Artificial intelligence-based behavior recognition method and apparatus, server, and storage medium

Abstract

Embodiments of the present application disclose a behavior recognition method and device based on artificial intelligence, a server and a storage medium. Embodiments of the present application obtain a plurality of object behavior features, the object behavior features being composed of a plurality of different types of sub-features; perform feature interaction on the plurality of object behavior features to obtain an initial feature group, the initial feature group including a plurality of target features, each target feature corresponding to a type of sub-feature; perform feature interaction on target features of different types to obtain a target feature group; and determine a recognition result of the plurality of object behavior features according to the target feature group, so as to determine an abnormal object according to the recognition result. In the embodiments of the present application, the correlation between object behaviors can be mined in multiple dimensions, and the accuracy of the recognition result is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computers, in particular to a behavior recognition method and device based on artificial intelligence, a server and a storage medium. BACKGROUND

[0002] With the development of computer technology, the application of object behavior recognition technology is more and more widely used, such as cloud technology, artificial intelligence, Internet and various other scenarios. Especially in the Internet application scenario, more and more objects carry out various activities through the Internet, and the demand for identification of abnormal behavior is increasing.

[0003] However, in the prior art, when identifying the behavior of an object, the specific behavior keywords of the object are generally obtained, and the keywords are judged to identify whether the behavior of the object is abnormal. The object behavior characteristics cannot be comprehensively identified, and the accuracy of the identification result is low. SUMMARY

[0004] The embodiments of the present application provide a behavior recognition method and device based on artificial intelligence, a server and a storage medium, which can mine the correlation of object behavior characteristics in multiple dimensions and improve the accuracy of the identification result.

[0005] The embodiments of the present application provide a behavior recognition method based on artificial intelligence, comprising: obtaining a plurality of object behavior characteristics, the object behavior characteristics being composed of a plurality of different types of sub-features; performing feature interaction on the plurality of object behavior characteristics to obtain an initial feature group, the initial feature group including a plurality of target features, each target feature corresponding to a type of sub-feature; performing feature interaction on different types of target features to obtain a target feature group; determining an identification result of the plurality of object behavior characteristics according to the target feature group, so as to determine an abnormal object according to the identification result.

[0006] The embodiments of the present application also provide a behavior recognition device based on artificial intelligence, comprising: an acquisition unit configured to acquire a plurality of object behavior characteristics, the object behavior characteristics being composed of a plurality of different types of sub-features; a first calculation unit configured to perform feature interaction on the plurality of object behavior characteristics to obtain an initial feature group, the initial feature group including a plurality of target features, each target feature corresponding to a type of sub-feature; a second calculation unit configured to perform feature interaction on different types of target features to obtain a target feature group; and an identification unit configured to determine an identification result of the plurality of object behavior characteristics according to the target feature group, so as to determine an abnormal object according to the identification result.

[0007] The embodiment of the present application further provides a server, comprising a memory storing a plurality of instructions; and a processor loading the instructions from the memory to perform the steps in any of the behavior recognition methods based on artificial intelligence provided by the embodiment of the present application.

[0008] The embodiment of the present application further provides a computer readable storage medium storing a plurality of instructions, which are loaded by a processor to perform the steps in any of the behavior recognition methods based on artificial intelligence provided by the embodiment of the present application.

[0009] The embodiment of the present application further provides a computer program product comprising computer programs / instructions, which, when executed by a processor, implement the steps in any of the behavior recognition methods based on artificial intelligence provided by the embodiment of the present application.

[0010] The embodiment of the present application can obtain a plurality of object behavior features, which are composed of a plurality of different types of sub-features; perform feature interaction on the plurality of object behavior features to obtain an initial feature group, wherein the initial feature group comprises a plurality of target features, and each target feature corresponds to a type of sub-feature; perform feature interaction on different types of target features to obtain a target feature group; and determine a recognition result of the plurality of object behavior features according to the target feature group, so as to determine an abnormal object according to the recognition result. In the present application, the correlation between object behavior features can be focused on by performing feature interaction on a plurality of object behavior features, and the context correlation of object behavior can be focused on by performing feature interaction between different types of features, so as to mine the correlation between object behaviors in multiple dimensions and improve the accuracy of the recognition result. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment 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 effort.

[0012] Figure 1 is a scene diagram of the behavior recognition method based on artificial intelligence provided by the embodiment of the present application;

[0013] Figure 2 is a flow diagram of the behavior recognition method based on artificial intelligence provided by one embodiment of the present application;

[0014] Figure 3 is a diagram of the method for calculating the adjusted object behavior features provided by the embodiment of the present application;

[0015] Figure 4 is a flowchart of a process for training a preset behavior recognition model provided by an embodiment of the present application.

[0016] Figure 5 is a flowchart of a process for training a preset behavior recognition model provided by an embodiment of the present application.

[0017] Figure 6 is a flowchart of a process for training a preset behavior recognition model provided by an embodiment of the present application.

[0018] Figure 7 is a flowchart of a process for training a preset behavior recognition model provided by an embodiment of the present application.

[0019] Figure 8 is a flowchart of a process for training a preset behavior recognition model provided by an embodiment of the present application.

[0020] Figure 9 is a flowchart of a process for training a preset behavior recognition model provided by an embodiment of the present application.

[0021] Figure 10 is a flowchart of a process for training a preset behavior recognition model provided by an embodiment of the present application.

[0022] Figure 11 is a flowchart of a process for training a preset behavior recognition model provided by an embodiment of the present application. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, any other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0024] The embodiments of the present application provide a behavior recognition method and device based on artificial intelligence, a server and a storage medium.

[0025] The behavior recognition device based on artificial intelligence can be integrated in an electronic device, which can be a terminal, a server, etc. The terminal can be a mobile phone, a tablet computer, a smart Bluetooth device, a notebook computer, or a personal computer (PC), etc. The server can be a single server or a server cluster composed of multiple servers.

[0026] In some embodiments, the artificial intelligence-based behavior recognition device can also be integrated in multiple electronic devices, for example, the artificial intelligence-based behavior recognition device can be integrated in multiple servers, and the artificial intelligence-based behavior recognition method of the present application can be implemented by the multiple servers.

[0027] In some embodiments, the server can also be implemented in the form of a terminal.

[0028] For example, referring to Figure 1 , the artificial intelligence-based behavior recognition device is integrated in a server, the server can obtain multiple object behavior features from a database, the object behavior features are composed of multiple different types of sub-features; the multiple object behavior features are interacted to obtain an initial feature group, the initial feature group includes multiple target features, each target feature corresponds to a sub-feature type; the different types of target features are interacted to obtain a target feature group; and the recognition result of the multiple object behavior features is determined according to the target feature group.

[0029] In the present embodiment, the electronic device can pay attention to the correlation between the object behavior features by interacting the multiple object behavior features, and pay attention to the context correlation of the object behavior by interacting the features between different types of features, so as to mine the correlation between the object behaviors in multiple dimensions and improve the accuracy of the recognition result.

[0030] It can be understood that in the specific embodiments of the present application, data related to object behavior features, object behavior information, object behavior sequences, etc. are involved, and when all the embodiments of the present application are applied to specific products or technologies, the object permission or consent needs to be obtained, and the collection, use and processing of the related data need to comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0031] The following will be described in detail respectively. It should be noted that the serial numbers of the following embodiments do not limit the preferred order of the embodiments.

[0032] Artificial intelligence (AI) is a technology that uses digital computers to simulate human perception of the environment, acquisition of knowledge and use of knowledge, which can enable machines to have functions similar to human perception, reasoning and decision-making. 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 technologies mainly include computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, autonomous driving, intelligent transportation, etc.

[0033] Machine Learning (ML) is a multi-disciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithmic complexity theory, etc. It is a specialized study of how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure, and continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent. It is applied in various fields of artificial intelligence. Machine learning and deep learning generally include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0034] With the research and progress of artificial intelligence technology, artificial intelligence technology is researched and applied in many fields, such as common smart home, smart wearable device, virtual assistant, smart speaker, smart marketing, unmanned vehicle, autonomous vehicle, unmanned aerial vehicle, robot, smart medical treatment, smart customer service, Internet of Vehicles, autonomous driving, intelligent transportation, etc. It is believed that with the development of technology, artificial intelligence technology will be applied in more fields and play an increasingly important role.

[0035] In this embodiment, an artificial intelligence-based behavior recognition method involving artificial intelligence is provided, as shown in Figure 2 The specific process of the artificial intelligence-based behavior recognition method can be as follows:

[0036] 110、Obtain a plurality of object behavior features, and the object behavior features are composed of a plurality of different types of sub-features.

[0037] The object refers to an object that can perform a behavior, such as a person or an intelligent electronic device, etc. The object behavior feature is feature information used to represent the behavior performed by the object, and the object behavior feature can be obtained by extracting the object behavior information. The sub-feature refers to a feature that constitutes the object behavior feature, and the sub-feature can be obtained by extracting different types of information in the object behavior information. The object behavior information refers to information about the behavior performed by the object, and the object behavior information can include a plurality of sub-information corresponding to a plurality of sub-features, such as the behavior performed, the time of performance, the content of performance, etc.

[0038] It should be noted that the present application can be applied to a plurality of object behavior recognition scenarios, such as social, financial, online consumption, etc. Different sub-feature types can be set for different application scenarios, such as time information, behavior data, and financial value for the financial scenario, to identify the illegal transaction behavior of the object, and time information, behavior data, and comment keywords for the social scenario (such as judging whether the object is a malicious comment).

[0039] Especially for the risk control scene containing financial behavior, the abnormal financial behavior of the object can be better identified. For example, when the application is applied in a financial scene, the object behavior information generated when the object performs a transfer behavior can be "transfers 1000 yuan on July 1", which contains three sub-information "July 1", "transfer" and "1000". The three sub-features corresponding to "July 1", "transfer" and "1000" are extracted from the information through natural language processing, and the object behavior feature "2021-7-1-transfer-1000" is obtained by combination, to represent the object's behavior of transferring 1000 yuan on July 1.

[0040] The object behavior information can be any one or a combination of multiple of text information, voice information, image information, etc. When it is text information, the sub-features can be obtained through natural language processing, and when it is voice information or image information, the sub-features can be obtained through voice recognition or image recognition, etc. For example, the voice information or image information is converted into text information through voice recognition or image recognition, and the sub-features are obtained through natural language processing.

[0041] The sub-features can be classified according to the sub-feature attributes, for example, the sub-features can be classified into time, amount, behavior, location or object attribute, etc. according to the attributes of the sub-features, or the sub-features can be classified according to the preset feature types, for example, the quantifiable features such as time and amount are classified as a first type, and the unquantifiable features such as behavior, location or object attribute are classified as a second type, etc. Since the sub-information of the object behavior information has a corresponding relationship with the sub-features, the types of the sub-features also correspond to the types of the sub-information.

[0042] In some embodiments, in order to retain as many features of the object behavior information as possible, different sub-information in the object behavior information is respectively subjected to feature extraction. Specifically, step 110 can include steps 1.1-1.3 as follows:

[0043] 1.1, obtaining a plurality of object behavior information, each object behavior information including a plurality of sub-information;

[0044] 1.2, performing feature extraction on the sub-information to obtain sub-features corresponding to the sub-information;

[0045] 1.3, for each object behavior information, combining all sub-features corresponding to the object behavior information to obtain object behavior features corresponding to each object behavior information.

[0046] In the feature extraction of the object behavior information, the object behavior information can be matched according to the sub-information types to obtain the sub-information, and the sub-features can be obtained by compiling or using natural language processing and the like. It should be noted that in the calculation of the object behavior features, the object behavior features can be represented in the form of vectors to facilitate the calculation process, for example, the sub-information is extracted from the object behavior information, and the sub-information is compiled and mapped to vectors to obtain the sub-features represented in the form of vectors, and then the sub-features are combined to obtain the object behavior features. Alternatively, the sub-information can be compiled to obtain the sub-features represented in the form of numbers, and the sub-features are combined as components to obtain the object behavior features. For example, for the object behavior information “transfer 1000 yuan on July 1”, the field information “July 1”, “transfer” and “1000” are compiled to obtain the vectors 701, 1 and 1000, and then the vectors (701, 1, 1000) are combined to obtain the object behavior features.

[0047] In some embodiments, in order to obtain the object behavior information that may have a correlation and increase the accuracy of the object behavior information identification, step 1.1 can include the step of obtaining an object behavior sequence, the object behavior sequence including a plurality of object behavior information. That is, the plurality of object behavior features is an object behavior feature sequence, and step 110 can include the step of obtaining an object behavior feature sequence, the object behavior feature sequence including a plurality of object behavior features.

[0048] The object behavior sequence refers to a sequence composed of a plurality of object behaviors, which can be sorted according to a preset sorting rule, such as time sequence or financial value size, or randomly sorted, etc. The object behavior sequence can be composed of a plurality of object behavior information corresponding to the same object, or a plurality of object behavior information corresponding to different objects that meet a preset condition, etc. For example, the object behavior information of a plurality of objects in the same chat group is obtained, such as a plurality of objects in the same chat group performing transfer, red envelope sending or red envelope grabbing behaviors, and the object name, behavior type, time and financial value corresponding to any object when performing the behavior (transfer, red envelope sending or red envelope grabbing behavior) are obtained as object behavior information. All or part of the object behavior information in the chat group is obtained as an object behavior sequence.

[0049] In some embodiments, in order to accurately identify the behavior of a specific object, the object behavior sequence can include a plurality of object behavior information corresponding to the same object.

[0050] In some embodiments, in order to facilitate the processing of the same type of sub-features, after step 110, the step of extracting a plurality of sub-feature sequences from the plurality of object behavior features according to the types of the sub-features can be further included, each sub-feature sequence corresponding to a sub-feature type.

[0051] In some embodiments, the plurality of object behavior features can be represented in various forms. In order to facilitate processing of the same type of sub-information, the plurality of object behavior features can be represented in the form of multiple sub-feature sequences. Step 110 can include steps 2.1-2.5 as follows.

[0052] 2.1, obtaining a plurality of object behavior information, each object behavior information including a plurality of sub-information;

[0053] 2.2, obtaining a plurality of sub-information sequences according to the type of sub-information;

[0054] 2.3, performing feature extraction on the sub-information to obtain sub-features corresponding to the sub-information;

[0055] 2.4, obtaining a plurality of sub-feature sequences according to the plurality of sub-information sequences and the sub-features;

[0056] 2.5, combining the plurality of sub-feature sequences to obtain the plurality of object behavior features.

[0057] When there are a plurality of object behavior features, after extracting the sub-information from the plurality of object behavior information, the sub-information can be sorted according to a preset sorting rule, for example, the sub-information can be sorted according to a preset sorting rule, or can be sorted in a random order, or the same type of sub-information can be sorted according to the sorting rule of the plurality of object behavior information to obtain a plurality of sub-information sequences. Each sub-information sequence corresponds to a type of sub-information, and a plurality of sub-feature sequences are obtained according to the plurality of sub-information sequences. For example, for "transferring 1000 yuan on July 1", "paying 500 yuan on August 10", and "receiving 800 yuan on September 15", three sub-information sequences "July 1-August 10-September 15", "transfer-payment-receipt", and "1000-500-800" can be obtained. Each object behavior feature corresponds to a component in the three sub-feature sequences to represent the plurality of object behavior features.

[0058] In some embodiments, in order to make the sub-features of the plurality of object behavior features correspond to each other and avoid confusion of feature relationships, the sorting rules of the plurality of sub-information sequences can be the same, i.e., the three sub-information sequences can be "July 1-September 15-August 10", "transfer-receipt-payment", and "1000-800-500" after sorting in descending order of the amount. The sorting rule of the sub-feature sequence can correspond to the sorting rule of the sub-information sequence.

[0059] In some embodiments, the sub-information includes time information, by calculating the time difference between the time information, quantifying the time value by the time difference component, the expression form of the feature corresponding to the time information can be simplified, the difficulty of the time dimension feature can be reduced, and the original characteristics of the time can be retained to a greater extent. Specifically, step 1.2 or step 2.3 can include steps 3.1-3.4 as follows:

[0060] 3.1, obtaining the time information in all object behavior information;

[0061] 3.2, calculating the time difference between any two time information;

[0062] 3.3, assigning the time information by the time difference to obtain the value corresponding to the time information;

[0063] 3.4, determining the value corresponding to the time information as the sub-feature corresponding to the time information.

[0064] Wherein, the time information refers to the time information related to the object behavior, which can include the time of the object behavior, the time of receiving the object behavior or the time of recording the object behavior by the system, etc. For example, the object behavior is object A transferring money to object B, and the time information can be the time of object A initiating the transfer, the time of object B receiving the transfer, or the time of the system recording / settling the transfer, etc. The time information can be represented in the form of a timestamp, or it can be the time recorded by a timer or a counter, etc.

[0065] Wherein, the time difference can be the difference between two adjacent time information, or the time difference between any one time information and the reference time information, etc. The reference time information is determined according to a predetermined condition, which can be any one of all time information, or the first time information among all time information, etc. The first time information can be the earliest one after sorting all time information, or the first one after sorting multiple object information. The time difference can be in units of days, hours, minutes, seconds or milliseconds, etc. as the time unit, which can be selected according to the application scenario or experience. For example, for the time information "July 1st-August 10th-September 15th", the time difference can be in units of days, then if August 10th is the reference time information, the time difference is 40, 0, -36, and if July 1st is the reference time information, the time difference is 0, 40, 76. After obtaining the time difference between the time information, the time difference can be used as the value of the time information, and it can be used as a sub-feature. For example, when the object behavior feature is a vector, when the time difference is 40, 0, -36, the 40, 0, -36 can be used as the components corresponding to July 1st, August 10th, and September 15th, respectively.

[0066] In some embodiments, in order to simplify the calculation of time difference and assignment process, and also facilitate the addition of new object behavior information corresponding to time information, step 3.2 can be: calculating the time difference between each time information and the previous time information. The first time information can be assigned as 0.

[0067] In some embodiments, the sub-information contains behavior data, in order to reasonably quantify the behavior data, the behavior data is encoded by behavior type, specifically, step 1.2 or step 2.3 can include steps 4.1-4.4 as follows:

[0068] 4.1, obtaining the behavior type of the behavior data;

[0069] 4.2, determining the encoding corresponding to the behavior type;

[0070] 4.3, determining the encoding corresponding to the behavior type of the behavior data as the encoding corresponding to the behavior data;

[0071] 4.4, determining the encoding corresponding to the behavior data as the sub-feature corresponding to the behavior data.

[0072] Wherein, the behavior data refers to the data used to represent the behavior performed by the object. For example, when the object performs a transfer behavior, the object behavior information generated can be "transfer 1000 yuan on July 1st", but the behavior performed is transfer, and the behavior data is "transfer".

[0073] The behavior type refers to the type corresponding to the behavior performed by the object. The behavior performed by the object can be divided into multiple types according to application scenarios or experience, for example, when the present application is applied in a social scenario, when the financial behavior of the object is targeted, the behavior of the object can be divided into types such as transfer out and transfer in, at this time, transfer out can include specific behaviors such as transfer and send red envelope, and transfer in can include specific behaviors such as receive money and receive red envelope. At this time, transfer out can be coded as 01, and transfer in can be coded as 02, when the behavior of the object is transfer, the encoding corresponding to the behavior can be determined as 01, at this time, 02 can be used as the sub-feature corresponding to the transfer behavior of the object. The object behavior information can also be preprocessed before being obtained, to represent the behavior data by the behavior type, that is, the behavior data in the object behavior information obtained in step 2.1 is the behavior type, so step 4.1 can be omitted, or the behavior data is directly used as the behavior type in step 1.

[0074] In some embodiments, in order to make the Euclidean distance of the behavior type corresponding feature more reasonable, increase the accuracy of the identification result, step 4.2 can include the step of: encoding the behavior type by one-hot encoding. For example, the behavior types that need to be encoded include types such as transfer, loan, refund, payment, etc. At this time, by one-hot encoding, they can be represented as (1, 0, 0, 0), (0, 1, 0, 0), (0, 0, 1, 0), and (0, 0, 0, 1) respectively. In this way, different behavior types are converted into 4-dimensional sparse vectors, making the distance calculation between features more reasonable.

[0075] In some embodiments, the sub-information includes financial values. Since different financial values differ greatly, by performing weight calculation on the financial values, the deviation between the financial values is reduced, and the compatibility of the processing process for financial values with large differences is improved. Specifically, step 1.2 or step 2.3 can include steps 5.1-5.4 as follows:

[0076] 5.1, obtaining the financial values in all object behavior information to obtain an initial vector;

[0077] 5.2, obtaining a preset weight matrix;

[0078] 5.3, performing weight calculation on the initial vector by the preset weight matrix to obtain a target vector;

[0079] 5.4, obtaining a sub-feature corresponding to the financial value according to the component corresponding to the financial value in the target vector.

[0080] In the application scenarios of the embodiments of the present application, such as social, financial, online consumption, etc., the object behavior involving financial behavior is an important behavior that needs to be identified and is also a common violation behavior. Therefore, when obtaining the object behavior information, the financial values in the object financial behavior are obtained. The financial values can include amounts, stocks, options, etc. that can be converted into cash values. In actual application scenarios, risk objects often use other types of words instead of financial values, so before step 110, these words can be converted into words corresponding to financial values according to experience.

[0081] The preset weight matrix can be set in advance according to the application scenario or experience. By taking all financial values as components of a vector to form an initial vector, and performing weight calculation on the initial vector according to the preset weight matrix, such as weighted summation, a target vector is obtained. For example, the initial vector can be an n-dimensional column vector, and the preset weight matrix can be an n*n matrix. By multiplying the preset weight matrix by the initial vector, the target vector is obtained, which is an n-dimensional column vector. The component corresponding to the financial value in the target vector can be used as the sub-feature corresponding to the financial value, or the component can be subjected to bias calculation to obtain the sub-feature corresponding to the financial value, etc.

[0082] In some embodiments, in order to improve the compatibility of the processing process for financial values with large differences, step 5.4 can include: performing bias calculation on the components in the target vector corresponding to the financial values to obtain sub-features corresponding to the financial values. For example, after obtaining the target vector, each component can be added to the corresponding bias value to obtain the sub-feature corresponding to the component.

[0083] In some embodiments, in order to identify the abnormal financial behavior of the object, the sub-information includes financial values. Since the amount of money values is represented by large data such as ten thousand, one million, etc., the calculation is large in the subsequent feature interaction process. Therefore, by performing feature enhancement on the financial values, the expression ability of the financial features is improved, and the calculation speed is improved. Step 1.2 or step 2.3 can include steps 6.1-6.2, as follows:

[0084] 6.1, performing feature enhancement on all the financial values by target operation to obtain a target financial value corresponding to each financial value;

[0085] 6.2, for each financial value, determining the target financial value corresponding to the financial value as the sub-feature corresponding to the financial value.

[0086] Wherein, the target operation refers to an operation for adjusting the size of all financial values, which can be linear operation such as addition, multiplication, etc., or nonlinear operation such as logarithmic operation, square root operation, exponential operation, trigonometric function operation, etc. By performing target operation on all financial values, each financial value can be adjusted to meet the preset numerical size, for example, by difference operation to adjust the financial value that is too large to a reasonable range.

[0087] In some implementation processes, in order to simplify the calculation process, the target operation is linear operation. For example, all financial values are divided by a coefficient to simultaneously reduce all financial values.

[0088] In some embodiments, in order to identify the abnormal financial behavior of the object, the sub-information includes financial values, and in order to reduce the deviation between the financial values while improving the expression ability of the financial features, step 1.2 or step 2.3 can include steps 7.1-7.5, as follows:

[0089] 7.1, performing feature enhancement on all the financial values by target operation to obtain a target financial value corresponding to each financial value;

[0090] 7.2, obtaining all target financial values to obtain an initial vector;

[0091] 7.3, obtaining a preset weight matrix;

[0092] 7.4, performing weight calculation on the initial vector by the preset weight matrix to obtain a target vector;

[0093] 7.5, obtaining a sub-feature corresponding to the financial value according to a component in the target vector corresponding to the financial value.

[0094] It should be noted that the specific implementation process of steps 7.1-7.5 can refer to the corresponding description of steps 5.1-5.4 and steps 6.1-6.2 described above, and will not be repeated here. By processing the financial value, the financial value feature is enhanced before the weight calculation, which can better preserve the original characteristics of the financial value and improve the accuracy of identifying abnormal financial behavior.

[0095] 120, performing feature interaction on the plurality of object behavior features to obtain an initial feature group, the initial feature group including a plurality of target features, each target feature corresponding to a sub-feature type.

[0096] Among them, feature interaction refers to mutual calculation of feature vectors to realize feature interaction fusion. RNN (Recurrent Neural Network, recurrent neural network) network, DNN (Deep Neural Network, deep neural network) network, FNN (Factorization-machine supported Neural Networks, feedforward neural network) network, etc. can be used to perform feature interaction, etc.

[0097] Among them, the target feature is the feature corresponding to the sub-feature type in the initial feature group. For example, when the preset sub-feature type is finance, the corresponding sub-feature obtained after the feature interaction of the corresponding feature of this sub-feature type is the target feature.

[0098] In some embodiments, the global attention weight is obtained by linearly transforming the plurality of object behavior features, and the object behavior feature with global attention is obtained by adjusting the object behavior feature according to the global attention feature, and step 120 can include steps 8.1-8.3, as follows:

[0099] 8.1, linearly transforming the plurality of object behavior features to obtain an attention weight;

[0100] 8.2, for each object behavior feature, performing weight calculation on the object behavior feature according to the attention weight to obtain an adjusted object behavior feature;

[0101] 8.3, combining all adjusted object behavior features to obtain an initial feature group.

[0102] The attention weights can include relation values ​​and feature values. Relation values ​​can be at least one of Q (Query) values ​​or K (Key) values, and feature values ​​can be V (Value) values. Q is used to learn the relationship between itself and other elements, K is used to learn the relationship between other elements and itself, and V represents specific information about each element. Each element corresponds to a behavioral feature of an object in this embodiment. The Q, K, and V values ​​can be obtained from the embedding results through different linear transformations, assuming X∈R. n×d It is an input sample sequence, where each sample is an image sub-feature, and the sample sequence is all the input image sub-features, where n is the number of samples (sequence length) and d is the dimension of a single sample. Query, Key, and Value are defined as follows: Query: Q = X × W Q W Q ∈R d×dq Key: K = X × W K W K ∈R d×dk Value: V = X × W K W K ∈R d×dv .like Figure 3 As shown, each object behavior feature vector is divided into 6 heads. Any object behavior feature vector is multiplied by the respective weight matrices (W). Q W K W V ), such as X multiplied by W0 Q Transform into Q0. Calculate attention using the Q, K, V matrices. For example, calculate the attention of the K and V values ​​of any object's behavior feature with the Q values ​​of other object's behavior features to obtain the Z matrix (Z0, ..., Z5). Concatenate all attention heads (Z0, ..., Z5) and multiply by the weight matrix W. 0 This yields the adjusted object behavior characteristics.

[0103] The initial feature group is a combination of features obtained by merging object behavior features. It can be represented as a set or a matrix, etc. For example, if object behavior features are represented by an n×1 column vector, combining the column vectors corresponding to m object behavior features yields an n×m matrix, which is the initial feature group.

[0104] In some implementations, multiple object behavior features can be convolved using a convolution operator to obtain an initial feature set. The length of the convolution operator can be 2 to 5. Specifically, step 120 may include steps 9.1 to 9.2, as follows:

[0105] 9.1. For each sub-feature sequence, perform convolution calculation on all sub-features to obtain the convolved sub-feature sequence;

[0106] 9.2, combine all the convolutional sub-feature sequences to obtain an initial feature group.

[0107] 130, perform feature interaction on different types of target features to obtain a target feature group.

[0108] Perform linear or nonlinear operation on different types of target features, or respectively establish a feature group for different types of target features, and perform interaction calculation between different types of target feature groups according to the method of steps 8.1-8.3, to obtain a target feature group, etc.

[0109] In some embodiments, the target features are enhanced by vector dot product calculation. Step 130 can include steps 10.1-10.2 as follows:

[0110] 10.1, obtain the vector corresponding to the target feature from the initial feature group;

[0111] 10.2, perform vector dot product calculation on the vectors corresponding to different types of target features to obtain a target feature group.

[0112] Wherein, the target feature group is the result of performing vector dot product calculation on the vectors corresponding to at least two target features of different types. For example, the target features can include financial and time types, and the sub-features corresponding to finance and time are represented as vectors respectively. The correlation between the sub-features corresponding to finance and time is learned by dot product calculation (i.e. correlation term) of the two vectors, to enhance the target features. The above steps 10.1-10.2 can also be implemented by the following formula:

[0113]

[0114] In the above formula, the first two terms are used to calculate linear regression, and the last term is used to calculate the correlation term between feature i and feature j (i.e. to implement step 10.1). i x j represents the combination of features x i and x j , v i represents the hidden vector of the ith dimension feature, v j represents the hidden vector of the jth dimension feature, and <> represents vector dot product. The length of the hidden vector is k (k << n), which contains k factors describing the features. The above formula quantifies the relationship between the features into the fitting of x and y by multiplying all the components of the input, to realize the feature interaction of the target features.

[0115] In step 10.1, the target feature is combined with all input components by multiplying the last term in the above formula, which is equivalent to vector dot multiplication, to obtain a correlation term. In step 10.2, the target feature group is calculated according to the above formula by linearly regressing the correlation term and the first two terms. The target feature group can be a one-dimensional vector.

[0116] When calculating different types of target features, different types of target features corresponding to the same object behavior feature can be interacted, and the calculation results corresponding to all object behavior features can be combined to obtain a target feature group. For example, different types of target features corresponding to the same object behavior feature can be calculated by vector dot product, and all calculation results can be combined. The combination of all target features can also be converted into a vector and then calculated, etc. For example, the sub-feature types can include finance, time, and behavior, which are sub-features corresponding to financial values, time information, and behavior data, respectively. All target feature groups can be combined and converted into a one-dimensional vector, which can be input into the above formula for feature interaction. For example, the initial feature group corresponding matrix can be converted into a one-dimensional vector by the Rshape function.

[0117] 140. Determine the identification result of the plurality of object behavior features according to the target feature group, so as to determine an abnormal object according to the identification result.

[0118] In some embodiments, before step 110, the preset behavior recognition model can also be trained to obtain a target behavior recognition model. The preset behavior recognition model can include a global attention network, a feature interaction network, and an identification result output network to perform steps 110-140 described above. Before step 110, steps 11.1-11.2 can also be included, as follows:

[0119] 11.1, obtain a training sample and a preset behavior recognition model;

[0120] 11.2, train the preset behavior recognition model using the training sample to obtain a target behavior recognition model, which is used to perform behavior recognition on the plurality of object behavior features.

[0121] The training sample can include positive and negative samples. For example, in the financial scenario, the positive sample can be a sample of an object with a violation behavior, and the negative sample can be a sample of an object without a violation behavior. The positive and negative samples are input into the preset behavior recognition model to output predicted values of the positive and negative samples, and the sample labels are used as true values. Through continuous iterative training, the target behavior recognition model is obtained until the preset loss function converges.

[0122] The specific implementation of training the preset behavior recognition model is not limited, which can include but is not limited to steps 10-40, such as Figure 4As shown, the method for training the preset behavior recognition model is as follows:

[0123] 10. inputting the training sample into the preset behavior recognition model, the training sample including a plurality of object behavior features, the object behavior features being composed of a plurality of different types of sub-features;

[0124] 20. performing feature interaction on the plurality of object behavior features through a global attention network to obtain an initial feature group, the initial feature group including a plurality of target features, each target feature corresponding to a type of sub-feature;

[0125] 30. performing feature interaction on different types of target features through a feature interaction network to obtain a target feature group;

[0126] 40. obtaining a target behavior recognition model according to the target feature group and a preset loss function through a recognition result output network, the target behavior recognition model being used to perform behavior recognition on the plurality of object behavior features.

[0127] The recognition result output network can include a fully connected network, the target feature group being input into the fully connected network to determine a probability of correct sample recognition, and the behavior recognition model parameters being adjusted reversely according to the probability until a preset convergence condition is met. It should be noted that the specific implementation method of steps 10-40 can refer to the corresponding description of the recognition process in the embodiments of the present application, which will not be described here.

[0128] When the target behavior recognition model is used to recognize the plurality of object behavior features, the plurality of object behavior features can be obtained, the plurality of object behavior features being input into the preset behavior recognition model, the object behavior features being composed of a plurality of different types of sub-features; feature interaction is performed on the plurality of object behavior features through a global attention network to obtain an initial feature group, the initial feature group including a plurality of target features, each target feature corresponding to a type of sub-feature; feature interaction is performed on different types of target features through a feature interaction network to obtain a target feature group; and a recognition result of the plurality of object behavior features is determined according to the target feature group through a recognition result output network.

[0129] The output recognition result can be a probability that the plurality of object behavior features is a target behavior or whether it is a target behavior, the target behavior can be a violation behavior or a non-violation behavior, etc., which can be set according to application scenarios or experience. For example, the target feature group is input into a fully connected network to output a probability that the plurality of object behavior features is a target behavior, and whether the plurality of object behavior features is a target behavior can be further determined according to the probability. It should be noted that the present application recognizes the plurality of object behavior features as a whole, that is, a recognition result is comprehensively output for the plurality of object behavior features by extracting the plurality of object behavior features.

[0130] The abnormal object is an object satisfying a preset recognition result. For example, when the output result is a probability that the behavior characteristics of the multiple objects are illegal behaviors, the preset recognition result can be a preset probability value. When the output result is greater than the preset probability value, the object performing the behavior characteristics of the multiple objects can be considered as an abnormal object. When the output result is less than or equal to the preset probability value, the object can be considered as a non-abnormal object. For another example, when the output result is that the behavior characteristics of the multiple objects are illegal behaviors, the object performing the behavior characteristics of the multiple objects can be considered as an abnormal object. When the output result is non-illegal behaviors, the object can be considered as a non-abnormal object, and so on.

[0131] In some embodiments, the identification result of the target behavior recognition model is optimized by adjusting the parameter of the loss function. Specifically, after step 11.2, steps 12.1-12.2 can be further included, as follows:

[0132] 12.1, when the target behavior recognition model does not satisfy a preset condition, adjusting the parameter value of the preset loss function to obtain an adjusted loss function;

[0133] 12.2, updating the target behavior recognition model according to the adjusted loss function.

[0134] The preset condition can be a condition corresponding to the accuracy of the prediction of the target behavior recognition model on the training sample, or other conditions set according to the scene or experience, and so on. For example, the preset condition is that the prediction probability of the positive sample is greater than 0.7, or the accuracy of the sample prediction is greater than 90%. When the preset condition is not satisfied, the behavior recognition model is continuously trained according to the adjusted loss function until the preset condition is satisfied.

[0135] In some embodiments, since the obtained training samples can be unbalanced in different application scenarios, the preset loss function can be a Focal Loss loss function. The parameter value of the preset loss function includes a class weight parameter and a sample difficulty weight parameter. By adjusting the class weight parameter and the sample difficulty weight parameter of the Focal Loss loss function, the problem of sample imbalance can be solved, and the identification result of the target behavior recognition model is optimized.

[0136] The behavior recognition scheme based on artificial intelligence provided by the embodiments of the present application can be applied in various object behavior recognition scenarios. For example, in the risk control scenario, a plurality of object behavior features can be obtained, the object behavior features are composed of a plurality of different types of sub-features; the plurality of object behavior features are interacted to obtain an initial feature group, the initial feature group includes a plurality of target features, and each target feature corresponds to a type of sub-feature; the target features of different types are interacted to obtain a target feature group; and a recognition result of the plurality of object behavior features is determined according to the target feature group, so as to determine an abnormal object according to the recognition result. By using the scheme provided in the embodiments of the present application, the object behavior sequence and the correlation between different dimensions can be learned in multiple dimensions, so as to recognize the abnormal financial behavior of the object and improve the accuracy of the recognition result.

[0137] As can be seen from the above, the embodiments of the present application can interact the plurality of object behavior features, focus on the correlation between the object behavior features, and then interact the features between different types of features, focus on the context correlation of the object behavior, mine the correlation between the object behaviors in multiple dimensions, and improve the accuracy of the recognition result.

[0138] Blockchain is a new application mode of distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm and other computer technologies. Blockchain, in essence, is a decentralized database, which is a series of data blocks associated using cryptographic methods, each data block contains information of a batch of network transactions, and is used to verify the validity (anti-fake) of the information and generate the next block. Blockchain can include a blockchain underlying platform, a platform product service layer, and an application service layer.

[0139] The underlying blockchain platform can include processing modules such as object management, basic services, smart contracts, and operations management. The object management module is responsible for managing the identity information of all blockchain participants, including maintaining public and private key generation (account management), key management, and maintaining the correspondence between object real identities and blockchain addresses (access management). Furthermore, under authorization, it monitors and audits transactions of certain real identities and provides risk control rule configuration (risk control audit). The basic services module is deployed on all blockchain node devices to verify the validity of business requests. After consensus is reached on valid requests, they are recorded in storage. For a new business request, the basic services first perform interface adaptation parsing and authentication (interface adaptation), and then encrypt the business information through a consensus algorithm (consensus management). After encryption, the data is transmitted completely and consistently to the shared ledger (network communication) and recorded and stored. The smart contract module is responsible for contract registration, issuance, triggering, and execution. Developers can define contract logic using a programming language and publish it to the blockchain (contract registration). According to the contract terms, the key or other events are invoked to trigger execution and complete the contract logic. It also provides functions for contract upgrades and cancellations. The operations management module is mainly responsible for deployment, configuration modification, contract settings, cloud adaptation, and real-time status visualization during product launch, such as alarms, network status monitoring, and node device health status monitoring.

[0140] The platform's product service layer provides the basic capabilities and implementation frameworks for typical applications. Developers can leverage these basic capabilities, along with the specific characteristics of their business needs, to implement blockchain-based business logic. The application service layer provides blockchain-based application services to business stakeholders.

[0141] In one embodiment, the server provided in this application can act as a node in a blockchain system. After acquiring multiple object behavior features, it processes the multiple object behavior features to obtain the identification results of the multiple object behavior features. After verifying the identification results, it stores them in the blockchain as a new block after the verification is passed, so as to ensure that these extraction results will not be tampered with.

[0142] The method described in the above embodiments will be further described in detail below.

[0143] In this embodiment, the method of this application embodiment will be described in detail using a risk control scenario for financial behavior as an example.

[0144] like Figure 5 and Figure 6 As shown, the specific process of an AI-based behavior recognition method is as follows:

[0145] 210. Preprocess the object behavior sequences used for training to obtain training samples, which include multiple object behavior features used for training.

[0146] The training sequence of object behaviors can be either sequences with no violations or sequences with violations. After preprocessing, positive and negative samples can be obtained respectively. Figure 7 As shown, the preprocessing method for the object behavior sequences used in training is as follows:

[0147] 211. Obtain the training object behavior sequence, which includes multiple training object behavior information, each of which includes time, behavior type and amount.

[0148] In risk control scenarios targeting financial activities, object behaviors have three attributes: time, behavior type, and amount. Time can be the time the behavior occurred, a Unix timestamp, and a natural number. Behavior type can represent different operations performed by the object; behavior types are discrete variables, not operable by addition or subtraction, but only used to determine equality or adjacency relationships, similar to text in natural language processing. Amount can represent the amount of funds involved in the object's operations; the amount is a natural number and can be used for various calculations, such as addition, subtraction, division, and other derivative calculations. Figure 6 As shown, the behavioral information of the three objects used for training includes "1-behavior A-3000", "19-behavior B-2999" and "40-behavior C-4", with the three data points corresponding to time, behavior type and amount, respectively.

[0149] 212. Extract the time series, behavior type series, and amount series from the object behavior series used for training.

[0150] like Figure 6 As shown, the three sequences are displayed in three rows: time series, behavior type series, and amount series. The values ​​or behaviors in the sequences are arranged in chronological order. Each column in the figure corresponds to the behavior information of an object.

[0151] 213. Perform feature enhancement on the amount in the amount sequence to obtain the enhanced amount sequence.

[0152] To ensure that the characteristics of this data sequence are not lost and that the model can learn the relationships such as the amount window between behavioral sequences, this embodiment of the application does not standardize the time and amount data. Instead, according to the characteristics of the data business, the amount is enhanced by multiplication and the time is differentially processed to remove non-standardized data of amount and time.

[0153] Since the proportion of the inflow and outflow of the object fund in the sequence is important, the absolute value of the amount is very large in the value domain of a large number of objects, and has little effect on the sequence, so it is assumed that the total amount of the object fund is multiplied and divided, which is equivalent to distinguishing black production. For example Figure 8 As shown in the formula, the amount in the object behavior sequence for training is multiplied by 1 / 3000, so the original (3000, 2999, 4) is adjusted to (300, 299.9, 0.4) to enhance the financial value and improve the financial feature expression ability.

[0154] For time, since the difference (time difference) is calculated, the time dimension is no longer enhanced.

[0155] 214, the weight matrix is calculated for the enhanced amount sequence to obtain the amount feature sequence.

[0156] Because the amount value domain span is relatively large (the amount of a single transaction of different objects can be large or small), it is found in actual application that the loss function fluctuates obviously when the amount value is directly input into the behavior recognition model, so the amount sequence is compatible by using a full connection layer with bias when the amount value is input into the behavior recognition model. For example, the formula y = W x is used for calculation, where W represents the weight matrix, y represents the amount feature sequence, and x represents the enhanced amount sequence. As shown in the formula, the original sequence (3000, 2999, 4) is calculated to obtain (2.95, 0.26, 0.001). Figure 6

[0157] 215, the time information in the time information sequence is assigned according to the time difference to obtain the time feature sequence.

[0158] By calculating the interval between the occurrence times of two behaviors, the time is calculated by subtraction, and the time difference is used to represent the time feature value, which increases the relationship between time and behavior sequence. In addition, in order to further reduce the difficulty of the time dimension feature, the time is normalized to 0 + time interval, that is, the first element is 0, and the rest are all time intervals from the previous time. As shown in the formula, the time feature sequence can be adjusted from (1, 19, 40) to (0, 18, 21). Figure 6

[0159] 216, the behavior type in the behavior type sequence is one-hot encoded to obtain the behavior type feature sequence.

[0160] Time and amount can be directly calculated by matrix, but behavior type cannot be calculated by coefficient multiplication, so embedding calculation is needed, which can be performed by One-hot Embedding operation in natural language processing field. For example Figure 9 ​​As shown, the behaviors A, B, and C are embedded, i.e., one-hot encoding is performed on the behaviors, and the encoding result vectors are mapped to a new space to obtain the behavior type feature sequence, where behavior A corresponds to (0.3, 0.8, 0.2, 0.1), behavior B corresponds to (0.1, 0.8, 0.6, 0.3), and behavior C corresponds to (0.1, 0.1, 0.7, 0.1). In Figure 6 the vectors corresponding to behaviors A, B, and C are represented as column vectors. After embedding the behavior types, all input data has been converted into decimal and natural number forms.

[0161] 217, combine the time feature sequence, the behavior type feature sequence, and the amount feature sequence to obtain an object behavior feature matrix for training, and use the matrix as a training sample.

[0162] The time feature sequence, the behavior type feature sequence, and the amount feature sequence obtained through the above process are represented in the form of vectors or matrices. As Figure 6 shown, all feature sequences can be combined in the form of rows to obtain a matrix, which can be used as a training sample to input a preset behavior recognition model. In the matrix, each column corresponds to an object behavior feature.

[0163] 220, input the training sample into the preset behavior recognition model, and the preset behavior recognition model includes a global attention network, a feature interaction network, and a recognition result output network.

[0164] 230, through the global attention network, the multiple object behavior features are subjected to feature interaction to obtain an initial feature group, and the initial feature group includes multiple target features, each target feature corresponding to a sub-feature type.

[0165] In actual applications, the irregular objects in the risk control scene mostly have certain behavior patterns, and there are also fixed patterns in behavior, so there is also relatively fixed context strong correlation in the behavior sequence. The above method can be used to use the proximity of adjacent behaviors, behavior sub-sequences (length 2-5), and amount difference / ratio to quantify the object behavior sequence, and a convolution operator with a length of 2-5 can also be used to perform convolution calculation on the three-way sequence. These methods are all for processing adjacent behaviors, and cannot perform cross-processing on the behavior sequence. However, in the object behavior sequence, the irregular behaviors are likely not directly adjacent, so a better processing method is needed.

[0166] To this end, a global attention network can be used, for example, a Transformer module is used as a global attention network, and the interaction between the behavior features is realized through the Transformer module, so as to realize the feature interaction between all object behavior features and identify the correlation between non-adjacent object behavior features. In the Transformer part, it includes Positional Embedding, Feed Forward and Self-attention. In the present scheme, when the training sample is input into the attention module, the matrix is split into a vector corresponding to each object behavior feature, so as to realize the interaction between multiple object behavior features through the attention module, and as much as possible to be compatible with the jumping violation behavior, and to identify the behavior sequence mixed with the violation behavior and the non-violation behavior.

[0167] 240. Through the feature interaction network, the feature interaction between the target features corresponding to time, amount and behavior is realized, and the target feature group is obtained.

[0168] Although the global attention network can realize the interaction between multiple object behavior features, the feature interaction in the time, behavior feature and amount dimensions is not realized. Therefore, the feature interaction network is used to interact the features of the time, amount and behavior feature sequence. This interaction can be mapped to the co-occurrence relationship of the context-related and operation time context-related on the object behavior operation event.

[0169] For example, the initial feature group obtained through the Transformer module is a matrix, the first row represents time, the second row represents behavior features, and the third row represents amount. Each column represents the correlation result in different dimensions, and the correlation between different dimensions can be further fused and strengthened through the feature interaction network.

[0170] For example, the feature interaction network can be an FM (Factorization Machine) model. The initial feature group can be converted into a one-dimensional vector through an Rshape function, and input into the FM model to enhance the signals on the time, amount and behavior sequence. The output target feature group is also a one-dimensional vector.

[0171] 250. Through the recognition result output network, the target behavior recognition model is obtained by iterative training according to the target feature group and the Focal Loss loss function.

[0172] The network outputting the recognition results can include fully connected layers. The target feature set (one-dimensional vector) is input into the fully connected layer to obtain the classification result. Since positive samples are scarce in risk control scenarios, especially when a risk scenario has just erupted, this embodiment uses the Focal Loss function to address the sample imbalance problem. The hyperparameter α (class weight parameter) is used to adjust the imbalance in the sample ratio, and the hyperparameter γ (sample difficulty weight parameter) is used to adjust the distribution of loss between easy and difficult samples. A value >1 increases the contribution of difficult samples to the loss during training. α and γ can be adjusted manually or through grid search until the model evaluation metric is optimal. The formula for the Focal Loss function is as follows:

[0173]

[0174] 260. Process the sequence of object behavior to be identified to obtain multiple object behavior features to be identified.

[0175] The object behavior sequence to be identified includes multiple object behavior information to be identified. The processing method for the object behavior sequence to be identified can be found in steps 211 to 216 above, to obtain the time feature sequence, behavior type feature sequence and amount feature sequence, and to combine the time feature sequence, behavior type feature sequence and amount feature sequence to obtain the matrix used for identification.

[0176] 270. Input the behavioral features of multiple objects to be identified into the target behavior recognition model to obtain the recognition results.

[0177] The process in which the target behavior recognition model recognizes the plurality of object behavior features to be recognized can be seen from steps 230-240. After obtaining the target feature group, the recognition result output network can be a full connection network. After inputting the target feature group into the full connection network for processing, the recognition result is obtained. For example, the recognition result can be that the probability of the object behavior sequence to be recognized being a violation behavior is 80%. Compared with the prior art, the model in this embodiment is used for object behavior recognition, and the effect is obviously improved by 10%. After obtaining the recognition result, the object can be further determined as an abnormal object according to the recognition result. For example, the object corresponding to the probability of the violation behavior being greater than a preset probability value (such as 70%) is determined as an abnormal object. As can be seen from the above, the embodiment of the application proposes a behavior recognition method based on artificial intelligence. The object behavior is recognized based on the behavior sequence. The global attention network is used for feature interaction of behaviors in the sequence to learn the relationship between different behaviors and pay attention to the relevance between the behaviors before and after. Then, the feature interaction network is used for interaction of features in different dimensions to increase the relevance between contexts, to mine the relevance between object behaviors in multiple dimensions, and to improve the accuracy of the recognition result. When the object behavior sequence is converted into object behavior features, the amount is not standardized, but is enhanced, and the time is processed by difference, so as to reduce the data dimension and increase the expression ability of the amount and time features.

[0178] In order to better implement the above method, the embodiment of the application further provides a behavior recognition device based on artificial intelligence. The behavior recognition device based on artificial intelligence can be integrated in an electronic device, which can be a terminal, a server, or the like. The terminal can be a mobile phone, a tablet computer, a smart Bluetooth device, a notebook computer, a personal computer, or the like. The server can be a single server or a server cluster composed of multiple servers.

[0179] For example, in this embodiment, the behavior recognition device based on artificial intelligence is integrated in a server.

[0180] For example, as shown in Figure 10 The behavior recognition device based on artificial intelligence can include an acquisition unit 310, a first calculation unit 320, a second calculation unit 330, and a recognition unit 340, as follows.

[0181] (I) Acquisition unit 310

[0182] The acquisition unit 310 is configured to acquire a plurality of object behavior features. The object behavior features are composed of a plurality of different types of sub-features.

[0183] In some embodiments, the acquisition unit 310 can be used for steps 13.1-13.3, as follows.

[0184] 13.1, obtaining a plurality of object behavior information, each object behavior information comprising a plurality of sub-information;

[0185] 13.2, performing feature extraction on the sub-information to obtain a sub-feature corresponding to the sub-information;

[0186] 13.3, for each object behavior information, combining all sub-features corresponding to the object behavior information to obtain an object behavior feature corresponding to each object behavior information.

[0187] In some embodiments, the sub-information includes time information, and step 13.2 can include steps 14.1-14.4 as follows:

[0188] 14.1, obtaining time information in all object behavior information;

[0189] 14.2, calculating the time difference between any two time information;

[0190] 14.3, assigning values to the time information through the time difference to obtain a value corresponding to the time information;

[0191] 14.4, determining the value corresponding to the time information as the sub-feature corresponding to the time information.

[0192] In some embodiments, the sub-information includes behavior data, and step 13.2 can include steps 15.1-15.4 as follows:

[0193] 15.1, obtaining a behavior type of the behavior data;

[0194] 15.2, determining an encoding corresponding to the behavior type;

[0195] 15.3, determining the encoding corresponding to the behavior type of the behavior data as the encoding corresponding to the behavior data;

[0196] 15.4, determining the encoding corresponding to the behavior data as the sub-feature corresponding to the behavior data.

[0197] In some embodiments, the sub-information includes financial value, and step 13.2 can include steps 16.1-16.4 as follows:

[0198] 16.1, obtaining a financial value in all object behavior information to obtain an initial vector;

[0199] 16.2, obtaining a preset weight matrix;

[0200] 16.3, performing weight calculation on the initial vector through the preset weight matrix to obtain a target vector;

[0201] 16.4、According to the component in the target vector corresponding to the financial value, the sub-feature corresponding to the financial value is obtained.

[0202] In some embodiments, the sub-information includes the financial value, and step 13.2 can include steps 17.1-17.2 as follows:

[0203] 17.1、Perform feature enhancement on all financial values through target operation to obtain a target financial value corresponding to each financial value;

[0204] 17.2、For each financial value, the target financial value corresponding to the financial value is determined as the sub-feature corresponding to the financial value.

[0205] (ii) The first calculation unit 320

[0206] For feature interaction on a plurality of object behavior features to obtain an initial feature group, the initial feature group includes a plurality of target features, and each target feature corresponds to a sub-feature type.

[0207] In some embodiments, the first calculation unit 320 can be specifically used for steps 18.1-18.3 as follows:

[0208] 18.1、Perform linear transformation on a plurality of object behavior features to obtain attention weights;

[0209] 18.2、For the object behavior features, perform weight calculation on the object behavior features according to the attention weights to obtain adjusted object behavior features;

[0210] 18.3、Combine all adjusted object behavior features to obtain an initial feature group.

[0211] (iii) The second calculation unit 330

[0212] For feature interaction on different types of target features to obtain a target feature group.

[0213] In some embodiments, the recognition unit 340 can be specifically used for steps 19.1-19.3 as follows:

[0214] 19.1、From the initial feature group, obtain a vector corresponding to each target feature;

[0215] 19.2、Perform vector dot product calculation on vectors corresponding to different types of target features to obtain a target feature group.

[0216] (iv) The recognition unit 340

[0217] For determining a recognition result of a plurality of object behavior features according to the target feature group, so as to determine an abnormal object according to the recognition result.

[0218] In some embodiments, the identifying unit 340 can also be used for steps 20.1-20.2 as follows:

[0219] 20.1, obtaining a training sample and a preset behavior recognition model;

[0220] 20.2, training the preset behavior recognition model using the training sample to obtain a target behavior recognition model, the target behavior recognition model being used to perform behavior recognition on the object behavior features.

[0221] In some embodiments, the preset behavior recognition model includes a global attention network, a feature interaction network, and a recognition result output network, and step 20.2 can include steps 20.2.1-20.2.4 as follows:

[0222] 20.2.1, inputting the training sample into the preset behavior recognition model, the training sample including the object behavior features, the object behavior features being composed of multiple different types of sub-features;

[0223] 20.2.2, performing feature interaction on the object behavior features by the global attention network to obtain an initial feature group, the initial feature group including multiple target features, each target feature corresponding to a type of sub-feature;

[0224] 20.2.3, performing feature interaction on the different types of target features by the feature interaction network to obtain a target feature group;

[0225] 20.2.4, obtaining the target behavior recognition model according to the target feature group and a preset loss function by the recognition result output network.

[0226] In some embodiments, after step 20.2, steps 20.3-20.4 can also be included as follows:

[0227] 20.3, when the target behavior recognition model does not meet a preset condition, adjusting a parameter value of the preset loss function to obtain an adjusted loss function;

[0228] 20.4, updating the target behavior recognition model according to the adjusted loss function.

[0229] In implementation, each of the above units can be implemented as an independent entity, or can be combined as the same or several entities, and the specific implementation of each of the above units can refer to the method embodiments above, which will not be repeated here.

[0230] Therefore, the embodiments of this application can improve the accuracy of recognition results by performing feature interaction on multiple object behavior features, focusing on the correlation between object behavior features, and then combining feature interaction between different types of features to focus on the contextual correlation of object behavior, so as to explore the correlation between object behaviors in multiple dimensions.

[0231] This application also provides an electronic device, which can be a terminal, a server, or other similar device. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers, etc.

[0232] In some embodiments, the AI-based behavior recognition device can also be integrated into multiple electronic devices, such as multiple servers, whereby the AI-based behavior recognition method of this application is implemented by multiple servers.

[0233] In this embodiment, a server will be used as an example for detailed description. For example, ... Figure 11 As shown, it illustrates a schematic diagram of the server structure involved in an embodiment of this application. Specifically:

[0234] The server may include components such as a processor 401 with one or more processing cores, a memory 402 with one or more computer-readable storage media, a power supply 403, an input module 404, and a communication module 405. Those skilled in the art will understand that... Figure 11 The SSS structure shown does not constitute a limitation on the SSS and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Wherein:

[0235] Processor 401 is the control center of the server, connecting various parts of the server via various interfaces and lines. It performs various server functions and processes data by running or executing software programs and / or modules stored in memory 402, and by calling data stored in memory 402. In some embodiments, processor 401 may include one or more processing cores; in some embodiments, processor 401 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, graphical interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into processor 401.

[0236] The memory 402 can be used to store software programs and modules, and the processor 401 can execute various function applications and data processing by running the software programs and modules stored in the memory 402. The memory 402 can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, application programs required by at least one function (such as a sound playing function, an image playing function, etc.), and the like; and the data storage area can store data created according to the use of the server, etc. In addition, the memory 402 can include a high-speed random access memory, and can also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device. Accordingly, the memory 402 can also include a memory controller to provide the processor 401 with access to the memory 402.

[0237] The server further includes a power supply 403 for powering the various components, and in some embodiments, the power supply 403 can be logically connected to the processor 401 through a power management system, so that the power management system can be used to manage charging, discharging, and power consumption management, etc. The power supply 403 can also include one or more direct current or alternating current power supplies, a recharging system, a power failure detection circuit, a power converter or inverter, a power status indicator, and the like.

[0238] The server can further include an input module 404, which can be used to receive input digital or character information, and to generate keyboard, mouse, joystick, optical or trackball signal inputs related to settings and function controls.

[0239] The server can further include a communication module 405, which in some embodiments can include a wireless module, and the server can use the wireless module of the communication module 405 for short-range wireless transmission, thereby providing wireless broadband Internet access. For example, the communication module 405 can be used to send and receive emails, browse web pages, and access streaming media, etc.

[0240] Although not shown, the server can also include a display unit, etc., which will not be described here. In particular, in the present embodiment, the processor 401 in the server will load the executable file corresponding to the process of one or more application programs into the memory 402 according to the following instructions, and the processor 401 will run the application programs stored in the memory 402 to implement various functions, as follows:

[0241] The plurality of object behavior features are obtained, and the object behavior features are composed of a plurality of different types of sub-features. The plurality of object behavior features are interacted to obtain an initial feature group, and the initial feature group includes a plurality of target features, each of which corresponds to a type of sub-feature. The target features of different types are interacted to obtain a target feature group. The identification result of the plurality of object behavior features is determined according to the target feature group, so as to determine an abnormal object according to the identification result.

[0242] The specific implementation of each operation can refer to the foregoing embodiments, and will not be described here.

[0243] As can be seen from the above, in the present application, the correlation between object behavior features is focused on by interacting the plurality of object behavior features, and the context correlation of object behavior is focused on by interacting the features between different types of features, so as to mine the correlation between object behaviors in multiple dimensions and improve the accuracy of the identification result.

[0244] Those skilled in the art can understand that all or part of the steps of the various methods of the above embodiments can be completed by instructions, or by related hardware controlled by the instructions, which can be stored in a computer readable storage medium and loaded and executed by a processor.

[0245] Therefore, the embodiments of the present application provide a computer readable storage medium, which stores a plurality of instructions. The instructions can be loaded by a processor to execute the steps in any of the behavior recognition methods based on artificial intelligence provided by the embodiments of the present application. For example, the instructions can execute the following steps:

[0246] The plurality of object behavior features are obtained, and the object behavior features are composed of a plurality of different types of sub-features. The plurality of object behavior features are interacted to obtain an initial feature group, and the initial feature group includes a plurality of target features, each of which corresponds to a type of sub-feature. The target features of different types are interacted to obtain a target feature group. The identification result of the plurality of object behavior features is determined according to the target feature group, so as to determine an abnormal object according to the identification result.

[0247] The storage medium can include a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0248] According to an aspect of the present application, a computer program product or computer program is provided, which includes computer programs / instructions stored in a computer readable storage medium. A processor of a computer device reads the computer programs / instructions from the computer readable storage medium, and the processor executes the computer programs / instructions, so that the computer device performs steps in various optional implementation manners of machine learning aspects or risk control aspects provided in the above embodiments.

[0249] Due to the instructions stored in the storage medium, steps in any of the behavior recognition methods based on artificial intelligence provided in the embodiments of the present application can be performed, thus the beneficial effects that can be achieved by any of the behavior recognition methods based on artificial intelligence provided in the embodiments of the present application can be achieved, which are described in detail in the foregoing embodiments and will not be described here again.

[0250] The above describes in detail a behavior recognition method, device, server and computer readable storage medium based on artificial intelligence provided in the embodiments of the present application. The principles and implementation manners of the present application are described by applying specific examples in this paper. The above embodiment descriptions are only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manners and application ranges will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. An artificial intelligence-based behavior recognition method, characterized by, include: Multiple object behavior features are acquired, which are composed of various types of sub-features; wherein, for financial scenarios, the sub-features include: time information, behavioral data and / or financial value; for social scenarios, the sub-features include: time information, behavioral data and / or comment keywords. The method involves performing feature interaction on the multiple object behavior features to obtain an initial feature group, including: performing a linear transformation on the multiple object behavior features to obtain attention weights; calculating the weights of each object behavior feature based on the attention weights to obtain adjusted object behavior features; and combining all the adjusted object behavior features to obtain the initial feature group. The initial feature group includes multiple target features, and each target feature corresponds to a sub-feature type. By performing feature interactions on different types of target features, target feature groups are obtained; Based on the target feature group, the identification results of the behavioral features of the multiple objects are determined so as to identify abnormal objects based on the identification results; wherein, for financial scenarios, the identification result is whether the object has engaged in illegal transaction behavior; for social scenarios, the identification result is whether the object has made malicious comments. 2.The behavior recognition method based on artificial intelligence according to claim 1, wherein, The acquisition of multiple object behavior characteristics includes: Obtain multiple object behavior information, each of which includes multiple sub-information; Feature extraction is performed on the sub-information to obtain the sub-features corresponding to the sub-information; For each object behavior information, all the sub-features corresponding to the object behavior information are combined to obtain the object behavior feature corresponding to each object behavior information. 3.The behavior recognition method based on artificial intelligence according to claim 2, wherein, The sub-information includes time information, and the step of extracting features from the sub-information to obtain the sub-features corresponding to the sub-information includes: Obtain the time information from all the object behavior information; Calculate the time difference between any two of the aforementioned time information; The time information is assigned a value by using the time difference to obtain the value corresponding to the time information; The value corresponding to the time information is determined as the sub-feature corresponding to the time information. 4.The behavior recognition method based on artificial intelligence according to claim 2, wherein, The sub-information includes behavioral data, and the step of extracting features from the sub-information to obtain the sub-features corresponding to the sub-information includes: The behavior type for obtaining the behavior data; Determine the encoding corresponding to the behavior type; The code corresponding to the behavior type of the behavior data is determined as the code corresponding to the behavior data; The code corresponding to the behavioral data is determined as the sub-feature corresponding to the behavioral data.

5. The behavior recognition method based on artificial intelligence as described in claim 2, characterized in that, The sub-information includes financial values, and the feature extraction of the sub-information to obtain the sub-features corresponding to the sub-information includes: Obtain the financial values ​​from all the object behavior information to obtain the initial vector; Obtain the preset weight matrix; The initial vector is weighted using a preset weight matrix to obtain the target vector; Based on the component in the target vector that corresponds to the financial value, the sub-feature corresponding to the financial value is obtained.

6. The behavior recognition method based on artificial intelligence as described in claim 2, characterized in that, The sub-information includes financial values, and the feature extraction of the sub-information to obtain the sub-features corresponding to the sub-information includes: By performing feature enhancement on all the financial values ​​through target operations, a target financial value corresponding to each financial value is obtained; For each of the financial values, the target financial value corresponding to the financial value is determined as the sub-feature corresponding to the financial value.

7. The behavior recognition method based on artificial intelligence as described in claim 1, characterized in that, The step of performing feature interaction on different types of target features to obtain target feature groups includes: From the initial feature set, obtain the vector corresponding to each of the target features; The vector dot product is calculated for the vectors corresponding to the different types of target features to obtain the target feature group.

8. The behavior recognition method based on artificial intelligence as described in claim 1, characterized in that, Before obtaining multiple object behavior characteristics, the method also includes: Obtain training samples and a pre-defined behavior recognition model; The preset behavior recognition model is trained using the training samples to obtain a target behavior recognition model, which is used to perform behavior recognition on the behavioral features of the multiple objects.

9. The behavior recognition method based on artificial intelligence as described in claim 8, characterized in that, The preset behavior recognition model includes a global attention network, a feature interaction network, and a recognition result output network. Training the preset behavior recognition model using the training samples to obtain the target behavior recognition model includes: The training samples are input into the preset behavior recognition model. The training samples include multiple object behavior features, and the object behavior features are composed of multiple different types of sub-features. The global attention network is used to perform feature interaction on the multiple object behavior features to obtain an initial feature group. The initial feature group includes multiple target features, and each target feature corresponds to a sub-feature type. Through the feature interaction network, different types of target features are interacted to obtain target feature groups; The target behavior recognition model is obtained by outputting the recognition results through the network, based on the target feature group and the preset loss function.

10. The behavior recognition method based on artificial intelligence as described in claim 9, characterized in that, After obtaining the target behavior recognition model through the recognition result output network based on the target feature group and a preset loss function, the method further includes: When the target behavior recognition model does not meet the preset conditions, the parameter values ​​of the preset loss function are adjusted to obtain the adjusted loss function; The target behavior recognition model is updated based on the adjusted loss function.

11. A behavior recognition device based on artificial intelligence, characterized in that, include: The acquisition unit is used to acquire multiple object behavioral features, which are composed of various types of sub-features; wherein, for a financial scenario, the sub-features include: time information, behavioral data, and / or financial value; for a social scenario, the sub-features include: time information, behavioral data, and / or comment keywords. The first calculation unit is used to perform feature interaction on the plurality of object behavior features to obtain an initial feature group, including: performing a linear transformation on the plurality of object behavior features to obtain attention weights; for each object behavior feature, calculating the weight of the object behavior feature according to the attention weights to obtain an adjusted object behavior feature; combining all the adjusted object behavior features to obtain the initial feature group; the initial feature group includes multiple target features, and each target feature corresponds to a sub-feature type; The second computing unit is used to perform feature interaction on different types of target features to obtain target feature groups; The identification unit determines the identification results of the behavioral characteristics of the multiple objects based on the target feature group, so as to identify abnormal objects based on the identification results; wherein, for financial scenarios, the identification result is whether the object has engaged in illegal transaction behavior; for social scenarios, the identification result is whether the object has made malicious comments.

12. The apparatus as claimed in claim 11, characterized in that, The acquisition unit is specifically used to acquire multiple object behavior information, each object behavior information including multiple sub-information; to extract features from the sub-information to obtain sub-features corresponding to the sub-information; and to combine all the sub-features corresponding to each object behavior information to obtain object behavior features corresponding to each object behavior information.

13. The apparatus as claimed in claim 12, characterized in that, The sub-information includes time information. The acquisition unit is specifically used to acquire the time information in all the object behavior information; calculate the time difference between any two pieces of time information; assign a value to the time information through the time difference to obtain the value corresponding to the time information; and determine the value corresponding to the time information as the sub-feature corresponding to the time information.

14. The apparatus as claimed in claim 12, characterized in that, The sub-information includes behavioral data. The acquisition unit is specifically used to acquire the behavioral type of the behavioral data; determine the code corresponding to the behavioral type; and determine the code corresponding to the behavioral type of the behavioral data as the code corresponding to the behavioral data. The code corresponding to the behavioral data is determined as the sub-feature corresponding to the behavioral data.

15. The apparatus as claimed in claim 12, characterized in that, The sub-information includes financial values. The acquisition unit is specifically used to acquire the financial values ​​from all the object behavior information to obtain an initial vector; acquire a preset weight matrix; and perform weight calculation on the initial vector using the preset weight matrix to obtain a target vector. Based on the component in the target vector that corresponds to the financial value, the sub-feature corresponding to the financial value is obtained.

16. The apparatus as claimed in claim 12, characterized in that, The sub-information includes financial values. The acquisition unit is specifically used to perform feature enhancement on all the financial values ​​through target operation to obtain a target financial value corresponding to each financial value; for each financial value, the target financial value corresponding to the financial value is determined as a sub-feature corresponding to the financial value.

17. The apparatus as claimed in claim 11, characterized in that, The second calculation unit is specifically used to obtain the vector corresponding to each target feature from the initial feature group; and to perform vector dot product calculation on the vectors corresponding to different types of target features to obtain the target feature group.

18. The apparatus as claimed in claim 11, characterized in that, The device further includes: The training unit is used to acquire training samples and a preset behavior recognition model; the preset behavior recognition model is trained using the training samples to obtain a target behavior recognition model, which is used to perform behavior recognition on the behavioral features of the multiple objects.

19. The apparatus as claimed in claim 18, characterized in that, The preset behavior recognition model includes a global attention network, a feature interaction network, and a recognition result output network; The training unit is specifically used to input the training samples into the preset behavior recognition model. The training samples include multiple object behavior features, which are composed of various types of sub-features. Through the global attention network, the multiple object behavior features are interacted to obtain an initial feature group. The initial feature group includes multiple target features, each of which corresponds to a sub-feature type. Through the feature interaction network, different types of target features are interacted to obtain a target feature group. The target behavior recognition model is obtained by outputting the recognition results through the network, based on the target feature group and the preset loss function.

20. The apparatus as claimed in claim 19, characterized in that, The device further includes: The update unit is used to adjust the parameter values ​​of the preset loss function when the target behavior recognition model does not meet the preset conditions, so as to obtain the adjusted loss function; and update the target behavior recognition model according to the adjusted loss function.

21. A server, characterized in that, It includes a processor and a memory, the memory storing multiple instructions; the processor loads instructions from the memory to execute the steps in the AI-based behavior recognition method as described in any one of claims 1 to 10.

22. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to execute the steps of the artificial intelligence-based behavior recognition method according to any one of claims 1 to 10.

23. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps in the AI-based behavior recognition method according to any one of claims 1 to 10.

Images