Data processing method, device and equipment based on recognition model, medium and product

Abstract

The invention provides a data processing method, device and equipment based on an identification model, a medium and a product, and the method comprises the steps: obtaining sample data used for training the identification model, the sample data comprising M groups of business data related to a sample object and an annotation label of the sample object; calling an identification model to perform multi-channel time sequence coding processing on service behavior data in the M groups of service data to obtain time sequence behavior embedding features; calling an identification model to perform knowledge coding processing on the business description text data in the M groups of business data to obtain business text knowledge features; calling a recognition model to perform exception recognition processing based on the time sequence behavior embedding features and the business text knowledge features to obtain an exception recognition result; and training an identification model based on the difference between the abnormal identification result and the annotation label, the trained identification model being used for identifying the object type of the business object. The accuracy of object type identification can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computers, in particular to the technical field of artificial intelligence, and specifically to a data processing method based on an identification model, a data processing device based on an identification model, a computer device, a computer readable storage medium, and a computer program product. BACKGROUND

[0002] With the deepening of the digitalization of businesses in the fields of finance, e-commerce, enterprise services, and the like, the demand for object type identification is increasingly urgent - for example, in the financial scenario, it is necessary to identify whether a user is a risk account, in the enterprise service scenario, it is necessary to classify whether a business subject is a compliant object, in the e-commerce scenario, it is necessary to judge whether a merchant is an abnormal operating subject, and the like. In the prior art, rules, supervised machine learning algorithms (such as logistic regression, decision tree, random forest, and the like) and deep learning techniques are used to determine the object type, but the accuracy of the existing object type identification method is low. SUMMARY

[0003] The embodiments of the present application provide a data processing method, device, and equipment based on an identification model, a medium, and a product, which can improve the accuracy of object type identification.

[0004] In one aspect, the embodiments of the present application provide a data processing method based on an identification model, which comprises: obtaining sample data used for training an identification model, the sample data comprising M groups of business data related to a sample object and a labeled label of the sample object; one group of business data is derived from one business performed by the sample object in a first business scenario, and one group of business data contains business behavior data and business description text data; the labeled label is used to indicate the labeled type of the sample object, and M is a positive integer; calling the identification model to perform multi-channel time sequence encoding processing on the business behavior data in the M groups of business data to obtain time sequence behavior embedding features, the time sequence behavior embedding features being used to represent the channel original characteristics of the business behavior data in the M groups of business data in time; calling the identification model to perform knowledge encoding processing on the business description text data in the M groups of business data to obtain business text knowledge features, the business text knowledge features being used to represent the business logic correlation between the business description text data in the M groups of business data; calling the identification model to perform abnormality identification processing based on the time sequence behavior embedding features and the business text knowledge features to obtain an abnormality identification result, the abnormality identification result being used to indicate a predicted type of the sample object; training the identification model based on the difference between the abnormality identification result and the labeled label, the trained identification model being used to identify the object type of a business object.

[0005] Correspondingly, the embodiment of the present application provides a data processing device based on an identification model, the device comprises: an acquisition unit, configured to acquire sample data used for training the identification model, the sample data comprising M groups of business data related to a sample object and a labeled label of the sample object, one group of business data being derived from one service performed by the sample object in a first service scenario, one group of business data containing service behavior data and service description text data, and the labeled label being used for indicating a labeling type of the sample object, M being a positive integer; a processing unit, configured to call the identification model to perform multi-channel time sequence encoding processing on the service behavior data in the M groups of business data, to obtain time sequence behavior embedding features, the time sequence behavior embedding features being used for representing channel original characteristics of the service behavior data in the M groups of business data in time; the processing unit is further configured to call the identification model to perform knowledge encoding processing on the service description text data in the M groups of business data, to obtain service text knowledge features, the service text knowledge features being used for representing service logic correlation between the service description text data in the M groups of business data; the processing unit is further configured to call the identification model to perform abnormality identification processing based on the time sequence behavior embedding features and the service text knowledge features, to obtain an abnormality identification result, the abnormality identification result being used for indicating a prediction type of the sample object; the processing unit is further configured to train the identification model based on a difference between the abnormality identification result and the labeled label, the trained identification model being used for identifying an object type of a service object.

[0006] In one embodiment, the identification model comprises a time sequence data dynamic quantization module and an embedding layer in a pre-trained language model. The processing unit calls the identification model to perform multi-channel time sequence encoding processing on the service behavior data in the M groups of business data, to obtain the time sequence behavior embedding features, which comprises: arranging the service behavior data in the M groups of business data according to an execution sequence of services corresponding to the M groups of business data, to obtain a time sequence service behavior sequence; calling the time sequence data dynamic quantization module in the identification model to perform data scaling and quantization processing on the time sequence service behavior sequence, to obtain a time sequence behavior quantization sequence; calling the embedding layer in the identification model to perform multi-channel embedding and splicing processing on the time sequence behavior quantization sequence, to obtain the time sequence behavior embedding features.

[0007] In one embodiment, the processing unit calls the time sequence data dynamic quantization module in the identification model to perform data scaling and quantization processing on the time sequence service behavior sequence, to obtain the time sequence behavior quantization sequence, which comprises: The time series data dynamic quantification module in the recognition model is called to perform data scaling processing on the continuous variable elements in the time series business behavior sequence according to a dispersion standardization manner, to obtain a time series behavior scaling sequence. The time series data dynamic quantification module in the recognition model is called to perform data quantification processing on the continuous variable elements in the time series behavior scaling sequence according to a binning index manner, to obtain a time series behavior quantization sequence. The binning index manner includes at least one of the following: an equal-width binning index manner, a clustering binning index manner, and a dynamic binning index manner.

[0008] In an embodiment, the embedding layer includes a word embedding layer and a position embedding layer; the time series behavior quantization sequence includes behavior quantization data corresponding to each business behavior data, and each behavior quantization data includes at least one element; The processing unit calls the embedding layer in the recognition model to perform multi-channel embedding splicing processing on the time series behavior quantization sequence, to obtain time series behavior embedding features, including: The word embedding layer in the recognition model is called to perform word embedding representation processing on each element in the time series behavior quantization sequence, to obtain word embedding features of each element in the time series behavior quantization sequence. The position embedding layer in the recognition model is called to perform position embedding representation processing on each element in the time series behavior quantization sequence, to obtain position embedding features of each element in the time series behavior quantization sequence. The word embedding features and the position embedding features of each element in the time series behavior quantization sequence are fused to obtain target embedding features of each element in the time series behavior quantization sequence. The target embedding features of all elements in the time series behavior quantization sequence are spliced based on the order of each behavior quantization data in the time series behavior quantization sequence and the order of each element in each behavior quantization data, to obtain time series behavior embedding features.

[0009] In an embodiment, the recognition model includes an encoding module in a pre-trained language model; the encoding module includes an embedding layer and an encoding layer. The processing unit calls the recognition model to perform knowledge encoding processing on the business description text data in the M groups of business data, to obtain business text knowledge features, including: The business description text data in the M groups of business data is arranged according to the execution order of the respective businesses of the M groups of business data, to obtain a time series business text sequence. The embedding layer in the recognition model is called to perform embedding representation processing on the time series business text sequence, to obtain time series text embedding features. The encoding layer in the recognition model is called to perform attention encoding processing on the time series text embedding features, to obtain business text knowledge features.

[0010] In an embodiment, the identification model comprises a classification module, the classification module comprising a nonlinear mapping layer, a normalization layer, and a classification layer, the nonlinear mapping layer comprising a cumulative distribution function of a standard normal distribution; The processing unit calls the identification model to perform abnormality identification processing based on the time-series behavior embedding feature and the business text knowledge feature, to obtain an abnormality identification result, comprising: The time-series behavior embedding feature and the business text knowledge feature are spliced to obtain a behavior text spliced feature; The nonlinear mapping layer in the identification model is called to perform nonlinear mapping processing on the behavior text spliced feature to obtain a nonlinear mapping feature; The normalization layer in the identification model is called to perform normalization processing on the nonlinear mapping feature to obtain a normalized feature; The classification layer in the identification model is called to perform classification processing on the normalized feature to obtain the abnormality identification result.

[0011] In an embodiment, the identification model comprises a normalization layer and a position embedding layer in a pre-trained language model; The processing unit trains the identification model based on the difference between the abnormality identification result and the labeled label, comprising: A binary classification loss function is obtained, and the abnormality identification result and the labeled label are input into the binary classification loss function to obtain a first loss parameter corresponding to the identification model; the first loss parameter is used to reflect the difference between the abnormality identification result and the labeled label; The first loss parameter is used to adjust the model parameters of the normalization layer and the position embedding layer in the identification model to train the identification model.

[0012] In an embodiment, the identification model further comprises a decoding module in the pre-trained language model; The processing unit adjusts the model parameters of the normalization layer and the position embedding layer in the identification model using the first loss parameter to train the identification model, comprising: When the labeled label indicates that the sample object is of an abnormal type, a real abnormal reason text is obtained; The time-series behavior embedding feature and the business text knowledge feature are spliced to obtain a behavior text spliced feature; The decoding module in the pre-trained language model is called to perform abnormal reason analysis processing based on the behavior text spliced feature to obtain a predicted abnormal reason text; Based on the difference between the real abnormal reason text and the predicted abnormal reason text, a second loss parameter corresponding to the identification model is determined; The first loss parameter and the second loss parameter are used to adjust the model parameters of the normalization layer, the position embedding layer, and the decoding module in the identification model to train the identification model.

[0013] In one embodiment, the acquisition unit is further configured to: acquire a general language model and business corpus text data in a first business scenario; the general language model comprises an encoding module and a decoding module; The processing unit is further configured to: mask processing the business corpus text data to obtain masked corpus text data; call the encoding module in the general language model to perform encoding processing on the masked corpus text data to obtain masked text encoding features; Call the decoding module in the general language model to perform mask prediction processing based on the masked text encoding features to obtain predicted mask content in the business corpus text data; Based on the difference between the predicted mask content and the real mask content in the business corpus text data, the general language model is trained to obtain a pre-trained language model.

[0014] In one embodiment, the acquisition unit acquires sample data for training the recognition model, comprising: Determine M items of business continuously performed by the sample object, and acquire behavior feature information of each item of business in P behavior feature channels and text feature information in Q text feature channels; P and Q are both positive integers; Use the behavior feature information of each item of business in the P behavior feature channels to construct initial behavior data of each item of business, and perform cleaning and filling processing on the initial behavior data of each item of business to obtain business behavior data of each item of business; Use the text feature information of each item of business in the Q text feature channels to construct business description text data of each item of business; Obtain the label tags obtained after expert evaluation processing of the M items of business; Based on the business behavior data and the business description text data of the M items of business, and the label tags, construct sample data for training the recognition model.

[0015] In one aspect, the embodiments of the present application provide a data processing method based on a recognition model, which comprises: Acquire N groups of target business data related to a business object, one group of target business data is derived from one item of business performed by the business object in a first business scenario, and one group of target business data comprises target business behavior data and target business description text data; N is a positive integer; Call the trained recognition model to perform multi-channel time sequence encoding processing on the target business behavior data in the N groups of target business data to obtain target time sequence behavior embedding features, the target time sequence behavior embedding features are used to represent the channel original characteristics of the target business behavior data in the N groups of target business data in time; The trained recognition model is called to perform knowledge coding processing on the target business description text data in the N groups of target business data, to obtain target business text knowledge features, and the target business text knowledge features are used to represent the business logic correlation between the target business description text data in the N groups of target business data. The trained recognition model is called to perform abnormality recognition processing based on the target time sequence behavior embedding features and the target business text knowledge features, to obtain the object type of the business object.

[0016] Correspondingly, an apparatus for data processing based on a recognition model is provided, and the apparatus comprises: An acquisition unit is configured to acquire N groups of target business data related to a business object, one group of target business data is derived from one business performed by the business object in a first business scenario, and one group of target business data comprises target business behavior data and target business description text data; N is a positive integer. A processing unit is configured to call the trained recognition model to perform multi-channel time sequence coding processing on the target business behavior data in the N groups of target business data, to obtain target time sequence behavior embedding features, and the target time sequence behavior embedding features are used to represent the channel original characteristics of the target business behavior data in the N groups of target business data in time. The processing unit is further configured to call the trained recognition model to perform knowledge coding processing on the target business description text data in the N groups of target business data, to obtain target business text knowledge features, and the target business text knowledge features are used to represent the business logic correlation between the target business description text data in the N groups of target business data. The processing unit is further configured to call the trained recognition model to perform abnormality recognition processing based on the target time sequence behavior embedding features and the target business text knowledge features, to obtain the object type of the business object.

[0017] In one embodiment, the trained recognition model comprises a time sequence data dynamic quantization module and an embedding layer. The processing unit calls the trained recognition model to perform multi-channel time sequence coding processing on the target business behavior data in the N groups of target business data, to obtain target time sequence behavior embedding features, which comprises: The target business behavior data in the N groups of target business data is arranged in the order of execution of the respective businesses, to obtain a target time sequence business behavior sequence; The time sequence data dynamic quantization module in the trained recognition model is called to perform data scaling and quantization processing on the target time sequence business behavior sequence, to obtain a target time sequence behavior quantization sequence; The embedding layer in the trained recognition model is called to perform multi-channel embedding and splicing processing on the target time sequence behavior quantization sequence, to obtain the target time sequence behavior embedding features.

[0018] In an embodiment, the trained recognition model comprises an encoding module, the encoding module comprising an embedding layer and an encoding layer; The processing unit calls the trained recognition model to perform knowledge coding processing on the target business description text data in the N groups of target business data, to obtain target business text knowledge features, comprising: According to the execution sequence of the respective businesses of the N groups of target business data, the target business description text data in the N groups of target business data is arranged to obtain a target time sequence business text sequence; The embedding layer in the trained recognition model is called to perform embedding representation processing on the target time sequence business text sequence to obtain target time sequence text embedding features; The encoding layer in the trained recognition model is called to perform attention coding processing on the target time sequence text embedding features to obtain target business text knowledge features.

[0019] In an embodiment, the trained recognition model comprises a classification module, the classification module comprising a nonlinear mapping layer, a normalization layer and a classification layer; The processing unit calls the trained recognition model to perform abnormality recognition processing based on the target time sequence behavior embedding features and the target business text knowledge features to obtain the object type of the business object, comprising: The target time sequence behavior embedding features and the target business text knowledge features are spliced to obtain target behavior text splicing features; The nonlinear mapping layer in the trained recognition model is called to perform nonlinear mapping processing on the target behavior text splicing features to obtain target nonlinear mapping features; The normalization layer in the trained recognition model is called to perform normalization processing on the target nonlinear mapping features to obtain target normalized features; The classification layer in the trained recognition model is called to perform classification processing on the target normalized features to obtain the object type of the business object.

[0020] In an embodiment, the trained recognition model comprises a decoding module, and the processing unit is further configured to: When the object type of the business object is an abnormal type, the target time sequence behavior embedding features and the target business text knowledge features are spliced to obtain target behavior text splicing features; The decoding module in the trained recognition model is called to perform abnormality reason analysis processing based on the target behavior text splicing features to obtain target abnormality reason text.

[0021] In one aspect, the embodiments of the present application provide a computer device, which comprises: A processor configured to load and execute a computer program; A computer readable storage medium, the computer readable storage medium storing a computer program, the computer program being executed by a processor to implement the data processing method based on the identification model.

[0022] In an aspect, an embodiment of the present application provides a computer readable storage medium, the computer readable storage medium storing a computer program, the computer program being loaded by a processor and executing the data processing method based on the identification model.

[0023] In an aspect, an embodiment of the present application provides a computer program product, the computer program product including a computer program, the computer program being executed by a processor to implement the data processing method based on the identification model.

[0024] In the embodiment of the present application, sample data for training the recognition model can be obtained, the sample data including M groups of business data related to a sample object and a labeled label of the sample object; one group of business data is derived from one service performed by the sample object in a first service scenario, and one group of business data contains service behavior data and service description text data; the labeled label is used to indicate the labeling type of the sample object, and M is a positive integer. The sample data is the raw material for model learning, the service behavior data provides a dynamic behavior pattern, the service description text data provides semantic logic, and the labeled label provides a correct answer, all of which together support the model learning to “how to identify the object type from the data”. The recognition model is called to perform multi-channel time sequence encoding processing on the service behavior data in the M groups of business data, to obtain time sequence behavior embedding features, which are used to represent the channel original characteristics of the service behavior data in the M groups of business data in time. In this way, the original behavior data can be converted into time pattern features that can be understood by the model, and the dynamic law behind the behavior (such as whether it deviates from the historical normal pattern) is revealed. The recognition model is called to perform knowledge encoding processing on the service description text data in the M groups of business data, to obtain service text knowledge features, which are used to represent the business logic correlation between the service description text data in the M groups of business data. In this way, the unstructured text can be converted into structured semantic features, and the logical motivation behind the service behavior (such as whether it is reasonable or whether it is driven by external factors) can be captured. The recognition model is called to perform abnormality recognition processing based on the time sequence behavior embedding features and the service text knowledge features, to obtain an abnormality recognition result, which is used to indicate a predicted type of the sample object. In this way, the time law of the behavior and the semantic logic of the text can be comprehensively considered, to realize accurate prediction of the object type. Based on the difference between the abnormality recognition result and the labeled label, the recognition model is trained, and the trained recognition model is used to identify the object type of the service object, that is, through the closed loop of data→model→prediction→feedback→optimization, the model gradually learns the ability of “how to accurately judge the object type from the behavior and the text”. As can be seen, by simultaneously analyzing the time pattern of the service behavior (multi-channel time sequence encoding) and the semantic logic of the text (knowledge encoding), the model can more comprehensively depict the object features, and combined with the supervision feedback of the labeled label, the model continuously optimizes the recognition ability of the object type, and finally realizes accurate classification of the object type. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 is an architecture schematic diagram of a data processing system provided by an embodiment of the present application; Figure 2 is a flow schematic diagram of a data processing method based on a recognition model provided by an embodiment of the present application Figure One ; Figure 3 is a structure schematic diagram of a general language model provided by an embodiment of the present application; Figure 4 is a flowchart of a data processing method based on an identification model provided by an embodiment of the present application Figure Two ; Figure 5 is a structural diagram of an identification model provided by an embodiment of the present application Figure One ; Figure 6 is a structural diagram of an identification model provided by an embodiment of the present application Figure Two ; Figure 7 is a flowchart of a data processing method based on an identification model provided by an embodiment of the present application Figure Three ; Figure 8 is a structural diagram of a data processing device based on an identification model provided by an embodiment of the present application Figure 9 is a structural diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0026] Object type identification refers to: based on the multi-channel data (such as the behavior characteristic information under the behavior characteristic channel and the text characteristic information under the text characteristic channel) generated by the object in the business execution process, mining the behavior rule and the business execution logic, and then accurately determining the object type. Exemplarily, the object type includes normal type (such as “normal” and “compliant”) and abnormal type (such as “abnormal” and “non-compliant”).

[0027] The present application provides a data processing method based on an identification model. The purpose of the data processing method is to train an identification model by fusing business behavior data and business description text data, to realize automatic identification of object type, and to support explainable analysis of abnormal reasons. The main challenges and solutions of the present application are as follows: (1) Modeling of time sequence characteristics of business behavior data: Business behavior data is structured multi-channel time sequence data, such as “transaction time + transaction location + transaction resource amount” in the financial scenario, and for example, “product listing time + product inventory change amount + product logistics status” in the e-commerce scenario, which has strong time sequence dependence (previous behavior easily affects subsequent behavior) and multi-channel independence (different channels carry different information). The traditional method easily ignores the independent characteristics between channels and the time sequence dynamic rule, which leads to inaccurate behavior pattern representation and affects the subsequent identification effect.

[0028] The application adopts multi-channel time sequence coding processing: independent coding of each channel, each group of service behavior data is split into independent channels (such as time channel, resource amount channel, space channel) according to dimensions, each channel is processed independently to avoid channel information cross interference; dynamic quantization standardization, through dispersion standardization, continuous variables (such as transaction resource amount, commodity inventory change amount) are scaled to a unified range, and then combined with binned index quantization to solve the dimension difference and abnormal value interference problem; multi-channel embedding splicing, for the quantized channels, the semantic feature of the elements in the channel (such as "early morning" corresponding to time sensitive attribute) is extracted by using the word embedding layer, and the time sequence position information (such as the order attribute of "the third transaction") is captured by using the position embedding layer, and finally the embedding representation of each element is spliced in the order of the channel to form a time sequence behavior embedding feature which retains the original characteristics of the multi-channel time sequence.

[0029] (2) Mining of service logic association of service description text data: Service description text data (such as "transaction remarks" and "contract terms" in the financial scenario, "commodity details" and "customer service dialogue" in the e-commerce scenario) is unstructured data, which implies business logic (such as transaction remarks "borrowing turnover" associated with fund chain risk, "original factory same model" may involve selling fake goods). Traditional methods cannot directly extract the logical association implied in the text, resulting in insufficient utilization of semantic information and inability to provide effective text support for identification.

[0030] The application realizes the structural transformation of text semantics through "pre-trained language model + knowledge coding": time sequence text sorting, arranging service description text data in the order of business execution, obtaining time sequence business text sequence; coding processing, using the coding module in the pre-trained language model to code process the time sequence business text sequence, and with the help of the business domain knowledge (such as domain knowledge in the financial scenario and domain knowledge in the e-commerce scenario) implied in the pre-trained language model, capturing the context dependency relationship between the service description text data in the coding process, and outputting business text knowledge features that can represent the business logic association.

[0031] (3) Fusion and alignment of multi-modal (behavior + text) features: The time sequence behavior embedding features (representing the behavior time sequence mode) and the business text knowledge features (representing the text logical association) belong to different modalities, which are prone to information discontinuity between modalities (such as "high-frequency transactions in the early morning" behavior and "loan remarks" text are not associated), which leads to the fact that the fused features cannot accurately reflect the true type of the object, affecting the identification accuracy.

[0032] The application adopts a dual strategy of "time alignment + feature fusion": time alignment, based on business execution time (such as transaction time, commodity listing time), aligning business behavior data and business description text data on the same time axis, ensuring that behavior and text correspond one by one in the time dimension; through the classification module, the time sequence behavior embedding features and the business text knowledge features are fused, and the association between the behavior mode and the text logic (such as "high-frequency transactions in the early morning" + "no credit record of counterpart account" jointly pointing to "high-risk account") is comprehensively mined, and the accurate recognition result is output.

[0033] (4) Domain adaptation of general language model: Although the general language model (such as T5 model) has general semantic understanding ability, there are a large number of domain-specific terms (such as "seven-day no-reason return" in e-commerce scenarios) and special logic (such as "overdue non-payment triggers risk control") in specific business scenarios. Direct application of the general language model may lead to business semantic representation bias and reduce the quality of business text knowledge features.

[0034] The application improves the task adaptability of the model through "domain pre-training + fine-tuning": domain mask pre-training, mask processing (covering part of the domain terms or key information) of the business corpus text data (such as financial transaction notes, e-commerce commodity descriptions) in the specific business scenario, training the encoding module and decoding module of the general language model to predict the masked content, so that the model can learn the business semantic association; specific task fine-tuning, using the labeled data (such as labeled labels, real abnormal reason text) to fine-tune the trained general language model (i.e. pre-trained language model) to adapt to the classification needs and abnormal reason analysis needs of specific business scenarios, forming a scenario-adapted identification model.

[0035] (5) Challenges of multi-task construction: The application needs to complete "object type recognition" (main task) and "abnormal reason generation" (secondary task) at the same time, facing target consistency conflict (the main task pursues classification accuracy, and the secondary task needs to generate explainable abnormal reason text) and feature coordination barrier (the main task requires features to have strong discriminability, and the secondary task requires features to have strong semantic association).

[0036] The application adopts the strategy of "loss coordination + feature interaction": loss coordination optimization, through the dual-loss mechanism of "first loss parameter (main task classification loss) + second loss parameter (secondary task generation loss)", the dual-task target consistency is ensured; feature interaction, the main task and the secondary task share part of the feature extraction layer (such as multi-channel time sequence encoding and knowledge encoding), but the output layer is independent (such as the classification module and the decoding module), to simultaneously support the object type recognition and the abnormal reason generation task.

[0037] The data processing method provided in the application can be executed by a computer device, which can be a terminal or a server, that is, the data processing method can be executed by a terminal or a server. Alternatively, the data processing method can also be executed by a terminal and a server together. The terminal mentioned above can include but is not limited to a terminal such as a smart phone (such as a smart phone deploying an Android system or a smart phone deploying an Internet Operating System (IOS)), a tablet computer, a portable personal computer, a Mobile Internet Device (MID), a vehicle-mounted device, a head-mounted device, a smart home and a smart robot. The server can be a stand-alone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, Content Delivery Network (CDN), and basic cloud computing services such as big data and artificial intelligence platforms.

[0038] Please refer to Figure 1 FIG. 1 is a schematic diagram of an architecture of a data processing system provided by an embodiment of the application. The data processing method mentioned above can be executed by a terminal and a server together, for example, the server can collect services executed by an object in a terminal, to construct sample data for training a recognition model and train the recognition model; and collect services executed by the object in the terminal, to implement object type recognition and abnormal reason analysis by using the trained recognition model.

[0039] The terminal and the server can be connected directly or indirectly through wired or wireless communication, which is not limited in the application. In addition, the number of terminals and servers is not limited in the embodiments of the application.

[0040] In the embodiments of the application, the collection and processing of related data should strictly comply with the requirements of relevant laws and regulations. The personal information needs to be obtained with the knowledge or consent of the personal subject (or with the legal basis for information acquisition), and the subsequent data use and processing behavior should be carried out within the scope of authorization of laws and regulations and personal information subjects. For example, when the embodiments of the application are applied to specific products or technologies, such as when sample data is acquired, the permission or consent of the sample object needs to be obtained, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards in the relevant region.

[0041] Next, the data processing method based on a recognition model provided by an embodiment of the application will be described in detail.

[0042] Please refer to Figure 2is a flowchart of a data processing method based on a recognition model provided by an embodiment of the present application Figure One The data processing method based on the recognition model can be executed by a server in an embodiment corresponding to the server Figure 1 The data processing method based on the recognition model involves a training process of a general language model, and includes but is not limited to steps S21-S25: S21, obtaining a general language model and business corpus text data in a first business scenario; the general language model includes an encoding module and a decoding module.

[0043] The general language model is trained based on a large amount of general corpus (such as Internet text, public books, news, etc.), and is a deep learning model with cross-scene basic semantic understanding and text generation capability. It is not optimized for a specific business scenario (such as the first business scenario), so there are natural limitations in understanding "business-specific terms and scene-specific logic".

[0044] Please refer to Figure 3FIG. 1 is a structural diagram of a general language model provided by an embodiment of the present application. The general language model comprises an encoding module and a decoding module. The encoding module is responsible for semantic understanding and feature extraction, and can convert input text (such as business corpus text data) into a structured semantic feature vector (such as a text encoding feature corresponding to the business corpus text data), capturing the context dependency and semantic logic in the text. For example, the encoding module can be composed of an embedding layer and an encoding layer. The embedding layer includes a word embedding layer and a position embedding layer. The word embedding layer is used to map discrete tokens (such as “transfer” and “urgent”) in the input text into continuous low-dimensional vectors (such as 768-dimensional vectors, referred to as word embedding features), representing the basic semantics of the tokens (such as the action semantics of the “transfer” corresponding to the financial scenario). The position embedding layer supplements position information (such as the position embedding feature of the third token) for each token, capturing the sequential dependency (such as the timing logic of “time + behavior”) in the text. The encoding layer can be stacked by multiple layers of Transformer encoders (such as 12 layers for a T5-base model and 24 layers for a T5-large model). Each layer of the Transformer encoder realizes global semantic modeling through a multi-head self-attention mechanism (Multi-Head Self-Attention) and a feed-forward network (FFN), and can convert the input text into a semantic vector rich in context information, providing a semantic basis for the input text for the decoding module. The decoding module can be responsible for feature decoding and text generation, and can generate target text (such as abnormal reason explanation and classification label) that conforms to grammar and semantic logic based on the semantic feature vector output by the encoding module. For example, the decoding model can be composed of a decoding layer and an output layer. The decoding layer is stacked by multiple layers of Transformer decoders (such as 12 layers for a T5-base model and 24 layers for a T5-large model). Each layer of the Transformer decoder completes fine processing of the input features through a combination of a self-attention mechanism (causal attention), a cross-attention mechanism, and a fully connected network. The self-attention mechanism only allows the features of the current generation position to focus on the previously processed features (without allowing to view the future information), avoiding future information leakage. The cross-attention mechanism can enable the features of the decoding layer to actively associate the text semantic features output by the encoding module, and filter and generate key information strongly related to the target. The fully connected network can perform nonlinear processing on the output features of the cross-attention mechanism, enhancing the richness of semantic expression. The output layer is the exit of the decoding module, responsible for directly converting the features output by the decoding layer into natural language text, which can be composed of a linear layer and a Softmax layer. The linear layer is used to map the features output by the decoding layer to the dimension of the word table (such as 50,000 token embedding features in the word embedding layer, and the output dimension is 50,000).The role of the Softmax layer is to convert the output of the linear layer into a probability distribution (ranging from [0, 1]) of each token, and then select the token at the current position according to the probability distribution. That is, the next position token can be predicted based on the previously generated tokens and the input text, realizing the text generation function. Among them, the token (token) is used to represent the basic unit of NLP task processing, usually 1 Chinese word, 1 English word, 1 number, 1 symbol is counted as 1 token.

[0045] For example, the general language model can be a T5 (Text-to-Text Transfer Transformer) model, and the core design concept of the T5 model is to unify all NLP (Natural Language Processing) tasks into a "text-to-text" conversion problem (such as classification, translation, summary, etc. can be regarded as "input text → output text" mapping).

[0046] The first business scenario can be any business scenario, such as a financial scenario or an e-commerce scenario. The business corpus text data is text data derived from the first business scenario and containing scenario-based semantics and logic. By fine-tuning the general language model with business corpus text data, the general language model can learn domain-specific vocabulary, thereby converting the "general knowledge" of the general language model into "domain knowledge" and solving the problem of "imprecise semantic understanding" of the general language model in vertical scenarios. For example, the business corpus text data can be text data strongly related to object type recognition, which can include but is not limited to: behavior-related text, i.e., descriptions directly related to business behavior (such as transaction notes "customer requires urgent processing" and device logs "temperature too high triggers protection"); rule text, i.e., business-related compliance requirements (such as the rule "transactions over 500,000 per day require secondary verification"); abnormal case text, i.e., expert-labeled abnormal case descriptions (such as "the account was marked as a risk account due to 'diversified transfers in + centralized transfers out'"); and dialogue interaction text, such as customer service dialogue "user feedback that the goods have not been received".

[0047] S22, performing mask processing on the business corpus text data to obtain masked corpus text data.

[0048] The preset mask symbol (such as [MASK], [BLANK], and [MASK] can be uniformly used as the mask symbol in this application) can be used to replace the content that is crucial to the business semantic understanding in the business corpus text data, and the mask corpus text data with “partial information missing but complete context” is generated. For example, when the business corpus text data is “the account is marked as a risk account due to ‘diversified transfer in + centralized transfer out’ ”, the mask corpus text data can be “the account is marked as a risk account due to ‘[MASK]’ + ‘[MASK]’ ”.

[0049] S23, calling the encoding module in the general language model to perform encoding processing on the mask corpus text data to obtain mask text encoding features.

[0050] In one embodiment, the embedding layer in the general language model can be called to perform embedding representation processing on the mask corpus text data to obtain mask text embedding features. For example, the embedding layer includes a word embedding layer and a position embedding layer, the word embedding layer includes a large number of word embedding features (such as a word embedding feature for representing token1, a word embedding feature for representing token2, and the like), and the position embedding layer includes a large number of position embedding features (such as a position embedding feature for representing the first position, a position embedding feature for representing the second position, and the like). The word embedding layer in the general language model can be called to find the word embedding features that can be used to represent each element (or each token) in the mask corpus text, and the position embedding layer in the general language model can be called to find the position embedding features that can be used to represent the position of each element in the mask corpus text. The word embedding features and the position embedding features of each element in the mask corpus text are fused (such as adding the word embedding features and the position embedding features element by element) to obtain the mask text embedding features. The encoding layer in the general language model is called to perform attention encoding processing on the mask text embedding features to obtain mask text encoding features, wherein if the mask corpus text data includes 10 elements, the mask text encoding features include 10 feature vectors of specific dimensions (such as 512 dimensions), each feature vector corresponds to the global semantic representation of an element, and contains the context semantic clues for inferring the [MASK] position element.

[0051] S24, calling the decoding module in the general language model to perform mask prediction processing based on the mask text encoding features to obtain the predicted mask content in the business corpus text data.

[0052] The decoding module in the general language model can take the masked text encoding feature as input, predict each masked position in sequence based on the existing context information (including the masked text encoding feature and the content of the predicted masked position), obtain the probability of each token being the content of each masked position, and take the token corresponding to the maximum probability as the content of the current masked position. The predicted mask content in the business corpus text data can be determined by using the content of all the masked positions.

[0053] S25, based on the difference between the predicted mask content and the real mask content in the business corpus text data, the general language model is trained to obtain a pre-trained language model.

[0054] The greater the difference between the predicted mask content and the real mask content in the business corpus text data, the less accurate the general language model prediction is; the smaller the difference between the predicted mask content and the real mask content in the business corpus text data, the more accurate the general language model prediction is.

[0055] In one embodiment, the probability that the content of each masked position is the real masked content can be obtained, and the mask loss parameter can be calculated using the mask loss function shown in formula 1. The greater the mask loss parameter, the greater the difference between the predicted mask content and the real mask content; the smaller the mask loss parameter, the smaller the difference between the predicted mask content and the real mask content. Formula 1 is: (1) Wherein, L m represents the mask loss parameter, represents the probability that the content of each masked position is the real masked content, represents the logarithmic function, identifies the sum of obtained for all masked positions.

[0056] Exemplarily, the model parameters of the general language model can be adjusted in a direction of reducing the difference between the predicted mask content and the real mask content in the business corpus text data, to train the general language model. For example, when training the general language model, a back propagation algorithm (such as a stochastic gradient descent (SGD) algorithm) can be used to adjust the model parameters of the general language model in a direction of reducing the mask loss parameter, so that the mask loss parameter calculated again by the general language model after each adjustment is less than the mask loss parameter calculated by the general language model before the adjustment. When the number of adjustments of the general language model reaches an upper limit of the number of mask adjustments (which can be set as needed), or the general loss parameter is less than a mask loss threshold (which can be set as needed), the general language model after the last adjustment can be determined as the pre-trained language model.

[0057] Exemplarily, assuming that the general language model includes M layers of Transformer encoders and M layers of Transformer decoders, only the first M1 layers of Transformer encoders and the first M1 layers of Transformer decoders can be used in the pre-trained language model, where M and M1 are positive integers, and M1 is less than M. Because the latter layers learn more specific patterns of the mask prediction task, this can preserve more general domain context semantics, and can reduce inference time and memory usage.

[0058] Batch (training batch) refers to a set of samples processed by the model at a time of training iteration (i.e., adjusting the model parameters once). In the actual training process of the general language model, the general language model usually iterates in units of Batch, that is, the general language model calculates the mask loss parameters corresponding to each business corpus text data in a Batch, and then adjusts the model parameters of the general language model by using the total mask loss parameter obtained by summing all the calculated mask loss parameters.

[0059] As can be seen, the application can mask the core elements (such as domain terms and business logic words) in the business domain exclusive corpus (such as financial transaction notes and e-commerce commodity details), and then use the training link of “coding module extracts context semantic association features of masked text + decoding module accurately predicts mask content” to make the general language model break away from the cognitive limitations of general corpus, actively learn and internalize the semantic logic specific to the business scenario, and effectively solve the problem of understanding deviation of general language model for vertical domain terms and cognitive deficiency of business rules.

[0060] See Figure 4 is a flowchart of a data processing method based on a recognition model provided by an embodiment of the application Figure TwoThe data processing method based on the recognition model can be executed by Figure 1 The server in the corresponding embodiment, the data processing method based on the recognition model involves the training process of the recognition model, and the data processing method based on the recognition model includes but is not limited to steps S41-S45: S41, obtaining sample data for training the recognition model, the sample data including M groups of business data related to a sample object and a labeled label of the sample object; one group of business data is derived from an item of business executed by the sample object in a first business scenario, and one group of business data contains business behavior data and business description text data; the labeled label is used to indicate the annotation type of the sample object, and M is a positive integer.

[0061] Among them, the sample data includes M groups of business data, each group of business data is derived from an item of business executed by the sample object (which can be any object, which can be a user, a device, an enterprise, etc.) in a first business scenario (such as a transaction in a financial scenario, an operation record in an enterprise service scenario, etc.), and M is a positive integer (usually greater than or equal to 2); the labeled label can be the real type (normal type or abnormal type) of the sample object determined by expert evaluation or historical detection.

[0062] Among them, each group of business data contains two types of sub-data, business behavior data and business description text data. The business behavior data is a quantifiable and structured behavior index in the business execution process, which records the structured behavior characteristics of the business. The business description text data is unstructured text information describing the business background, purpose, rules, etc.

[0063] In one embodiment, the sample data for training the recognition model is obtained, including: Step 1, determine M items of continuous business of the sample object, such as 100 transactions of the user in the past 30 days in the financial scenario, or 72 hours of continuous operation record of the device. The time sequence of the continuous business (i.e. M items of business) such as "transfer at midnight → refund at dawn" can reflect the dynamic behavior pattern (such as the persistence of abnormal operation) of the object. Obtain the behavior feature information of each item of business in P behavior feature channels and the text feature information in Q text feature channels; P and Q are both positive integers. The behavior feature information is the objective quantitative record of the business execution, which is extracted through P behavior feature channels, and is used to reflect the dynamic behavior pattern of the sample object. For example, in the financial scenario, the behavior feature channels can include but are not limited to: time channel, and the corresponding behavior feature information in the time channel can include at least one of the following: transaction time, time interval from the previous transaction; resource amount channel, and the corresponding behavior feature information in the resource amount channel can include at least one of the following: single transaction resource amount, today's cumulative transaction resource amount; device channel, and the corresponding behavior feature information in the device channel can include at least one of the following: operation device model, device operating system type, device address (such as IP address), device region. The text feature information is the natural language description of the business execution, which is extracted through Q text feature channels, and is used to reflect the subjective intention or context logic of the sample object. For example, in the financial scenario, the text feature channels can include but are not limited to: remark channel, and the corresponding text feature information in the remark channel can include transaction remark information; scene description channel, and the corresponding text feature information in the scene description channel can include at least one of the following: business compliance rules corresponding to the transaction, such as quota limit, verification requirement, permission range, etc.; inherent attribute characteristics of transaction parties, such as credit rating, historical behavior label, etc.; business intention description of the transaction, such as consumption, investment, repayment, order payment, order refund, etc.

[0064] Step 2, using the behavior feature information of each service in P behavior feature channels, constructing the initial behavior data of each service. For example, the 3 behavior feature information of a transaction is "time: xxxx-05-01 03:15, location: region 1, transaction resource amount: 1000", and after integration, the initial behavior data of the transaction is formed [xxxx-05-01 03:15, region 1, 1000]. The initial behavior data of each service is cleaned and filled to obtain the service behavior data of each service. The initial behavior data is the original record of the behavior feature, but there may be missing values, outliers or redundant information (such as repeated timestamps, incorrect device identifiers), which need to be cleaned and filled to obtain high-quality service behavior data. For example, the transaction resource amount of a transaction is marked as "10000000000" (far beyond the normal range), which is determined as an outlier, and can be filled with the mean value of the transaction resource amount. For example, the "location" information of a transaction is missing, which can be filled by combining the device address to determine the belonging region. For the missing behavior feature information in the initial behavior data, a placeholder can be selected to replace it to ensure the integrity of the modeling of the service behavior data.

[0065] Step 3, using the text feature information of each service in Q text feature channels, constructing the service description text data of each service. For example, the 2 text feature information of a transaction is "remark: short-term loan turnover, counterparty account name: XX trade", and after integration, the service description text data of the transaction is formed: "short-term loan turnover, counterparty account is 'XX trade'".

[0066] Step 4, obtaining the labeled labels of M services after expert evaluation. That is, the labeled labels are given by the business domain experts (such as financial risk control experts, e-commerce compliance experts) based on the M service data related to the sample object, and the real type of the sample object is given after comprehensive evaluation.

[0067] Step 5, based on the service behavior data and service description text data of M services, and the label labels, constructing sample data for training the recognition model.

[0068] As can be seen, the present application converts the original business record into data containing behavior-text-type that can be learned by the model through the process of "continuous business selection → multi-dimensional feature extraction → data cleaning and filling → text structuring → label labeling".

[0069] S42, calling the recognition model to perform multi-channel time sequence coding processing on the service behavior data in the M group of service data, obtaining time sequence behavior embedding features, and the time sequence behavior embedding features are used to represent the channel original characteristics of the service behavior data in the M group of service data in time.

[0070] See Figure 5is a structural schematic diagram of an identification model provided by an embodiment of the present application, and the identification model comprises a time series data dynamic quantification module, an encoding module in a pre-trained language model, and a classification module; the encoding module in the pre-trained language model comprises an embedding layer (including a word embedding layer and a position embedding layer) and an encoding layer; and the classification module comprises a nonlinear mapping layer, a normalization layer, and a classification layer. The following will be described in combination with Figure 5 The steps S42-S45 are described.

[0071] The multi-channel time series encoding processing of the service behavior data in the M groups of service data in the present application is a key link for converting the structured service behavior features into time series semantic features understandable by the model. This process generates embedding features (i.e., time series behavior embedding features) that can accurately represent the time series rules of the service behavior by preserving the time order of the service behavior and the original characteristics of the channel, thereby providing a basis for the subsequent identification model to judge the type of the sample object at the behavior sequence level.

[0072] In one embodiment, the identification model is called to perform multi-channel time series encoding processing on the service behavior data in the M groups of service data, to obtain time series behavior embedding features, including: Step 1. The service behavior data in the M groups of service data is arranged in the order of execution of the respective corresponding services, to obtain a time series service behavior sequence.

[0073] For example, the sample object has three transactions: the service behavior data of transaction 1 is [xxxx-10-01 09:00, 2000, device identifier 1], the service behavior data of transaction 2 is [xxxx-10-01 23:59, 1000, device identifier 1], and the service behavior data of transaction 3 is [xxxx-10-02 00:05, 500, device identifier 1], and the time series service behavior sequence is It can be seen that the size of the time series service behavior sequence is [M, P].

[0074] Step 2. The time series data dynamic quantification module in the identification model is called to perform data scaling and quantification processing on the time series service behavior sequence, to obtain a time series behavior quantization sequence.

[0075] The data scaling and quantization processing can be used to solve the problem of inconsistent feature scales of different channels and the difficulty of directly encoding continuous variables. In an implementation, a time series data dynamic quantization module in the recognition model can be called to perform data scaling processing on the continuous variable elements in the time series business behavior sequence according to a dispersion standardization manner, to obtain a time series behavior scaling sequence. The continuous variable element refers to a variable that can take any value within a certain interval, such as a transaction resource amount. The calculation formula of the dispersion standardization manner is: scaled value = (original value - minimum value) / (maximum value - minimum value). For example, the maximum transaction resource amount and the minimum transaction resource amount in M groups of businesses are calculated, and then each transaction resource amount in the time series business behavior sequence is processed by data scaling according to the dispersion standardization manner. Through data scaling, the dimensional differences of different channels can be eliminated, and the weights of the features of each channel can be balanced in subsequent processing, so as to avoid the bias caused by large value features dominating the model training.

[0076] The time series data dynamic quantization module in the recognition model is called to perform data quantization processing on the continuous variable elements in the time series behavior scaling sequence according to a binning index manner, to obtain a time series behavior quantization sequence. The scaled continuous variable element is still a floating point type, which needs to be converted into a discrete integer index through binning index, so as to facilitate subsequent embedding layer processing. The core logic of the binning index is to divide the value interval of the scaled continuous variable element into a plurality of bins, each bin corresponds to a unique integer index, and the scaled continuous variable element is mapped to the index of the bin to which it belongs. For example, the calculation logic of the binning index manner is shown in the following formula (2): (2) wherein x represents the scaled continuous variable element, represents the index of the bin to which the scaled continuous variable element is mapped.

[0077] For example, the binning index manner includes at least one of the following: an equal-width binning index manner, a clustering binning index manner, and a dynamic binning index manner. The equal-width binning index manner refers to dividing the bins at a fixed interval (such as every 10 bins), which is suitable for scenarios where the feature distribution is relatively uniform. The clustering binning index manner refers to dividing the data into clusters (bins) with similar densities through a clustering algorithm, and the bin boundary is determined by the cluster center, which is suitable for scenarios where the data distribution is uneven. The dynamic binning index manner refers to dynamically adjusting the bin width according to the data distribution (such as a smaller bin width for a high-frequency transaction interval and a larger bin width for a low-frequency interval), which is suitable for scenarios where the time series fluctuates greatly.

[0078] The present application discretizes the continuous variable, which not only preserves the numerical size relationship in the time series behavior quantization sequence, but also reduces the processing complexity of the subsequent embedding layer.

[0079] In addition, the time sequence behavior quantization sequence can also include a category attribute element (such as a device operating system type), and the category attribute element usually has multiple different values. One value of the category attribute element can be represented by a symbol in the time sequence behavior quantization sequence (for example, 1 represents a first operating system type, 2 represents a second operating system type, and so on).

[0080] In step 3, an embedding layer in the identification model is called to perform multi-channel embedding splicing processing on the time sequence behavior quantization sequence, to obtain time sequence behavior embedding features.

[0081] The multi-channel embedding splicing processing is used to convert the dynamic quantized time sequence behavior data (i.e., the time sequence behavior quantization sequence) into semantic time sequence features that can be efficiently processed by the model. The time sequence behavior quantization sequence includes behavior quantization data corresponding to each business behavior data, and each behavior quantization data includes at least one element (such as a value obtained by performing data scaling quantization processing on behavior feature information).

[0082] In one implementation method, a word embedding layer in the recognition model can be called to perform word embedding representation processing on each element in the time-series behavior quantization sequence to obtain a word embedding feature of each element in the time-series behavior quantization sequence. The word embedding layer is a learnable embedding layer and contains a large number of word embedding features. Therefore, the word embedding feature representing each element can be found in the word embedding layer. A position embedding layer in the recognition model can be called to perform position embedding representation processing on each element in the time-series behavior quantization sequence to obtain a position embedding feature of each element in the time-series behavior quantization sequence. The position embedding layer is also a learnable embedding layer and contains a large number of position embedding features. Therefore, the position embedding feature representing the position of each element can be found in the position embedding layer. The word embedding feature and the position embedding feature of each element in the time-series behavior quantization sequence are fused (such as splicing fusion, weighted summation fusion, attention fusion, etc.) to obtain a target embedding feature of each element in the time-series behavior quantization sequence. Based on the sequence of each behavior quantization data in the time-series behavior quantization sequence and the sequence of each element in each behavior quantization data, the target embedding features of all elements in the time-series behavior quantization sequence are spliced to obtain a time-series behavior embedding feature. For example, the time-series behavior embedding feature can be [target embedding feature corresponding to the transaction time of the first transaction, target embedding feature corresponding to the transaction resource amount of the first transaction, target embedding feature corresponding to the transaction location of the first transaction, target embedding feature corresponding to the transaction time of the second transaction, target embedding feature corresponding to the transaction resource amount of the second transaction, target embedding feature corresponding to the transaction location of the second transaction]. In this way, on the one hand, the business execution sequence and the internal sequence of the channel are retained, so that the model can capture the correlation rules across businesses and channels; on the other hand, the target embedding feature can retain the original characteristics of the behavior feature information channel by means of the domain knowledge in the pre-trained language model, that is, the original business meaning of the behavior feature information is retained.

[0083] S43, calling the recognition model to perform knowledge coding processing on the business description text data in the M groups of business data to obtain business text knowledge features, and the business text knowledge features are used to represent the business logic association between the business description text data in the M groups of business data.

[0084] In the recognition model, the knowledge coding processing on the business description text data in the M groups of business data is a key link for converting unstructured text information into semantic features that can be understood by the model. This process captures the business logic association between texts to generate features (i.e., business text knowledge features) that can accurately represent the semantics and logical relationship of the texts, providing a judgment basis at the text level for the recognition model.

[0085] In an implementation, the service description text data in the M groups of service data can be arranged in the order of execution of the respective services to obtain a time sequence service text sequence. The execution logic of the arrangement processing is consistent with the execution logic of the arrangement processing of the service behavior data in the M groups of service data, which can ensure one-to-one correspondence of the behavior and the text in the time dimension, and ensure that the subsequent processing can capture the time sequence logic association.

[0086] The embedding layer (including the word embedding layer and the position embedding layer) in the recognition model is called to perform embedding representation processing on the time sequence service text sequence to obtain time sequence text embedding features. The embedding representation processing includes position embedding representation processing and word embedding representation processing. The word embedding representation processing refers to calling the word embedding layer in the recognition model to map each word in the time sequence service text sequence to a pre-trained word embedding feature (which retains the meaning of the business terminology) to capture the semantics of the word. The position embedding representation processing refers to calling the position embedding layer in the recognition model to fuse the position embedding features for the word embedding features of each word, so that the time sequence text embedding features retain the semantics of the words and the time sequence positions, laying a foundation for subsequent capture of the logical association. The encoding layer in the recognition model is called to perform attention encoding processing on the time sequence text embedding features to obtain business text knowledge features. The core of the encoding layer is the self-attention mechanism, which can mine the time sequence dependencies, causal associations, semantic complementarities and other business logics among the texts, and integrate the dispersed text information into business text knowledge features rich in business logical associations.

[0087] S44, the recognition model is called to perform abnormality recognition processing based on the time sequence behavior embedding features and the business text knowledge features to obtain an abnormality recognition result, which is used to indicate the predicted type of the sample object.

[0088] The abnormality recognition processing is a link that fuses the time sequence rules of the business behavior data (i.e., the time sequence behavior embedding features) and the semantic logic of the business description text data (i.e., the business text knowledge features) to finally determine the type of the sample object (normal type or abnormal type). This process converts the double features into an explicit prediction result through the three-step processing of nonlinear mapping, normalization and classification by the classification module, solves the problem of one-sidedness of single feature judgment, and ensures the accuracy of abnormality recognition.

[0089] The recognition model includes a classification module, which includes a nonlinear mapping layer, a normalization layer and a classification layer. In an implementation, the time sequence behavior embedding features and the business text knowledge features can be spliced to obtain behavior text spliced features. For example, the time sequence behavior embedding features are [f1, f2,..., f M×P ], the business text knowledge features are f Q , and the behavior text spliced features are F=[f1, f2,..., f M×P , fQ The nonlinear mapping layer in the recognition model is called to perform nonlinear mapping processing on the behavior text splicing feature, to obtain nonlinear mapping features. The behavior text splicing feature may have problems of linear inseparability or distribution deviation (such as the scale difference between the time sequence behavior embedding feature and the business text knowledge feature), and the nonlinear mapping layer can transform the behavior text splicing feature and map it to a standard normal distribution space, to enhance the model's ability to capture nonlinear relationships.

[0090] The normalization layer in the recognition model is called to perform normalization processing on the nonlinear mapping features, to obtain normalized features. The nonlinear mapping features may have problems of excessively large variance or unstable distribution (such as differences in mean and variance of different batches of data), which can cause difficulties in model training. The normalization layer adjusts the feature distribution to a stable distribution with a mean of 0 and a variance of 1 through standardization processing (such as Layer Normalization), to accelerate model convergence and improve generalization ability. The classification layer in the recognition model is called to perform classification processing on the normalized features, to obtain an abnormality recognition result. The normalized features already have a stable distribution and rich semantic information, and the classification layer can map them to a probability distribution of abnormal types (i.e., the probability of belonging to an abnormal type) through a fully connected layer + sigmoid function. The probability of the sample object belonging to an abnormal type can be determined as the abnormality recognition result. When the probability of the sample object belonging to an abnormal type is greater than a preset probability threshold (which can be set as needed), the abnormality recognition result can be determined to indicate that the predicted type of the sample object is an abnormal type. When the probability of the sample object belonging to an abnormal type is less than or equal to the preset probability threshold (which can be set as needed), the abnormality recognition result can be determined to indicate that the predicted type of the sample object is a normal type.

[0091] For example, the GELU (Gaussian Error Linear Unit) function can be used to implement nonlinear mapping processing. The GELU function is a nonlinear activation function widely used in deep learning models, which combines the probability characteristics of the normal distribution and linear transformation, and is more consistent with the distribution law of real data, thereby improving the model's ability to model complex patterns. The GELU function is shown in the following formula (3): (3) where x represents an input value, is the cumulative distribution function of the standard normal distribution.

[0092] S45, based on the difference between the abnormality recognition result and the labeled label, the recognition model is trained, and the trained recognition model is used to identify the object type of the business object.

[0093] The smaller the difference between the difference between the abnormal identification result and the labeled label, the more accurate the model prediction is. The greater the difference between the difference between the abnormal identification result and the labeled label, the less accurate the model prediction is. In an embodiment, the object type identification can be a binary classification task of identifying whether the object type is a normal type or an abnormal type, and training the identification model based on the difference between the abnormal identification result and the labeled label can include: obtaining a binary classification loss function, inputting the abnormal identification result and the labeled label into the binary classification loss function to obtain a first loss parameter corresponding to the identification model. The binary classification loss function can be a commonly used cross-entropy loss function. The first loss parameter is used to reflect the difference between the abnormal identification result and the labeled label; the greater the first loss parameter, the greater the difference between the difference between the abnormal identification result and the labeled label; the smaller the first loss parameter, the smaller the difference between the difference between the abnormal identification result and the labeled label. The first loss parameter is used to adjust the model parameters of the normalization layer and the position embedding layer in the identification model to train the identification model. For example, when training the identification model, a back propagation algorithm (such as a stochastic gradient descent (SGD) algorithm) can be used to adjust the model parameters of the normalization layer and the position embedding layer in the identification model in the direction of reducing the first loss parameter, so that the first loss parameter calculated by the identification model after each adjustment is less than the first loss parameter calculated by the identification model before the adjustment. When the number of adjustments of the normalization layer and the position embedding layer in the identification model reaches a first preset upper limit of the number of adjustments (which can be set as needed), or the first loss parameter is less than a first preset loss threshold (which can be set as needed), the identification model after the last adjustment can be determined as the trained identification model.

[0094] In the identification model, the encoding layer and the word embedding layer in the pre-trained language model have learned the business domain knowledge in the first business scenario, and if the model parameters of the encoding layer and the word embedding layer in the pre-trained language model are adjusted, the learned business semantic representation can be easily destroyed. Therefore, by adjusting the model parameters of the normalization layer and the position embedding layer, the object type identification task can be adapted, the distribution adaptability and sequence position perception ability of the model to the input data can be enhanced, and the accuracy of the object type identification can be improved.

[0095] In the actual training process of the identification model, iteration is usually performed in batches, that is, the identification model calculates the first loss parameter corresponding to each sample data in a batch, and then adjusts the model parameters of the identification model by summing all the calculated first loss parameters to obtain a first total loss parameter.

[0096] See Figure 6 is a structure of an identification model provided by an embodiment of the present applicationFigure Two The decoding module in the pre-trained language model is the core of converting the input encoding features into target sequences (such as natural language texts) through a self-attention mechanism and a feedforward network. In the present application, the decoding module can be extended for anomaly reason analysis, i.e., generating anomaly reason explanations in natural language form according to input features (i.e., behavior-text splicing features) that fuse behavior patterns and text semantics.

[0097] In another embodiment, the model parameters of the normalization layer and the position embedding layer in the identification model are adjusted using the first loss parameter to train the identification model, including: when the label indicates that the sample object is of an abnormal type, a real abnormal reason text is obtained, which can be an abnormal reason description annotated by a business expert according to the sample data. The time series behavior embedding feature and the business text knowledge feature are spliced to obtain a behavior text splicing feature. The decoding module in the pre-trained language model is called to perform abnormal reason analysis and processing based on the behavior text splicing feature to obtain a predicted abnormal reason text. The time series behavior embedding feature provides a quantitative representation of the behavior trajectory; the business text knowledge feature provides a semantic representation of the subjective intention. After splicing, the decoding module can generate a more comprehensive abnormal reason by utilizing the behavior rule and the semantic logic. Based on the difference between the real abnormal reason text and the predicted abnormal reason text, the second loss parameter corresponding to the identification model is determined. The greater the difference between the real abnormal reason text and the predicted abnormal reason text, the greater the second loss parameter, indicating that the model prediction is less accurate; the smaller the difference between the real abnormal reason text and the predicted abnormal reason text, the smaller the second loss parameter, indicating that the model prediction is more accurate. For example, the probability that the content of each position in the predicted abnormal reason text is the content of the corresponding position in the real abnormal reason text can be obtained, and the negative of the sum of the logarithms of all position probabilities is determined as the second loss parameter, which is consistent with the calculation logic of formula (1) above. The normalization layer, the position embedding layer, and the decoding module in the identification model are adjusted using the first loss parameter and the second loss parameter to train the identification model. In an exemplary implementation, the normalization layer and the position embedding layer in the identification model can be adjusted a first number of times in the direction of decreasing the first loss parameter, and then the decoding module in the identification model can be adjusted a second number of times in the direction of decreasing the second loss parameter. The first number can be the number of adjustments when the first loss parameter is less than a first preset loss threshold (which can be set as needed) or a first preset upper limit of the number of adjustments. The second number can be the number of adjustments when the second loss parameter is less than a second preset loss threshold (which can be set as needed) or a second preset upper limit of the number of adjustments (which can be set as needed), so as to ensure the accuracy of the object type identification of the model first, and then ensure the accuracy of the abnormal reason analysis of the model. As can be seen, the training of the identification model can increase the supervision signal of the abnormal reason analysis on the original basis, and the model parameters are optimized by the double loss (identification error + reason generation error) to make the model not only able to judge the object type, but also to accurately explain the abnormal reason.

[0098] The trained identification model is used to identify the object type of the business object, and the detailed implementation process is described in the embodiment corresponding to Figure 7 .

[0099] Therefore, this application acquires sample data and constructs a recognition model that includes a dynamic quantization module for time-series data, an embedding and encoding layer of a pre-trained language model, a classification module, and a decoding module. It then processes business behavior data through multi-channel time-series encoding to obtain time-series behavior embedding features, processes business description text data through knowledge encoding to obtain business text knowledge features, fuses the dual features, and outputs anomaly recognition results through a classification module. Finally, based on the difference between the anomaly recognition results and the labeled tags, and combining the difference between the actual and predicted anomaly causes, a dual loss is constructed. The model parameters are then specifically optimized to complete training, achieving a closed-loop process from data to model. This not only accurately identifies the anomaly types of business objects but also generates anomaly causes that conform to business logic. This improves recognition accuracy and generalization ability, enhances the interpretability of results, effectively supports business risk assessment and handling decisions, reduces labor costs, and adapts to the needs of multi-scenario business implementation.

[0100] Please see Figure 7 This is a flowchart illustrating a data processing method based on a recognition model provided in an embodiment of this application. Figure Three This data processing method based on the recognition model can be derived from... Figure 1 The server execution in the corresponding embodiment involves the application process of the trained recognition model in the data processing scheme. The data processing method based on the recognition model includes, but is not limited to, steps S71-S74: S71. Obtain N sets of target business data related to the business object. Each set of target business data comes from a business operation performed by the business object in the first business scenario. Each set of target business data includes target business behavior data and target business description text data; N is a positive integer.

[0101] In one embodiment, acquiring N sets of target business data related to a business object includes: determining N consecutively executed business operations of the target object; acquiring behavioral feature information for each business operation under P behavioral feature channels and text feature information under Q text feature channels; where P and Q are both positive integers. Using the behavioral feature information for each business operation under the P behavioral feature channels, initial behavioral data for each business operation is constructed; this initial behavioral data is then cleaned and filled to obtain target business behavioral data for each business operation. Using the text feature information for each business operation under the Q text feature channels, target business description text data for each business operation is constructed. Finally, using the target business behavioral data and target business description text data of the N business operations, N sets of target business data related to the business object are constructed. The detailed implementation process is the same as described above. Figure 4 The execution logic described in S41 of the corresponding embodiment is the same, and will not be repeated here.

[0102] S72, calling the trained recognition model to perform multi-channel time sequence coding processing on the target business behavior data in the N groups of target business data, to obtain target time sequence behavior embedding features, the target time sequence behavior embedding features being used to represent the channel original characteristics of the target business behavior data in the N groups of target business data in time.

[0103] In one embodiment, the trained recognition model comprises a time sequence data dynamic quantization module and an embedding layer. Calling the trained recognition model to perform multi-channel time sequence coding processing on the target business behavior data in the N groups of target business data, to obtain target time sequence behavior embedding features, comprises: arranging the target business behavior data in the N groups of target business data according to the execution order of the respective corresponding businesses of the N groups of target business data, to obtain a target time sequence business behavior sequence. Calling the time sequence data dynamic quantization module in the trained recognition model to perform data scaling and quantization processing on the target time sequence business behavior sequence, to obtain a target time sequence behavior quantization sequence. Calling the embedding layer in the trained recognition model to perform multi-channel embedding splicing processing on the target time sequence behavior quantization sequence, to obtain the target time sequence behavior embedding features. The detailed implementation process is consistent with the foregoing Figure 4 The execution logic described in S42 in the corresponding embodiment is consistent, and will not be repeated here.

[0104] S73, calling the trained recognition model to perform knowledge coding processing on the target business description text data in the N groups of target business data, to obtain target business text knowledge features, the target business text knowledge features being used to represent the business logic association between the target business description text data in the N groups of target business data.

[0105] In one embodiment, the trained recognition model comprises an encoding module, which includes an embedding layer and an encoding layer. Calling the trained recognition model to perform knowledge coding processing on the target business description text data in the N groups of target business data, to obtain target business text knowledge features, comprises: arranging the target business description text data in the N groups of target business data according to the execution order of the respective corresponding businesses of the N groups of target business data, to obtain a target time sequence business text sequence; calling the embedding layer in the trained recognition model to perform embedding representation processing on the target time sequence business text sequence, to obtain target time sequence text embedding features. Calling the encoding layer in the trained recognition model to perform attention coding processing on the target time sequence text embedding features, to obtain the target business text knowledge features. The detailed implementation process is consistent with the foregoing Figure 4 The execution logic described in S43 in the corresponding embodiment is consistent, and will not be repeated here.

[0106] S74, calling the trained recognition model to perform abnormality recognition processing based on the target time sequence behavior embedding features and the target business text knowledge features, to obtain the object type of the business object.

[0107] In an embodiment, the trained recognition model comprises a classification module, and the classification module comprises a nonlinear mapping layer, a normalization layer, and a classification layer. The trained recognition model is invoked to perform abnormality recognition processing based on the target time-series behavior embedding feature and the target business text knowledge feature, to obtain the object type of the business object, including: performing splicing processing on the target time-series behavior embedding feature and the target business text knowledge feature to obtain a target behavior text splicing feature; invoking the nonlinear mapping layer in the trained recognition model to perform nonlinear mapping processing on the target behavior text splicing feature to obtain a target nonlinear mapping feature; invoking the normalization layer in the trained recognition model to perform normalization processing on the target nonlinear mapping feature to obtain a target normalized feature; and invoking the classification layer in the trained recognition model to perform classification processing on the target normalized feature to obtain the object type of the business object. The detailed implementation process is the same as that described above Figure 4 The execution logic described in S44 in the corresponding embodiment is consistent, and thus is not described here again.

[0108] In a feasible implementation manner, the trained recognition model comprises a decoding module. When the object type of the business object is an abnormal type, the target time-series behavior embedding feature and the target business text knowledge feature are spliced to obtain a target behavior text splicing feature; and the decoding module in the trained recognition model is invoked to perform abnormality reason analysis processing based on the target behavior text splicing feature to obtain a target abnormality reason text. The detailed implementation process is the same as that described above Figure 4 The execution logic described in S45 in the corresponding embodiment is consistent, and thus is not described here again.

[0109] As can be seen, in the present application, the trained recognition model can obtain a target time-series behavior embedding feature through multi-channel time-series encoding processing of target business behavior data, obtain a target business text knowledge feature through knowledge encoding processing of target business description text data, fuse the two features, and output the object type of a business object through a classification module and output abnormality reason analysis through a decoding module, to realize a full-process closed loop from data to model, not only accurately identify the abnormal type of a business object, but also generate abnormality reasons conforming to business logic, to not only improve recognition accuracy and generalization ability, but also enhance result interpretability, effectively support business risk judgment and disposal decision, reduce manual cost, and adapt to multi-scenario business landing needs.

[0110] Please refer to Figure 8 is a structural schematic diagram of a data processing apparatus based on a recognition model provided by an embodiment of the present application. The data processing apparatus based on the recognition model can be a computer program (including program code) running in the computer device in the foregoing. The data processing apparatus based on the recognition model can be used to execute corresponding steps in the method provided by the embodiments of the present application.

[0111] In an embodiment, the data processing apparatus 80 based on the recognition model can comprise: an acquisition unit 81 configured to acquire sample data for training the recognition model, the sample data comprising M groups of service data related to a sample object and a labeled label of the sample object, wherein each group of service data is derived from an item of service performed by the sample object in a first service scenario, and each group of service data comprises service behavior data and service description text data, and the labeled label is used to indicate a labeling type of the sample object, and M is a positive integer; a processing unit 82 configured to invoke the recognition model to perform multi-channel time series encoding processing on the service behavior data in the M groups of service data to obtain time series behavior embedding features, wherein the time series behavior embedding features are used to represent channel original characteristics of the service behavior data in the M groups of service data in time; The processing unit 82 is further configured to invoke the recognition model to perform knowledge encoding processing on the service description text data in the M groups of service data to obtain service text knowledge features, wherein the service text knowledge features are used to represent business logic correlation between the service description text data in the M groups of service data; The processing unit 82 is further configured to invoke the recognition model to perform abnormality recognition processing based on the time series behavior embedding features and the service text knowledge features to obtain an abnormality recognition result, wherein the abnormality recognition result is used to indicate a predicted type of the sample object; The processing unit 82 is further configured to train the recognition model based on a difference between the abnormality recognition result and the labeled label, and the trained recognition model is used to identify an object type of a service object.

[0112] In an embodiment, the recognition model comprises a time series data dynamic quantization module and an embedding layer in a pre-trained language model. The processing unit 82 invokes the recognition model to perform multi-channel time series encoding processing on the service behavior data in the M groups of service data to obtain time series behavior embedding features, which comprises: arranging the service behavior data in the M groups of service data according to an execution sequence of the respective services of the M groups of service data to obtain a time series service behavior sequence; invoking the time series data dynamic quantization module in the recognition model to perform data scaling and quantization processing on the time series service behavior sequence to obtain a time series behavior quantization sequence; invoking the embedding layer in the recognition model to perform multi-channel embedding and splicing processing on the time series behavior quantization sequence to obtain the time series behavior embedding features.

[0113] In an embodiment, the processing unit 82 invokes the time series data dynamic quantization module in the recognition model to perform data scaling and quantization processing on the time series service behavior sequence to obtain a time series behavior quantization sequence, which comprises: The time series data dynamic quantification module in the recognition model is called to perform data scaling processing on the continuous variable elements in the time series business behavior sequence according to a dispersion standardization manner, to obtain a time series behavior scaling sequence. The time series data dynamic quantification module in the recognition model is called to perform data quantification processing on the continuous variable elements in the time series behavior scaling sequence according to a binning index manner, to obtain a time series behavior quantization sequence. The binning index manner includes at least one of the following: an equal-width binning index manner, a clustering binning index manner, and a dynamic binning index manner.

[0114] In an embodiment, the embedding layer includes a word embedding layer and a position embedding layer; the time series behavior quantization sequence includes behavior quantization data corresponding to each business behavior data, and each behavior quantization data includes at least one element. The processing unit 82 calls the embedding layer in the recognition model to perform multi-channel embedding splicing processing on the time series behavior quantization sequence, to obtain time series behavior embedding features, including: The word embedding layer in the recognition model is called to perform word embedding representation processing on each element in the time series behavior quantization sequence, to obtain word embedding features of each element in the time series behavior quantization sequence. The position embedding layer in the recognition model is called to perform position embedding representation processing on each element in the time series behavior quantization sequence, to obtain position embedding features of each element in the time series behavior quantization sequence. The word embedding features and the position embedding features of each element in the time series behavior quantization sequence are fused to obtain target embedding features of each element in the time series behavior quantization sequence. The target embedding features of all elements in the time series behavior quantization sequence are spliced based on the order of each behavior quantization data in the time series behavior quantization sequence and the order of each element in each behavior quantization data, to obtain time series behavior embedding features.

[0115] In an embodiment, the recognition model includes an encoding module in a pre-trained language model; the encoding module includes an embedding layer and an encoding layer. The processing unit 82 calls the recognition model to perform knowledge encoding processing on the business description text data in the M groups of business data, to obtain business text knowledge features, including: The business description text data in the M groups of business data is arranged according to the execution order of the respective businesses of the M groups of business data, to obtain a time series business text sequence. The embedding layer in the recognition model is called to perform embedding representation processing on the time series business text sequence, to obtain time series text embedding features. The encoding layer in the recognition model is called to perform attention encoding processing on the time series text embedding features, to obtain business text knowledge features.

[0116] In one embodiment, the recognition model includes: a classification module, which includes a nonlinear mapping layer, a normalization layer and a classification layer, wherein the nonlinear mapping layer includes a cumulative distribution function of a standard normal distribution; Processing unit 82 invokes the recognition model to perform anomaly recognition processing based on temporal behavior embedding features and business text knowledge features, and obtains anomaly recognition results, including: The temporal behavior embedding features and business text knowledge features are concatenated to obtain behavior text concatenation features; The non-linear mapping layer in the recognition model is invoked to perform non-linear mapping processing on the behavior text splicing features to obtain non-linear mapping features; The normalization layer in the recognition model is invoked to normalize the nonlinear mapping features, resulting in normalized features. The classification layer in the recognition model is invoked to classify the normalized features, thereby obtaining the anomaly recognition result.

[0117] In one embodiment, the recognition model includes: a normalization layer and a position embedding layer from a pre-trained language model; Processing unit 82 trains the recognition model based on the difference between the anomaly recognition results and the labeled data, including: Obtain the binary classification loss function, input the anomaly identification result and the label into the binary classification loss function, and obtain the first loss parameter corresponding to the identification model; the first loss parameter is used to reflect the difference between the anomaly identification result and the label. The model parameters of the normalization layer and the location embedding layer in the recognition model are adjusted using the first loss parameter in order to train the recognition model.

[0118] In one embodiment, the recognition model further includes: a decoding module in a pre-trained language model; Processing unit 82 uses the first loss parameter to adjust the model parameters of the normalization layer and the location embedding layer in the recognition model to train the recognition model, including: When the label indicates that the sample object is an anomaly, retrieve the text of the actual anomaly cause. The temporal behavior embedding features and business text knowledge features are concatenated to obtain behavior text concatenation features; The decoding module in the pre-trained language model is called to perform anomaly cause analysis based on behavioral text concatenation features, and the predicted anomaly cause text is obtained. Based on the difference between the actual anomaly cause text and the predicted anomaly cause text, the second loss parameter corresponding to the recognition model is determined. The model parameters of the normalization layer, the location embedding layer, and the decoding module in the recognition model are adjusted using the first loss parameter and the second loss parameter to train the recognition model.

[0119] In one embodiment, the obtaining unit 81 is further configured to: Acquire a general language model and business corpus text data in the first business scenario; the general language model includes an encoding module and a decoding module; Processing unit 82 is also used for: The business corpus text data is masked to obtain masked corpus text data; The encoding module in the general language model is called to encode the masked corpus text data to obtain the masked text encoding features. The decoding module in the general language model is called to perform mask prediction processing based on the mask text encoding features to obtain the predicted mask content in the business corpus text data. Based on the difference between the predicted mask content and the actual mask content in the business corpus text data, a general language model is trained to obtain a pre-trained language model.

[0120] In one embodiment, the acquiring unit 81 acquires sample data for training the recognition model, including: Determine the M consecutive business operations executed by the sample object, and obtain the behavioral feature information of each business operation under P behavioral feature channels and the text feature information under Q text feature channels; P and Q are both positive integers; Using the behavioral feature information of each business under P behavioral feature channels, we construct the initial behavioral data of each business, clean and fill the initial behavioral data of each business to obtain the business behavioral data of each business. By utilizing the text feature information of each business under Q text feature channels, construct the business description text data for each business; Obtain the labels and annotations for M business items after expert evaluation and processing; The data consists of business behavior data and business description text data based on M business items, as well as sample data constructed from tags to train the recognition model.

[0121] In another embodiment, the data processing device 80 based on the recognition model may include: Acquisition unit 81 is used to acquire N sets of target business data related to the business object. Each set of target business data comes from a business operation performed by the business object in the first business scenario. Each set of target business data includes target business behavior data and target business description text data; N is a positive integer. Processing unit 82 is used to call the trained recognition model to perform multi-channel temporal coding processing on the target business behavior data in N sets of target business data to obtain target temporal behavior embedding features. The target temporal behavior embedding features are used to characterize the original channel characteristics of the target business behavior data in N sets of target business data in time. Processing unit 82 is also used to call the trained recognition model to perform knowledge encoding processing on the target business description text data in N sets of target business data to obtain target business text knowledge features. The target business text knowledge features are used to characterize the business logic relationship between the target business description text data in N sets of target business data. The processing unit 82 is also used to call the trained recognition model to perform anomaly recognition processing based on the target temporal behavior embedding features and the target business text knowledge features, so as to obtain the object type of the business object.

[0122] In one embodiment, the trained recognition model includes: a time-series data dynamic quantization module and an embedding layer; Processing unit 82 calls the trained recognition model to perform multi-channel temporal encoding on the target business behavior data in N sets of target business data to obtain target temporal behavior embedding features, including: According to the execution order of the corresponding business in each of the N sets of target business data, the target business behavior data in the N sets of target business data are arranged and processed to obtain the target time-series business behavior sequence. The time-series data dynamic quantization module in the trained recognition model is invoked to perform data scaling and quantization processing on the target time-series business behavior sequence, resulting in the target time-series behavior quantization sequence. The embedding layer in the trained recognition model is invoked to perform multi-channel embedding and splicing processing on the target temporal behavior quantization sequence to obtain the target temporal behavior embedding features.

[0123] In one embodiment, the trained recognition model includes an encoding module, which comprises an embedding layer and an encoding layer; Processing unit 82 calls the trained recognition model to perform knowledge encoding processing on the target business description text data in N sets of target business data, to obtain the target business text knowledge features, including: According to the execution order of the corresponding business in each of the N sets of target business data, the target business description text data in the N sets of target business data are arranged to obtain the target time-series business text sequence. The embedding layer in the trained recognition model is invoked to perform embedding representation processing on the target time-series business text sequence, thereby obtaining the target time-series text embedding features; The encoding layer in the trained recognition model is invoked to perform attention encoding on the target temporal text embedding features, thereby obtaining the target business text knowledge features.

[0124] In one embodiment, the trained recognition model includes: a classification module, which includes a nonlinear mapping layer, a normalization layer, and a classification layer; Processing unit 82 invokes the trained recognition model to perform anomaly recognition processing based on the target temporal behavior embedding features and the target business text knowledge features, obtaining the object type of the business object, including: The target temporal behavior embedding features and target business text knowledge features are concatenated to obtain the target behavior text concatenation features; The non-linear mapping layer in the trained recognition model is invoked to perform non-linear mapping processing on the target behavior text splicing features to obtain the target non-linear mapping features. The normalization layer in the trained recognition model is invoked to normalize the nonlinear mapping features of the target, thus obtaining the normalized features of the target. The classification layer in the trained recognition model is invoked to classify the normalized features of the target, thereby obtaining the object type of the business object.

[0125] In one embodiment, the trained recognition model includes: a decoding module, and the processing unit 82 is further configured to: When the object type of the business object is an exception type, the target temporal behavior embedding feature and the target business text knowledge feature are concatenated to obtain the target behavior text concatenation feature. The decoding module in the trained recognition model is invoked to perform anomaly cause analysis based on the target behavior text splicing features, thereby obtaining the target anomaly cause text.

[0126] According to one embodiment of this application, Figure 8 The data processing device based on the recognition model shown can be constructed by combining each unit into one or more other units, or by further dividing one or more units into multiple functionally smaller units. This achieves the same operation without affecting the technical effect of the embodiments of this application. The above units are based on logical function division. In practical applications, the function of one unit can be implemented by multiple units, or the function of multiple units can be implemented by one unit. In other embodiments of this application, the data management device for the space adjustment object can also include other units. In practical applications, these functions can also be implemented with the assistance of other units, and can be implemented by multiple units working together. According to another embodiment of this application, a computer program (including program code) capable of performing the steps involved in the corresponding methods described above can be run on a general-purpose computing device, such as a computer, which includes processing elements and storage elements such as a central processing unit (CPU), random access storage medium (RAM), and read-only storage medium (ROM). Figure 8The data processing apparatus based on the recognition model shown herein, and the data processing method based on the recognition model for implementing the embodiments of this application, are described. The computer program may be recorded on, for example, a computer-readable recording medium, loaded onto the aforementioned computing device via the computer-readable recording medium, and run therein.

[0127] In this embodiment, sample data for training the recognition model can be obtained. The sample data includes M sets of business data related to the sample object and annotation labels for the sample object. One set of business data originates from a business activity performed by the sample object in a first business scenario. This set of business data includes business behavior data and business description text data. The annotation labels indicate the annotation type of the sample object, where M is a positive integer. The sample data is the raw material for model learning. Business behavior data provides dynamic behavior patterns, business description text data provides semantic logic, and annotation labels provide the correct answers. All three together support the model in learning "how to identify object types from data." The recognition model is invoked to perform multi-channel temporal encoding processing on the business behavior data in the M sets of business data to obtain temporal behavior embedding features. These features characterize the original channel characteristics of the business behavior data in the M sets of business data over time. This transforms the original behavior data into temporal pattern features that the model can understand, revealing the dynamic patterns behind the behavior (such as whether it deviates from historical normal patterns). The recognition model is then invoked to perform knowledge encoding processing on the business description text data in the M sets of business data to obtain business text knowledge features. These features characterize the business logic relationships between the business description text data in the M sets of business data. This transforms unstructured text into structured semantic features, capturing the underlying logical motivations of business behaviors (such as rationality and external driving factors). An anomaly detection model is then invoked, based on temporal behavioral embedding features and business text knowledge features, to obtain anomaly detection results. These results indicate the predicted type of the sample object. This approach integrates the temporal patterns of behavior with the semantic logic of the text, achieving accurate prediction of object types. Based on the differences between the anomaly detection results and the labeled tags, the detection model is trained. The trained model is then used to identify the object type of business objects. In other words, through a closed loop of data → model → prediction → feedback → optimization, the model gradually learns the ability to accurately determine object types from behavior and text. Therefore, this application, by simultaneously analyzing the temporal patterns of business behavior (multi-channel temporal encoding) and the semantic logic of text (knowledge encoding), enables the model to more comprehensively characterize object features. Combined with supervised feedback from labeled tags, the model continuously optimizes its ability to identify object types, ultimately achieving accurate classification of object types.

[0128] Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Please refer to [link / reference]. Figure 9The computer device 90 includes a processor 91, a communication interface 92, and a computer-readable storage medium 93. The processor 91, communication interface 92, and computer-readable storage medium 93 can be connected via a bus or other means. The communication interface 92 is used to receive and transmit data, and may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), controlled by the processor 91 for sending and receiving data. The computer-readable storage medium 93 can be stored in the memory of the computer device 90 (including the computer device 90's built-in memory, and may also include extended memory supported by the computer device 90). The computer-readable storage medium 93 is used to store computer programs, including program instructions, and the processor 91 is used to execute the program instructions stored in the computer-readable storage medium 93. The processor 91 (or CPU (Central Processing Unit)) is the computing and control core of the computer device 90. It can parse various instructions within the computer device 90 and process various data of the computer device 90. It is suitable for implementing one or more instructions, specifically for loading and executing one or more instructions to achieve the corresponding method flow or corresponding function.

[0129] This application embodiment also provides a computer-readable storage medium (Memory), which is a memory device in a computer device 90 for storing programs and data. It is understood that the computer-readable storage medium here may include the built-in storage medium in the computer device 90, or it may include extended storage media supported by the computer device 90. The computer-readable storage medium provides storage space that stores the processing system of the computer device 90. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by the processor 91. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here may be high-speed RAM memory, or non-volatile memory, such as at least one disk storage device; optionally, it may also be at least one computer-readable storage medium located remotely from the aforementioned processor.

[0130] In one embodiment, the computer-readable storage medium stores one or more instructions; the processor 91 loads and executes one or more instructions stored in the computer-readable storage medium to implement the corresponding steps in the above-described embodiment of the data processing method based on the recognition model.

[0131] In one implementation, one or more instructions in a computer-readable storage medium are loaded by processor 91 and executed as follows: Obtain sample data for training the recognition model. The sample data includes M sets of business data related to the sample object and the annotation labels of the sample object. One set of business data comes from a business performed by the sample object in the first business scenario. The set of business data includes business behavior data and business description text data. The annotation labels are used to indicate the annotation type of the sample object, and M is a positive integer. The recognition model is invoked to perform multi-channel temporal coding on the business behavior data in the M groups of business data to obtain temporal behavior embedding features. The temporal behavior embedding features are used to characterize the original channel characteristics of the business behavior data in the M groups of business data in terms of time. The recognition model is invoked to perform knowledge encoding processing on the business description text data in the M groups of business data to obtain business text knowledge features. The business text knowledge features are used to characterize the business logic relationship between the business description text data in the M groups of business data. The identification model is invoked to perform anomaly identification processing based on temporal behavior embedding features and business text knowledge features to obtain anomaly identification results, which are used to indicate the prediction type of sample objects. Based on the difference between the anomaly identification results and the labeled tags, the identification model is trained, and the trained identification model is used to identify the object type of business objects.

[0132] In one embodiment, the recognition model includes: a time-series data dynamic quantization module and an embedding layer in a pre-trained language model; Processor 91 invokes the recognition model to perform multi-channel time-series encoding on the business behavior data in the M groups of business data, obtaining time-series behavior embedding features, including: According to the execution order of the business corresponding to each of the M groups of business data, the business behavior data in the M groups of business data are arranged and processed to obtain the time-series business behavior sequence. The time-series data dynamic quantization module in the recognition model is invoked to perform data scaling and quantization processing on the time-series business behavior sequence, resulting in a time-series behavior quantization sequence; The embedding layer in the recognition model is invoked to perform multi-channel embedding and splicing processing on the time-series behavior quantization sequence to obtain the time-series behavior embedding features.

[0133] In one embodiment, the processor 91 invokes the time-series data dynamic quantization module in the recognition model to perform data scaling and quantization processing on the time-series business behavior sequence, obtaining a time-series behavior quantization sequence, including: The dynamic quantization module for time-series data in the recognition model is invoked to perform data scaling on the continuous variable elements in the time-series business behavior sequence according to the deviation standardization method, so as to obtain the time-series behavior scaling sequence. The dynamic quantization module for time-series data in the recognition model is invoked to perform data quantization processing on the continuous variable elements in the time-series behavior scaling sequence according to the binning index method, so as to obtain the time-series behavior quantization sequence. Among them, binning indexing methods include at least one of the following: equal-width binning indexing method, clustered binning indexing method, and dynamic binning indexing method.

[0134] In one embodiment, the embedding layer includes a word embedding layer and a location embedding layer; the temporal behavior quantization sequence contains behavior quantization data corresponding to each business behavior data, and each behavior quantization data contains at least one element; Processor 91 calls the embedding layer in the recognition model to perform multi-channel embedding and concatenation processing on the quantized sequence of temporal behavior, obtaining temporal behavior embedding features, including: The word embedding layer in the recognition model is invoked to perform word embedding representation processing on each element in the temporal behavior quantization sequence, thereby obtaining the word embedding features of each element in the temporal behavior quantization sequence; The location embedding layer in the recognition model is invoked to perform location embedding representation processing on each element in the temporal behavior quantization sequence, thereby obtaining the location embedding feature of each element in the temporal behavior quantization sequence; The word embedding features and position embedding features of each element in the temporal behavior quantization sequence are fused to obtain the target embedding features of each element in the temporal behavior quantization sequence. Based on the sequential order of each behavior quantization data in the time-series behavior quantization sequence and the sequential order of each element in each behavior quantization data, the target embedding features of all elements in the time-series behavior quantization sequence are concatenated to obtain the time-series behavior embedding features.

[0135] In one embodiment, the recognition model includes: an encoding module in a pre-trained language model; the encoding module includes an embedding layer and an encoding layer; Processor 91 invokes the recognition model to perform knowledge encoding processing on the business description text data in the M groups of business data, obtaining business text knowledge features, including: According to the execution order of the business corresponding to each of the M groups of business data, the business description text data in the M groups of business data are arranged to obtain the time-series business text sequence. The embedding layer in the recognition model is invoked to perform embedding representation processing on the time-series business text sequence to obtain the time-series text embedding features; The encoding layer in the recognition model is invoked to perform attention encoding on the temporal text embedding features to obtain business text knowledge features.

[0136] In one embodiment, the recognition model includes: a classification module, which includes a nonlinear mapping layer, a normalization layer and a classification layer, wherein the nonlinear mapping layer includes a cumulative distribution function of a standard normal distribution; Processor 91 invokes the recognition model to perform anomaly recognition processing based on temporal behavior embedding features and business text knowledge features, obtaining anomaly recognition results, including: The temporal behavior embedding features and business text knowledge features are concatenated to obtain behavior text concatenation features; The non-linear mapping layer in the recognition model is invoked to perform non-linear mapping processing on the behavior text splicing features to obtain non-linear mapping features; The normalization layer in the recognition model is invoked to normalize the nonlinear mapping features, resulting in normalized features. The classification layer in the recognition model is invoked to classify the normalized features, thereby obtaining the anomaly recognition result.

[0137] In one embodiment, the recognition model includes: a normalization layer and a position embedding layer from a pre-trained language model; Processor 91 trains the recognition model based on the difference between the anomaly detection results and the labeled information, including: Obtain the binary classification loss function, input the anomaly identification result and the label into the binary classification loss function, and obtain the first loss parameter corresponding to the identification model; the first loss parameter is used to reflect the difference between the anomaly identification result and the label. The model parameters of the normalization layer and the location embedding layer in the recognition model are adjusted using the first loss parameter in order to train the recognition model.

[0138] In one embodiment, the recognition model further includes: a decoding module in a pre-trained language model; Processor 91 uses the first loss parameter to adjust the model parameters of the normalization layer and the location embedding layer in the recognition model to train the recognition model, including: When the label indicates that the sample object is an anomaly, retrieve the text of the actual anomaly cause. The temporal behavior embedding features and business text knowledge features are concatenated to obtain behavior text concatenation features; The decoding module in the pre-trained language model is called to perform anomaly cause analysis based on behavioral text concatenation features, and the predicted anomaly cause text is obtained. Based on the difference between the actual anomaly cause text and the predicted anomaly cause text, the second loss parameter corresponding to the recognition model is determined. The model parameters of the normalization layer, the location embedding layer, and the decoding module in the recognition model are adjusted using the first loss parameter and the second loss parameter to train the recognition model.

[0139] In one embodiment, processor 91 is further configured to: Acquire a general language model and business corpus text data in the first business scenario; the general language model includes an encoding module and a decoding module; The business corpus text data is masked to obtain masked corpus text data; The encoding module in the general language model is called to encode the masked corpus text data to obtain the masked text encoding features. The decoding module in the general language model is called to perform mask prediction processing based on the mask text encoding features to obtain the predicted mask content in the business corpus text data. Based on the difference between the predicted mask content and the actual mask content in the business corpus text data, a general language model is trained to obtain a pre-trained language model.

[0140] In one embodiment, the processor 91 acquires sample data for training the recognition model, including: Determine the M consecutive business operations executed by the sample object, and obtain the behavioral feature information of each business operation under P behavioral feature channels and the text feature information under Q text feature channels; P and Q are both positive integers; Using the behavioral feature information of each business under P behavioral feature channels, we construct the initial behavioral data of each business, clean and fill the initial behavioral data of each business to obtain the business behavioral data of each business. By utilizing the text feature information of each business under Q text feature channels, construct the business description text data for each business; Obtain the labels and annotations for M business items after expert evaluation and processing; The data consists of business behavior data and business description text data based on M business items, as well as sample data constructed from tags to train the recognition model.

[0141] In another implementation, one or more instructions in the computer-readable storage medium are loaded by processor 91 and executed as follows: Obtain N sets of target business data related to the business object. Each set of target business data comes from a business operation performed by the business object in the first business scenario. Each set of target business data includes target business behavior data and target business description text data; N is a positive integer. The trained recognition model is invoked to perform multi-channel temporal coding on the target business behavior data in N sets of target business data to obtain target temporal behavior embedding features. The target temporal behavior embedding features are used to characterize the original channel characteristics of the target business behavior data in N sets of target business data in time. The trained recognition model is invoked to perform knowledge encoding on the target business description text data in N sets of target business data to obtain target business text knowledge features. The target business text knowledge features are used to characterize the business logic relationship between the target business description text data in N sets of target business data. The trained recognition model is invoked to perform anomaly recognition processing based on the target temporal behavior embedding features and the target business text knowledge features, thereby obtaining the object type of the business object.

[0142] In one embodiment, the trained recognition model includes: a time-series data dynamic quantization module and an embedding layer; Processor 91 calls the trained recognition model to perform multi-channel temporal encoding on the target business behavior data in N sets of target business data to obtain target temporal behavior embedding features, including: According to the execution order of the corresponding business in each of the N sets of target business data, the target business behavior data in the N sets of target business data are arranged and processed to obtain the target time-series business behavior sequence. The time-series data dynamic quantization module in the trained recognition model is invoked to perform data scaling and quantization processing on the target time-series business behavior sequence, resulting in the target time-series behavior quantization sequence. The embedding layer in the trained recognition model is invoked to perform multi-channel embedding and splicing processing on the target temporal behavior quantization sequence to obtain the target temporal behavior embedding features.

[0143] In one embodiment, the trained recognition model includes an encoding module, which comprises an embedding layer and an encoding layer; Processor 91 calls the trained recognition model to perform knowledge encoding processing on the target business description text data in N sets of target business data, obtaining target business text knowledge features, including: According to the execution order of the corresponding business in each of the N sets of target business data, the target business description text data in the N sets of target business data are arranged to obtain the target time-series business text sequence. The embedding layer in the trained recognition model is invoked to perform embedding representation processing on the target time-series business text sequence, thereby obtaining the target time-series text embedding features; The encoding layer in the trained recognition model is invoked to perform attention encoding on the target temporal text embedding features, thereby obtaining the target business text knowledge features.

[0144] In one embodiment, the trained recognition model includes: a classification module, which includes a nonlinear mapping layer, a normalization layer, and a classification layer; Processor 91 invokes the trained recognition model to perform anomaly recognition processing based on the target temporal behavior embedding features and the target business text knowledge features, obtaining the object type of the business object, including: The target temporal behavior embedding features and target business text knowledge features are concatenated to obtain the target behavior text concatenation features; The non-linear mapping layer in the trained recognition model is invoked to perform non-linear mapping processing on the target behavior text splicing features to obtain the target non-linear mapping features. The normalization layer in the trained recognition model is invoked to normalize the nonlinear mapping features of the target, thus obtaining the normalized features of the target. The classification layer in the trained recognition model is invoked to classify the normalized features of the target, thereby obtaining the object type of the business object.

[0145] In one embodiment, the trained recognition model includes a decoding module, and the processor 91 is further configured to: When the object type of the business object is an exception type, the target temporal behavior embedding feature and the target business text knowledge feature are concatenated to obtain the target behavior text concatenation feature. The decoding module in the trained recognition model is invoked to perform anomaly cause analysis based on the target behavior text splicing features, thereby obtaining the target anomaly cause text.

[0146] Based on the same inventive concept, the principle and beneficial effects of the computer device 90 provided in the embodiments of this application in solving the problem are similar to the principle and beneficial effects of the data processing method based on the recognition model in the method embodiments of this application in solving the problem. Please refer to the principle and beneficial effects of the implementation of the method. For the sake of brevity, they will not be repeated here.

[0147] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0148] This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the aforementioned data processing method based on the recognition model.

[0149] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0150] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a management device that integrates one or more available media, such as a server, data center, or other space-managed object. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0151] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data processing method based on a recognition model, characterized in that, The method includes: Obtain sample data for training the recognition model. The sample data includes M sets of business data related to the sample object and the annotation labels of the sample object. One set of the business data comes from a business performed by the sample object in a first business scenario. The set of business data includes business behavior data and business description text data. The annotation labels are used to indicate the annotation type of the sample object, and M is a positive integer. The recognition model is invoked to perform multi-channel temporal coding on the business behavior data in the M groups of business data to obtain temporal behavior embedding features. The temporal behavior embedding features are used to characterize the original channel characteristics of the business behavior data in the M groups of business data in time. The recognition model is invoked to perform knowledge encoding processing on the business description text data in the M groups of business data to obtain business text knowledge features. The business text knowledge features are used to characterize the business logic correlation between the business description text data in the M groups of business data. The recognition model is invoked to perform anomaly recognition processing based on the temporal behavior embedding features and the business text knowledge features to obtain anomaly recognition results, which are used to indicate the prediction type of the sample object. Based on the difference between the anomaly identification result and the label, the identification model is trained, and the trained identification model is used to identify the object type of the business object.

2. The method as described in claim 1, characterized in that, The recognition model includes: a time-series data dynamic quantization module and an embedding layer in a pre-trained language model; The step of calling the recognition model to perform multi-channel temporal coding processing on the business behavior data in the M groups of business data to obtain temporal behavior embedding features includes: According to the execution order of the business corresponding to each of the M groups of business data, the business behavior data in the M groups of business data are arranged to obtain a time-series business behavior sequence. The time-series data dynamic quantization module in the recognition model is invoked to perform data scaling and quantization processing on the time-series business behavior sequence to obtain a time-series behavior quantization sequence. The embedding layer in the recognition model is invoked to perform multi-channel embedding and splicing processing on the time-series behavior quantization sequence to obtain the time-series behavior embedding features.

3. The method as described in claim 2, characterized in that, The step of calling the time-series data dynamic quantization module in the recognition model to perform data scaling and quantization processing on the time-series business behavior sequence to obtain a time-series behavior quantization sequence includes: The time-series data dynamic quantization module in the recognition model is invoked to perform data scaling processing on the continuous variable elements in the time-series business behavior sequence according to the deviation standardization method, so as to obtain the time-series behavior scaling sequence; The time-series data dynamic quantization module in the recognition model is invoked to perform data quantization processing on the continuous variable elements in the time-series behavior scaling sequence according to the binning index method, so as to obtain the time-series behavior quantization sequence; The binning indexing method includes at least one of the following: equal-width binning indexing method, clustered binning indexing method, and dynamic binning indexing method.

4. The method as described in claim 2, characterized in that, The embedding layer includes a word embedding layer and a location embedding layer; the temporal behavior quantization sequence contains behavior quantization data corresponding to each of the business behavior data, and each of the behavior quantization data contains at least one element; The step of calling the embedding layer in the recognition model to perform multi-channel embedding and concatenation processing on the time-series behavior quantization sequence to obtain time-series behavior embedding features includes: The word embedding layer in the recognition model is invoked to perform word embedding representation processing on each element in the temporal behavior quantization sequence, thereby obtaining the word embedding features of each element in the temporal behavior quantization sequence; The location embedding layer in the recognition model is invoked to perform location embedding representation processing on each element in the temporal behavior quantization sequence, thereby obtaining the location embedding feature of each element in the temporal behavior quantization sequence; The word embedding features and position embedding features of each element in the temporal behavior quantization sequence are fused to obtain the target embedding features of each element in the temporal behavior quantization sequence. Based on the sequential order of each behavior quantization data in the time-series behavior quantization sequence and the sequential order of each element in each behavior quantization data, the target embedding features of all elements in the time-series behavior quantization sequence are concatenated to obtain the time-series behavior embedding features.

5. The method as described in claim 1, characterized in that, The recognition model includes: an encoding module in a pre-trained language model; the encoding module includes an embedding layer and an encoding layer; The process of calling the recognition model to perform knowledge encoding on the business description text data in the M groups of business data to obtain business text knowledge features includes: According to the execution order of the corresponding services in the M groups of service data, the service description text data in the M groups of service data are arranged to obtain a time-series service text sequence. The embedding layer in the recognition model is invoked to perform embedding representation processing on the time-series business text sequence to obtain time-series text embedding features; The encoding layer in the recognition model is invoked to perform attention encoding on the temporal text embedding features to obtain business text knowledge features.

6. The method as described in claim 1, characterized in that, The recognition model includes: a classification module, which includes a nonlinear mapping layer, a normalization layer and a classification layer, wherein the nonlinear mapping layer includes the cumulative distribution function of a standard normal distribution; The step of calling the recognition model to perform anomaly recognition processing based on the temporal behavior embedding features and the business text knowledge features, and obtaining anomaly recognition results, includes: The temporal behavior embedding features and the business text knowledge features are concatenated to obtain behavior text concatenation features; The non-linear mapping layer in the recognition model is invoked to perform non-linear mapping processing on the behavior text splicing features to obtain non-linear mapping features; The normalization layer in the recognition model is invoked to normalize the nonlinear mapping features, thereby obtaining normalized features; The classification layer in the recognition model is invoked to classify the normalized features, thereby obtaining the anomaly recognition result.

7. The method as described in claim 1, characterized in that, The recognition model includes: a normalization layer and a position embedding layer from a pre-trained language model; The step of training the recognition model based on the difference between the anomaly recognition result and the labeled data includes: A binary classification loss function is obtained, and the anomaly identification result and the label are input into the binary classification loss function to obtain the first loss parameter corresponding to the identification model; the first loss parameter is used to reflect the difference between the anomaly identification result and the label. The model parameters of the normalization layer and the location embedding layer in the recognition model are adjusted using the first loss parameter in order to train the recognition model.

8. The method as described in claim 7, characterized in that, The recognition model also includes: the decoding module in the pre-trained language model; The step of adjusting the model parameters of the normalization layer and the location embedding layer in the recognition model using the first loss parameter to train the recognition model includes: When the label indicates that the sample object is an anomaly, obtain the text of the actual anomaly cause. The temporal behavior embedding features and the business text knowledge features are concatenated to obtain behavior text concatenation features; The decoding module in the pre-trained language model is invoked to perform anomaly cause analysis based on the behavioral text concatenation features, thereby obtaining the predicted anomaly cause text. Based on the difference between the actual anomaly cause text and the predicted anomaly cause text, the second loss parameter corresponding to the recognition model is determined; Using the first loss parameter and the second loss parameter, the model parameters of the normalization layer, the location embedding layer and the decoding module in the recognition model are adjusted to train the recognition model.

9. The method according to any one of claims 1-8, characterized in that, The method further includes: Acquire a general language model and business corpus text data in the first business scenario; the general language model includes an encoding module and a decoding module; The business corpus text data is masked to obtain masked corpus text data; The encoding module in the general language model is invoked to encode the masked corpus text data to obtain the masked text encoding features. The decoding module in the general language model is invoked to perform mask prediction processing based on the mask text encoding features, thereby obtaining the predicted mask content in the business corpus text data. Based on the difference between the predicted mask content and the actual mask content in the business corpus text data, the general language model is trained to obtain a pre-trained language model.

10. The method as described in claim 1, characterized in that, The step of obtaining sample data for training the recognition model includes: Determine the M consecutive business operations executed by the sample object, and obtain the behavioral feature information of each business operation under P behavioral feature channels and the text feature information under Q text feature channels; P and Q are both positive integers; Using the behavioral feature information of each service under P behavioral feature channels, the initial behavioral data of each service is constructed. The initial behavioral data of each service is then cleaned and filled to obtain the business behavioral data of each service. Using the text feature information of each service under Q text feature channels, construct the service description text data for each service; Obtain the labels obtained after expert evaluation of the M service items; The business behavior data and business description text data of the M business items, as well as the tag data, are used to construct sample data for training the recognition model.

11. A data processing method based on a recognition model, characterized in that, The method includes: Obtain N sets of target business data related to the business object. Each set of target business data originates from a business operation performed by the business object in a first business scenario. Each set of target business data includes target business behavior data and target business description text data; N is a positive integer. The trained recognition model is invoked to perform multi-channel temporal coding on the target business behavior data in the N sets of target business data to obtain target temporal behavior embedding features. The target temporal behavior embedding features are used to characterize the original channel characteristics of the target business behavior data in the N sets of target business data in time. The trained recognition model is invoked to perform knowledge encoding processing on the target business description text data in the N sets of target business data to obtain target business text knowledge features. The target business text knowledge features are used to characterize the business logic correlation between the target business description text data in the N sets of target business data. The trained recognition model is invoked to perform anomaly recognition processing based on the target temporal behavior embedding features and the target business text knowledge features to obtain the object type of the business object.

12. The method as described in claim 11, characterized in that, The trained recognition model includes: a time-series data dynamic quantization module and an embedding layer; The trained recognition model is invoked to perform multi-channel temporal encoding on the target business behavior data in the N sets of target business data to obtain target temporal behavior embedding features, including: According to the execution order of the corresponding services in the N sets of target service data, the target service behavior data in the N sets of target service data are arranged to obtain the target time-series service behavior sequence. The time-series data dynamic quantization module in the trained recognition model is invoked to perform data scaling and quantization processing on the target time-series business behavior sequence, thereby obtaining the target time-series behavior quantization sequence. The embedding layer in the trained recognition model is invoked to perform multi-channel embedding and splicing processing on the target temporal behavior quantization sequence to obtain the target temporal behavior embedding features.

13. The method as described in claim 11, characterized in that, The trained recognition model includes an encoding module, which comprises an embedding layer and an encoding layer; The trained recognition model is invoked to perform knowledge encoding processing on the target business description text data in the N sets of target business data to obtain target business text knowledge features, including: According to the execution order of the corresponding services in the N sets of target service data, the target service description text data in the N sets of target service data are arranged to obtain the target time-series service text sequence. The embedding layer in the trained recognition model is invoked to perform embedding representation processing on the target time-series business text sequence to obtain the target time-series text embedding features; The encoding layer in the trained recognition model is invoked to perform attention encoding on the target temporal text embedding features to obtain the target business text knowledge features.

14. The method as described in claim 11, characterized in that, The trained recognition model includes a classification module, which comprises a nonlinear mapping layer, a normalization layer, and a classification layer. The trained recognition model is invoked to perform anomaly recognition processing based on the target temporal behavior embedding features and the target business text knowledge features, to obtain the object type of the business object, including: The target temporal behavior embedding features and the target business text knowledge features are concatenated to obtain the target behavior text concatenation features; The non-linear mapping layer in the trained recognition model is invoked to perform non-linear mapping processing on the target behavior text splicing features to obtain the target non-linear mapping features. The normalization layer in the trained recognition model is invoked to normalize the nonlinear mapping features of the target, thereby obtaining the target normalized features; The classification layer in the trained recognition model is invoked to classify the normalized features of the target to obtain the object type of the business object.

15. The method according to any one of claims 11-14, characterized in that, The trained recognition model includes a decoding module, and the method further includes: When the object type of the business object is an abnormal type, the target temporal behavior embedding feature and the target business text knowledge feature are concatenated to obtain the target behavior text concatenation feature. The decoding module in the trained recognition model is invoked to perform anomaly cause analysis based on the target behavior text splicing features, thereby obtaining the target anomaly cause text.

16. A data processing device based on a recognition model, characterized in that, The device includes: The acquisition unit is used to acquire sample data for training the recognition model. The sample data includes M sets of business data related to the sample object and the annotation labels of the sample object. One set of the business data comes from a business performed by the sample object in a first business scenario. The set of business data includes business behavior data and business description text data. The annotation labels are used to indicate the annotation type of the sample object, and M is a positive integer. The processing unit is used to call the recognition model to perform multi-channel temporal coding processing on the business behavior data in the M groups of business data to obtain temporal behavior embedding features. The temporal behavior embedding features are used to characterize the original channel characteristics of the business behavior data in the M groups of business data in time. The processing unit is further configured to call the recognition model to perform knowledge encoding processing on the business description text data in the M groups of business data to obtain business text knowledge features, which are used to characterize the business logic correlation between the business description text data in the M groups of business data. The processing unit is further configured to call the recognition model to perform anomaly recognition processing based on the temporal behavior embedding features and the business text knowledge features, and obtain anomaly recognition results, wherein the anomaly recognition results are used to indicate the prediction type of the sample object; The processing unit is further configured to train the recognition model based on the difference between the anomaly recognition result and the label, and the trained recognition model is used to identify the object type of the business object.

17. A data processing device based on a recognition model, characterized in that, The device includes: The acquisition unit is used to acquire N sets of target business data related to the business object. Each set of target business data originates from a business operation performed by the business object in a first business scenario. Each set of target business data includes target business behavior data and target business description text data; N is a positive integer. The processing unit is used to call the trained recognition model to perform multi-channel temporal coding processing on the target business behavior data in the N sets of target business data to obtain target temporal behavior embedding features. The target temporal behavior embedding features are used to characterize the original channel characteristics of the target business behavior data in the N sets of target business data in time. The processing unit is further configured to call the trained recognition model to perform knowledge encoding processing on the target business description text data in the N sets of target business data to obtain target business text knowledge features. The target business text knowledge features are used to characterize the business logic correlation between the target business description text data in the N sets of target business data. The processing unit is further configured to invoke the trained recognition model to perform anomaly recognition processing based on the target temporal behavior embedding features and the target business text knowledge features, so as to obtain the object type of the business object.

18. A computer device, characterized in that, The computer device includes: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the data processing method based on a recognition model as described in any one of claims 1-10, or the data processing method based on a recognition model as described in any one of claims 11-15.

19. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed by the data processing method based on the recognition model as described in any one of claims 1-10, or the data processing method based on the recognition model as described in any one of claims 11-15.

20. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the data processing method based on the recognition model as described in any one of claims 1-10, or the data processing method based on the recognition model as described in any one of claims 11-15.

Images