Multi-dimensional errand running platform risk early warning system for delivery services

Abstract

The application relates to the technical field of risk early warning of running platforms, and discloses a multi-dimensional risk early warning system for a running platform of a distribution service, which comprises the following modules: a space-time index modeling module, which is used for collecting original data of the distribution service, mapping positions to spatial grid units, and constructing a heterogeneous graph; a reachable domain construction module, which is used for calculating the shortest path based on a road network and generating a distribution reachable domain; a path consistency module, which is used for obtaining an observation track and generating a path consistency deviation index; a collusion loop module, which is used for extracting an induced subgraph and generating a collusion loop index; a two-way reinforcement module, which is used for performing conditional tightening when the path deviation or collusion risk reaches a threshold value; a hierarchical treatment module, which is used for generating a risk level and outputting a deterministic treatment instruction; and a solidification playback module, which is used for solidifying multi-dimensional data to form evidence data blocks. The application realizes accurate identification, hierarchical early warning and traceable verification of multi-agent and multi-level risks in the distribution service.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of risk early warning of running legs platform, and particularly relates to a multi-dimensional risk early warning system for a running legs platform of a delivery service. BACKGROUND

[0002] With the popularity of instant delivery and running legs services, delivery platforms are increasingly important in urban logistics systems. Currently, platforms form a complex multi-agent business network through a large number of registered riders, merchants and users. However, in the high-frequency, dispersed and dynamic delivery scenarios, platform operation faces many risk problems, including fake orders, rider detours, account sharing, device sharing, malicious collusion and abnormal order settlement, etc. These risk behaviors are often hidden in massive spatio-temporal data and business interactions, and traditional monitoring methods based on rules or single-dimensional features cannot effectively identify and warn.

[0003] In the prior art, some delivery platforms attempt to identify suspicious orders through path deviation detection or abnormal trajectory analysis, but such methods usually rely on fixed thresholds or single spatial features, and cannot adapt to the multi-dimensional changes of delivery tasks under different time, space and network conditions. At the same time, for complex risk behaviors such as multi-account collusion, device sharing and fake task concatenation, there is a lack of means to depict the implicit relationship between agents from the global graph structure level. In addition, existing systems mostly stay in the post-check stage, lacking a mechanism for real-time early warning and closed-loop tracing in the risk generation stage, resulting in lagging risk identification and low disposal efficiency. SUMMARY

[0004] The application provides a multi-dimensional risk early warning system for a running legs platform of a delivery service, which solves the technical problems of single risk identification dimension, inability to simultaneously consider path abnormalities and multi-agent collusion relationships, fixed early warning threshold and lack of self-adaptive mechanism, and lack of risk result traceable verification capability in related technologies.

[0005] The application provides a multi-dimensional risk early warning system for a running legs platform of a delivery service, which includes:

[0006] A spatio-temporal index modeling module for collecting delivery service raw data, mapping locations to spatial grid cells, discretizing time into time buckets, building a heterogeneous graph, and labeling pickup and delivery grids; wherein the delivery service raw data includes order information, user account, rider account, merchant account, device and network identifier, location and time;

[0007] A reachable domain construction module for calculating the shortest path between the pickup and delivery grids based on the road network, projecting the shortest path as a grid sequence, and expanding the delivery reachable domain according to the neighborhood layers;

[0008] a path consistency module configured to obtain an observation trajectory, form an observation trajectory grid set and an actual trajectory length, and generate a path consistency deviation index within a delivery reachable domain;

[0009] a collusion loop module configured to extract an induced subgraph from the heterogeneous graph, enumerate simple closed loops, filter effective loops, and generate a collusion loop index;

[0010] a bidirectional reinforcement module configured to tighten an effective loop filtering condition when the path consistency deviation index reaches a first preset threshold, and tighten a trajectory verification condition when the collusion loop index reaches a second preset threshold;

[0011] a hierarchical treatment module configured to grade orders, generate a risk level, and output a deterministic process treatment instruction for performing order settlement, account and device verification, and trajectory playback verification;

[0012] a solidification playback module configured to solidify the delivery reachable domain, the observation trajectory grid set, the shortest path, the path consistency deviation index, the collusion loop index, the effective loop, and the number of adjacent layers, and form a solidified evidence data block.

[0013] Further, the spatiotemporal index modeling module comprises:

[0014] mapping the location to a spatial grid unit at a fixed resolution, discretizing the time information into time buckets at a fixed time granularity, and establishing a spatial grid index and a time bucket index corresponding to each data record;

[0015] constructing a heterogeneous graph with user accounts, rider accounts, merchant accounts, device and network identifiers, spatial grid units, and time buckets as nodes, and the association relationship between orders, the binding relationship between accounts and devices or networks, and the residence relationship of accounts in a specific spatial grid unit as edges, and labeling the pickup grid and the delivery grid corresponding to each order in the heterogeneous graph.

[0016] Further, the generation of the delivery reachable domain comprises:

[0017] Step 11. Selecting road network anchor points that are spatially adjacent to the pickup grid and the delivery grid in the road network as the starting point and the ending point, taking the weighted sum of the road segment length, the turning cost, and the road level cost as the path weight, calculating the total path weight of all candidate paths using the shortest path algorithm, and selecting the path with the smallest total path weight as the shortest path;

[0018] Step 12. Sampling the shortest path at a fixed interval related to the spatial grid unit resolution to obtain an ordered coordinate point list, mapping each coordinate point to a spatial grid unit, and generating a grid sequence by removing duplicates in the order of appearance, labeling the pickup grid and the delivery grid at the beginning and end of the grid sequence, and establishing an index binding between each spatial grid unit and the corresponding road segment identifier;

[0019] Step 13, centering each spatial grid cell in the grid sequence, performing a hexagonal grid-based neighborhood query according to a preset neighborhood layer number, expanding layer by layer to obtain an adjacent cell set, and performing a union operation on all adjacent cell sets to form a delivery reachable domain.

[0020] Further, an observation trajectory is obtained, an observation trajectory grid set is formed, and an actual trajectory length is obtained, a path consistency deviation index is generated within the delivery reachable domain, including:

[0021] Step 21, time sequence verification and sampling frequency unification are performed on the observation trajectory, map matching is performed according to the road network, the observation trajectory points are matched to the road network on which the shortest path is located according to the shortest distance and continuous mileage principle, a matched trajectory sequence is obtained, and a matching mark is recorded for each trajectory segment;

[0022] Step 22, the matched trajectory points are mapped to spatial grid cells one by one, an observation trajectory grid set is formed by removing duplicates according to the order of appearance, and an actual trajectory length is obtained by accumulating the lengths of the matching road segments formed by each trajectory point;

[0023] Step 23, under the constraint of the delivery reachable domain, the coverage deficiency ratio, the off-road segment ratio, and the abnormal stay time ratio are calculated, wherein the coverage deficiency ratio is obtained by the complementary value of the intersection area ratio of the observation trajectory grid set and the delivery reachable domain, the off-road segment ratio is the ratio of the number of non-shortest path segments to the total number of trajectory segments, and the abnormal stay time ratio is the ratio of the low-speed and same-cell stay time to the promised delivery time;

[0024] Step 24, the over-distance ratio is calculated according to the shortest path length and the actual trajectory length, and the over-distance ratio, the coverage deficiency ratio, the off-road segment ratio, and the abnormal stay time ratio are linearly weighted according to fixed weights to obtain the path consistency deviation index.

[0025] Further, an induced subgraph is extracted from the heterogeneous graph, simple closed loops are enumerated and effective loops are screened, and a collusion loop index is generated, including:

[0026] Step 31, a sliding window is set in time sequence, covering a fixed number of time buckets, nodes and edges with time attributes belonging to the sliding window are screened in the heterogeneous graph to form an induced subgraph;

[0027] Step 32, a simple closed loop enumeration algorithm is executed in the induced subgraph, a closed loop set is enumerated, and whether each closed loop contains at least two types of business subject nodes is determined according to the node type, and device sharing identifiers and spatiotemporal co-occurrence identifiers are searched in the edge attribute and are marked as device sharing identifier values and spatiotemporal co-occurrence identifier values, respectively; wherein the business subject nodes include: user account, rider account and merchant account;

[0028] Step 33, validity judgment is performed on the enumerated closed loop set, and only when the closed loop contains at least two types of business subject nodes and the device sharing identifier value and the space-time co-occurrence identifier value are both 1, the closed loop is defined as an effective ring, and an effective ring weight inversely proportional to the ring length is set for each effective ring to form an effective ring set and a corresponding weight set;

[0029] Step 34, the effective ring weights in the effective ring set are weighted and summed, and the collusion loop index is calculated by taking the order quantity in the sliding window as the normalization coefficient.

[0030] Further, the collusion loop module further includes loop coupling degree calculation on the effective ring set, including: selecting any two effective rings from the effective ring set, respectively counting the number of shared nodes and the number of shared edges of the two effective rings, and recording the number of nodes and edges of each pair of effective rings; the ratio of the number of shared nodes to the average of the number of nodes of the two effective rings is taken as the loop coupling degree;

[0031] When the two effective rings have shared edges, the loop coupling degree is corrected according to the number of shared edges, the ratio of the number of shared edges of the two rings to the smaller value of the number of edges in the two rings is taken as the edge overlap ratio, and a correction amount is obtained by multiplying a preset correction coefficient, the correction amount is added to the original loop coupling degree to obtain the corrected loop coupling degree, and if the corrected loop coupling degree exceeds a preset coupling threshold, the two rings are determined as strongly associated rings, and the shared nodes are marked as high association nodes.

[0032] Further, it further includes a regional aggregation analysis module, which is used to aggregate multiple spatial grid units to form a geographical region group based on spatial adjacency relationship, and calculate the regional average values of the path consistency deviation index and the collusion loop index in each geographical region group, and calculate the average values of all geographical region groups as global average values, and when the difference between the average value of the path consistency deviation index or the average value of the collusion loop index of any geographical region group and the corresponding global average value exceeds a preset regional deviation threshold, the geographical region group is marked as a regional high-risk area.

[0033] Further, the bidirectional reinforcement module includes:

[0034] Step 41, respectively compare the path consistency deviation index with the first preset threshold and the collusion loop index with the second preset threshold to generate trajectory side tightening trigger flag and graph side tightening trigger flag, and when the trajectory side tightening trigger flag is true, execute the effective ring screening tightening process, and when the graph side tightening trigger flag is true, execute the trajectory verification tightening process;

[0035] Step 42, under the trajectory side tightening trigger, reduce the neighborhood layer number of the delivery reachable domain by a fixed step and recalculate the delivery reachable domain, limit the effective ring screening condition to the device sharing identifier value being equal to 1 and the space-time co-occurrence identifier value being equal to 1, and re-screen the effective ring set and update the weight set under the constraint of the updated delivery reachable domain;

[0036] Step 43, under the graph side tightening trigger, limit the space boundary of the trajectory verification condition to the delivery reachable domain after the neighborhood layer number is reduced, increase the weight of the coverage deficiency proportion under the fixed weight framework of the path consistency deviation index and correspondingly scale the remaining weights, recalculate the path consistency deviation index in the updated delivery reachable domain, and compare it with the first preset threshold to generate an updated trajectory side tightening trigger flag.

[0037] Further, a risk level is generated, and a deterministic process handling instruction is output, including:

[0038] Step 51, in each sliding window, count the number of abnormal orders with a path consistency deviation index greater than or equal to a first preset threshold;

[0039] Step 52, according to the trajectory side tightening trigger flag, the graph side tightening trigger flag, the number of abnormal orders, and the order quantity threshold in the window, generate three risk levels of high, medium, and low according to Boolean logic rules, wherein when the graph side abnormal flag is true and the effective ring set is non-empty and the number of abnormal orders is greater than or equal to the order quantity threshold, it is determined as a high risk level, when the graph side abnormal flag is true and the effective ring set is non-empty, or when the trajectory side abnormal flag is true and the number of abnormal orders is less than the order quantity threshold, it is determined as a medium risk level, and the remaining cases are determined as a low risk level;

[0040] Step 53, according to the risk level, construct a deterministic process handling instruction, and for each deterministic handling instruction, generate a unique instruction identifier and a corresponding risk level, sliding window, order primary key, and effective ring set reference relationship.

[0041] Further, the solidification playback module further includes a time trend comparison mechanism for calculating the average values of the path consistency deviation index and the collusion loop index corresponding to two adjacent sliding windows on the time sequence, taking the difference between the two as the trend change amount, and dividing by the time interval of the two windows to obtain the path consistency deviation index change rate and the collusion loop index change rate. When the absolute value of any index change rate exceeds the preset trend threshold, the playback interface call is triggered to retrieve the corresponding solidification evidence data block and perform risk level review.

[0042] The application has the beneficial effects that: the application realizes joint modeling of the delivery task at the path layer and the graph structure layer by constructing a heterogeneous graph that fuses orders, accounts, devices, networks, locations, and time; realizes synchronous identification of abnormal trajectory behavior and multi-agent collaborative behavior of the rider by collaborative calculation of the path consistency deviation index and the collusion loop index; establishes a feedback coupling mechanism between the trajectory side and the graph side through the bidirectional reinforcement module, so that the risk judgment has the characteristics of self-adaptive convergence; realizes dynamic perception and continuous monitoring of the risk in the spatial and temporal dimensions through regional aggregation analysis and time series trend comparison; and builds a traceable evidence chain structure through the solidification playback module to support risk review and result verification. Overall, the application realizes accurate identification, graded early warning, and closed-loop disposal of multi-dimensional risks in delivery services, effectively improving the intelligent level of risk control, data interpretability, and operational safety of the running platform. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a module schematic diagram of the multi-dimensional running platform risk early warning system for delivery services of the application. DETAILED DESCRIPTION

[0044] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can be changed without departing from the scope of the content of this specification. Various processes or components can be omitted, replaced, or added according to needs. In addition, features described with respect to some examples can also be combined in other examples.

[0045] As shown in Figure 1 , the multi-dimensional running platform risk early warning system for delivery services includes:

[0046] The spatiotemporal index modeling module 1 is used to collect delivery service original data, map the location to a spatial grid unit, discretize the time into a time bucket, build a heterogeneous graph, and label the pickup grid and the delivery grid; wherein the delivery service original data includes: order information, user account, rider account, merchant account, device and network identifier, location, and time;

[0047] The reachable domain construction module 2 is used to calculate the shortest path between the pickup grid and the delivery grid based on the road network, project the shortest path as a grid sequence, and expand the delivery reachable domain according to the neighborhood layers;

[0048] The path consistency module 3 is used to obtain the observed trajectory, form an observed trajectory grid set and an actual trajectory length, and generate a path consistency deviation index within the delivery reachable domain;

[0049] The collusion loop module 4 is used to extract the induced subgraph from the heterogeneous graph, enumerate simple closed loops and filter effective loops, and generate the collusion loop index.

[0050] The bidirectional reinforcement module 5 is used to tighten the effective loop screening conditions when the path consistency deviation index reaches the first preset threshold, and to tighten the trajectory verification conditions when the collusion loop index reaches the second preset threshold.

[0051] The graded handling module 6 is used to grade orders, generate risk levels, and output deterministic process handling instructions for executing order settlement, account and device verification, and trajectory playback verification.

[0052] The solidified playback module 7 is used to solidify the delivery reachability domain, the observation trajectory grid set, the shortest path, the path consistency deviation index, the collusion loop index, the effective loop and the number of neighboring layers, forming a solidified evidence data block.

[0053] In one embodiment of the present invention, the system collects raw delivery service data from the data interface of the delivery platform and maps the location data into spatial grid cells to achieve discretization and regional management of spatial coordinates. The spatial grid cell refers to the smallest geographical unit divided within the target delivery area according to a fixed spatial resolution, used to carry the geographical location mapping results of delivery events. The time data is discretized, dividing the continuous time axis into time buckets of fixed length. Each time bucket is a time grouping unit used to aggregate delivery events within the same time interval during the delivery process. Each time bucket corresponds to a defined time range, such as five minutes or ten minutes, used to represent orders, trajectories, and interactions within that time period.

[0054] After spatial and temporal discretization, the system constructs a heterogeneous graph based on the collected raw delivery service data. This heterogeneous graph is a graph structure containing multiple types of nodes and edges to represent the complex relationships between multiple entities in the delivery process. Nodes in the heterogeneous graph include user account nodes, rider account nodes, merchant account nodes, device nodes, network identifier nodes, spatial grid cell nodes, and time bucket nodes, with different node types representing different entities in the delivery ecosystem. Edges represent the types of relationships between nodes, including: association edges constructed based on order relationships, binding edges constructed based on account-device or network binding relationships, and dwell edges constructed based on account dwell relationships within specific spatial grid cells. Through heterogeneous graph modeling, the interaction characteristics of multi-source data and potential risk propagation paths during the delivery process can be displayed within a unified topological framework.

[0055] In addition, the pickup grid and delivery grid are labeled in the heterogeneous graph, corresponding to the starting spatial unit and the target spatial unit of the order, respectively, to establish the spatial start and end mapping relationship of the delivery task.

[0056] In an embodiment of the present application, the generation of the delivery reachable domain comprises:

[0057] Step 11, selecting the road network anchor points adjacent to the pickup grid and the delivery grid in the road network as the starting point and the ending point of the path search, taking the weighted sum of the road segment length, the turning cost and the road grade cost as the path weight, calculating the total path weight of all candidate paths by using the shortest path algorithm, and selecting the path with the minimum total path weight as the shortest path, which corresponds to the optimal driving path of the rider from the pickup location to the delivery location in geography; wherein the road network anchor point refers to a key node with spatial positioning function in the road network, such as a road intersection or a road segment endpoint, which is used to connect the topological nodes of different road segments; the shortest path algorithm preferably uses Dijkstra algorithm to ensure the uniqueness and reproducibility of the path calculation results under the same input conditions;

[0058] Step 12, sampling along the shortest path at a fixed interval related to the resolution of the spatial grid unit to obtain an ordered coordinate point list, mapping each coordinate point to a spatial grid unit and generating a grid sequence by removing duplicates in the order of appearance, marking the pickup grid and the delivery grid at the beginning and the end of the grid sequence respectively, and establishing an index binding between each spatial grid unit and the corresponding road segment, which ensures the correspondence between the path space expression and the actual road network;

[0059] Step 13, taking each spatial grid unit in the grid sequence as the center, performing a neighborhood query based on the hexagonal grid according to the preset number of neighborhood layers, expanding layer by layer to obtain a set of adjacent units, and performing a set union operation on all the sets of adjacent units to form a delivery reachable domain. Wherein, the number of neighborhood layers is used to define the level range of the outward expansion from the center unit in the grid topology; the hexagonal grid structure is used to improve the integrity and directional continuity of the spatial neighborhood query due to its equidistant adjacency characteristics and uniform coverage capability; the delivery reachable domain represents the spatial range that the rider may pass through from the pickup location to the delivery location under normal delivery conditions.

[0060] This embodiment realizes the accurate modeling of the delivery task by introducing multi-factor weighted shortest path calculation at the road network level; realizes the continuous expression of the delivery path in discrete space by using fixed interval sampling and grid mapping at the spatial level; realizes the construction and boundary determination of the delivery space reachable area based on the neighborhood expansion of the hexagonal topology, which provides an accurate spatial constraint framework for the subsequent calculation of the path consistency deviation index and the detection of abnormal trajectories.

[0061] In an embodiment of the present application, the observation trajectory is obtained, the observation trajectory grid set and the actual trajectory length are formed, and the path consistency deviation index is generated in the delivery reachable domain, comprising:

[0062] Step 21, the system first acquires the observation trajectory of the rider in the whole process of the delivery task, that is, a time sequence coordinate point set collected based on terminal device positioning information, carries out time sequence checking and sampling frequency unification on the observation trajectory, eliminates time out-of-order points and abnormal sampling intervals, carries out map matching according to the road network, matches the observation trajectory points to the shortest path road network according to the principle of shortest distance and continuous mileage, obtains the matched trajectory sequence, and records the matching mark for each trajectory segment to distinguish whether the trajectory point is located on the shortest path calculated by the system; the map matching takes the shortest distance and continuous mileage as constraint conditions to ensure that the trajectory matching result is unique and reproducible; specifically, a path weight function is established by calculating the vertical distance, direction angle and mileage continuity cost between the observation trajectory point and the road segment, the path weight minimum is taken as the judgment criterion to determine the matching road segment of the trajectory point;

[0063] Step 22, the matched trajectory points are mapped into spatial grid cells point by point according to spatial position, and the observation trajectory grid set is formed by removing duplicates according to the appearance order, each grid cell represents the projection position of the trajectory in the spatial discrete structure, realizes the grid expression of the continuous trajectory, and the actual trajectory length is obtained by accumulating the length of the matching road segment formed by each trajectory point;

[0064] Step 23, under the constraint of the delivery reachable domain, the insufficient coverage ratio, the off-road segment ratio and the abnormal stay time ratio are calculated, wherein the insufficient coverage ratio is obtained by the complementary value of the intersection area ratio of the observation trajectory grid set and the delivery reachable domain, which is used to represent the degree of trajectory not covering the theoretical reachable region, the off-road segment ratio is the ratio of the number of matching marks for non-shortest path segments to the total number of trajectory segments, which is used to reflect the frequency of trajectory deviation from the road baseline, and the abnormal stay time ratio is the ratio of the low-speed and same-cell stay time to the promised delivery time, which is used to quantify the abnormal waiting or staying behavior; the three ratios together constitute the spatial, topological and time dimension features of the path deviation;

[0065] Step 24, the over-distance ratio is calculated according to the shortest path length and the actual trajectory length, and the over-distance ratio, the insufficient coverage ratio, the off-road segment ratio and the abnormal stay time ratio are linearly weighted according to fixed weights to obtain the path consistency deviation index. The index is used to quantify the deviation degree between the actual driving path of the rider and the theoretical optimal path, and the larger the value is, the higher the deviation degree of the delivery behavior is.

[0066] The embodiment realizes the time sequence integrity of the distribution track by trajectory checking and sampling unification in the time dimension; realizes the spatial consistency expression of the distribution track by performing deterministic map matching and grid mapping in the spatial dimension; realizes the quantitative representation of the deviation degree of the distribution behavior by constructing multi-dimensional proportional features and generating a path consistency deviation index by weighting, provides a high-precision and interpretable path deviation analysis mechanism for a multi-dimensional running platform risk early warning system, enables the system to automatically identify and early warn abnormal detours, false trips and abnormal stay behaviors, and thus improves the risk identification capability and operation safety in the distribution service scene.

[0067] In an embodiment of the present application, an induced subgraph is extracted from a heterogeneous graph, simple closed loops are enumerated and effective loops are screened to generate a collusion loop index, including:

[0068] Step 31, a sliding window is set on a time sequence, covering a fixed number of time buckets, nodes and edges with time attributes belonging to the sliding window are screened in the heterogeneous graph to form an induced subgraph; the induced subgraph reflects the dynamic interaction structure between multiple subjects within the sliding window;

[0069] Step 32, a simple closed loop enumeration algorithm is performed in the induced subgraph, a closed loop set is enumerated, whether each closed loop contains at least two types of business subject nodes is determined according to the node type, and device sharing identifiers and space-time co-occurrence identifiers are searched in the edge attribute, the former is used to determine whether different nodes use the same device or network environment, and the latter is used to determine whether different nodes have position overlap or time overlap in the same time bucket and adjacent space grid, and are marked as device sharing identifier value and space-time co-occurrence identifier value respectively, the value 1 represents the existence of device sharing or space-time co-occurrence relationship, and the value 0 represents the non-existence of the corresponding relationship; wherein the business subject nodes include: user account, rider account and merchant account;

[0070] Step 33, the closed loop set obtained by enumeration is determined for effectiveness, only when the closed loop contains at least two types of business subject nodes and the device sharing identifier value and the space-time co-occurrence identifier value are both 1, the closed loop is defined as an effective loop, and an effective loop weight inversely proportional to the loop length is set for each effective loop to form an effective loop set and a corresponding weight set; the effective loop weight is inversely proportional to the loop length, so as to reflect the feature that a shorter path loop represents a closer association relationship;

[0071] Step 34, the effective loop weights in the effective loop set are weighted and summed, and the collusion loop index is calculated by taking the number of orders in the sliding window as the normalization coefficient, which is used to represent the potential cooperation behavior strength between the distribution business subjects in a specific time window, the greater the value, the more intensive the co-occurrence relationship between the business subjects in the device, time and space layers, and the higher the collusion risk.

[0072] The embodiment realizes dynamic capture of the delivery business association relationship by introducing a sliding window mechanism in the time dimension; realizes accurate identification of the multi-subject collaborative mode by performing simple closed-loop enumeration and multi-attribute retrieval in the structure layer; realizes quantification of the potential collusion strength in the business network based on weighted normalization processing of the ring length and the order quantity, provides the system with the ability of abnormal detection and risk quantification in the graph structure layer, enables the system to identify and early warn the risk scenarios of collaborative behaviors such as account sharing, device sharing or false orders, and improves the risk control accuracy and business security of the platform.

[0073] In an embodiment of the application, the collusion loop module further comprises loop coupling degree calculation on the effective ring set, comprising: selecting any two effective rings from the effective ring set, respectively counting the number of shared nodes and the number of shared edges of the two effective rings, and recording the number of nodes and the number of edges of each pair of effective rings; taking the ratio of the number of shared nodes to the average of the number of nodes of the two effective rings as the loop coupling degree, that is, the higher the proportion of shared nodes of the two effective rings, the greater the loop coupling degree, indicating that the overlap between the two rings in the business subject layer is stronger, and the connection tightness between different risk structures in the business network is quantified; wherein the shared node refers to the same business subject node existing in the two effective rings, which may be the same user account, rider account or merchant account; the shared edge represents the edge with the same connection relationship in the two effective rings, such as the same device sharing edge or the same time and space co-occurrence edge;

[0074] When the two effective rings have shared edges, the loop coupling degree is corrected according to the number of shared edges, the ratio of the number of shared edges of the two rings to the smaller value of the number of edges in the two rings is taken as the edge overlap ratio, and a predetermined correction coefficient is multiplied to obtain a correction amount, the correction amount and the original loop coupling degree are added to obtain the corrected loop coupling degree, if the corrected loop coupling degree exceeds the preset coupling threshold, the two rings are determined as strong association rings, and the shared nodes are marked as high association nodes. The high association node represents a key business subject participating in multiple strong coupling rings at the same time, which may be a potential collusion hub node. The value range of the correction coefficient is 0 to 1.

[0075] The embodiment realizes basic association identification between different risk rings by counting the number of shared nodes and the size of the structure at the node level; realizes quantification of the ring structure overlap by introducing edge overlap ratio correction at the edge level; realizes deep identification of the potential collusion relationship between the delivery business subjects by setting the coupling threshold and marking the high association nodes, so that the system can find the implicit collaborative mode across accounts, devices or time and space in the business network, thereby improving the accuracy and interpretation ability of the platform in abnormal collusion detection and risk tracking.

[0076] In an embodiment of the present application, a regional aggregation analysis module is further included, configured to aggregate multiple spatial grid cells to form geographical area groups based on spatial adjacency relationship, wherein the spatial adjacency relationship is used to determine the spatial contact and topological relevance between adjacent grid cells; the regional average of the path consistency deviation index and the collusion loop index is calculated in each geographical area group, and the average of all geographical area groups is calculated as the global average; when the difference between the average of the path consistency deviation index or the average of the collusion loop index of any geographical area group and the corresponding global average exceeds the preset regional deviation threshold, the geographical area group is marked as a regional high-risk area.

[0077] The present embodiment realizes the spatial clustering expression of distribution risks by constructing geographical area groups; realizes the identification of abnormal aggregation of risks in the geographical dimension by calculating the deviation between the regional average and the global average; realizes the automatic identification and spatial early warning of regional risks by setting the deviation threshold and marking the high-risk areas, so that the system can accurately identify and dynamically monitor the regional abnormal activities, risk aggregation points and potential collaborative behavior concentration areas in the distribution network, thereby effectively improving the intelligent level of the platform in the aspects of geographical risk prevention and control and strategy decision-making.

[0078] In an embodiment of the present application, the bidirectional reinforcement module comprises:

[0079] Step 41, respectively compare the path consistency deviation index with the first preset threshold and the collusion loop index with the second preset threshold, generate a trajectory side tightening trigger flag and a graph side tightening trigger flag, when the trajectory side tightening trigger flag is true, execute the effective loop screening tightening process, when the graph side tightening trigger flag is true, execute the trajectory verification tightening process; the first preset threshold is used to determine whether the path consistency deviation index is in an abnormal range, and the value can be determined according to the statistical distribution of the path deviation index in the historical distribution tasks, the second preset threshold is used to determine whether the collusion loop index reaches the collaborative risk level, and the value can be determined according to the risk distribution of the historical loop index sample; this step ensures the linkage regulation and control of the trajectory side deviation and the graph side structure risk, thereby avoiding misjudgment caused by single index abnormality;

[0080] Step 42, under the trajectory side tightening trigger, reduce the neighborhood layer number of the distribution reachable domain by a fixed step and recalculate the distribution reachable domain, limit the effective loop screening condition to the device sharing identifier value equal to 1 and the spatiotemporal co-occurrence identifier value equal to 1, and re-screen the effective loop set and update the weight set with the updated distribution reachable domain as the constraint, so that the effective loop screening is more strict and the loop weight distribution is more concentrated; this process strengthens the identification condition of the graph structure risk in the case of trajectory abnormality, effectively reducing the interference of irrelevant loops;

[0081] Step 43, under the graph side tightening trigger, the spatial boundary of the trajectory checking condition is limited to the delivery reachable domain after the reduction of the neighborhood layer number, the weight of the insufficient coverage ratio is increased in the fixed weight framework of the path consistency deviation index, and the remaining weights are scaled accordingly, so that the path coverage becomes the main factor in the deviation evaluation, the path consistency deviation index is recalculated in the updated delivery reachable domain, and an updated trajectory side tightening trigger flag is generated by comparing with the first preset threshold. This process realizes the re-correction of trajectory deviation analysis under high collusion risk conditions, and makes the trajectory side calculation sensitive to changes in the risk of the graph structure.

[0082] The embodiment realizes adaptive convergence of the graph structure identification condition under path deviation anomaly through the trajectory side tightening trigger; realizes dynamic reinforcement of the trajectory deviation evaluation parameter when the collusion risk increases through the graph side tightening trigger; through mutual feedback and parameter adjustment of both sides, a closed-loop reinforcement judgment mechanism between the path layer and the graph layer is realized, so that the system can maintain the stability and high sensitivity of the early warning judgment in the complex scene where the path deviation and the collusion risk exist at the same time, thereby improving the identification accuracy and prevention and control ability of the platform to abnormal delivery behavior and collusion risk.

[0083] In an embodiment of the present application, the orders are classified, the risk level is generated, and the deterministic process disposal instruction for executing order settlement, account and device verification and trajectory playback verification is output, including:

[0084] Step 51, in each sliding window, the number of abnormal orders whose path consistency deviation index is greater than or equal to the first preset threshold is counted;

[0085] Step 52, according to the trajectory side tightening trigger flag, the graph side tightening trigger flag, the number of abnormal orders and the order quantity threshold in the window, three risk levels of high, medium and low are generated according to the Boolean logic rule, wherein when the graph side abnormal flag is true and the effective ring set is non-empty and the number of abnormal orders is greater than or equal to the order quantity threshold, it is determined as high risk level, when the graph side abnormal flag is true and the effective ring set is non-empty, or when the trajectory side abnormal flag is true and the number of abnormal orders is less than the order quantity threshold, it is determined as medium risk level, and the rest is determined as low risk level; in this step, the graph side abnormal flag reflects the risk state of the system in the collusion structure identification layer, the trajectory side abnormal flag reflects the abnormal situation in the path deviation layer, and the logical combination of the two ensures the double reliability and interpretability of the risk level determination;

[0086] In step 53, the deterministic flow treatment instruction is constructed according to the risk level, wherein the high risk level corresponds to the output order settlement suspension instruction, the account and device verification instruction and the track playback verification instruction, the medium risk level corresponds to the output account and device verification instruction and the track playback verification instruction, the low risk level corresponds to the output track playback verification instruction, and a unique instruction identifier, a corresponding risk level, a sliding window, an order primary key and an effective ring set reference relationship are generated for each deterministic treatment instruction.

[0087] The embodiment realizes dynamic quantitative division of the risk level by combining the statistics of the path deviation and the graph structure anomaly in the sliding window; realizes explicit hierarchical expression of the system risk judgment by generating high, medium and low three-level risk levels based on the Boolean logic rules; realizes executable and closed-loop response of the risk judgment result by generating the deterministic flow treatment instruction, so that the system can accurately distinguish the risk level and quickly execute the corresponding control strategy in the complex business scenario, thereby improving the automation level of the platform and the abnormal response efficiency.

[0088] In an embodiment of the present application, the solidification playback module is used to structure and solidify the key indicators and results in the delivery risk analysis process, and to realize risk review and evidence playback based on time series change trend, so as to ensure the traceability and consistency of the risk identification result. The solidified evidence data block is indexed by time stamp and order primary key, realizing consistent storage of all core indicators and graph structure results.

[0089] The time series trend comparison mechanism is used to calculate the average value of the path consistency deviation index and the collusion loop index corresponding to the two adjacent sliding windows on the time series, and the difference between the two is used as the trend change amount, and divided by the time interval of the two windows to obtain the path consistency deviation index change rate and the collusion loop index change rate, which are used to measure the change speed and direction of the risk indicators in time. When the absolute value of any index change rate exceeds the preset trend threshold, the playback interface call is triggered to call the corresponding solidified evidence data block to perform risk level review. The playback interface is used to call the solidified data block to perform index recalculation and level reevaluation, realizing dynamic self-correction of the system to the risk judgment.

[0090] The embodiment realizes traceability and verifiability of the delivery risk judgment result by solidifying the multi-dimensional data, realizes dynamic change monitoring and trend detection of the risk indicators by performing trend comparison on the path consistency deviation index and the collusion loop index on the time series, so that the system can not only accurately identify the risk event, but also continuously supervise and automatically review the risk evolution process in the time series dimension, thereby improving the explainability, reliability and long-term stability of the platform risk warning.

[0091] It should be noted that the interval and the threshold size are set for the convenience of comparison, wherein the size of the threshold depends on how much sample data and the base number set by the person skilled in the art for each group of sample data, as long as it does not affect the proportional relationship between the parameters and the quantized values. And the above formula is the calculation of the dimensionless value, the formula is obtained by collecting a large amount of data to simulate the formula of the nearest real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation.

[0092] The above describes the embodiments of the present application, but the present application is not limited to the above specific embodiments, and the above specific embodiments are only illustrative but not restrictive, and the person skilled in the art can make many forms under the inspiration of the present embodiment, which all belong to the protection of the present embodiment.

Claims

1. A multi-dimensional errand platform risk warning system for a delivery service, characterized in that, The method comprises the following steps: The spatiotemporal index modeling module is used to collect delivery service raw data, map locations to spatial grid cells, discretize time into time buckets, construct a heterogeneous graph, and label pickup grid and delivery grid; wherein the delivery service raw data comprises order information, user account, rider account, merchant account, device and network identifier, location and time; The reachable domain construction module is used to calculate the shortest path between the pickup grid and the delivery grid based on the road network, project the shortest path into a grid sequence, and expand the grid sequence according to the number of neighborhood layers to generate a delivery reachable domain; The path consistency module is used to obtain an observed trajectory, form an observed trajectory grid set and an actual trajectory length, and generate a path consistency deviation index within the delivery reachable domain; The collusion loop module is used to extract an induced subgraph from the heterogeneous graph, enumerate simple closed loops and filter effective loops, and generate a collusion loop index, comprising: Step 31: Set a sliding window on the time sequence, covering a fixed number of time buckets, filter nodes and edges in the heterogeneous graph whose time attribute belongs to the sliding window, and form an induced subgraph; Step 32: In the induced subgraph, execute a simple closed loop enumeration algorithm to enumerate a closed loop set, and for each closed loop, determine whether it contains at least two types of business subject nodes according to the node type, and retrieve device sharing identifiers and spatiotemporal co-occurrence identifiers in the edge attribute, and mark them as device sharing identifier values and spatiotemporal co-occurrence identifier values respectively; wherein the business subject nodes include user account, rider account and merchant account; Step 33: Determine the effectiveness of the enumerated closed loop set, and only when the closed loop contains at least two types of business subject nodes and the device sharing identifier value and the spatiotemporal co-occurrence identifier value are both 1, define the closed loop as an effective loop, and set an effective loop weight for each effective loop in inverse proportion to the loop length, form an effective loop set and a corresponding weight set; Step 34: Weighted sum the effective loop weights in the effective loop set, and calculate the collusion loop index by taking the number of orders in the sliding window as the normalization coefficient; The bidirectional reinforcement module is used to tighten the effective loop filtering condition when the path consistency deviation index reaches a first preset threshold, and to tighten the trajectory verification condition when the collusion loop index reaches a second preset threshold; The hierarchical disposal module is used to grade orders, generate risk levels, and output deterministic process disposal instructions for executing order settlement, account and device verification, and trajectory playback verification; The solidification playback module is used to solidify the delivery reachable domain, the observed trajectory grid set, the shortest path, the path consistency deviation index, the collusion loop index, the effective loop and the number of neighborhood layers, and form a solidified evidence data block.

2. The multi-dimensional runner platform risk warning system for delivery services of claim 1, wherein, The spatiotemporal index modeling module comprises: Map locations to spatial grid cells according to a fixed resolution, discretize time information into time buckets according to a fixed time granularity, and establish the spatial grid index and the time bucket index corresponding to each data record; A heterogeneous graph is constructed by taking a user account, a rider account, a merchant account, a device and network identifier, a spatial grid unit and a time bucket as nodes, and taking the association relationship between orders, the binding relationship between accounts and devices or networks, and the residence relationship of accounts in a specific spatial grid unit as edges, and each order is labeled with the pickup grid and the delivery grid in the heterogeneous graph.

3. The multi-dimensional runner platform risk warning system for delivery services of claim 1, wherein, The generation of the delivery reachable domain includes: Step 11, selecting the road network anchor points adjacent to the pickup grid and the delivery grid in the road network as the starting point and the ending point, taking the weighted sum of the road segment length, the turning cost and the road level cost as the path weight, using the shortest path algorithm to calculate the total path weight of all candidate paths, and selecting the path with the minimum total path weight as the shortest path; Step 12, sampling along the shortest path at a fixed interval related to the spatial grid unit resolution to obtain an ordered coordinate point list, mapping each coordinate point to a spatial grid unit and generating a grid sequence by removing duplicates in order, labeling the pickup grid and the delivery grid at the beginning and the end of the grid sequence, and establishing an index binding between each spatial grid unit and the corresponding road segment; Step 13, taking each spatial grid unit in the grid sequence as the center, performing a hexagonal grid-based neighborhood query according to the preset number of neighborhood layers, and expanding layer by layer to obtain a set of adjacent units, and performing a set union operation on all sets of adjacent units to form a delivery reachable domain.

4. The multi-dimensional runner platform risk warning system for delivery services of claim 1, wherein, Obtaining an observation trajectory, forming an observation trajectory grid set and an actual trajectory length, generating a path consistency deviation index within the delivery reachable domain, including: Step 21, time sequence verification and sampling frequency unification are performed on the observation trajectory, map matching is performed according to the road network, the observation trajectory points are matched to the road network where the shortest path is located according to the shortest distance and continuous mileage principle, a matched trajectory sequence is obtained, and a matching mark is recorded for each trajectory segment; Step 22, mapping the matched trajectory points to spatial grid units one by one, removing duplicates in order to form an observation trajectory grid set, and obtaining an actual trajectory length by accumulating the lengths of the matching road segments formed by each trajectory point; Step 23, calculating the coverage deficiency ratio, the off-road segment ratio and the abnormal stay time ratio under the constraint of the delivery reachable domain, wherein the coverage deficiency ratio is obtained by the complement of the intersection area ratio of the observation trajectory grid set and the delivery reachable domain, the off-road segment ratio is the ratio of the number of non-shortest path segments to the total number of trajectory segments, and the abnormal stay time ratio is the ratio of the low-speed and same-unit stay time to the promised delivery time; Step 24, calculating the over-distance ratio according to the shortest path length and the actual trajectory length, and linearly weighting the over-distance ratio, the coverage deficiency ratio, the off-road segment ratio and the abnormal stay time ratio according to the fixed weight to obtain the path consistency deviation index.

5. The multi-dimensional runner platform risk alert system for delivery services of claim 1, wherein, The collusion loop module further includes performing loop coupling degree calculation on the effective loop set, including: selecting any two effective loops from the effective loop set, respectively counting the number of shared nodes and the number of shared edges, and recording the number of nodes and the number of edges of each pair of effective loops; taking the ratio of the number of shared nodes to the average of the number of nodes of the two effective loops as the loop coupling degree; When two valid loops exist sharing edges, the loop coupling degree is corrected according to the number of shared edges, the ratio of the number of shared edges of the two loops to the smaller number of edges in the two loops is taken as the edge overlap ratio, and the correction amount is obtained by multiplying the preset correction coefficient, the corrected loop coupling degree is obtained by adding the correction amount to the original loop coupling degree, and if the corrected loop coupling degree exceeds the preset coupling threshold, the two loops are determined to be strongly associated loops, and the shared nodes thereof are marked as high-association nodes.

6. The multi-dimensional runner platform risk alert system for delivery services of claim 1, wherein, The regional aggregation analysis module is further included, which is configured to aggregate multiple spatial grid units to form geographical area groups according to spatial adjacency relationship on the basis of spatial grid units, calculate regional average values of the path consistency deviation index and the collusion loop index in each geographical area group respectively, and calculate the average values of all geographical area groups as global average values, and when the difference between the average value of the path consistency deviation index or the average value of the collusion loop index of any geographical area group and the corresponding global average value exceeds the preset regional deviation threshold, the geographical area group is marked as a regional high-risk area.

7. The multi-dimensional runner platform risk alert system for delivery services of claim 1, wherein, The bidirectional reinforcement module includes: Step 41, respectively comparing the path consistency deviation index with the first preset threshold and the collusion loop index with the second preset threshold to generate a trajectory side tightening trigger flag and a graph side tightening trigger flag, performing an effective loop screening tightening process when the trajectory side tightening trigger flag is true, and performing a trajectory verification tightening process when the graph side tightening trigger flag is true; Step 42, under the trajectory side tightening trigger, reducing the neighborhood layer number of the delivery reachable domain by a fixed step and recalculating the delivery reachable domain, limiting the effective loop screening condition to the device sharing identification value being equal to 1 and the spatiotemporal co-occurrence identification value being equal to 1, and re-screening the effective loop set and updating the weight set under the constraint of the updated delivery reachable domain; Step 43, under the graph side tightening trigger, limiting the spatial boundary of the trajectory verification condition to the delivery reachable domain after the neighborhood layer number is reduced, increasing the weight of the coverage deficiency ratio under the fixed weight framework of the path consistency deviation index and correspondingly scaling the remaining weights, recalculating the path consistency deviation index in the updated delivery reachable domain, and comparing it with the first preset threshold to generate an updated trajectory side tightening trigger flag.

8. The multi-dimensional runner platform risk alert system for delivery services of claim 1, wherein, A risk level is generated, and a deterministic process handling instruction is output, including: Step 51, in each sliding window, the number of abnormal orders with the path consistency deviation index greater than or equal to the first preset threshold is counted; Step 52, according to the trajectory side tightening trigger flag, the graph side tightening trigger flag, the number of abnormal orders, and the order quantity threshold in the window, three risk levels of high, medium and low are generated according to Boolean logic rules, wherein when the graph side abnormal flag is true, the effective loop set is not empty, and the number of abnormal orders is greater than or equal to the order quantity threshold, it is determined as a high-risk level, when the graph side abnormal flag is true and the effective loop set is not empty, or when the trajectory side abnormal flag is true and the number of abnormal orders is less than the order quantity threshold, it is determined as a medium-risk level, and the remaining cases are determined as a low-risk level; Step 53, according to the risk level structure determination flow treatment instruction, and for each deterministic treatment instruction generation unique instruction identification and corresponding risk level, sliding window, order primary key and effective ring set reference relationship.

9. The multi-dimensional runner platform risk alert system for delivery services of claim 1, wherein, The solidification playback module further comprises a timing trend comparison mechanism for calculating the corresponding path consistency deviation index average and collusion loop index average according to the two adjacent sliding windows on the time sequence, taking the difference between the two as the trend change, and dividing by the time interval of the two windows to obtain the path consistency deviation index change rate and the collusion loop index change rate. When the absolute value of any index change rate exceeds the preset trend threshold, the playback interface call is triggered to call the corresponding solidification evidence data block and perform risk level review.

Images