logic for the sample object feature information so that feature calculation can be performed automatically offline at regular intervals (i.e., offline feature processing).

[0175] Further, the computer device can perform model fusion preparation, divide the sample object feature information to obtain a training set and a verification set, and read in the training set in the sample feature information. Based on the training set, the initial linear state recognition model and the initial cross state recognition model are independently trained. When the linear state recognition model and the cross state recognition model are trained (i.e., the convergence condition is met), the verification set in the sample object feature information is read in, the verification set is input into the linear state recognition model and the cross state recognition model respectively, the output results of the linear state recognition model and the cross state recognition model on the verification set are obtained, and the output results corresponding to the linear state recognition model and the cross state recognition model are stored respectively. Further, the computer device can perform model fusion, read in the training set in the sample object feature information, read in the output results output by the linear state recognition model and the cross state recognition model respectively, splice the training set in the sample object feature information and the output results output by the linear state recognition model and the cross state recognition model respectively to obtain target fusion feature information. The initial combined state recognition model is trained based on the target sample fusion feature information, the model of the initial combined state recognition model is optimized to obtain a combined state recognition model, and the combined state recognition model is solidified for subsequent calling. When the computer device needs to predict the target processing state information reflecting the target object about the target business, the linear state recognition model, the cross state recognition model and the combined state recognition model can be directly called to execute the contents of steps s101-s105 in the above embodiment, and the target processing state information reflecting the target object about the target business is obtained. Figure 3

[0176] Further, when the computer device needs to determine the target processing state information of the target object about the target business, the model (i.e., the linear state recognition model, the cross state recognition model and the combined state recognition model) can be called online. The computer device can pull the latest feature set of the target object from the online storage engine at a fixed time, obtain feature calculation logic, derive original data from the online real-time log, perform feature calculation on the original data based on the feature calculation logic to obtain object feature information of the target object. The obtained object feature information is input into the linear state recognition model to output first processing state information, the object feature information is input into the cross state recognition model to output second processing state information, the first processing state information, the second processing state information and the object feature are spliced to obtain target fusion feature information, and the target fusion feature information is input into the combined state recognition model to output target processing state information. The target object is processed according to the target processing state information, such as business recommendation to the target object according to the target processing state information.

[0177] As shown in Figure 11 Figure 11 ​​This is a schematic diagram illustrating the training of a combined state recognition model provided in an embodiment of this application, as shown below. Figure 11 As shown, the training of the combined state recognition model by the computer equipment includes three parts: training data preparation 110a, feature processing 110b, and model fusion training 110c. Among them, as... Figure 11 As shown, during the training data preparation phase, computer equipment can acquire labeled sample objects based on manual annotation or pre-defined business logic. The computer equipment can then create a basic profile (i.e., basic object feature information) of these labeled sample objects, calculate evaluation metrics for abnormal object types, and filter the sample objects based on the distribution anomaly theorem. This means filtering out abnormal sample objects and determining whether each sample object meets the criteria. If it does not meet the criteria, the computer equipment reacquires the sample objects; if it does, the filtered normal sample objects are stored offline. For example... Figure 11 As shown, in the feature processing 110b stage, the computer device can construct the basic profile features of the sample object (i.e., Figure 3 The general feature information in step S101, and the construction of business vertical features ( Figure 3 In step S101, the target object's business characteristic information pertains to multiple business operations. Secondly, the computer equipment can aggregate the sample object's feature information by incorporating the time dimension, obtaining basic profile features and business-specific vertical features across different practice spans. Then, the computer equipment can perform feature preprocessing and feature encoding on the basic profile features and business-specific vertical features across different practice spans to obtain processed sample feature information. For details, please refer to the above. Figure 3As described in step S101, the embodiment of the present application will not be described again. Then, the computer device can merge the processed sample object feature information and store it offline to HDFS for subsequent reading. The computer device can solidify the feature processing logic of the sample object feature information to automatically perform feature calculation at regular intervals offline (i.e., offline feature processing). In the model fusion training 110c stage, the computer device can divide the sample object feature information to obtain a training set and a validation set, and read the training set in the sample feature information. Based on the training set, the initial linear state recognition model and the initial cross state recognition model are independently trained. When the linear state recognition model and the cross state recognition model are trained (i.e., the convergence condition is met), the validation set in the sample object feature information is read, and the validation set is input into the linear state recognition model and the cross state recognition model, respectively, to obtain the validation results of the linear state recognition model and the cross state recognition model on the validation set. The first validation result and the second validation result corresponding to the linear state recognition model and the cross state recognition model are stored. Further, the computer device can splice the first validation result, the second validation result and the sample object feature information to obtain target fusion feature information. The initial combined state recognition model is trained based on the target sample fusion feature information, the initial combined state recognition model is model optimized, whether the optimized combined linear state recognition model meets the standard is judged, if it meets the standard, the combined state recognition model obtained by training is solidified for subsequent calling. If it does not meet the standard, the initial combined state recognition model is retrained based on the target fusion feature information.

[0178] As Figure 12 shown, Figure 12 is a schematic diagram of the model prediction effect provided by the embodiment of the present application, Figure 12 shows the prediction mode, and the prediction effect of the public accumulation payment state of the target object. As Figure 12As shown in the table, three prediction methods are provided, i.e., an artificially specified strong rule, a non-deep learning scheme, and a multi-layer fusion scheme (i.e., the multi-model fusion scheme in the present application). When the prediction method is the artificially specified strong rule, the online AUC (i.e., area under the curve, a model evaluation index) and the offline AUC are lower than those of the non-deep learning scheme and the multi-layer fusion scheme. When the prediction method is the non-deep learning scheme, the online AUC (i.e., area under the curve, a model evaluation index) and the offline AUC are lower than those of the multi-layer fusion scheme. When the prediction method is the multi-layer fusion scheme, the online AUC (i.e., area under the curve, a model evaluation index) and the offline AUC are higher than those of the non-deep learning scheme and the artificially specified strong rule. As can be seen, the multi-layer fusion scheme in the present application can improve the prediction accuracy of the model, and thus improve the prediction accuracy of the target processing state information of the target object about the target business.

[0179] As shown in the table, Figure 13 As shown in the table, Figure 13 is a schematic diagram of a model business effect provided by an embodiment of the present application, Figure 13 demonstrates a comparison of business processing effects (such as advertisement click rate and advertisement conversion rate) when a target object about a target business is processed according to target processing state information obtained by different prediction methods. As shown in the table, Figure 13 As shown in the table, three prediction methods are provided, i.e., an artificially specified strong rule, a non-deep learning scheme, and a multi-layer fusion scheme (i.e., the multi-model fusion scheme in the present application). When the prediction method is the artificially specified strong rule, the advertisement click rate and the advertisement conversion rate are lower than those of the non-deep learning scheme and the multi-layer fusion scheme. When the prediction method is the non-deep learning scheme, the advertisement click rate and the advertisement conversion rate are lower than those of the multi-layer fusion scheme. When the prediction method is the multi-layer fusion scheme, the advertisement click rate and the advertisement conversion rate are higher than those of the non-deep learning scheme and the artificially specified strong rule. As can be seen, the multi-layer fusion scheme in the present application can improve the prediction accuracy of the target processing state information of the target object about the target business, and thus improve the business processing effect of the related business processing of the target object.

[0180] When the above embodiments of the present application are applied to specific products or technologies, the user's permission or consent needs to be obtained, and the collection, use and processing of the object feature information of the target object, the sample object feature information of the sample object, and the log data of the target object and the sample object about the multimedia data (such as an application) need to comply with relevant laws, regulations and standards of the relevant countries and regions. That is, the computer device can obtain the object feature information of the target object, the sample object feature information of the sample object, and the log data of the target object and the sample object about the multimedia data only when the computer device obtains the authorization information of the user for the above information. That is, the object feature information of the target object, the sample object feature information of the sample object, and the log data of the target object and the sample object about the multimedia data are obtained only after the user's authorization.

[0181] For example, the computer device displays a permission prompt interface in the multimedia interface of the multimedia platform, the permission prompt interface is used to prompt the user that the object feature information of the target object, the sample object feature information of the sample object, and the log data of the target object and the sample object about the multimedia data are currently being collected, and after obtaining the confirmation operation of the user to the permission prompt interface, the step of obtaining the object feature information of the target object, the sample object feature information of the sample object, and the log data of the target object and the sample object about the multimedia data is started, otherwise it is ended.

[0182] In the embodiment of the present application, the sample object feature information about the target service of a sample object and the target processing state information label about the target service of the sample object are acquired. A linear state recognition model is called to perform linear state recognition on the sample object according to the sample object feature information, to obtain first recognition processing state information reflecting the target service of the sample object. A cross state recognition model is called to perform cross state recognition on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, to obtain second recognition processing state information reflecting the target service of the sample object. Further, an initial combined state recognition model is called to fuse the sample object feature information, the first recognition processing state information and the second recognition processing state information, to obtain target sample fusion feature information, and perform combined state prediction on the sample object according to the target sample fusion feature information, to obtain predicted target processing state information reflecting the target service of the sample object. The initial combined state recognition model is adjusted according to the target processing state information label and the predicted target processing state information, to obtain a combined state recognition model, which is used to determine the target processing state information reflecting the target service of a target object according to the object feature information of the target object. It can be seen that the present application can fuse the inputs of the linear state recognition model and the cross state recognition model with the sample object feature information to obtain target sample fusion feature information, and further obtain sample object feature information of the sample object in multiple dimensions, and train the initial combined state recognition model according to the target sample fusion feature information. In this way, the output of the explicit state recognition model and the output of the cross state recognition model can be supervised and optimized in weight, the prediction accuracy of the combined state recognition model obtained by training can be improved, and further the recognition accuracy of the target processing state information reflecting the target service of the target object can be improved.

[0183] Please refer to Figure 14 , Figure 14 is a structural schematic diagram of a state recognition device provided by an embodiment of the present application. The state recognition device can be a computer program (including program code) running in a computer device, for example, the state recognition device is an application software; the state recognition device can be used to execute corresponding steps in the state recognition method provided by the embodiments of the present application. As shown in Figure 14 , the state recognition device can include a first acquisition module 11, a linear state recognition module 12, a cross state recognition module 13, a fusion module 14, a combined state recognition module 15, a second acquisition module 16, a first calling module 17, a second calling module 18, a third calling module 19, a first adjustment module 20, a fourth calling module 21, a second adjustment module 22, a fifth calling module 23 and a third adjustment module 24.

[0184] The first obtaining module 11 is configured to obtain object characteristic information of a target object about a target service and an information amount contained in the object characteristic information.

[0185] The linear state recognition module 12 is configured to perform linear state recognition on the target object according to the object characteristic information, to obtain first processing state information reflecting the target object about the target service.

[0186] The cross state recognition module 13 is configured to perform cross state recognition on the target object according to the information amount contained in the object characteristic information and the object characteristic information, to obtain second processing state information reflecting the target object about the target service.

[0187] The fusion module 14 is configured to fuse the object characteristic information, the first processing state information and the second processing state information, to obtain target fusion characteristic information.

[0188] The combined state recognition module 15 is configured to perform combined state recognition on the target object according to the target fusion characteristic information, to obtain target processing state information reflecting the target object about the target service.

[0189] The linear state recognition module 12 includes:

[0190] The deep feature extraction unit 1201 is configured to call a feature extraction layer of a linear state recognition model, to perform deep feature extraction on the object characteristic information, to obtain deep characteristic information of the target object about the target service.

[0191] The linear state recognition unit 1202 is configured to call a state recognition layer of the linear state recognition model, to perform linear state recognition on the target object according to the deep characteristic information, to obtain the first processing state information reflecting the target object about the target service.

[0192] The deep feature extraction unit 1201 is specifically configured to:

[0193] Call the feature extraction layer of the linear state recognition model, to perform correlation relationship recognition on the object characteristic information in each dimension of the object characteristic information, to obtain object correlation characteristic information.

[0194] Perform linear processing on the object correlation characteristic information, to obtain linear characteristic information.

[0195] According to the linear characteristic information, determine the deep characteristic information of the target object about the target service.

[0196] The cross state recognition module 13 includes:

[0197] The determination unit 1301 is configured to invoke an influence degree determination layer of the cross-state recognition model, and determine an influence degree of the object feature information according to an information amount contained in the object feature information.

[0198] The cross-state recognition unit 1302 is configured to invoke a state recognition layer of the cross-state recognition model, and perform cross-state recognition on the target object according to the influence degree and the object feature information, to obtain second processing state information reflecting the target object about the target service.

[0199] The cross-state recognition unit 1302 is specifically configured to:

[0200] invoke the state recognition layer of the cross-state recognition model, and determine a weighting parameter of the object feature information according to the influence degree;

[0201] perform weighting processing on the object feature information according to the weighting parameter, to obtain weighted object feature information;

[0202] perform feature cross combination on the weighted object feature information and the object feature information, to obtain cross combination feature information;

[0203] perform state recognition on the cross combination feature information, to obtain the second processing state information reflecting the target object about the target service.

[0204] The combination state recognition module 15 includes:

[0205] The acquisition unit 1501 is configured to invoke a weight prediction layer of the combination state recognition model, and acquire a combination weight between feature information of any two dimensions in the target fusion feature information.

[0206] The feature combination unit 1502 is configured to invoke a feature combination layer of the combination state recognition model, and perform feature combination on the feature information of any two dimensions in the target fusion feature information based on the combination weight, to obtain feature combined target fusion feature information.

[0207] The combination state recognition unit 1503 is configured to invoke a state recognition layer of the combination state recognition model, and perform state recognition on the feature combined target fusion feature information, to obtain the target processing state information reflecting the target object about the target service.

[0208] The state recognition apparatus further includes:

[0209] The second acquisition module 16 is configured to acquire sample object feature information of a sample object about a target service and a target processing state information label of the sample object about the target service.

[0210] The first calling module 17 is configured to call the linear state recognition model, perform linear state recognition on the sample object according to the sample object feature information, and obtain first recognition processing state information reflecting the sample object with respect to the target business.

[0211] The second calling module 18 is configured to call the cross state recognition model, perform cross state recognition on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, and obtain second recognition processing state information reflecting the sample object with respect to the target business.

[0212] The third calling module 19 is configured to call the initial combined state recognition model, fuse the sample object feature information, the first recognition processing state information, and the second recognition processing state information, obtain target sample fusion feature information, perform combined state prediction on the sample object according to the target sample fusion feature information, and obtain predicted target processing state information reflecting the sample object with respect to the target business.

[0213] The first adjusting module 20 is configured to adjust the initial combined state recognition model according to the target processing state information label and the predicted target processing state information, and obtain a combined state recognition model. The combined state recognition model is configured to determine target processing state information reflecting a target object with respect to a target business according to object feature information of the target object.

[0214] The first adjusting module 20 is configured to determine a combined state prediction error of the initial combined state recognition model according to the target processing state information label and the predicted target processing state information. If the combined state prediction error is not in a convergence state, the initial combined state recognition model is adjusted according to the combined state prediction error. The adjusted initial combined state recognition model is determined as the combined state recognition model.

[0215] The state recognition device further includes:

[0216] The fourth calling module 21 is configured to call the initial linear state recognition model, perform linear state prediction on the sample object according to the sample object feature information, and obtain first predicted processing state information reflecting the sample object with respect to the target business.

[0217] The second adjusting module 22 is configured to determine a first loss function according to the target processing state information label and the first predicted processing state information, adjust parameters of the initial linear state recognition model according to the first loss function, and obtain a linear state recognition model. The linear state recognition model is configured to perform linear state prediction on an object and obtain first processing state information of the object with respect to the business.

[0218] The second adjustment module is configured to determine a linear state prediction error of the initial linear state recognition model according to the target processing state information label and the first predicted processing state information, adjust the initial linear state recognition model according to the linear state prediction error if the linear state prediction error is not in a convergent state, and determine the adjusted initial linear state recognition model as the linear state recognition model.

[0219] The state recognition device further includes:

[0220] The fifth calling module 23 is configured to call the initial cross state recognition model, perform cross state prediction on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, and obtain second predicted processing state information of the sample object about the target service.

[0221] The third adjustment module 24 is configured to determine a second loss function according to the target processing state information label and the second predicted processing state information, perform parameter adjustment on the initial cross state recognition model according to the second loss function, and obtain the cross state recognition model. The cross state recognition model is configured to perform cross state prediction on the object to obtain second processing state information of the object about the service.

[0222] The third adjustment module is configured to determine a cross state prediction error of the initial cross state recognition model according to the target processing state information label and the second predicted processing state information, adjust the initial cross state recognition model according to the cross state prediction error if the cross state prediction error is not in a convergent state, and determine the adjusted initial cross state recognition model as the cross state recognition model.

[0223] Optionally, the obtaining module obtains sample object feature information of a sample object about a target service, and the obtaining includes:

[0224] The obtaining module obtains initial sample object feature information of a sample object about a target service, filters the initial sample object feature information based on an abnormality detection criterion, and obtains first candidate object feature information.

[0225] The feature encoding module is configured to perform feature encoding on the first candidate object feature information, and obtain second candidate object feature information.

[0226] The vector conversion module is configured to call a feature vector conversion model, perform vector conversion on the second candidate object feature information, obtain vector-converted second candidate object feature information, and determine the vector-converted second candidate object feature information as the sample object feature information of the sample object about the target service.

[0227] Optionally, the acquisition module encodes the first candidate object feature information to obtain second candidate object feature information, including:

[0228] acquiring a feature type of the first candidate object feature information;

[0229] determining a target encoding mode for encoding the first candidate object feature information from a set of encoding modes according to the feature type;

[0230] encoding the first candidate object feature information by using the target encoding mode to obtain second candidate object feature information.

[0231] According to an embodiment of the present application, Figure 3 The steps involved in the state recognition method shown can be executed by Figure 14 each module in the state recognition apparatus shown. For example, Figure 3 Step S101 shown in the method can be executed by Figure 14 the first acquisition module 11 in the apparatus, Figure 3 Step S102 shown in the method can be executed by Figure 14 the linear state recognition module 12 in the apparatus, Figure 3 Step S103 shown in the method can be executed by Figure 14 the cross state recognition module 13 in the apparatus, Figure 3 Step S104 shown in the method can be executed by Figure 14 the fusion module 14 in the apparatus, and so on.

[0232] According to an embodiment of the present application, Figure 14 The modules in the state recognition apparatus shown can be respectively or all combined into one or several units to constitute, or some of the units can be further split into a plurality of sub-units with smaller functions, and the same operations can be realized without affecting the realization of the technical effects of the embodiments of the present application. The above modules are divided based on logical functions, and in actual applications, the functions of one module can also be realized by a plurality of units, or the functions of a plurality of modules can be realized by one unit. In other embodiments of the present application, the test apparatus can also include other units, and in actual applications, these functions can also be realized by other units, and can be realized by a plurality of units in cooperation.

[0233] In the embodiments of the present application, the first processing state information reflecting the target object about the target service is obtained by linear state recognition of the target object according to the object characteristic information of the target object. Meanwhile, the second processing state information reflecting the target object about the target service is obtained by cross state recognition of the target object according to the information amount contained in the object characteristic information and the object characteristic information. Further, the target fusion characteristic information is obtained by fusing the object characteristic information of the target object and the state information (i.e. the first processing state information and the second processing state information) obtained by using different state recognition manners. The target processing state information reflecting the target object about the target service is obtained by combined state recognition of the target object according to the target fusion characteristic information. That is, the target processing state information reflecting the target object about the target service is obtained by optimized recognition of the processing state of the target object about the target service according to the target fusion characteristic information, which can improve the recognition accuracy of the target processing state information reflecting the target object about the target service. In addition, the object characteristic information of the target object is extracted according to two different feature extraction manners, i.e. linear state recognition and linear state recognition, and then the processing state information of the target object under different angles is obtained. The first processing state information and the second processing state information of the target object under different angles are fused with the object characteristic information, and then the object characteristic information of the target object in multiple dimensions is obtained. The recognition accuracy of the target processing state information reflecting the target object about the target service can be improved by combined state recognition of the object characteristic information of the target object in multiple dimensions. In addition, the first processing state information of the target object about the target service is output by the linear state recognition model, and the second processing state information of the target object about the target service is output by the cross state recognition model. Then, the first processing state information input by the linear state recognition model, the second processing state information output by the cross state recognition model and the object characteristic information are fused by the combined state recognition model to obtain the target fusion characteristic information of the target object in multiple dimensions. The target processing state information reflecting the target object about the target service is recognized according to the target fusion characteristic information, which can improve the recognition accuracy of the target processing state information reflecting the target object about the target service.

[0234] Please refer to Figure 15 , Figure 15 is a structural schematic diagram of a computer device provided by the embodiments of the present application. As shown in Figure 15As shown, the computer device 1000 can include a processor 1001, a network interface 1004 and a memory 1005, and can further include a user interface 1003 and at least one communication bus 1002. The communication bus 1002 is configured to realize the connection and communication between the components. The user interface 1003 can include a display, a keyboard, and optionally a standard wired interface, a wireless interface. The network interface 1004 can optionally include a standard wired interface, a wireless interface (such as a Wi-Fi interface). The memory 1005 can be a high-speed RAM memory, or a non-volatile memory such as at least one disk storage. The memory 1005 can be optionally at least one storage device located away from the aforementioned processor 1001. For example Figure 15 As shown, the memory 1005 as a computer readable storage medium can include an operating system, a network communication module, a user interface module and a device control application.

[0235] In Figure 15 As shown in the computer device 1000, the network interface 1004 can provide network communication functions; the user interface 1003 is mainly used to provide an input interface for the user; and the processor 1001 can be used to call the device control application stored in the memory 1005 to realize:

[0236] Obtaining object characteristic information of a target object about a target service, and an information amount contained in the object characteristic information;

[0237] Performing linear state recognition on the target object according to the object characteristic information, to obtain first processing state information for reflecting the target object about the target service;

[0238] Performing cross state recognition on the target object according to the information amount contained in the object characteristic information and the object characteristic information, to obtain second processing state information for reflecting the target object about the target service;

[0239] Fusing the object characteristic information, the first processing state information and the second processing state information to obtain target fusion characteristic information;

[0240] Performing combined state recognition on the target object according to the target fusion characteristic information, to obtain target processing state information for reflecting the target object about the target service.

[0241] It should be understood that the computer device 1000 described in the embodiments of the present application can perform the foregoing Figure 3 Or Figure 9The description of the state recognition method in the corresponding embodiment can also execute the foregoing Figure 14 The description of the state recognition device in the corresponding embodiment will not be repeated here. In addition, the beneficial effects of using the same method will not be repeated.

[0242] In addition, it should be noted here that the embodiments of the present application also provide a computer readable storage medium, and the computer readable storage medium stores the computer program executed by the state recognition device mentioned above, and the computer program includes program instructions, which can execute the foregoing Figure 3 and Figure 9 The description of the state recognition method in the corresponding embodiment will not be repeated here.

[0243] In addition, the beneficial effects of using the same method will not be repeated. For technical details not disclosed in the computer readable storage medium embodiments involved in the present application, please refer to the description of the method embodiments of the present application. As an example, the program instructions can be deployed on one computing device for execution, or on multiple computing devices located in one place for execution, or on multiple computing devices distributed in multiple places and interconnected through a communication network for execution. The multiple computing devices distributed in multiple places and interconnected through a communication network can constitute a blockchain system.

[0244] In addition, it should be noted that the embodiments of the present application also provide a computer program product or a computer program, which can include computer instructions that can be stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor can execute the computer instructions to make the computer device execute the foregoing Figure 3 and Figure 9 The description of the state recognition method in the corresponding embodiment will not be repeated here. In addition, the beneficial effects of using the same method will not be repeated. For technical details not disclosed in the computer program product or computer program embodiments involved in the present application, please refer to the description of the method embodiments of the present application.

[0245] It should be noted that for each of the foregoing method embodiments, in order to simply describe, it is expressed as a combination of a series of actions, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, some steps can be performed in other order or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0246] The steps in the method embodiments of the present application can be adjusted in sequence, combined and reduced according to actual needs.

[0247] The modules in the device embodiments of the present application can be combined, divided and reduced according to actual needs.

[0248] Those skilled in the art can understand that all or part of the processes in the above-mentioned method embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned method embodiments. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.

[0249] Those skilled in the art can understand that all or part of the processes in the above-mentioned method embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned method embodiments. The storage medium can be a magnetic disc, an optical disc, a read-only memory (ROM) or a random access memory (RAM), etc.

[0250] The above only describes the preferred embodiments of the present application, and of course cannot limit the scope of the present application. Therefore, equivalent changes made according to the claims of the present application are still within the scope of the present application.

Claims

1. A state recognition method characterized by, The method comprises the following steps: acquiring sample object feature information of a sample object about a target business and target processing state information labels of the sample object about the target business; calling a linear state recognition model to perform linear state recognition on the sample object according to the sample object feature information, to obtain first recognition processing state information reflecting the sample object about the target business; calling a cross state recognition model to perform cross state recognition on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, to obtain second recognition processing state information reflecting the sample object about the target business; calling an initial combined state recognition model to fuse the sample object feature information, the first recognition processing state information and the second recognition processing state information, to obtain target sample fusion feature information, and performing combined state prediction on the sample object according to the target sample fusion feature information, to obtain predicted target processing state information reflecting the sample object about the target business; adjusting the initial combined state recognition model according to the target processing state information labels and the predicted target processing state information, to obtain a combined state recognition model; acquiring object feature information of a target object about a target business and an information amount contained in the object feature information; performing linear state recognition on the target object according to the object feature information, to obtain first processing state information reflecting the target object about the target business; performing cross state recognition on the target object according to the information amount contained in the object feature information and the object feature information, to obtain second processing state information reflecting the target object about the target business; fusing the object feature information, the first processing state information and the second processing state information, to obtain target fusion feature information; calling the combined state recognition model to perform combined state recognition on the target object according to the target fusion feature information, to obtain target processing state information reflecting the target object about the target business.

2. The method of claim 1, wherein, The linear state recognition on the target object according to the object feature information to obtain first processing state information reflecting the target object about the target business comprises the following steps: calling a feature extraction layer of a linear state recognition model to perform deep feature extraction on the object feature information, to obtain deep feature information of the target object about the target business; calling a state recognition layer of the linear state recognition model to perform linear state recognition on the target object according to the deep feature information, to obtain first processing state information reflecting the target object about the target business.

3. The method of claim 2, wherein, The deep feature extraction on the object feature information by calling the feature extraction layer of the linear state recognition model to obtain deep feature information of the target object about the target business comprises the following steps: The feature extraction layer of the linear state recognition model is called to recognize the correlation of the object feature information in each dimension of the object feature information, and object correlation feature information is obtained. The object correlation feature information is linearly processed to obtain linear feature information. According to the linear feature information, the deep feature information of the target object about the target business is determined.

4. The method of claim 1, wherein, According to the information amount contained in the object feature information and the object feature information, cross-state recognition of the target object is performed to obtain second processing state information reflecting the target object about the target business, including: An influence degree determination layer of the cross-state recognition model is called to determine the influence degree of the object feature information according to the information amount contained in the object feature information. A state recognition layer of the cross-state recognition model is called to perform cross-state recognition of the target object according to the influence degree and the object feature information to obtain second processing state information reflecting the target object about the target business.

5. The method of claim 4, wherein, The state recognition layer of the cross-state recognition model is called to perform cross-state recognition of the target object according to the influence degree and the object feature information to obtain second processing state information reflecting the target object about the target business, including: The state recognition layer of the cross-state recognition model is called to determine the weighting parameter of the object feature information according to the influence degree. The object feature information is weighted processed according to the weighting parameter to obtain weighted processed object feature information. The weighted processed object feature information and the object feature information are combined to obtain cross-combination feature information. The cross-combination feature information is recognized to obtain second processing state information reflecting the target object about the target business.

6. The method of claim 1, wherein, The combination state recognition model is called to perform combination state recognition of the target object according to the target fusion feature information to obtain target processing state information reflecting the target object about the target business, including: The weight prediction layer of the combination state recognition model is called to obtain the combination weight between the feature information of any two dimensions in the target fusion feature information. The feature combination layer of the combination state recognition model is called to combine the feature information of any two dimensions in the target fusion feature information based on the combination weight to obtain feature combined target fusion feature information. The state recognition layer of the combination state recognition model is called to perform combination state recognition on the feature combined target fusion feature information to obtain target processing state information reflecting the target object about the target business.

7. The method of claim 6, wherein, According to the target processing state information label and the predicted target processing state information, the initial combination state recognition model is adjusted to obtain a combination state recognition model, including: According to the target processing state information label and the predicted target processing state information, the combination state prediction error of the initial combination state recognition model is determined. If the combined state prediction error is not in a convergent state, adjusting the initial combined state recognition model according to the combined state prediction error; determining the adjusted initial combined state recognition model as the combined state recognition model.

8. The method of claim 1, wherein, The method further comprises: calling an initial linear state recognition model, performing linear state prediction on the sample object according to the sample object feature information, to obtain first predicted processing state information of the sample object about the target business; adjusting the initial linear state recognition model according to the target processing state information label and the first predicted processing state information, to obtain a linear state recognition model; the linear state recognition model is used to perform linear state prediction on an object, to obtain first processing state information of the object about a business.

9. The method of claim 8, wherein, The adjusting the initial linear state recognition model according to the target processing state information label and the first predicted processing state information, to obtain a linear state recognition model, comprises: determining a linear state prediction error of the initial linear state recognition model according to the target processing state information label and the first predicted processing state information; if the linear state prediction error is not in a convergent state, adjusting the initial linear state recognition model according to the linear state prediction error; determining the adjusted initial linear state recognition model as the linear state recognition model.

10. The method of claim 1, wherein, The method further comprises: calling an initial cross state recognition model, performing cross state prediction on the sample object according to the information amount contained in the sample object feature information and the sample object feature information, to obtain second predicted processing state information of the sample object about the target business; adjusting the initial cross state recognition model according to the target processing state information label and the second predicted processing state information, to obtain a cross state recognition model; the cross state recognition model is used to perform cross state prediction on an object, to obtain second processing state information of the object about a business.

11. The method of claim 10, wherein, The adjusting the initial cross state recognition model according to the target processing state information label and the second predicted processing state information, to obtain a cross state recognition model, comprises: determining a cross state prediction error of the initial cross state recognition model according to the target processing state information label and the second predicted processing state information; if the cross state prediction error is not in a convergent state, adjusting the initial cross state recognition model according to the cross state prediction error; determining the adjusted initial cross state recognition model as the cross state recognition model.

12. The method of claim 1, wherein, The obtaining sample object feature information of a sample object about a target business comprises: obtaining initial sample object feature information of a sample object about a target business, performing filtering processing on the initial sample object feature information based on an anomaly detection criterion, to obtain first candidate object feature information; performing feature coding on the first candidate object feature information, to obtain second candidate object feature information; The feature vector conversion model is called to perform vector conversion on the second candidate object feature information, to obtain second candidate object feature information after vector conversion, and the second candidate object feature information after vector conversion is determined as the sample object feature information of the sample object about the target service.

13. The method of claim 12, wherein, The feature encoding of the first candidate object feature information to obtain the second candidate object feature information comprises: obtaining a feature type of the first candidate object feature information; determining a target encoding mode for feature encoding of the first candidate object feature information from an encoding mode set according to the feature type; performing feature encoding on the first candidate object feature information by using the target encoding mode to obtain the second candidate object feature information.

14. A state recognition apparatus characterized by comprising: comprises: a second obtaining module configured to obtain sample object feature information of a sample object about a target service and target processing state information labels of the sample object about the target service; a first calling module configured to call a linear state recognition model, perform linear state recognition on the sample object according to the sample object feature information, and obtain first recognition processing state information reflecting the sample object about the target service; a second calling module configured to call a cross state recognition model, perform cross state recognition on the sample object according to an information amount contained in the sample object feature information and the sample object feature information, and obtain second recognition processing state information reflecting the sample object about the target service; a third calling module configured to call an initial combined state recognition model, fuse the sample object feature information, the first recognition processing state information, and the second recognition processing state information to obtain target sample fusion feature information, perform combined state prediction on the sample object according to the target sample fusion feature information, and obtain predicted target processing state information reflecting the sample object about the target service; and adjust the initial combined state recognition model according to the target processing state information labels and the predicted target processing state information to obtain a combined state recognition model; a first obtaining module configured to obtain object feature information of a target object about a target service and an information amount contained in the object feature information; a linear state recognition module configured to perform linear state recognition on the target object according to the object feature information to obtain first processing state information reflecting the target object about the target service; a cross state recognition module configured to perform cross state recognition on the target object according to the information amount contained in the object feature information and the object feature information to obtain second processing state information reflecting the target object about the target service; a fusion module configured to fuse the object feature information, the first processing state information, and the second processing state information to obtain target fusion feature information; and a combined state prediction module configured to perform combined state prediction on the target object according to the target fusion feature information to obtain predicted target processing state information reflecting the target object about the target service. The combination state recognition module is configured to invoke the combination state recognition model, and perform combination state recognition on the target object according to the target fusion feature information, to obtain target processing state information reflecting the target object with respect to the target business.

15. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the method in any one of claims 1 to 13.

16. A computer storage medium, comprising, The computer storage medium stores a computer program, and the computer program is executed by the processor to implement the method in any one of claims 1 to 13.

17. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the method in any one of claims 1 to 13. The computer program is executed by the processor to implement the method in any one of claims 1 to 13.

Images