Transaction risk evaluation processing method and device, storage medium and electronic equipment

Abstract

The invention discloses a transaction risk evaluation processing method and device, a storage medium and electronic equipment, and the method comprises the steps: obtaining an original multi-dimensional behavior time sequence of a target object in a target transaction scene, inputting the multi-dimensional behavior time sequence into a target transaction processing model, performing multi-time scale analysis on the original multi-dimensional behavior time sequence through the target transaction processing model to obtain a plurality of reference scale behavior time sequence characteristics, and fusing all the reference scale behavior time sequence characteristics to obtain multi-scale fusion behavior representation; and performing position sensing coding on the original multi-dimensional behavior time sequence through a target transaction processing model to obtain global time sequence behavior representation, and performing representation fusion on the multi-scale fusion behavior representation and the global time sequence behavior representation through the target transaction processing model to obtain object comprehensive behavior representation. And performing transaction risk evaluation processing based on the object comprehensive behavior representation to obtain an object transaction risk evaluation result for the target object.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to a transaction risk assessment and processing method, device, storage medium, and electronic device. Background Art

[0002] With the widespread application of big data technology and intelligent transaction risk control methods, service platforms are increasingly relying on in-depth analysis of behavioral data when handling transaction risks. Especially in transaction scenarios involving supply chain management, industrial chain risk assessment, daily consumption, and transportation, platforms often need to identify, model, and evaluate the dynamic behavior of transaction subjects to determine their potential risk level. This allows them to better provide transaction services for their users. Summary of the Invention

[0003] This specification provides a transaction risk assessment and processing method, device, storage medium, and electronic device. The technical solution is as follows:

[0004] In a first aspect, this specification provides a transaction risk assessment and processing method, the method comprising:

[0005] In a target transaction scenario, obtaining an original multidimensional behavior time series of a target object, and inputting the original multidimensional behavior time series into a target transaction processing model;

[0006] Performing multi-time-scale analysis on the original multi-dimensional behavior time series using the target transaction processing model to obtain multiple reference-scale behavior time series features, and fusing all of the reference-scale behavior time series features to obtain a multi-scale fused behavior representation; and performing position-aware encoding on the original multi-dimensional behavior time series using the target transaction processing model to obtain a global time series behavior representation;

[0007] The multi-scale fusion behavior representation and the global temporal behavior representation are characterized and fused through the target transaction processing model to obtain an object comprehensive behavior representation, and transaction risk assessment processing is performed based on the object comprehensive behavior representation to obtain an object transaction risk assessment result for the target object.

[0008] In a second aspect, this specification provides a transaction risk assessment and processing device, the device comprising:

[0009] A data input module is used to obtain the original multidimensional behavior time series of the target object in the target transaction scenario, and input the original multidimensional behavior time series into the target transaction processing model;

[0010] a model processing module configured to perform multi-time-scale analysis on the original multi-dimensional behavior time series using the target transaction processing model to obtain multiple reference-scale behavior time series features, fuse all of the reference-scale behavior time series features to obtain a multi-scale fused behavior representation; and perform position-aware encoding on the original multi-dimensional behavior time series using the target transaction processing model to obtain a global time series behavior representation;

[0011] The risk assessment module is used to fuse the multi-scale fusion behavior representation and the global temporal behavior representation through the target transaction processing model to obtain an object comprehensive behavior representation, and perform transaction risk assessment processing based on the object comprehensive behavior representation to obtain an object transaction risk assessment result for the target object.

[0012] In a third aspect, the present specification provides a computer storage medium storing at least one instruction, wherein the instruction is suitable for being loaded by a processor and executing the method steps of one or more embodiments of the present specification.

[0013] In a fourth aspect, the present specification provides a computer program product, wherein the computer program product stores at least one instruction, wherein the instruction is suitable for being loaded by a processor and executing the method steps of one or more embodiments of the present specification.

[0014] In a fifth aspect, this specification provides an electronic device, which may include: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method steps of one or more embodiments of this specification.

[0015] The beneficial effects of the technical solutions provided by some embodiments of this specification include at least:

[0016] In one or more embodiments of the present specification, in a target transaction scenario, the electronic device first obtains the original multi-dimensional behavior time series of the target object within a preset time range, introduces a target transaction processing model, and performs multi-time scale parsing processing and location-aware coding processing on the behavior time series, respectively generating a multi-scale fusion behavior representation for reflecting the local trend of the behavior, and a global temporal behavior representation for modeling time structure information; and jointly models the above two types of representations through representation fusion to obtain an object comprehensive behavior representation that can comprehensively characterize the behavior characteristics of the target object, and further performs transaction risk assessment processing based on the comprehensive behavior representation, and outputs the transaction risk assessment result of the target object. Through the above processing flow, the system can effectively improve the modeling accuracy and risk identification ability when processing highly sparse, multi-dimensional, and highly time-dependent transaction behavior data, realize fine-grained behavior dynamic perception, multi-scale risk feature extraction and structured transaction assessment modeling, and enhance the model's ability to identify abnormal patterns, periodic signals and potential risks while ensuring the fidelity of behavior information, which has high practical value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in this specification or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 This is a scenario diagram of a transaction risk assessment and processing system provided in this specification;

[0019] Figure 2 This is a flowchart of a transaction risk assessment and processing method provided in this manual;

[0020] Figure 3 This is a flow chart of a characterization process provided in this specification;

[0021] Figure 4 This is a schematic diagram of a scenario processed by a target transaction processing model provided in this specification;

[0022] Figure 5 This is a flowchart of an absolute position embedded coding process provided in this specification;

[0023] Figure 6 This is a flowchart of another embodiment of the transaction risk assessment processing method provided in this specification;

[0024] Figure 7 This is a flowchart for constructing a behavioral causal relationship provided in this manual;

[0025] Figure 8 This is a structural diagram of a transaction risk assessment and processing device provided in this specification;

[0026] Figure 9 This is a structural diagram of a model processing module provided in this specification;

[0027] Figure 10 This is a structural diagram of an electronic device provided in this manual. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in this specification in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments in this specification, not all of them. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this specification.

[0029] In the description of this specification, it should be understood that the terms "first", "second", etc. are used for descriptive purposes only and should not be understood as indicating or implying relative importance. In the description of this specification, it should be noted that, unless otherwise expressly specified and limited, "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units that are not listed, or may optionally include other steps or units inherent to these processes, methods, products or devices. For those of ordinary skill in the art, the specific meanings of the above terms in this specification can be understood according to the specific circumstances. In addition, in the description of this specification, unless otherwise specified, "multiple" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0030] With the accelerated development of industrial digitization and supply chain finance, inter-enterprise credit relationships, contract performance capabilities, capital flows, and transaction behaviors are becoming increasingly visible. Against this backdrop, financial institutions, platform companies, and risk management systems urgently need to conduct continuous, granular risk assessments of various transaction entities within the supply chain. In particular, when granting credit or conducting access reviews for suppliers, distributors, service providers, and other stakeholders, modeling and analysis of their behavioral data within platforms, systems, or network environments is often required to quantify their contract performance risks, behavioral risks, or credit volatility risks.

[0031] At present, in the field of risk modeling of industrial chain transactions, the widely used technical means are based on traditional statistical learning methods, especially extracting features through artificial statistical feature engineering, and using tree models (such as XGBoost, LightGBM) for risk prediction. This type of method performs well for static variables and statistical indicators within a fixed time window, and has a certain degree of interpretability. However, in actual transactions, the risks of transaction objects often come from their dynamic behavior evolution, especially the existence of fine-grained time series features such as short-term fluctuations, periodic bursts, and periodic anomalies, which are difficult to be effectively captured by static modeling.

[0032] Furthermore, transactional behavior characteristics recorded on the platform, such as daily transaction volume, login behavior, invoice activity, complaint responses, and payment responses, naturally possess time series properties. These sequences are often sparse and high-dimensional. Directly expanding them at a daily granularity as model input can easily lead to dimensionality explosion (for example, 90 days x 10-dimensional behavior = 900-dimensional input). Furthermore, they contain a large number of zero values, resulting in reduced model inference efficiency and weakened generalization capabilities.

[0033] Furthermore, while traditional methods like monthly aggregation and sliding window averaging can alleviate the sparsity problem to some extent, they can severely compromise key risk cues such as cyclical trends and extreme anomalies inherent in time series signals. Manually extracting these time series patterns is both costly and subjective, and it struggles to adapt to large-scale, high-dimensional, and rapidly changing behavioral data streams.

[0034] Against the backdrop of the rapid development of neural networks, although some time series classification models have attempted to introduce convolutional networks, recurrent neural networks, or Transformer structures to model behavioral sequences, these models are mostly oriented towards continuous and dense time series data, and are unable to effectively cope with the high sparsity, multi-dimensional interactions, and structural weak dependencies that are common in platform-based transaction data. They still cannot meet the complex requirements of industry chain transaction risk control for real-time, accuracy, and structural interpretability.

[0035] Therefore, how to design a transaction risk assessment processing method that adapts to the characteristics of transaction behavior data, has multi-time scale analysis capabilities, can integrate global time perception and structural behavior associations, and can automatically extract highly sparse behavior time series features for risk assessment is one of the key issues that urgently need to be solved in the current transaction risk assessment field.

[0036] The present specification is described in detail below with reference to specific embodiments.

[0037] See Figure 1 , is a scenario diagram of a transaction risk assessment and processing system provided in this specification. Figure 1As shown, the transaction risk assessment and processing system may include at least a client cluster and a service platform 100 .

[0038] The client cluster may include at least one client, such as Figure 1 As shown, it specifically includes client 1 corresponding to user 1, client 2 corresponding to user 2, ..., client n corresponding to user n, where n is an integer greater than 0.

[0039] Each client in the client cluster can be an electronic device with communication capabilities, including but not limited to wearable devices, handheld devices, personal computers, tablet computers, in-vehicle devices, smartphones, computing devices, or other processing devices connected to a wireless modem. Electronic devices may be called different names in different networks, such as user equipment, access terminals, subscriber units, subscriber stations, mobile stations, mobile stations, remote stations, remote terminals, mobile devices, user terminals, electronic devices, wireless communication devices, user agents or user devices, cellular phones, cordless phones, personal digital assistants (PDAs), and electronic devices in 5G networks or future evolution networks.

[0040] The service platform 100 can be a separate server device, such as a rack-mounted, blade, tower, or cabinet-mounted server device, or a workstation, mainframe computer, or other hardware device with strong computing capabilities; it can also be a server cluster composed of multiple servers. The servers in the service cluster can be symmetrically composed, wherein each server has equivalent functions and status in the transaction link, and each server can provide services to the outside world independently. The independent service can be understood as not requiring the assistance of other servers.

[0041] In one or more embodiments of the present specification, the service platform 100 may establish a communication connection with at least one client in the client cluster, and complete data interaction during the transaction risk assessment process, such as online transaction data interaction, based on the communication connection.

[0042] It should be noted that the service platform 100 and at least one client in the client cluster establish a communication connection through a network for interactive communication, wherein the network can be a wireless network or a wired network, the wireless network includes but is not limited to a cellular network, a wireless local area network, an infrared network or a Bluetooth network, and the wired network includes but is not limited to Ethernet, a universal serial bus (USB) or a controller area network. In one or more embodiments of the specification, technologies and / or formats including Hypertext Markup Language (HTML) and Extensible Markup Language (XML) are used to represent data exchanged over the network (such as a target compressed package). In addition, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Network (VPN), and Internet Protocol Security (IPsec) can also be used to encrypt all or some links. In other embodiments, customized and / or dedicated data communication technologies can also be used to replace or supplement the above-mentioned data communication technologies.

[0043] The transaction risk assessment and processing system embodiments provided in this specification share the same concept as the transaction risk assessment and processing methods described in one or more embodiments. The execution entity corresponding to the transaction risk assessment and processing methods described in one or more embodiments of this specification may be the aforementioned service platform 100; the execution entity corresponding to the transaction risk assessment and processing methods described in one or more embodiments of this specification may also be the electronic device corresponding to the client, depending on the specific application environment. The implementation process of the transaction risk assessment and processing system embodiments can be found in the following method embodiments and will not be further elaborated here.

[0044] based on Figure 1 The scenario diagram shown is as follows, and the transaction risk assessment and processing method provided by one or more embodiments of this specification is introduced in detail.

[0045] See Figure 2 , provides a flow chart of a transaction risk assessment and processing method for one or more embodiments of this specification. This method can be implemented using a computer program and run on a transaction risk assessment and processing device based on the von Neumann architecture. The computer program can be integrated into an application or run as a standalone tool application. The transaction risk assessment and processing device can be an electronic device.

[0046] Specifically, the transaction risk assessment and processing method includes:

[0047] S102: In a target transaction scenario, obtaining an original multidimensional behavior time series of a target object, and inputting the original multidimensional behavior time series into a target transaction processing model;

[0048] Target transaction scenarios: These refer to the transaction scenarios in which this solution is applied, such as risk review in supply chain services, merchant performance assessment on e-commerce platforms, user credit risk assessment on user platforms, user behavior risk assessment on transportation platforms, and abnormal user operation behavior detection. Each transaction scenario involves different behavioral characteristics and target risk categories, but all focus on the "risk status of a specific object during the transaction process" to conduct transaction risk assessment and processing.

[0049] Target object: refers to the unit that needs to be evaluated in the transaction scenario, such as an enterprise, user, service provider, a subsystem or process link, etc.

[0050] Raw multi-dimensional behavior time series refers to the platform's original behavior records of multiple behavioral dimensions of the target object over a period of time (this can be understood as unprocessed, such as normalization, aggregation, feature extraction, and retaining the original structure). It is organized in chronological order and is a type of time-ordered sequential data collected from the target object's historical behavior. "Multi-dimensional" refers to the variety of behaviors, each of which constitutes a characteristic dimension, such as clicks, payments, complaints, logins, invoice anomalies, etc.; "time series" refers to the fact that the records of each dimension evolve over time and are arranged at granularity such as days, hours, and minutes.

[0051] In practical applications, raw multidimensional behavioral time series are usually sparse data. For example, if 10-dimensional time series features are expanded to a daily granularity with each dimension length of 90, the data length of the raw multidimensional behavioral time series will reach 900, and these 900 features contain a high proportion of zeros.

[0052] In an optional embodiment of the present invention, for a set target transaction scenario, behavioral information data of a target object is collected from a preset data source, and the behavioral information data is aggregated and processed based on the time dimension to construct an original multidimensional behavioral time series associated with the target object. The original multidimensional behavioral time series is used to represent the behavioral performance status of the target object along multiple behavioral feature dimensions in a preset time interval.

[0053] For example, in a risk assessment scenario for a business object within a specific enterprise chain, the multiple behavioral feature dimensions include, but are not limited to, transaction behavior, access behavior, fund flows, complaint records, authentication operations, risk control feedback, and other transaction-related behavioral indicators. The time series can be granular with days, hours, minutes, etc., preferably with days as the minimum time unit to ensure that the dynamic changes in the behavior of the business object over a continuous natural time period are captured.

[0054] Furthermore, the system can filter the behavioral data based on the target object identifier, arrange it in ascending order by timestamp and organize it into a T×D dimensional data structure, where T represents the length of the time series and D represents the number of behavioral dimensions; after completing time alignment and formatting, the original multi-dimensional behavioral time series is input as structured input data into the target transaction processing model for subsequent feature analysis and risk assessment processing.

[0055] S104: performing multi-time-scale analysis on the original multi-dimensional behavior time series using the target transaction processing model to obtain multiple reference-scale behavior time series features, fusing all of the reference-scale behavior time series features to obtain a multi-scale fused behavior representation; and performing position-aware encoding on the original multi-dimensional behavior time series using the target transaction processing model to obtain a global time series behavior representation.

[0056] Multi-timescale feature extraction: This refers to the process of modeling the original behavioral time series using multiple parsing methods with different time observation windows to summarize and structurally model the behavioral sequence from different time span dimensions. This feature extraction process can capture the dynamic characteristics of behavior at different time scales, providing hierarchical structural information for subsequent fusion representation.

[0057] Optionally, the multiple time scales include, but are not limited to: raw behavioral expression at a single-day granularity (for maintaining detail fidelity), local trend modeling over a multi-day window (for extracting short-term behavioral fluctuations), and periodic aggregation over a cross-week or cross-month window (for capturing repetitive behavioral patterns).

[0058] Reference-scale behavioral temporal features: These are intermediate features generated by extracting behavioral features at different time scales. They are used to reflect the behavioral evolution characteristics of the target object within a specific time scale. These reference-scale behavioral temporal features do not directly output evaluation results, but serve as the fundamental building blocks for multi-scale fused behavioral representation.

[0059] Optionally, each reference scale behavior time series feature may include: behavior trend direction (such as increasing, decreasing, sudden change), fluctuation pattern within the time period (such as high frequency, high amplitude, intermittent), behavior density or activity (such as event frequency, proportion change).

[0060] Multi-scale fused behavioral representation: This is achieved by fusing behavioral temporal features generated at multiple reference time scales to form a structured, high-dimensional representation vector that uniformly represents the target object's behavioral pattern. This multi-scale fused behavioral representation, outputted by the model's intermediate layer, is used to characterize the target object's transactional characteristics by combining temporal structure and behavioral morphology.

[0061] Optionally, the fusion method may include but is not limited to: feature concatenation, weighted summation, and weighted combination of attention mechanisms;

[0062] In an optional embodiment of the present invention, the target transaction processing model includes multiple structured processing parts for performing behavior feature analysis. The target transaction processing model performs multi-time scale feature extraction processing and time series position aware encoding processing on the input original multi-dimensional behavior time series.

[0063] Specifically, the multi-time scale feature extraction process includes: performing time scale analysis on the original multi-dimensional behavior time series. In this process, the original behavior sequence feature extraction, short-term trend modeling and cross-period feature aggregation are performed in sequence. The original behavior sequence feature extraction is used to maintain the original expression of the behavior ontology at the daily granularity, the short-term trend modeling is used to characterize the local fluctuation pattern of the behavior in the continuous time segment, and the cross-period feature aggregation is performed by aggregating and analyzing the behavior dynamics across time windows to characterize its periodic change rules. The behavior time series features generated under the above multiple reference scales are aggregated for feature fusion, and a multi-scale fused behavior representation is generated by feature dimension splicing or weighted superposition.

[0064] At the same time, the target transaction processing model also performs temporal position-aware encoding processing in parallel. During this processing, each time step of the original multidimensional behavior time series is embedded and fused with its corresponding absolute time position identifier to form an enhanced behavior sequence representation containing temporal position information. The position embedding information can be represented by a fixed code, a learnable parameter matrix, or a position function based on trigonometric functions. The target transaction processing model further globally models the behavior sequence based on the enhanced behavior sequence through a position-aware self-attention mechanism, thereby obtaining a global temporal behavior representation that reflects the evolutionary structure of the target object's behavior.

[0065] Through the above two processing paths, the target transaction processing model can deeply model the behavioral information of the target object from the two dimensions of local temporal dynamics and global behavioral sequence structure, providing a multi-level and combinable behavioral representation basis for subsequent risk assessment.

[0066] S106: The multi-scale fusion behavior representation and the global temporal behavior representation are characterized and fused by the target transaction processing model to obtain an object comprehensive behavior representation, and transaction risk assessment processing is performed based on the object comprehensive behavior representation to obtain an object transaction risk assessment result for the target object.

[0067] Representation fusion processing involves combining, interacting, or integrating multiple behavioral representation vectors from different sources or structural pathways to generate a high-dimensional, comprehensive feature vector that simultaneously captures multiple aspects of information. Representation fusion is a key step in achieving "information synthesis," enabling the model to possess holistic understanding.

[0068] Optionally, the representation fusion processing may have the following technical features: semantic dimension alignment: ensuring the superposition of features from different sources in the vector space; channel weighting mechanism: setting learnable channel weights for information sources (such as multi-scale features, global features); nonlinear interaction strategy: modeling the nonlinear combination relationship between features through attention mechanisms, adaptive fusion layers, MLP, etc.

[0069] Comprehensive object behavior representation: This refers to a high-dimensional feature vector formed by fusing multiple substructure representations to comprehensively represent the target object's behavioral characteristics across the entire behavioral sequence. This representation vector has the ability to express information across scales, time periods, and behavior types.

[0070] The technical role of comprehensive object behavior representation is that it can support subsequent evaluation models to make overall risk judgments on target objects, can be used as feature abstraction results for reuse in other upper-level tasks (such as risk interpretation and behavior prediction), and can improve the model's discrimination and generalization capabilities in highly sparse time series tasks.

[0071] Transaction risk assessment and processing: This refers to the process by which the system predicts or determines the potential risk level of a target object in the current transaction scenario based on the object's comprehensive behavioral representation. The assessment results can be used for transaction decision-making, approval and control, or risk warning and response mechanisms within the system.

[0072] Optionally, the implementation of risk assessment processing may include but is not limited to: risk prediction based on supervised learning parts (such as neural network classifiers, scorers); risk mapping based on a predefined rule system; outputting classification labels and risk confidence scores based on a multi-objective discrimination mechanism.

[0073] Object transaction risk assessment results: refers to the result identifier output by the risk assessment module, which represents the risk status of the target object in the transaction scenario and is in a form that can be judged by a machine or interpreted by the transaction system.

[0074] Optionally, the evaluation results may be one or more of the following: structured risk levels (such as "low risk / medium risk / high risk"), numerical risk scores (such as "risk score is 0.83"), risk classification labels (such as "loan rejection", "enhanced review", "recommended credit"), multi-indicator joint results (such as "stable in the short term, but with medium- and long-term risk tendencies"), and the evaluation results may be combined with system strategies as the basis for the final transaction response.

[0075] In an optional embodiment of the present invention, the control target transaction processing model performs feature-level representation fusion processing based on the previous multi-scale fusion behavior representation and the global temporal behavior representation to generate a comprehensive behavior representation of the target object.

[0076] Specifically, the representation fusion structure in the target transaction processing model is used to receive the behavioral feature information of the above two dimensions, and adopts the multi-channel attention fusion mechanism, feature transformation mechanism or other representation joint strategies within the structure to mine the correlation between various behavioral representations and adjust the feature weights, thereby generating a unified representation vector with high expressiveness.

[0077] The comprehensive behavior representation structurally reflects the behavior characteristics of the target object at different time scales, different time positions, and cross-behavioral interaction perspectives, and is a deep semantic expression of the input behavior sequence.

[0078] After obtaining the comprehensive behavioral representation of the object, the target transaction processing model further performs transaction risk assessment. This process includes: performing transaction assessment based on the comprehensive behavioral representation. The environment performs risk assessment on the target object based on a preset risk assessment strategy, training model, or assessment rule system, and outputs an assessment result representing its transaction risk level, risk propensity, or risk probability score.

[0079] In one or more embodiments of the present specification, in a target transaction scenario, the electronic device first obtains the original multi-dimensional behavior time series of the target object within a preset time range, introduces a target transaction processing model, and performs multi-time scale parsing processing and location-aware coding processing on the behavior time series, respectively generating a multi-scale fusion behavior representation for reflecting the local trend of the behavior, and a global temporal behavior representation for modeling time structure information; and jointly models the above two types of representations through representation fusion to obtain an object comprehensive behavior representation that can comprehensively characterize the behavior characteristics of the target object, and further performs transaction risk assessment processing based on the comprehensive behavior representation, and outputs the transaction risk assessment result of the target object. Through the above processing flow, the system can effectively improve the modeling accuracy and risk identification ability when processing highly sparse, multi-dimensional, and highly time-dependent transaction behavior data, realize fine-grained behavior dynamic perception, multi-scale risk feature extraction and structured transaction assessment modeling, and enhance the model's ability to identify abnormal patterns, periodic signals and potential risks while ensuring the fidelity of behavior information, which has high practical value and promotion prospects.

[0080] Optionally, the target transaction processing model may include a first representation path and a second representation path, such as Figure 3 As shown, Figure 3 This is a flow chart of characterization processing, specifically performing multi-time-scale analysis on the original multi-dimensional behavior time series through the target transaction processing model to obtain multiple reference-scale behavior time series features, fusing all the reference-scale behavior time series features to obtain a multi-scale fused behavior representation; and performing position-aware encoding on the original multi-dimensional behavior time series through the target transaction processing model to obtain a global time series behavior representation, and fusing the multi-scale fused behavior representation and the global time series behavior representation through the target transaction processing model to obtain a comprehensive behavior representation of the object. The following methods can be used:

[0081] like Figure 4 As shown, Figure 4 A schematic diagram of a scenario processed by a target transaction processing model, wherein the target transaction processing model may include a first representation path and a second representation path; and may also include a transformer network;

[0082] The first characterization pathway refers to a modeling branch structure set in the target transaction processing model for performing multi-time-scale behavior feature extraction tasks. It is configured to extract multi-level behavior characterization information covering different time window scales based on the original multi-dimensional behavior time series. The first characterization pathway can be called the LCT pathway.

[0083] The second representation path refers to the feature processing path set in the target transaction processing model for performing time position information modeling and global behavior sequence structure modeling. It is configured to combine the time position information in the original multidimensional behavior time series to generate a global temporal behavior representation with time perception capabilities; the second representation path can be called the PositionEmbedding path.

[0084] The transformer network refers to an attention mechanism structural unit set in the target transaction processing model for performing multi-path feature fusion and temporal structure modeling. Its core function is to perform feature interaction processing, weight adjustment and deep information fusion on the behavior representation results from different feature paths, thereby generating a more expressive comprehensive behavior representation of the object. The transformer network can adopt the Transformer architecture.

[0085] S202: performing multi-time scale analysis on the original multi-dimensional behavior time series through the first characterization pathway to obtain original behavior sequence characteristics, short-term behavior trend characteristics, and cross-period behavior pattern characteristics;

[0086] Original behavior sequence features: refers to the set of basic behavior features that are formed based on the observation values ​​of each behavior dimension in a single time slice in the original multidimensional behavior time series without time window transformation or high-order processing. This feature is used to retain the actual state information of the behavior at the finest time granularity and constitutes the primary input basis for behavior modeling. The original behavior sequence features can be expressed as a multivariate time series (MCT), such as Figure 4 The original behavior sequence features generated by the portion marked 101 shown in FIG; for example, a sequence feature extraction module may be used to generate the original behavior sequence features;

[0087] Short-term behavioral trend features: refers to the intermediate feature results that can reflect the trend or directional pattern of behavioral changes by modeling the behavior changes in local continuous time segments. The feature modeling process can use sliding window processing, sequence recursion mechanism, etc. to capture short-term behavioral change patterns such as upward trends, downward trends, sudden increases and decreases. Its generation helps to characterize whether the target object's recent behavior is stable, abnormal, or has a tendency to sudden risks, such as Figure 4 The short-term behavior trend features produced by the portion labeled 102 shown in FIG. 1 , for example, a trend modeling module can be used here to produce short-term behavior trend features.

[0088] Cross-cycle behavior pattern characteristics: refers to the characteristic results obtained after analyzing the repetitive, periodic, and rhythmic structure of the target object's behavior in a larger time span dimension such as across weeks and months. Its modeling mechanism may include time segment aggregation, cross-cycle comparison, periodic signal extraction, etc., to reveal whether the behavior has repetitive patterns or seasonal risk signs in a fixed time interval. This feature provides a core reference for the "structural stability" or "stage abnormality" of the modeled object, such as Figure 4 The cross-period behavior pattern features generated by the portion labeled 103 shown in FIG. , for example, the cross-period behavior pattern features can be generated using a pattern extraction module.

[0089] In an optional implementation, the first characterization path in the control target transaction processing model executes the extraction processing of multi-time-scale behavioral features, specifically including: performing fine-grained analysis on the behavioral observation values ​​of each time slice in the original multidimensional behavioral time series, and extracting the original behavioral sequence features that characterize the original behavioral state of the target object; based on the set sliding window strategy and time series modeling structure, performing local continuity modeling operations on the original behavioral sequence features to obtain short-term behavioral trend features that reflect the short-term behavioral fluctuation trend; further, performing cross-period behavioral pattern modeling operations on the behavioral changes of short-term trend features under a larger time span, and extracting cross-period behavioral pattern features that characterize the periodic evolution of behavior.

[0090] Through the processing in S202 above, various dynamic behavioral features at different time scales can be extracted from daily behavior records. The original behavioral data is expanded along the time dimension and decomposed into three feature layers: "current behavior state," "short-term behavioral changes," and "periodic behavior structure" through a window mechanism. This provides the time perception depth and multi-window behavior representation capabilities required for time series modeling. This provides multi-level, cross-window behavioral structure information support for subsequent fusion modeling.

[0091] In a feasible implementation, S202 may be performed in the following manner:

[0092] The first characterization pathway is used to perform feature change processing on the daily behavior data in the original multidimensional behavior time series to obtain original behavior sequence features, and multiple long-short-term memory neural networks (such as Figure 4 The LSTM shown in FIG1 is used to analyze the temporal dependency relationship of the original behavior sequence features of the daily time dimension to obtain the short-term behavior trend features, and the convolutional neural network (such as Figure 4 The CNN shown in the figure (Cross Time means to discover the "cross-cycle" behavioral pattern) analyzes the short-term behavioral trend characteristics into cross-month behavioral pattern characteristics to obtain cross-cycle behavioral pattern characteristics;

[0093] In an optional embodiment, the first characterization pathway is configured to perform multi-time scale structural modeling on the original multidimensional behavior time series, wherein, for daily granularity behavior data, the first characterization pathway sequentially performs original behavior feature extraction, short-term behavior trend modeling, and cross-period behavior pattern analysis to extract structured behavior representations for characterizing the behavior evolution characteristics at different time levels.

[0094] Specifically, in this implementation, the first characterization pathway first performs feature processing on the daily behavioral data in the original multidimensional behavioral time series, extracting time-step-by-time features of the original behavioral sequence. These features reflect the original observed state of each behavioral dimension of the target subject in a single time slice. Feature processing may include operations such as behavioral normalization, differential transformation, and sliding smoothing to improve the modelability of the original data in terms of behavioral distribution and numerical scale.

[0095] Subsequently, the original behavior sequence features are input into multiple long short-term memory neural networks (LSTMs). Figure 4 The structural unit shown in the figure is used to analyze the temporal dependencies in the behavior sequence based on the time dimension, and then extract the behavioral change trend of the target object within the sliding time window, forming a short-term behavioral trend feature. This feature can be used to characterize the magnitude, direction, and fluctuation of recent behavior changes, enhancing the model's ability to perceive local temporal behavior changes.

[0096] Furthermore, the above short-term behavior trend features are input into the convolutional neural network structure, and the convolutional neural network can be Figure 4 The Cross Time CNN module is shown in the figure. This module is used to perform convolution sliding processing along the time axis to identify periodic patterns and structural changes in behavior across time periods (Cross Time). The Cross Time CNN module not only supports local pattern recognition in a single behavioral dimension, but also can jointly model multiple behavioral channels to achieve cross-perception between behavioral dimensions. The cross-period behavioral pattern features modeled and output by this module can effectively capture the repetitive patterns, rhythmic fluctuations, and abnormal aggregation trends of behaviors within monthly or larger cycles.

[0097] The above original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features will serve as the input basis for the subsequent feature fusion stage to generate a unified behavior representation with multi-time scale expression capabilities.

[0098] S204: Generate a multi-scale fusion behavior representation based on the original behavior sequence characteristics, short-term behavior trend characteristics, and cross-period behavior pattern characteristics;

[0099] Multi-scale fusion behavior representation: refers to the structured behavior representation vector generated by fusing original features, short-term trend features and periodic pattern features. It has dynamic information covering multiple time levels and can take into account global changes and rhythmic trends without losing detailed information. It is an important representation input for comprehensive modeling.

[0100] In an optional embodiment, the first representation path of the control target transaction processing model performs multi-scale feature fusion processing based on the multiple time-scale behavioral features extracted in the previous step, specifically including: taking the original behavior sequence features, short-term behavior trend features and cross-cycle behavior pattern features as targets for feature fusion, the feature fusion link performs structural consistency alignment processing on features at different time scales, and combines multiple groups of behavioral features according to a preset fusion strategy or adaptive weight mechanism to generate a multi-scale fusion behavior representation for uniformly describing the behavioral state of the target object.

[0101] Optionally, the feature fusion process may include, but is not limited to, concatenating feature dimensions to form a long vector representation, using an attention mechanism to determine the importance of features at different temporal levels and then performing weighted combination, or employing a feedforward network structure for nonlinear mapping enhancement. The resulting multi-scale fused behavioral representation has the ability to comprehensively express behavior across time scales and can serve as a high-level semantic input for joint representation modeling in subsequent model layers.

[0102] In a possible implementation, S204 may be performed as follows:

[0103] The original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features are fused and spliced ​​through the feature fusion layer of the first characterization path to obtain a multi-scale fused behavior representation.

[0104] Specifically, the feature fusion layer of the first characterization path receives the feature output results from the original behavior sequence feature extraction part, the short-term trend modeling part and the cross-period behavior pattern analysis part, and fuses and splices the above-mentioned original behavior sequence features, short-term behavior trend features and cross-period behavior pattern features according to a preset fusion method to form a multi-scale fusion behavior representation with a unified dimensional structure.

[0105] Optionally, the fusion and splicing operation may include vector merging along the feature dimension direction, and may also include feature normalization processing, dimension alignment operation or position embedding enhancement processing to ensure that the outputs of each feature path are comparable and combinable.

[0106] Through the above method, the feature fusion layer can unify and integrate the behavioral representation information covering different time levels at the structural level, thereby generating a multi-scale fused behavioral representation with complete semantics and rich feature information, providing a stable and high-quality representation basis for subsequent feature fusion modeling and risk assessment.

[0107] S206: Performing absolute position embedding coding on the original multi-dimensional behavior time series through the second representation path to obtain a global temporal behavior representation.

[0108] Absolute Positional Encoding (APE) maps the position of each time step in a behavior sequence (e.g., day 1, day 2, etc.) into vector-based temporal position information using a preset function or a trainable method. This information is then integrated with the original behavior data to enhance the model's ability to identify the temporal position of the behavior. Encoding methods include: fixed position functions (e.g., sin / cos), learnable embedding vector matrices, and timestamp normalization mapping.

[0109] Global temporal behavior representation: This refers to a unified behavior representation vector generated by modeling the complete behavior sequence structure after incorporating temporal position information. It is used to characterize the target object's behavioral state, evolutionary path, and its relationship to the temporal structure throughout the entire analysis cycle. This representation has the following characteristics: it represents the state changes of behavior distributed along the time axis, perceives the causal relationship and development trend of the behavior sequence, and supplements the lack of temporal semantics in local-scale modeling.

[0110] In an optional embodiment, the second representation path in the control target transaction processing model introduces time position information into the original multidimensional behavior time series for enhanced modeling, specifically including: generating a corresponding absolute time position signal based on the sequential index of each time step in the original behavior time series, and vectorizing the time position signal to obtain a corresponding absolute position embedding vector.

[0111] Subsequently, the absolute position embedding vector is combined with the behavior feature vector corresponding to each time step to form an enhanced behavior time slice representation sequence that incorporates temporal and positional semantics. This enhanced sequence, as the intermediate representation result of the second representation pathway, is input into the global temporal modeling structure. Global behavior sequence modeling is performed using a cross-time step information association mechanism, ultimately outputting a global temporal behavior representation that covers the entire timeline and is time-aware.

[0112] This processing enables the above target transaction processing model to have the ability to identify "when the behavior occurs" and "how the time structure affects the behavior performance", providing a time position alignment basis for subsequent behavior fusion and risk reasoning.

[0113] In one possible embodiment, Figure 5 As shown, Figure 5 This is a flow chart of absolute position embedding coding. The specific implementation of S206 can refer to the following method:

[0114] S3002: Absolute position encoding is performed on each time step in the original multidimensional behavior time series through the second representation pathway to obtain an absolute position encoding feature;

[0115] First, the second representation pathway performs absolute position encoding on each time step in the original multidimensional behavior time series to obtain corresponding absolute position encoding features. The absolute position encoding process can map the time step index to a fixed-dimensional vector representation using a preset position encoding function or a learnable embedding matrix based on the index value of each time step in the complete timeline.

[0116] Specifically, the position encoding method can be: position encoding based on sine and cosine functions, a learnable position embedding matrix based on training parameters, or a numerical embedding representation constructed using normalized timestamps;

[0117] The absolute position encoding feature is used to indicate the position semantics of each behavior time slice in the global time structure, which facilitates the model to perceive the specific time position information of the behavior.

[0118] S3004: Perform embedded coding based on the original multi-dimensional behavior time series and the absolute position coding features to obtain a global temporal behavior representation.

[0119] After obtaining the absolute position encoding features, an embedded coding process is performed based on the original multidimensional behavior time series and the position encoding features to generate a global temporal behavior representation with temporal structure awareness. The embedded coding process involves concatenating or fusing the behavior features of each time step with their corresponding absolute position encoding features and passing them as input to a global sequence modeling network (such as a time-aware Transformer or a temporal convolutional network).

[0120] Based on this embedded representation, the model performs global sequence modeling operations, models cross-time slice dependencies and performs semantic aggregation processing on the behavior-position composite features of all time steps, and outputs a global temporal behavior representation of unified dimensions.

[0121] Global temporal behavior representation can comprehensively reflect the time-sensitive behavior pattern of the target object within the complete behavior cycle, including the behavior development path, behavior mutations at key time points, delayed fluctuation structure, etc. It is an important behavior expression structure that complements multi-scale local features.

[0122] In this manual, by introducing an absolute position encoding mechanism, mapping the time step index into a structured position vector and jointly embedding it with the original behavior features, the model's ability to recognize and utilize temporal semantics can be significantly enhanced, so that the model not only focuses on the content of the behavior, but also perceives the time point of the behavior and its relative importance in the global time structure; further, through the global modeling structure, the fused behavior-position sequence is modeled across time slices, which can effectively capture the trend, stage and mutation characteristics of the behavior evolution over time, thereby improving the time perception ability and global structure understanding ability of behavior modeling, and providing a more stable, accurate and semantically complete global temporal behavior representation for subsequent behavior fusion and transaction risk assessment.

[0123] S208: Performing multi-channel attention fusion processing on the multi-scale fusion behavior representation and the global temporal behavior representation through the transformer network of the target transaction processing model to obtain an object comprehensive behavior representation.

[0124] Optionally, the transformer network has the following main structures: multi-head attention mechanism module, channel feature mapping and alignment structure, feedforward network and residual connection mechanism, hierarchical normalization and activation function processing unit

[0125] Multi-channel attention fusion processing: This refers to taking multiple feature sources (channels) as input, dynamically learning the contribution weight of each channel information to the fused representation through the attention mechanism, and achieving deep combination and structural fusion of inter-channel features under the guidance of attention. This multi-channel attention fusion processing enables the model to adaptively enhance the semantic expression of key pathways based on the actual distribution of behavioral features and suppress redundant or low-relevant features.

[0126] Optionally, fusion methods may include: unified attention calculation after splicing, channel-wise cross-attention modeling, channel-weighted averaging, or selective fusion using a gating mechanism;

[0127] The comprehensive behavior representation of an object refers to a unified high-dimensional feature vector generated by integrating the feature representations output by multiple structural modeling pathways and fusion processing by a transformer network to comprehensively describe the behavior state of the target object.

[0128] In an optional embodiment, a converter network module is configured in the control target transaction processing model, and the converter network is used to perform fusion modeling processing on the multi-scale fusion behavior representation output by the first representation path and the global temporal behavior representation output by the second representation path to generate a comprehensive behavior representation of the object.

[0129] Specifically, the transformer network is based on a multi-channel attention fusion mechanism, taking the multi-scale fusion behavior representation and the global temporal behavior representation as input. By constructing the attention association relationship between features, the feature components from different modeling paths are dynamically weighted and interactively modeled to achieve feature alignment and semantic fusion between information channels.

[0130] Optionally, the fusion process may include channel mapping, cross-attention calculation, self-attention enhancement, residual connections, and feature normalization. Ultimately, a unified, high-dimensional, comprehensive behavior representation of the object is output. This representation integrates local behavior pattern features with global temporal structure features, providing enhanced behavior discrimination and contextual understanding capabilities, serving as direct input for subsequent transaction risk assessment.

[0131] In this specification, by inputting the original multi-dimensional behavior time series into the first representation pathway and the second representation pathway respectively, and sequentially performing multi-time scale analysis, absolute position embedding coding, and attention fusion modeling between pathways, the temporal behavior characteristics of the target object can be comprehensively extracted from multiple dimensions including local behavior details, short-term trend changes, periodic patterns, and global time structures. The multi-path behavior representations are further fused through a transformer network to generate a comprehensive behavior representation of the object with high semantic expression capability, thereby significantly improving the modeling accuracy of the behavior evolution law and the perception of risk anomaly signals, and realizing higher-dimensional, stronger temporal dependence, and more explainable transaction risk assessment processing.

[0132] Optionally, for better feature modeling, the transaction behavior causal graph representation is also incorporated into the modeling process, thereby further enhancing the model's behavior understanding ability and transaction risk assessment performance. Figure 6 , Figure 6 This is a flow chart of another embodiment of a transaction risk assessment method proposed in one or more embodiments of this specification. Specifically:

[0133] In an improved embodiment of the present invention, the target transaction processing model not only includes a multi-time scale modeling pathway and a time-aware pathway for extracting temporal behavioral features, but also further introduces a transaction behavior causal relationship modeling mechanism to construct a graph representation reflecting the causal dependency structure between behavioral events. This is integrated into the comprehensive behavioral modeling process to enhance the modeling capabilities of risk causal structures. The processing flow includes the following steps:

[0134] S402: In a target transaction scenario, obtaining an original multidimensional behavior time series of a target object, and inputting the original multidimensional behavior time series into a target transaction processing model;

[0135] For details, please refer to the method steps in other embodiments of this specification, which will not be repeated here.

[0136] S404: Performing multi-time-scale analysis on the original multi-dimensional behavior time series using the target transaction processing model to obtain multiple reference-scale behavior time series features, fusing all of the reference-scale behavior time series features to obtain a multi-scale fused behavior representation; and performing location-aware encoding on the original multi-dimensional behavior time series using the target transaction processing model to obtain a global time series behavior representation.

[0137] For details, please refer to the method steps in other embodiments of this specification, which will not be repeated here.

[0138] S406: constructing a behavior causal relationship for the original multi-dimensional behavior time series using the target transaction processing model to obtain a transaction behavior causal graph representation;

[0139] In one feasible implementation, to enhance the target transaction processing model's ability to understand the causes of transaction behavior risks, the target transaction processing model includes a behavior causal graph construction module for performing causal relationship modeling. This module is used to construct a transaction behavior causal graph representation of the target object in the current transaction scenario based on the original multidimensional behavior time series. Specifically, the module performs dependency analysis on the behavioral event features in the original multidimensional behavior time series to construct a transaction behavior causal graph representation of the target object in the current transaction context. This process may include the following processing steps:

[0140] First, the behavioral event feature extraction and graph candidate generation are performed through the behavioral causal graph construction module. Specifically, each behavioral dimension in the original multidimensional behavioral time series is event-processed, and the behavioral features of each day or time step are parsed into structured behavioral event nodes. Each behavioral event node may include: behavior type (such as clicks, complaints, payment failures, etc.), timestamp, intensity (such as behavior value, frequency), trigger context (such as anomalies, login status, external signals). Based on the above behavioral event nodes, the initial candidate event graph structure is generated, and the candidate node set and initial edge set are constructed in the initial candidate event graph structure to describe the possible causal relationship paths.

[0141] Then, based on the initial event graph, the behavior causal graph construction module analyzes and models the causal relationship between behavioral events through the following processing flow:

[0142] (1) Timing dependency analysis process

[0143] Based on the time sequence, we analyze whether a certain type of event frequently triggers another type of event within a fixed lag time window (such as "the payment failure rate increases within 2 days after the information is modified") to generate a lag association relationship.

[0144] (2) Mutation response analysis and processing flow

[0145] Identify significant change points in behavioral events (such as a sudden increase in the number of complaints or a sudden drop in transaction frequency), trace the preceding events, and infer whether there is a behavioral inducement relationship.

[0146] (3) Interactive pattern extraction process

[0147] Combined with the frequency of event combinations within the sliding window, behavioral paths with high co-occurrence probability are identified, such as "continuous login failures → customer service consultation → complaint", and the behavioral causal chain is expressed in the form of a path diagram.

[0148] Furthermore, the transaction behavior causal graph structure is obtained by structural calibration based on the transaction scenario knowledge graph through the behavior causal graph construction module; specifically, in order to ensure the business rationality of the behavior causal graph, the transaction behavior domain knowledge base can be further introduced to filter, complete or adjust the node and edge relationships in the initial causal graph structure, including: deleting edges that do not meet the scenario constraints (such as non-strongly related behaviors), merging repeated events or equivalent nodes, introducing missing but key reasoning edges (such as indirect causal chains indicated by rules), and enhancing knowledge by introducing the transaction behavior domain knowledge base to ensure that the generated graph has transaction relevance, semantic consistency and explanation capabilities.

[0149] Finally, the behavior causal graph construction module converts the processed transaction behavior causal graph structure into a structured feature representation that can be used for subsequent representation fusion to obtain a transaction behavior causal graph representation. The transaction behavior causal graph representation can be: a graph embedding vector generated by a graph neural network, a path sequence encoding based on structure traversal, and a multi-perspective behavior path attention representation.

[0150] The transaction behavior causal graph representation is input into the subsequent representation fusion module as the supplementary pathway output and jointly modeled with the output of the time scale modeling path.

[0151] Through the above-mentioned construction and processing of behavioral causal relationships, the model can supplement the construction and learning of the logic of behavioral event occurrence, triggering mechanism and potential cause chain in the feature modeling process, make up for the limitation of traditional time series modeling that cannot reveal "why it happened", enhance the model's interpretability, reasoning ability and structural recognition ability, and lay a causal logic foundation for high-accuracy transaction risk assessment.

[0152] In one possible implementation, please refer to Figure 7 , Figure 7 This is a flowchart of establishing a causal relationship between behaviors. Specifically, the following methods can be used to implement S406:

[0153] S5002: Performing behavior change analysis on each event behavior feature of the original multidimensional behavior time series using the target transaction processing model to obtain multiple change trend interaction patterns, and determining temporal dependency information, hysteresis correlation information, and mutation response effect information between each event behavior feature based on the change trend interaction patterns;

[0154] Event behavior characteristics: refers to the evolution trajectory of a specific type of behavior (such as login, payment, complaint) in the time series.

[0155] Behavioral change analysis: Analyze the trends, fluctuations, and mutations of behavioral values ​​in time series.

[0156] Change trend interaction pattern: refers to the joint evolution relationship pattern of multiple behavioral characteristics over time.

[0157] Temporal dependency: A predecessor or response relationship between behavioral events in a temporal order.

[0158] Lag association: There is a significant delay in the impact of a previous behavioral event on a subsequent event.

[0159] Mutation response effect: collective synchronous changes in other behaviors caused by a behavioral mutation.

[0160] In an optional implementation, to accurately model the behavior causal structure, the behavior causal graph construction module in the control target transaction processing model first analyzes the behavioral changes of each event behavior feature in the original multidimensional behavior time series to explore the interactive dependency patterns between event behaviors and provide data support for subsequent causal structure construction. This step includes the target transaction processing model executing the following processing:

[0161] (1) The target transaction processing model uses a behavioral causal graph to build a module to analyze the behavioral changes of event behavior characteristics. "Event behavior characteristics" refers to the discrete behavioral observations collected on the time series dimension for a specific transaction object in the original multidimensional behavior time series. Each type of behavior can be regarded as a dimension, such as click volume, login status, number of transaction failures, user feedback, account status, etc.

[0162] Behavior change analysis can be specifically performed by constructing a behavior causal graph module based on the changing trend of each event behavior feature in the time series, and extracting the structure of its evolutionary features in the time series, including but not limited to:

[0163] Trend direction identification: determine whether a certain interval in the behavior sequence shows a continuous rise, continuous fall or oscillation state;

[0164] Mutation point detection: Identify abnormal surges or decreases in behavior values ​​within a short time window;

[0165] Behavior intensity change rate: Calculates the change ratio between adjacent time steps to measure the intensity of the behavior;

[0166] Stability analysis: Identify periods of time during which behavior maintains a stable state, used to mark reference behavior stages.

[0167] The behavior change analysis can be implemented through the behavior causal graph construction module based on the sliding window method, weighted difference method, moving average analysis, Z-score detection or other time series change detection algorithms.

[0168] (2) The target transaction processing model further realizes the formation of the change trend interaction pattern through the behavior causal graph construction module. This link is based on the single variable change trend of each event behavior feature. The behavior causal graph construction module further analyzes the interaction pattern between different behavior features in the time dimension and constructs multiple change trend interaction patterns, including but not limited to:

[0169] Synchronous change pattern: such as "click volume increases □ complaint volume increases";

[0170] Sequential trigger mode: such as "increase in failed logins → increase in customer service inquiries → increase in complaints";

[0171] Reverse change pattern: such as "decreased active days □ decreased payment behavior";

[0172] Delayed evolution pattern: For example, "the probability of transaction anomalies increases three days after the status changes."

[0173] The above interaction patterns are used to assist in identifying potential causal directions, time-dependent paths, and logical trigger chains.

[0174] (3) Based on the change trend interaction pattern, three types of causal dependency information are extracted, namely, temporal dependency relationship information, hysteresis correlation information, and mutation response effect information between each of the event behavior features. Specifically, based on the above multiple change trend interaction patterns, causal structure candidate information is extracted from the following three dimensions:

[0175] a. Temporal dependency information: This refers to the temporal sequence between two events, where the probability of the preceding event significantly increases the probability of the subsequent event. This information is obtained through behavior timestamp comparison and joint probability analysis, and is used to identify potential causal leading paths.

[0176] b. Lagged Correlation Information: This refers to the time delay between the impact of prior behavioral characteristics on subsequent behaviors, i.e., a delayed response. For example, the probability of "user authentication information modification" triggering "fund freeze" increases significantly two days later. This can be identified through methods such as sliding window cross-correlation analysis and Granger causality analysis.

[0177] c. Mutation-response effect information: This refers to the pattern of simultaneous short-term responses from multiple other behavioral variables triggered by a sudden change in a specific behavioral trait. This information is extracted through statistical methods based on the aggregation of mutation points and accompanying events, and is often used to identify key triggering events or abnormal triggers.

[0178] In this manual, by executing the above-mentioned behavior change analysis and trend interaction modeling, it is possible to accurately identify the potential temporal logical relationship and trigger mechanism between behavioral events, providing structural prior information for the construction of causal graphs. This not only improves the depth of behavior understanding, but also significantly enhances the accuracy of causal modeling and event correlation reasoning capabilities, laying a high-quality graph structure foundation for subsequent transaction risk causal analysis.

[0179] S5004: Constructing an initial behavior causal graph structure based on the temporal dependency information, hysteresis correlation information, and mutation response effect information among the event behavior features;

[0180] Initial behavior causal graph structure: structured graph data constructed based on causal candidate information between behaviors, reflecting the potential causal chain of each behavior node of the target object in the transaction behavior sequence.

[0181] Node: A unit entity in a graph used to represent a specific behavior event.

[0182] Directed Edge: A line that indicates that a certain behavioral event has a causal or triggering effect on another event and has directionality.

[0183] Causal edge attributes: optional parameters such as edge weight, delay time, causal strength, and confidence.

[0184] In one optional implementation, the target transaction processing model constructs an initial behavioral causal graph structure for the target object within the current transaction cycle based on the causal candidate information extracted during the behavior change analysis. This processing step aims to explicitly express the potential causal relationships between behaviors in the form of a graph structure, serving as the basis for subsequent graph optimization and feature embedding. This step includes the following processing:

[0185] (1) The target transaction processing model first initializes the graph structure and defines nodes. Specifically, based on the characteristics of various behavioral events in the original multidimensional behavioral time series, a node set of the causal graph is constructed. Each node corresponds to a behavioral event type with causal potential. The node definition may include but is not limited to:

[0186] User behavior events: such as "login failed", "data modification", and "payment successful";

[0187] System status events: such as "freeze notification" and "credit rating change";

[0188] Risk events: such as "complaint submission", "fraud determination", and "blacklist marking".

[0189] Each node may contain attribute characteristics such as its behavior type identification, average occurrence frequency, and abnormal trigger probability.

[0190] (2) The target transaction processing model then constructs causal edge relationships to obtain an initial behavioral causal graph structure. Based on the three types of causal relationship information identified in step S5002, namely, timing dependency information, hysteresis correlation information, and mutation response effect information, the causal edges between nodes are defined and connected.

[0191] The construction principles are as follows:

[0192] If the occurrence time of node A is always earlier than that of node B, and the occurrence of A significantly increases the probability of the occurrence of B, then a directed edge from A to B is established and marked as "temporal dependency";

[0193] If the trigger probability of node B increases within a specific time delay window after the occurrence of node A, a "hysteresis edge" is established from A to B, which can have a time delay parameter.

[0194] If node A is a mutation trigger point and is activated simultaneously with multiple nodes B, C, and D, then a "mutation response edge" is established from A to multiple response nodes.

[0195] At the same time, the attribute fields of the edge may include: causal weight (such as based on correlation score), lag time, response amplitude, etc.

[0196] Furthermore, the target transaction processing model determines the graph structure organization form:

[0197] The initial causal graph is a directed graph structure that supports the following graph organization features:

[0198] Single-starting point, multi-trigger path structure: supports a single action triggering multi-path diffusion (e.g., "data modification" → "account abnormality" + "authentication failure");

[0199] Multi-node common cause path structure: supports the aggregation of multiple behaviors to trigger a single behavior (e.g., "high-frequency login + short-term payment failure" → "risk flag");

[0200] Circular dependency graph structure: Weak cyclic structures may exist in complex behaviors, and the system can set a maximum loop depth to prevent information leakage.

[0201] Furthermore, after the initial causal graph structure is constructed, it can be used as an intermediate modeling product for: graph optimization processing in subsequent S5006, behavior graph embedding representation generation, and graph neural network causal reasoning modeling.

[0202] Furthermore, the graph structure can be stored using an adjacency matrix, adjacency list, or graph database structure, supporting multi-path mapping, weighted edge mapping, and semantic annotation mapping.

[0203] S5006: Calling the transaction behavior domain knowledge base to adjust the node relationship of the candidate behavior causal path in the initial behavior causal graph structure to obtain a transaction behavior causal graph representation.

[0204] Transaction behavior domain knowledge base: refers to a knowledge system composed of expert rules, scenario specifications and behavior chain templates, which is used to calibrate nodes, paths and dependencies in the causal modeling process.

[0205] Candidate causal behavior path: refers to the set of behavioral causal edges generated in the initial causal graph and not verified by business semantics.

[0206] Calibrated causal modeling: refers to the process of correcting, completing, adjusting and semantically unifying the causal graph structure based on knowledge resources.

[0207] Transaction behavior causal graph representation: refers to the feature representation that can be processed by machines after the causal graph structure after structural optimization is embedded or modeled.

[0208] In an optional implementation, to further improve the business accuracy, semantic consistency, and transaction relevance of behavior causal modeling, the target transaction processing model, after constructing the initial behavior causal graph structure, calls the transaction behavior domain knowledge base to perform structural optimization and semantic calibration on the candidate behavior causal paths in the graph structure, thereby generating a transaction behavior causal graph representation with high credibility and scenario constraint consistency. This processing step may include the following operations:

[0209] (1) Pre-construct a transaction behavior domain knowledge base, which is a pre-defined structured knowledge resource library used to store common behavior node types, legal behavior trigger paths, constraint rules, and causal dependency templates in business scenarios. Its contents may include: behavior type specifications and classification systems (such as "authentication type", "risk type", and "transaction type" events); legal behavior flow rules (such as "data modification" cannot directly trigger "freeze" but can trigger "authentication retry"); risk trigger templates and response mechanisms (such as "continuous abnormal login → customer service consultation → complaint" is a high-risk combination); node semantic ontology structure, used to identify node synonym relationships, hierarchical relationships, etc.

[0210] The transaction behavior domain knowledge base can be established based on manual construction by business experts, rule mining, graph learning, etc.

[0211] (2) In practical applications, after obtaining the initial behavior causal graph structure, the model conducts structural review and relationship adjustment on the candidate behavior causal paths in the graph based on the above-mentioned knowledge base content, including:

[0212] a. Node legitimacy verification and correction operations: Identify node pairs with semantic conflicts in the graph (e.g., "Payment successful" → "Blacklist mark"); invoke the constraint rules defined in the knowledge base to delete edges that do not conform to the business process logic; and unify nodes with irregular naming and ambiguous semantics (e.g., unifying "identity re-verification" and "authentication retry" into a "secondary authentication" node).

[0213] b. Causal edge weight adjustment operation: Update the causal strength parameters of the edges based on the business priority, risk score, etc. of different paths in the knowledge base; if some paths are marked as "high-confidence causal chains" by the knowledge base, their edge weights are increased or their priorities are retained.

[0214] c. Behavior path completion operation: If a key intermediate node is missing in the initial diagram, the system can call the standard behavior chain template provided by the knowledge base to complete the path; for example, if the path from "data modification" directly to "freeze" is incomplete, intermediate nodes such as "secondary authentication" or "transaction limit adjustment" can be inserted.

[0215] (3) Finally, a causal graph representation of transaction behavior is generated. After the above structural calibration and semantic enhancement, the causal graph structure constitutes the final transaction behavior causal graph. To support subsequent feature fusion and risk assessment modeling, the system converts the graph structure into an embeddable representation form, including but not limited to: graph embedding vectors (generated by graph neural networks such as GCN and GAT), path attention representation (generated by weighted learning of high-frequency causal paths), and graph structure summary (such as graph structure statistical features, key subgraph pattern encoding, etc.).

[0216] Furthermore, the obtained transaction behavior causal graph representation will be used as the third type of feature input to participate in the joint modeling of multi-scale fusion behavior representation and global temporal behavior representation, thereby improving the causal explanation ability and structural controllability in transaction modeling.

[0217] In this specification, by performing behavioral change analysis, causal dependency information mining, graph structure construction and domain knowledge enhancement processing on the original multi-dimensional behavioral time series, the method can dynamically construct a transaction behavior causal graph with time dependency, delayed responsiveness and trigger mechanism explainability based on the behavior of the target object; then, in the behavior modeling process, graph structure expression and business knowledge fusion are introduced, so that the model can not only perceive "what happened" and "when it happened", but also identify "why it happened" and "how it evolved", which significantly improves the causal reasoning ability, structural understanding ability and risk identification accuracy of behavior modeling, and provides more explanatory and discriminative structural feature support for subsequent behavior representation fusion and transaction risk assessment.

[0218] S408: The multi-scale fusion behavior representation, the global temporal behavior representation, and the transaction behavior causal graph representation are fused through the target transaction processing model to obtain an object comprehensive behavior representation.

[0219] Schematically, through the transformer fusion module, the model feeds multi-scale fused behavior representations, global temporal behavior representations, and newly added transaction behavior causal graph representations into a multi-channel attention mechanism. This performs cross-pathway feature interaction modeling and semantic fusion processing to generate a unified structure for comprehensive object behavior representations. This fusion structure dynamically adjusts the importance weights of different path features based on semantic dependencies, effectively integrating the temporal evolution path of the behavior, the location of the behavior, and the underlying causal structure.

[0220] S410: Performing transaction risk assessment based on the object comprehensive behavior representation to obtain an object transaction risk assessment result for the target object.

[0221] For details, please refer to the method steps in other embodiments of this specification, which will not be repeated here.

[0222] In this manual, the transaction behavior causal graph modeling structure is introduced to enable the target transaction processing model to construct a risk cause landscape from the event relationship structure level, making up for the limitations of pure time series feature modeling in behavior interpretability and trigger mechanism understanding; by integrating the three types of representation information of "behavior evolution process + time position structure + behavior causal mechanism", the model's risk characterization ability, judgment accuracy and interpretability in high-dimensional complex transaction scenarios are greatly improved.

[0223] The following will be combined Figure 8 , the transaction risk assessment processing device provided in this specification is introduced in detail. It should be noted that, Figure 8 The transaction risk assessment processing device shown is used to execute this instruction Figures 1 to 7 For the convenience of explanation, only the parts related to this specification are shown. For the specific technical details not disclosed, please refer to this specification. Figures 1 to 7 The embodiment shown.

[0224] See Figure 8 , which shows a schematic diagram of the structure of the transaction risk assessment processing device of this specification. The transaction risk assessment processing device 1 can be implemented as all or part of the user electronic device through software, hardware, or a combination of both. According to some embodiments, the transaction risk assessment processing device 1 includes a data input module 11, a model processing module 12, and a risk assessment module 13, which are specifically used to:

[0225] The data input module 11 is used to obtain the original multi-dimensional behavior time series of the target object in the target transaction scenario, and input the original multi-dimensional behavior time series into the target transaction processing model;

[0226] The model processing module 12 is configured to perform multi-time-scale analysis on the original multi-dimensional behavior time series using the target transaction processing model to obtain multiple reference-scale behavior time series features, fuse all the reference-scale behavior time series features to obtain a multi-scale fused behavior representation; and perform position-aware encoding on the original multi-dimensional behavior time series using the target transaction processing model to obtain a global time series behavior representation.

[0227] The risk assessment module 13 is used to fuse the multi-scale fusion behavior representation and the global temporal behavior representation through the target transaction processing model to obtain an object comprehensive behavior representation, and perform transaction risk assessment processing based on the object comprehensive behavior representation to obtain an object transaction risk assessment result for the target object.

[0228] In a feasible implementation, the target transaction processing model includes a first representation path and a second representation path, see Figure 9 , Figure 9 This is a structural diagram of a model processing module, wherein the model processing module 12 includes:

[0229] A first characterization unit 121 is configured to perform multi-time-scale analysis on the original multidimensional behavior time series through the first characterization path to obtain original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features, and generate a multi-scale fused behavior representation based on the original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features; and

[0230] The second representation unit 122 is configured to perform absolute position embedding coding on the original multi-dimensional behavior time series through the second representation path to obtain a global temporal behavior representation.

[0231] In a feasible implementation, performing multi-time-scale analysis on the original multidimensional behavior time series through the first representation pathway to obtain original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features, and generating a multi-scale fused behavior representation based on the original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features, includes:

[0232] Performing feature change processing on the daily behavior data in the original multidimensional behavior time series through the first characterization pathway to obtain original behavior sequence features, performing temporal dependency analysis on the original behavior sequence features of the daily time dimension through multiple long short-term memory neural networks in the first characterization pathway to obtain short-term behavior trend features, and performing cross-month behavior pattern analysis on the short-term behavior trend features through the convolutional neural network in the first characterization pathway to obtain cross-cycle behavior pattern features;

[0233] The original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features are fused and spliced ​​through the feature fusion layer of the first characterization path to obtain a multi-scale fused behavior representation.

[0234] In a feasible implementation manner, performing absolute position embedding encoding on the original multidimensional behavior time series through the second representation path to obtain a global temporal behavior representation includes:

[0235] Absolute position encoding is performed on each time step in the original multidimensional behavior time series through the second representation pathway to obtain an absolute position encoding feature;

[0236] Embedding coding is performed based on the original multi-dimensional behavior time series and the absolute position coding features to obtain a global temporal behavior representation.

[0237] In a feasible implementation, the step of fusing the multi-scale fusion behavior representation and the global temporal behavior representation through the target transaction processing model to obtain the object comprehensive behavior representation includes:

[0238] The transformer network of the target transaction processing model performs multi-channel attention fusion processing on the multi-scale fusion behavior representation and the global temporal behavior representation to obtain the object comprehensive behavior representation.

[0239] In a feasible embodiment, the device is further used for:

[0240] Constructing a behavioral causal relationship for the original multi-dimensional behavior time series through the target transaction processing model to obtain a transaction behavior causal graph representation;

[0241] The step of fusing the multi-scale fusion behavior representation and the global temporal behavior representation through the target transaction processing model to obtain the object comprehensive behavior representation includes:

[0242] The multi-scale fusion behavior representation, the global temporal behavior representation and the transaction behavior causal graph representation are fused through the target transaction processing model to obtain the object comprehensive behavior representation.

[0243] In a feasible implementation, performing behavior causal relationship analysis on the original multi-dimensional behavior time series using the target transaction processing model to obtain multiple pieces of behavior causal dependency information, and constructing a transaction behavior causal graph representation based on the behavior causal dependency information, includes:

[0244] Performing behavioral change analysis on each event behavior feature of the original multidimensional behavior time series using the target transaction processing model to obtain multiple change trend interaction patterns, and determining temporal dependency relationship information, hysteresis association information, and mutation response effect information between each event behavior feature based on the change trend interaction patterns;

[0245] Constructing an initial behavior causal graph structure based on the temporal dependency information, hysteresis correlation information, and mutation response effect information among the event behavior characteristics;

[0246] The transaction behavior domain knowledge base is called to adjust the node relationship of the candidate behavior causal path in the initial behavior causal graph structure to obtain a transaction behavior causal graph representation.

[0247] It should be noted that the transaction risk assessment and processing device provided in the above embodiment, when executing the transaction risk assessment and processing method, only uses the division of the above functional modules as an example. In actual application, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the transaction risk assessment and processing device provided in the above embodiment and the transaction risk assessment and processing method embodiment are based on the same concept. The implementation process is detailed in the method embodiment and will not be repeated here.

[0248] The above serial numbers in this specification are for description only and do not represent the advantages or disadvantages of the embodiments.

[0249] This specification also provides a computer storage medium, which can store multiple instructions, which are suitable for being loaded and executed by a processor as described above. Figures 1 to 7 The transaction risk assessment and processing method of the embodiment shown in the figure can be found in the specific execution process. Figures 1 to 7 The detailed description of the illustrated embodiment will not be repeated here.

[0250] This specification also provides a computer program product, which stores at least one instruction, and the at least one instruction is loaded and executed by the processor as described above. Figures 1 to 7 The transaction risk assessment and processing method of the embodiment shown in the figure can be found in the specific execution process. Figures 1 to 7 The detailed description of the illustrated embodiment will not be repeated here.

[0251] Please refer to Figure 10 , which is a block diagram of the structure of an electronic device provided in an embodiment of this specification. The electronic device described in this specification may include one or more of the following components: a processor 1010, a memory 1020, an input device 1030, an output device 1040, and a bus 1050. The processor 1010, memory 1020, input device 1030, and output device 1040 may be connected via a bus 1050.

[0252] The processor 1010 may include one or more processing cores. The processor 1010 utilizes various interfaces and circuits to connect various components within the electronic device. It executes instructions, programs, code sets, or instruction sets stored in the memory 1020, as well as accesses data stored in the memory 1020, to perform various functions of the electronic device and process data. Optionally, the processor 1010 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field-programmable gate array (FPGA), or a programmable logic array (PLA). The processor 1010 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and application programs; the GPU is responsible for rendering and drawing display content; and the modem handles wireless communications. It is understood that the modem may not be integrated into the processor 1010 and may be implemented separately via a communications chip.

[0253] The memory 1020 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory 1020 includes a non-transitory computer-readable storage medium. The memory 1020 may be used to store instructions, programs, codes, code sets, or instruction sets.

[0254] The input device 1030 is used to receive input commands or data and includes, but is not limited to, a keyboard, mouse, camera, microphone, or touch-screen device. The output device 1040 is used to output commands or data and includes, but is not limited to, a display device and a speaker. In the embodiments of this specification, the input device 1030 may be a temperature sensor for obtaining the operating temperature of the electronic device. The output device 1040 may be a speaker for outputting audio signals.

[0255] In addition, those skilled in the art will understand that the structures of the electronic devices shown in the above figures do not limit the electronic devices. The electronic devices may include more or fewer components than shown, or may combine certain components or arrange the components differently. For example, the electronic devices may also include radio frequency circuits, input units, sensors, audio circuits, wireless fidelity (WIFI) modules, power supplies, Bluetooth modules, and other components, which are not described in detail here.

[0256] In the embodiments of this specification, the execution entity of each step can be the electronic device described above. Optionally, the execution entity of each step is the operating system of the electronic device. The operating system can be Android, iOS, or other operating systems, and this embodiment of this specification does not limit this.

[0257] exist Figure 10 In the electronic device, the processor 1010 can be used to call the program stored in the memory 1020 and execute it to implement the transaction risk assessment and processing method as described in the various method embodiments of this specification.

[0258] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory, or a random access memory.

[0259] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, storage, and display, etc.), and signals involved in the embodiments of this specification are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the original multi-dimensional behavioral time series involved in this specification were all obtained with full authorization.

[0260] The above disclosure is only a preferred embodiment of this specification, and certainly cannot be used to limit the scope of rights of this specification. Therefore, equivalent changes made according to the claims of this specification are still within the scope covered by this specification.

Claims

1. A transaction risk assessment and processing method, the method comprising: In a target transaction scenario, obtaining an original multidimensional behavior time series of a target object, and inputting the original multidimensional behavior time series into a target transaction processing model; Performing multi-time-scale analysis on the original multi-dimensional behavior time series using the target transaction processing model to obtain multiple reference-scale behavior time series features, and fusing all the reference-scale behavior time series features to obtain a multi-scale fused behavior representation; and, performing position-aware encoding on the original multidimensional behavior time series through the target transaction processing model to obtain a global temporal behavior representation; The multi-scale fusion behavior representation and the global temporal behavior representation are characterized and fused through the target transaction processing model to obtain an object comprehensive behavior representation, and transaction risk assessment processing is performed based on the object comprehensive behavior representation to obtain an object transaction risk assessment result for the target object.

2. The method according to claim 1, wherein the target transaction processing model comprises a first representation path and a second representation path, performing multi-time-scale analysis on the original multi-dimensional behavior time series using the target transaction processing model to obtain multiple reference-scale behavior time series features, and fusing all the reference-scale behavior time series features to obtain a multi-scale fused behavior representation; Furthermore, performing position-aware encoding on the original multidimensional behavior time series through the target transaction processing model to obtain a global temporal behavior representation, including: Performing multi-time-scale analysis on the original multidimensional behavior time series through the first representation pathway to obtain original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features, and generating a multi-scale fused behavior representation based on the original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features; as well as, The original multi-dimensional behavior time series is subjected to absolute position embedding encoding through the second representation path to obtain a global temporal behavior representation.

3. The method according to claim 2, wherein the first characterization pathway is used to perform multi-time-scale analysis on the original multidimensional behavior time series to obtain original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features, and the multi-scale fused behavior representation is generated based on the original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features, including: Performing feature change processing on the daily behavior data in the original multidimensional behavior time series through the first characterization pathway to obtain original behavior sequence features, performing temporal dependency analysis on the original behavior sequence features of the daily time dimension through multiple long short-term memory neural networks in the first characterization pathway to obtain short-term behavior trend features, and performing cross-month behavior pattern analysis on the short-term behavior trend features through the convolutional neural network in the first characterization pathway to obtain cross-cycle behavior pattern features; The original behavior sequence features, short-term behavior trend features, and cross-period behavior pattern features are fused and spliced ​​through the feature fusion layer of the first characterization path to obtain a multi-scale fused behavior representation.

4. The method according to claim 2, wherein the step of performing absolute position embedding encoding on the original multidimensional behavior time series through the second representation path to obtain a global temporal behavior representation comprises: Absolute position encoding is performed on each time step in the original multidimensional behavior time series through the second representation pathway to obtain an absolute position encoding feature; Embedding coding is performed based on the original multi-dimensional behavior time series and the absolute position coding features to obtain a global temporal behavior representation.

5. The method according to claim 1, wherein the step of fusing the multi-scale fusion behavior representation and the global temporal behavior representation using the target transaction processing model to obtain the object comprehensive behavior representation comprises: The transformer network of the target transaction processing model performs multi-channel attention fusion processing on the multi-scale fusion behavior representation and the global temporal behavior representation to obtain the object comprehensive behavior representation.

6. The method according to claim 1, further comprising: Constructing a behavioral causal relationship for the original multi-dimensional behavior time series through the target transaction processing model to obtain a transaction behavior causal graph representation; The step of fusing the multi-scale fusion behavior representation and the global temporal behavior representation through the target transaction processing model to obtain the object comprehensive behavior representation includes: The multi-scale fusion behavior representation, the global temporal behavior representation and the transaction behavior causal graph representation are fused through the target transaction processing model to obtain the object comprehensive behavior representation.

7. The method according to claim 6, wherein the step of performing behavioral causal relationship analysis on the original multidimensional behavior time series using the target transaction processing model to obtain multiple pieces of behavioral causal dependency information, and constructing a transaction behavior causal graph representation based on the behavioral causal dependency information, comprises: Performing behavioral change analysis on each event behavior feature of the original multidimensional behavior time series using the target transaction processing model to obtain multiple change trend interaction patterns, and determining temporal dependency relationship information, hysteresis association information, and mutation response effect information between each event behavior feature based on the change trend interaction patterns; Constructing an initial behavior causal graph structure based on the temporal dependency information, hysteresis correlation information, and mutation response effect information among the event behavior characteristics; The transaction behavior domain knowledge base is called to adjust the node relationship of the candidate behavior causal path in the initial behavior causal graph structure to obtain a transaction behavior causal graph representation.

8. A transaction risk assessment and processing device, comprising: A data input module is used to obtain the original multidimensional behavior time series of the target object in the target transaction scenario, and input the original multidimensional behavior time series into the target transaction processing model; a model processing module configured to perform multi-time-scale analysis on the original multidimensional behavior time series using the target transaction processing model to obtain multiple reference-scale behavior time series features, and fuse all the reference-scale behavior time series features to obtain a multi-scale fused behavior representation; and, performing position-aware encoding on the original multidimensional behavior time series through the target transaction processing model to obtain a global temporal behavior representation; The risk assessment module is used to fuse the multi-scale fusion behavior representation and the global temporal behavior representation through the target transaction processing model to obtain an object comprehensive behavior representation, and perform transaction risk assessment processing based on the object comprehensive behavior representation to obtain an object transaction risk assessment result for the target object.

9. A computer storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the method steps according to any one of claims 1 to 7.

10. A computer program product, wherein the computer program product stores at least one instruction, wherein the at least one instruction is loaded by a processor and executes the method steps according to any one of claims 1 to 7.

11. An electronic device comprising: A processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method steps according to any one of claims 1 to 7.

Images