Electronic bidding transaction platform monitoring method and system
By combining multivariate time series and behavioral topology analysis with minimum energy path deduction, a causal chain of abnormal behavior in the electronic bidding and tendering transaction platform is generated, which solves the problem of difficulty in identifying hidden collaborative behavior in existing technologies and achieves high-precision and real-time supervision.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN JI NENG INVITE TENDERS
- Filing Date
- 2026-03-02
- Publication Date
- 2026-06-02
AI Technical Summary
Existing electronic bidding and tendering platforms are unable to promptly identify covert collaborative behaviors such as bid rigging, collusion, and manipulation of bid evaluation. They also lack the technical ability to infer the causes of anomalies and the relationship chains between the main parties, resulting in lagging supervision and inconsistent conclusions.
By employing multivariate time series anomaly detection, behavioral topology analysis, and multidimensional regulatory space deformation analysis techniques, a structured model of bidding behavior is constructed. Causal chains of abnormal behavior are generated through minimum energy potential path deduction, enabling automated identification and explainable regulation.
It achieves high-precision identification of bid rigging, collusion, and manipulation of bid evaluation, provides highly traceable and real-time regulatory results, and improves the regulatory adaptability and review efficiency of complex collaborative behaviors.
Smart Images
Figure CN122133096A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electronic bidding and tendering information supervision technology, and in particular to methods and systems for supervising electronic bidding and tendering transaction platforms. Background Technology
[0002] Electronic bidding and tendering platforms digitize and traceably manage bidding processes through online publication of tender documents, submission of bids, bid opening and evaluation, and announcement of winning bids. Existing platforms typically include monitoring functions, controlling access to key nodes, retaining logs, and verifying processes. They also provide alerts for anomalies through statistical reports, threshold rules, or post-audit methods. For example, they can compare or randomly check indicators such as bidding time, number of bids, price fluctuations, and deviations in expert scores to support regulatory authorities' compliance oversight of bidding and tendering activities.
[0003] However, existing technologies mostly focus on single projects, single entities, or single indicators, relying primarily on pre-set rules or simple statistical features. This makes it difficult to promptly identify covert collaborative behaviors such as bid rigging, collusion, and manipulation of bid evaluation. Because related entities often circumvent these restrictions through multi-role cooperation, resource sharing, synchronized behavior, and cross-stage coordination, traditional methods struggle to establish unified temporal relationships within multi-dimensional operational data. Furthermore, they find it difficult to extract stable structural features from the relationship between behavioral events and resource usage, resulting in a lag and uncertainty in the identification of abnormal patterns.
[0004] Current regulatory outputs are mostly risk scores or anomaly alerts, lacking the technical ability to infer the causes of anomalies and the chain of relationships between stakeholders, making it difficult to form an interpretable, verifiable, and traceable chain of evidence. Regulatory personnel often need to rely on manually retrieving logs and comparing operation records and interaction relationships to reconstruct the course of events, which is labor-intensive and results in inconsistent conclusions, making it difficult to meet the needs of real-time, interpretable, and traceable intelligent supervision of complex bidding processes.
[0005] Therefore, how to provide methods and systems for supervising electronic bidding and tendering transaction platforms is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0006] One objective of this invention is to propose a method and system for supervising electronic bidding and tendering transaction platforms. This invention comprehensively utilizes multivariate time series anomaly detection, behavioral topology analysis, and multidimensional regulatory space deformation analysis technologies to perform structured modeling and anomaly identification of the operational behaviors of bidding entities, evaluation experts, and related resources during the bidding process. Furthermore, it generates causal chains of abnormal behaviors through minimum energy potential path deduction, thereby achieving automated identification and interpretable supervision of covert collaborative behaviors such as bid rigging, collusion, and manipulation of bid evaluation. It has the advantages of high identification accuracy, strong traceability, good real-time performance, and strong adaptability to complex collaborative behaviors.
[0007] The electronic bidding and tendering transaction platform supervision method according to embodiments of the present invention includes:
[0008] Collect operational data from the bidding entities, bidding companies, evaluation experts and agencies in the electronic bidding platform during the entire bidding process, and construct a multivariate bidding behavior time series in chronological order.
[0009] Based on the behavioral events and resource usage relationships of the bidding entities in the bidding process, a behavioral topology sequence is constructed, behavioral nodes and relational edges are generated, forming the topological structure of bidding behavior, calculating topological invariants and generating behavioral topological phase information;
[0010] A multidimensional regulatory space is constructed based on regulatory indicators. The state of the regulatory space is updated based on input behavior. Curvature calculation, density calculation, torsion rate calculation and local volume change analysis are performed to obtain the deformation results of the regulatory space.
[0011] An improved TranAD model was used to detect anomalies in the time series of multivariate bidding behavior, and information on the time points of abnormal behavior and the corresponding anomaly dimensions was obtained.
[0012] Based on the time points of abnormal behavior, combined with the behavioral topology phase information and the deformation results of the regulatory space, a potential path space for bidding behavior is constructed. Migration energy is set for different types of behavioral changes. Based on the state migration relationship, a set of potential behavioral paths is generated and the minimum energy path from the normal state to the abnormal state is solved.
[0013] The abnormal behavior causal chain is generated based on the minimum energy path, and the abnormal behavior causal chain is provided to the regulatory authorities as a regulatory output for review and verification.
[0014] Optionally, the operation data includes download behavior data, file modification behavior data, file upload behavior data, price change data, rating record data, user login data, device fingerprint data, network address data, deposit payment data, deposit refund data, and association data between bidding entities and resources.
[0015] Optionally, constructing a time series of multivariate bidding behavior in chronological order includes:
[0016] Download behavior data, file modification behavior data, file upload behavior data, price change data, scoring record data, user login data, device fingerprint data, network address data, and bid security deposit data are time-aligned, and various types of behavior data are mapped to corresponding time series variables using a unified time axis as an index, thereby forming a multivariate bidding behavior time series containing multiple behavioral dimensions.
[0017] Optionally, the process of forming the topology of bidding behavior, calculating topological invariants, and generating behavioral topological phase information includes:
[0018] A discrete time axis is defined, and the behavioral events of each bidding entity are sorted in chronological order to form a behavioral sequence of the bidding entity. The behavioral events include behavioral type identifiers and timestamps.
[0019] Construct a set of nodes and a set of edges for a behavioral topology. The set of nodes includes nodes of bidding companies, nodes of evaluation experts, resource nodes, and event nodes. The resource nodes include nodes of network addresses, nodes of device fingerprints, and nodes of fund accounts. The set of edges includes sequential edges, resource-related edges, time-coupled edges, and role-related edges, and records the corresponding time information for each edge.
[0020] The topology of the bidding behavior is constructed using the set of nodes and edges and time information, and a sequence of topology slices is generated according to preset discrete time points. The existence relationship of various edges at time points is recorded in each time slice.
[0021] Calculate topological feature quantities on the topological structure. The topological feature quantities include the number of simple rings formed by the combined action of sequential edges and resource-related edges, the average shortest path length of node pairs obtained based on time-coupled edges, the behavior synchronization index obtained based on the ratio of the longest common subsequence length between different behavior sequences, and the number of nested levels formed by the combined action of role-related edges and resource-related edges.
[0022] The topological features are determined based on preset weights and thresholds to generate behavioral topological phase information, which includes topological features and phase identifiers.
[0023] Optionally, obtaining the deformation result of the regulatory space includes:
[0024] Based on operational data, normalization and time alignment are performed at a fixed time granularity, and six regulatory indicators are generated through preset deterministic mapping rules. The six regulatory indicators are time consistency indicator, resource sharing density indicator, role coupling strength indicator, document change manifold indicator, fund path curvature indicator, and operation rhythm indicator.
[0025] A multi-dimensional regulatory space is constructed, which includes a basic indicator layer and a policy constraint layer. The basic indicator layer uses the six regulatory indicators mentioned above as coordinate axes, and the policy constraint layer uses project type and amount range as constraint dimensions. Anchor units are generated based on the compliance template library, and the multi-dimensional regulatory space is divided into rules. Each anchor unit has a defined boundary and compliance range.
[0026] The six regulatory indicators and policy constraint parameters corresponding to each bidding entity at each time point are mapped to state points in a multi-dimensional regulatory space. State trajectories are formed in chronological order. The neighborhood density, directional change and distance change of adjacent intervals are calculated in the fixed neighborhood of each state point to obtain local compression indication, local expansion indication and trajectory bending indication.
[0027] A standardized disturbance test is performed on the state point. The disturbance includes making small positive and negative adjustments to each regulatory indicator and policy constraint parameter with a fixed amplitude. The displacement amplitude and displacement direction of the state point in the multidimensional regulatory space are recorded to obtain sensitivity indication and stability indication. The template deviation indication is generated using the boundary of the anchoring unit as a constraint.
[0028] The neighborhood density indicator, direction and distance change indicator, sensitivity indicator, stability indicator and template deviation indicator are weighted and summarized to generate deformation score and deformation identifier, forming the deformation result of the regulatory space and outputting it together with the corresponding timestamp and bidding entity identifier.
[0029] Optionally, obtaining the time point of the abnormal behavior and the corresponding abnormal dimension information includes:
[0030] The six regulatory indicators are normalized and sliced according to fixed time windows to form a continuous and non-overlapping window sequence, and the original values of all dimensions are retained in each window.
[0031] Constructing an improved TranAD model:
[0032] A parallel surveillance mask path is added between the original encoder and decoder. The surveillance mask path receives the surveillance space deformation result and generates a set of weight matrices.
[0033] During the encoding phase, both the temporal embedding vector and the weight matrix are introduced to adjust the attention allocation of multi-head attention.
[0034] In the decoding stage, a dual-branch structure of sequential reconstruction and reverse reconstruction is adopted, and the reconstruction results of the two branches are output through the residual fusion layer.
[0035] Unsupervised training of the improved TranAD model was performed within a data range containing only normal behavior. The training objective was to minimize the window-level reconstruction error. The TranAD model parameters were frozen after the TranAD model training was completed.
[0036] The continuous window sequence is input into the improved TranAD model to obtain the corresponding reconstructed window sequence. The window-level error vector is calculated by comparing the input window and the reconstructed window. The window anomaly score is determined according to the magnitude of the error values of each dimension in the window error vector. Then, the time point anomaly score and dimension-level anomaly score are generated according to the time points covered by the window.
[0037] The abnormal behavior time point is determined from the abnormal score of the time point according to the preset quantile threshold, and the corresponding abnormal dimension information is obtained by combining the abnormal score of the dimension level.
[0038] Optionally, the step of generating a set of potential behavioral paths based on state transition relationships and solving for the minimum energy path from the normal state to the abnormal state includes:
[0039] Obtain the time point of abnormal behavior and the corresponding abnormal dimension information, obtain the behavior topology phase information and the deformation result of the regulatory space; generate a state set by discretizing the bid entity identifier, timestamp, behavior type and resource identifier, and establish the state reachability relationship according to the time sequence to form an initial state transition set.
[0040] A potential path space for bidding behavior is constructed, which consists of an event layer and a constraint layer. The event layer carries the temporal directed structure of states and transitions. The constraint layer generates gating rules based on the behavioral topology phase information and the deformation results of the regulatory space. A gating mark is set on each transition to give the result of passing, restricting or eliminating, thus obtaining a set of restricted state transitions.
[0041] For each migration in the restricted state migration set, migration energy is configured. The migration energy is obtained by linear combination of time interval deviation component, resource sharing intensity component, role coupling intensity component, file change magnitude component, fund path change amount component, and operation rhythm mutation component according to preset weights.
[0042] For each time point of abnormal behavior, two levels of pruning are performed in the restricted state transition set based on the corresponding abnormal dimension information. Time coverage pruning only retains the transitions to reachable time points, and dimension matching pruning only retains the transitions that have a positive correlation with the abnormal dimension, generating a set of candidate potential paths with time points as the target.
[0043] Dynamic programming is used to determine the minimum energy potential path for the candidate potential path set. The key bidding entities, key resources and key behavior sequences involved in the potential path are extracted, and the abnormal behavior time points, corresponding abnormal dimension information, key bidding entities, key resources and key behavior sequences are output.
[0044] Optionally, the generation of the causal chain of anomalous behavior based on the minimum energy path includes:
[0045] Obtain the minimum energy potential path determined for each abnormal behavior time point. The minimum energy potential path includes a state transition sequence arranged in chronological order, as well as the key bidding entities, key resources, and key behavior sequence corresponding to the state transition sequence.
[0046] Based on the state transition order in the minimum energy potential path, the original operation data is time-aligned and rearranged to generate an event chain that corresponds one-to-one with the minimum energy potential path. Each event in the event chain includes a bidder identifier, behavior type identifier, timestamp, resource identifier, and corresponding original operation data index.
[0047] An abnormal behavior causal chain is constructed based on an event chain. The abnormal behavior causal chain consists of multiple causal nodes and causal edges connecting the causal nodes. The causal nodes are generated by events in the event chain, and the causal edges are generated by the state transition relationship between adjacent events. Each causal edge is associated with and recorded with the corresponding migration energy, behavior topology phase information identifier, and regulatory space deformation result identifier.
[0048] An evidence chain package is generated based on the causal chain of the abnormal behavior. The evidence chain package includes the time point of the abnormal behavior, the corresponding abnormal dimension information, the list of key bidding entities, the list of key resources, the sequence of key behaviors, the details of the event chain, and the original operation data index and timestamp information corresponding to the event chain.
[0049] The abnormal behavior causal chain and evidence chain package are output as the regulatory result of the electronic bidding and tendering transaction platform, and the abnormal behavior causal chain and evidence chain package are stored in the regulatory database.
[0050] The electronic bidding and tendering transaction platform supervision system according to an embodiment of the present invention includes the following modules:
[0051] The data acquisition and sequence construction module is used to collect operational data throughout the entire bidding process and construct a multivariate time series of bidding behavior in chronological order.
[0052] The behavior topology construction module is used to construct a behavior topology structure based on the relationship between behavior events and resource usage, and to calculate topological invariants to generate behavior topology phase information;
[0053] The multidimensional regulatory space analysis module is used to construct a multidimensional regulatory space based on regulatory indicators and perform curvature, density, twist rate and volume change analysis to obtain the deformation results of the regulatory space.
[0054] Anomaly detection module is used to detect anomalies in multivariate bidding behavior time series using the improved TranAD model, and output the time points of abnormal behavior and the corresponding anomaly dimension information.
[0055] The latent path and minimum energy solution module is used to construct the latent path space by combining the time points of abnormal behavior, the topological phase information of behavior, and the deformation results of the regulatory space, and to solve for the minimum energy path.
[0056] The causal chain output module is used to generate causal chains of abnormal behavior based on the minimum energy path and output them for regulatory review and verification.
[0057] The beneficial effects of this invention are:
[0058] This invention collects operational data from the entire bidding process and constructs a multivariate bidding behavior time series in chronological order. It incorporates multi-source data such as downloading, modifying, uploading, quoting, scoring, logging in, device fingerprinting, network address, deposit processing, and entity-resource association into a unified time series expression. This enables continuous monitoring of behavioral rhythm, key node changes, and multi-dimensional linkages, transforming supervision from traditional post-event spot checks to time-series supervision throughout the entire process, and improving the ability to detect abnormal signs and the timeliness of triggering them.
[0059] This invention further constructs a behavioral topology based on the relationship between behavioral events and resource usage, calculates topological invariants and generates behavioral topological phase information, constructs a multidimensional regulatory space based on regulatory indicators and outputs deformation results, and constrains and cross-verifies concealed collaborative behaviors from both structural and spatial links; on this basis, an improved TranAD model is used to detect anomalies in multivariate time series, outputting the time points of abnormal behaviors and corresponding anomaly dimension information, realizing multidimensional anomaly localization of complex collaborative behaviors such as bid rigging, collusion, and bid manipulation, and reducing false alarms and missed detections caused by relying on a single rule or single indicator.
[0060] This invention constructs a potential path space for bidding behavior by combining behavioral topological phase information and regulatory space deformation results after an anomaly is triggered. It sets migration energy for different types of behavioral changes and solves the minimum energy path from the normal state to the abnormal state, thereby generating a causal chain of abnormal behavior as a regulatory output. This transforms the regulatory result from a risk warning into a verifiable behavioral chain and subject relationship chain, providing regulatory authorities with interpretable and traceable evidence chain support, improving the efficiency and consistency of reviewing complex transaction behaviors, and meeting the real-time, interpretable, and traceable intelligent regulatory needs of electronic bidding platforms. Attached Figure Description
[0061] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0062] Figure 1 The flowchart is a process for supervising the electronic bidding and tendering transaction platform proposed in this invention.
[0063] Figure 2 This is a schematic diagram of the electronic bidding and tendering transaction platform supervision system proposed in this invention. Detailed Implementation
[0064] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0065] refer to Figure 1 The regulatory methods for electronic bidding and tendering transaction platforms include:
[0066] Collect operational data from the bidding entities, bidding companies, evaluation experts and agencies in the electronic bidding platform during the entire bidding process, and construct a multivariate bidding behavior time series in chronological order.
[0067] Based on the behavioral events and resource usage relationships of the bidding entities in the bidding process, a behavioral topology sequence is constructed, behavioral nodes and relational edges are generated, forming the topological structure of bidding behavior, calculating topological invariants and generating behavioral topological phase information;
[0068] A multidimensional regulatory space is constructed based on regulatory indicators. The state of the regulatory space is updated based on input behavior. Curvature calculation, density calculation, torsion rate calculation and local volume change analysis are performed to obtain the deformation results of the regulatory space.
[0069] An improved TranAD model was used to detect anomalies in the time series of multivariate bidding behavior, and information on the time points of abnormal behavior and the corresponding anomaly dimensions was obtained.
[0070] Based on the time points of abnormal behavior, combined with the behavioral topology phase information and the deformation results of the regulatory space, a potential path space for bidding behavior is constructed. Migration energy is set for different types of behavioral changes. Based on the state migration relationship, a set of potential behavioral paths is generated and the minimum energy path from the normal state to the abnormal state is solved.
[0071] The abnormal behavior causal chain is generated based on the minimum energy path, and the abnormal behavior causal chain is provided to the regulatory authorities as a regulatory output for review and verification.
[0072] In this embodiment, the operation data includes download behavior data, file modification behavior data, file upload behavior data, price change data, rating record data, user login data, device fingerprint data, network address data, deposit payment data, deposit refund data, and association data between bidding entities and resources.
[0073] In this embodiment, the step of constructing a multivariate bidding behavior time series in chronological order includes:
[0074] Download behavior data, file modification behavior data, file upload behavior data, price change data, scoring record data, user login data, device fingerprint data, network address data, and bid security deposit data are time-aligned, and various types of behavior data are mapped to corresponding time series variables using a unified time axis as an index, thereby forming a multivariate bidding behavior time series containing multiple behavioral dimensions.
[0075] In this embodiment, the process of forming the topology of bidding behavior, calculating topological invariants, and generating behavioral topological phase information includes:
[0076] A discrete time axis is defined, and the behavioral events of each bidding entity are sorted in chronological order to form a behavioral sequence of the bidding entity. The behavioral events include behavioral type identifiers and timestamps.
[0077] Construct a set of nodes and a set of edges for a behavioral topology. The set of nodes includes nodes of bidding companies, nodes of evaluation experts, resource nodes, and event nodes. The resource nodes include nodes of network addresses, nodes of device fingerprints, and nodes of fund accounts. The set of edges includes sequential edges, resource-related edges, time-coupled edges, and role-related edges, and records the corresponding time information for each edge.
[0078] The topology of the bidding behavior is constructed using the set of nodes and edges and time information. A topology sequence of time slices is generated according to the preset discrete time points. The existence relationship of various edges at the time points is recorded in each time slice. The preset discrete time points are a set of time points divided by a fixed time granularity, which is 1 minute to 10 minutes and remains unchanged within the same bidding project cycle.
[0079] Calculate topological feature quantities on the topological structure. The topological feature quantities include the number of simple rings formed by the combined action of sequential edges and resource-related edges, the average shortest path length of node pairs obtained based on time-coupled edges, the behavior synchronization index obtained based on the ratio of the longest common subsequence length between different behavior sequences, and the number of nested levels formed by the combined action of role-related edges and resource-related edges.
[0080] The topological features are determined based on preset weights and thresholds to generate behavioral topological phase information, which includes topological features and phase identifiers.
[0081] In this embodiment, obtaining the deformation result of the regulatory space includes:
[0082] Based on operational data, normalization and time alignment are performed at a fixed time granularity, and six regulatory indicators are generated through preset deterministic mapping rules. These six indicators are: time consistency indicator, resource sharing density indicator, role coupling strength indicator, document change manifold indicator, fund path curvature indicator, and operational rhythm indicator.
[0083] The time consistency index is calculated by the dispersion of the time interval sequence of adjacent behavioral events of the bidding entity within the time granularity window. The dispersion is the ratio of the standard deviation to the mean of the time interval sequence, and is normalized by comparing it with the historical baseline of the bidding entity within the project cycle.
[0084] The resource sharing density index is calculated by the number of times a bidding entity shares resources with other bidding entities within the time granularity window. Resource sharing includes network address sharing, device fingerprint sharing, and fund account sharing. The number of times is the sum of the number of times network address sharing, device fingerprint sharing, and fund account sharing, and is normalized according to the total number of actions of the bidding entity within the window.
[0085] The role coupling strength index is calculated by the number of associated events between the bidding entity and the evaluation experts and agents within a time granularity window. Associated events include the number of time-proximity co-occurrences of scoring events, scoring modification events, evaluation login events and evaluation result submission events within the same project. The time proximity means that the time difference between the two types of events does not exceed a preset proximity window. The index is obtained by normalizing the number of all associated events within the project.
[0086] The document change manifold index is calculated by combining the modification intensity and similarity of the bid documents within the time granularity window. The modification intensity is the ratio of the number of document modifications to the number of document uploads, and the document similarity is the proportion of consistent hash segments of the bid documents. The modification intensity and document similarity are then weighted according to preset weights.
[0087] The curvature index of the funding path is calculated by the degree of change in the flow of funds during the deposit payment and refund. The degree of change is a weighted combination of the similarity and difference markers of the payment account and the refund account, the number of times the funding account is reused, and the dispersion of the time interval between payment and refund. It is then normalized by comparing with the historical baseline within the project period.
[0088] The operation rhythm index is calculated by the occurrence frequency and switching frequency of five types of behaviors (download, modify, upload, quote, and login) within the time granularity window. The occurrence frequency is the number of each behavior divided by the window length, and the switching frequency is the number of times two adjacent behavior types are different divided by the total number of behaviors within the window. The occurrence frequency vector and the switching frequency are weighted according to preset weights.
[0089] A multi-dimensional regulatory space is constructed, which includes a basic indicator layer and a policy constraint layer. The basic indicator layer uses the six regulatory indicators mentioned above as coordinate axes, and the policy constraint layer uses project type and amount range as constraint dimensions. Anchor units are generated based on the compliance template library, and the multi-dimensional regulatory space is divided into rules. Each anchor unit has a defined boundary and compliance range.
[0090] The six regulatory indicators and policy constraint parameters corresponding to each bidding entity at each time point are mapped to state points in a multi-dimensional regulatory space. State trajectories are formed in chronological order. The neighborhood density, directional change and distance change of adjacent intervals are calculated in the fixed neighborhood of each state point to obtain local compression indication, local expansion indication and trajectory bending indication.
[0091] A standardized disturbance test is performed on the state point. The disturbance includes making small positive and negative adjustments to each regulatory indicator and policy constraint parameter with a fixed amplitude. The displacement amplitude and displacement direction of the state point in the multidimensional regulatory space are recorded to obtain sensitivity indication and stability indication. The template deviation indication is generated using the boundary of the anchoring unit as a constraint.
[0092] The neighborhood density indicator, direction and distance change indicator, sensitivity indicator, stability indicator and template deviation indicator are weighted and summarized to generate deformation score and deformation identifier, forming the deformation result of the regulatory space and outputting it together with the corresponding timestamp and bidding entity identifier.
[0093] In this embodiment, obtaining the time point of the abnormal behavior and the corresponding abnormal dimension information includes:
[0094] The six regulatory indicators are normalized and sliced according to fixed time windows to form a continuous and non-overlapping window sequence, and the original values of all dimensions are retained in each window.
[0095] Constructing an improved TranAD model:
[0096] A parallel surveillance mask path is added between the existing encoder and decoder. This surveillance mask path receives the surveillance space deformation result and generates a set of weight matrices, specifically:
[0097] The deformation results of the regulatory space are aligned according to the time window to form a deformation feature sequence. The deformation feature sequence is transformed into an attention weight vector through a linear mapping layer. The attention weight vector is then converted into a set of weight matrices corresponding to each attention head of the multi-head attention through a normalization layer and a reshaping layer.
[0098] Each weight matrix in the weight matrix set is multiplied element-wise with the attention score matrix of the corresponding attention head to complete the weighted modulation of attention allocation; the weight matrix set is updated once in each time window.
[0099] During the encoding phase, both the temporal embedding vector and the weight matrix are introduced to adjust the attention allocation of multi-head attention.
[0100] In the decoding stage, a dual-branch structure of sequential reconstruction and reverse reconstruction is adopted, and the reconstruction results of the two branches are output through a residual fusion layer, wherein:
[0101] The sequential reconstruction branch encodes the input sequential reconstruction branch, and generates the reconstruction output of each time step in the order of time of the samples within the time window, so as to obtain the sequential reconstruction result that is consistent with the original input time order;
[0102] The reverse reconstruction branch encodes the input to the reverse reconstruction branch. It first arranges the samples in the time window in reverse time order and then generates the reconstruction output of each time step in sequence to obtain the reverse reconstruction result. Then, it performs reverse restoration processing on the reverse reconstruction result to make the time order of the reverse reconstruction result consistent with the original input time window.
[0103] Unsupervised training of the improved TranAD model was performed within a data range containing only normal behavior. The training objective was to minimize the window-level reconstruction error. The TranAD model parameters were frozen after the TranAD model training was completed.
[0104] The continuous window sequence is input into the improved TranAD model to obtain the corresponding reconstructed window sequence. The window-level error vector is calculated by comparing the input window and the reconstructed window. The window anomaly score is determined according to the magnitude of the error values of each dimension in the window error vector. Then, the time point anomaly score and dimension-level anomaly score are generated according to the time points covered by the window.
[0105] Abnormal behavior time points are determined from the abnormal time point scores according to a preset quantile threshold, and the corresponding abnormal dimension information is obtained by combining the dimension-level abnormal scores. Specifically, determining abnormal behavior time points from the abnormal time point scores according to the preset quantile threshold involves:
[0106] Within a preset verification data range, all abnormal scores at each time point are aggregated, sorted from smallest to largest, and an abnormal score threshold corresponding to a preset percentile threshold is determined. The abnormal score at each time point during the operation is compared with the abnormal score threshold. When the abnormal score at a time point is not less than the abnormal score threshold, the time point is determined as an abnormal behavior time point and recorded.
[0107] In this embodiment, the step of generating a set of potential behavioral paths based on state transition relationships and solving for the minimum energy path from the normal state to the abnormal state includes:
[0108] Obtain the time point of abnormal behavior and the corresponding abnormal dimension information, obtain the behavior topology phase information and the deformation result of the regulatory space; generate a state set by discretizing the bid entity identifier, timestamp, behavior type and resource identifier, and establish the state reachability relationship according to the time sequence to form an initial state transition set.
[0109] A potential path space for bidding behavior is constructed, which consists of an event layer and a constraint layer. The event layer carries the temporal directed structure of states and transitions. The constraint layer generates gating rules based on the behavioral topology phase information and the deformation results of the regulatory space. A gating mark is set on each transition to give the result of passing, restricting or eliminating, thus obtaining a set of restricted state transitions.
[0110] For each migration in the restricted state migration set, migration energy is configured. The migration energy is obtained by linear combination of time interval deviation component, resource sharing intensity component, role coupling intensity component, file change magnitude component, fund path change amount component, and operation rhythm mutation component according to preset weights.
[0111] For each time point of abnormal behavior, two levels of pruning are performed in the restricted state transition set based on the corresponding abnormal dimension information. Time coverage pruning only retains the transitions to reachable time points, and dimension matching pruning only retains the transitions that have a positive correlation with the abnormal dimension, generating a set of candidate potential paths with time points as the target.
[0112] Dynamic programming is used to determine the minimum energy potential path for the candidate potential path set. The key bidding entities, key resources and key behavior sequences involved in the potential path are extracted, and the abnormal behavior time points, corresponding abnormal dimension information, key bidding entities, key resources and key behavior sequences are output.
[0113] In this embodiment, the generation of the causal chain of abnormal behavior based on the minimum energy path includes:
[0114] Obtain the minimum energy potential path determined for each abnormal behavior time point. The minimum energy potential path includes a state transition sequence arranged in chronological order, as well as the key bidding entities, key resources, and key behavior sequence corresponding to the state transition sequence.
[0115] Based on the state transition order in the minimum energy potential path, the original operation data is time-aligned and rearranged to generate an event chain that corresponds one-to-one with the minimum energy potential path. Each event in the event chain includes a bidder identifier, behavior type identifier, timestamp, resource identifier, and corresponding original operation data index.
[0116] An abnormal behavior causal chain is constructed based on an event chain. The abnormal behavior causal chain consists of multiple causal nodes and causal edges connecting the causal nodes. The causal nodes are generated by events in the event chain, and the causal edges are generated by the state transition relationship between adjacent events. Each causal edge is associated with and recorded with the corresponding migration energy, behavior topology phase information identifier, and regulatory space deformation result identifier.
[0117] An evidence chain package is generated based on the causal chain of the abnormal behavior. The evidence chain package includes the time point of the abnormal behavior, the corresponding abnormal dimension information, the list of key bidding entities, the list of key resources, the sequence of key behaviors, the details of the event chain, and the original operation data index and timestamp information corresponding to the event chain.
[0118] The abnormal behavior causal chain and evidence chain package are output as the regulatory result of the electronic bidding and tendering transaction platform, and the abnormal behavior causal chain and evidence chain package are stored in the regulatory database.
[0119] refer to Figure 2 The electronic bidding and tendering transaction platform supervision system includes the following modules:
[0120] The data acquisition and sequence construction module is used to collect operational data throughout the entire bidding process and construct a multivariate time series of bidding behavior in chronological order.
[0121] The behavior topology construction module is used to construct a behavior topology structure based on the relationship between behavior events and resource usage, and to calculate topological invariants to generate behavior topology phase information;
[0122] The multidimensional regulatory space analysis module is used to construct a multidimensional regulatory space based on regulatory indicators and perform curvature, density, twist rate and volume change analysis to obtain the deformation results of the regulatory space.
[0123] Anomaly detection module is used to detect anomalies in multivariate bidding behavior time series using the improved TranAD model, and output the time points of abnormal behavior and the corresponding anomaly dimension information.
[0124] The latent path and minimum energy solution module is used to construct the latent path space by combining the time points of abnormal behavior, the topological phase information of behavior, and the deformation results of the regulatory space, and to solve for the minimum energy path.
[0125] The causal chain output module is used to generate causal chains of abnormal behavior based on the minimum energy path and output them for regulatory review and verification.
[0126] Example 1:
[0127] To verify the feasibility of this invention in practice, it was applied to an electronic bidding and tendering platform. This platform handles online bidding for engineering construction and service procurement projects. The platform employs unified identity authentication, electronic bid document management, online bid evaluation, and electronic management of bid security deposits. Regulatory authorities supervise the transaction process through the platform's backend. Previously, supervision relied primarily on rule comparison and manual spot checks, such as checking for concentrated bidding times, abnormal pricing, and deviations of expert scores from the average. However, in actual operation, it was found that some bid-rigging, collusion, and bid evaluation manipulation behaviors circumvented rule triggers through decentralized operations, resource reuse, and rhythm control, leading to delayed anomaly detection and difficulty in establishing a clear causal chain of behavior.
[0128] In this embodiment, the method of the present invention is integrated into the platform backend as a regulatory analysis module to synchronously analyze existing platform data without altering the original business processes. The system first collects operational data from the bidding entity, bidding companies, evaluation experts, and agents throughout the entire bidding process, and constructs a multivariate bidding behavior time series in chronological order, covering download behavior, file modification behavior, file upload behavior, price changes, scoring records, user login, device fingerprints, network addresses, and information such as deposit payment and refund. All of the above data originates from existing platform logs and business records, ensuring completeness and traceability.
[0129] The system constructs a behavioral topology sequence based on the behavioral events and resource usage relationships of bidding entities during the bidding process, generates behavioral nodes and relationship edges, forms a bidding behavioral topology structure, and calculates topological invariants to generate behavioral topology phase information. In multiple projects, the system identified that some bidding companies formed highly similar behavioral paths in key stages and shared network addresses or device resources, and the topology structure exhibited stable and closed characteristics.
[0130] The system constructs a multi-dimensional regulatory space based on six regulatory indicators: time consistency, resource sharing density, role coupling strength, document change manifold, fund path curvature, and operational rhythm. It updates the space state based on input behavior and outputs the regulatory space deformation results. In some projects, multiple bidding companies exhibited a phenomenon of simultaneous convergence and rapid shift of multi-dimensional indicators within a short period, demonstrating a clear trend of collaborative behavior.
[0131] Next, the system uses an improved TranAD model to detect anomalies in the time series of multivariate bidding behavior, obtaining the time points of abnormal behavior and corresponding anomaly dimension information. In a certain engineering project, the system identified the time points of abnormal behavior approximately one hour after the bid opening, with the anomalies concentrated in network address, device fingerprint, and file modification rhythm.
[0132] The system uses the time points of abnormal behavior as anchor points, combines behavioral topological phase information with the deformation results of the regulatory space to construct a potential path space for bidding behavior. It sets migration energies for different types of behavioral changes and solves for the minimum energy path from the normal state to the abnormal state. The minimum energy path explicitly points to a small number of key bidding entities and key resource combinations, avoiding the traversal of a large number of irrelevant logs.
[0133] Finally, the system maps the minimum energy path to a causal chain of abnormal behavior and provides this causal chain as a regulatory output for review and verification by regulators. Regulators can directly retrieve the corresponding original data based on the causal chain to complete verification, reducing manual workload compared to traditional methods.
[0134] Table 1. Statistical Table of Anomaly Identification and Supervision Effectiveness in Trial Operation Projects
[0135]
[0136] As shown in Table 1, in the six bidding projects selected for the trial operation, the number of bidding companies ranged from 6 to 11, and the overall scale is consistent with the common level of engineering construction and service procurement projects, which can realistically reflect the business scenarios of the electronic bidding platform in daily operation. Under different project types, the method of this invention can effectively supervise bidding activities involving multiple entities, indicating that the method is not dependent on project scale or a single business type, and has good versatility and adaptability.
[0137] Regarding the identification of anomalies, all six projects identified and confirmed anomalies, showing a reasonable positive correlation between the number of confirmed anomalies and the number of bidding companies. In construction projects with a larger number of bidding companies, the number of confirmed anomalies was relatively higher; for example, 3 anomalies were confirmed in project T-06, which had 11 bidding companies. Conversely, in service procurement projects with a smaller number of bidding companies, the number of confirmed anomalies was relatively lower. This result indicates that the present invention can identify potential collaborative behaviors in scenarios with more complex multi-entity interactions without diminishing its effectiveness as the number of companies increases.
[0138] The average time to anomaly detection for each project ranged from 0.9 to 2.0 hours, with an overall average of approximately 1.45 hours, demonstrating good timeliness. Particularly in some construction projects, abnormal behavior was identified and reported by the system approximately one hour after the bid opening, significantly advancing the detection time compared to traditional methods relying on post-event complaints or manual spot checks. This indicates that the present invention, through multivariate time series analysis, behavioral topological constraints, and minimum energy path deduction, can promptly capture abnormal signs during bidding activities, providing regulatory authorities with an earlier intervention window.
[0139] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for supervising electronic bidding and tendering transaction platforms, characterized in that: include: Collect operational data from the bidding entities, bidding companies, evaluation experts and agencies in the electronic bidding platform during the entire bidding process, and construct a multivariate bidding behavior time series in chronological order. Based on the behavioral events and resource usage relationships of the bidding entities in the bidding process, a behavioral topology sequence is constructed, behavioral nodes and relational edges are generated, forming the topological structure of bidding behavior, calculating topological invariants and generating behavioral topological phase information; A multidimensional regulatory space is constructed based on regulatory indicators. The state of the regulatory space is updated based on input behavior. Curvature calculation, density calculation, torsion rate calculation and local volume change analysis are performed to obtain the deformation results of the regulatory space. An improved TranAD model was used to detect anomalies in the time series of multivariate bidding behavior, and information on the time points of abnormal behavior and the corresponding anomaly dimensions was obtained. Based on the time points of abnormal behavior, combined with the behavioral topology phase information and the deformation results of the regulatory space, a potential path space for bidding behavior is constructed. Migration energy is set for different types of behavioral changes. Based on the state migration relationship, a set of potential behavioral paths is generated and the minimum energy path from the normal state to the abnormal state is solved. The abnormal behavior causal chain is generated based on the minimum energy path, and the abnormal behavior causal chain is provided to the regulatory authorities as a regulatory output for review and verification.
2. The method for supervising an electronic bidding and tendering transaction platform according to claim 1, characterized in that, The operational data includes download behavior data, file modification behavior data, file upload behavior data, price change data, rating record data, user login data, device fingerprint data, network address data, deposit payment data, deposit refund data, and association data between bidding entities and resources.
3. The method for supervising an electronic bidding and tendering transaction platform according to claim 1, characterized in that, The construction of a multivariate bidding behavior time series in chronological order includes: Download behavior data, file modification behavior data, file upload behavior data, price change data, scoring record data, user login data, device fingerprint data, network address data, and bid security deposit data are time-aligned, and various types of behavior data are mapped to corresponding time series variables using a unified time axis as an index, thereby forming a multivariate bidding behavior time series containing multiple behavioral dimensions.
4. The method for supervising an electronic bidding and tendering transaction platform according to claim 1, characterized in that, The topological structure that forms the bidding behavior, calculating topological invariants and generating behavioral topological phase information, includes: A discrete time axis is defined, and the behavioral events of each bidding entity are sorted in chronological order to form a behavioral sequence of the bidding entity. The behavioral events include behavioral type identifiers and timestamps. Construct a set of nodes and a set of edges for a behavioral topology. The set of nodes includes nodes of bidding companies, nodes of evaluation experts, resource nodes, and event nodes. The resource nodes include nodes of network addresses, nodes of device fingerprints, and nodes of fund accounts. The set of edges includes sequential edges, resource-related edges, time-coupled edges, and role-related edges, and records the corresponding time information for each edge. The topology of the bidding behavior is constructed using the set of nodes and edges and time information, and a sequence of topology slices is generated according to preset discrete time points. The existence relationship of various edges at time points is recorded in each time slice. Calculate topological feature quantities on the topological structure. The topological feature quantities include the number of simple rings formed by the combined action of sequential edges and resource-related edges, the average shortest path length of node pairs obtained based on time-coupled edges, the behavior synchronization index obtained based on the ratio of the longest common subsequence length between different behavior sequences, and the number of nested levels formed by the combined action of role-related edges and resource-related edges. The topological features are determined based on preset weights and thresholds to generate behavioral topological phase information, which includes topological features and phase identifiers.
5. The method for supervising an electronic bidding and tendering transaction platform according to claim 1, characterized in that, The obtained regulatory space deformation results include: Based on operational data, normalization and time alignment are performed at a fixed time granularity, and six regulatory indicators are generated through preset deterministic mapping rules. The six regulatory indicators are time consistency indicator, resource sharing density indicator, role coupling strength indicator, document change manifold indicator, fund path curvature indicator, and operation rhythm indicator. A multi-dimensional regulatory space is constructed, which includes a basic indicator layer and a policy constraint layer. The basic indicator layer uses the six regulatory indicators mentioned above as coordinate axes, and the policy constraint layer uses project type and amount range as constraint dimensions. Anchor units are generated based on the compliance template library, and the multi-dimensional regulatory space is divided into rules. Each anchor unit has a defined boundary and compliance range. The six regulatory indicators and policy constraint parameters corresponding to each bidding entity at each time point are mapped to state points in a multi-dimensional regulatory space. State trajectories are formed in chronological order. The neighborhood density, directional change and distance change of adjacent intervals are calculated in the fixed neighborhood of each state point to obtain local compression indication, local expansion indication and trajectory bending indication. A standardized disturbance test is performed on the state point. The disturbance includes making small positive and negative adjustments to each regulatory indicator and policy constraint parameter with a fixed amplitude. The displacement amplitude and displacement direction of the state point in the multidimensional regulatory space are recorded to obtain sensitivity indication and stability indication. The template deviation indication is generated using the boundary of the anchoring unit as a constraint. The neighborhood density indicator, direction and distance change indicator, sensitivity indicator, stability indicator and template deviation indicator are weighted and summarized to generate deformation score and deformation identifier, forming the deformation result of the regulatory space and outputting it together with the corresponding timestamp and bidding entity identifier.
6. The method for supervising an electronic bidding and tendering transaction platform according to claim 1, characterized in that, The process of obtaining the time point of the abnormal behavior and the corresponding abnormal dimension information includes: The six regulatory indicators are normalized and sliced according to fixed time windows to form a continuous and non-overlapping window sequence, and the original values of all dimensions are retained in each window. Constructing an improved TranAD model: A parallel surveillance mask path is added between the original encoder and decoder. The surveillance mask path receives the surveillance space deformation result and generates a set of weight matrices. During the encoding phase, both the temporal embedding vector and the weight matrix are introduced to adjust the attention allocation of multi-head attention. In the decoding stage, a dual-branch structure of sequential reconstruction and reverse reconstruction is adopted, and the reconstruction results of the two branches are output through the residual fusion layer. Unsupervised training of the improved TranAD model was performed within a data range containing only normal behavior. The training objective was to minimize the window-level reconstruction error. The TranAD model parameters were frozen after the TranAD model training was completed. The continuous window sequence is input into the improved TranAD model to obtain the corresponding reconstructed window sequence. The window-level error vector is calculated by comparing the input window and the reconstructed window. The window anomaly score is determined according to the magnitude of the error values of each dimension in the window error vector. Then, the time point anomaly score and dimension-level anomaly score are generated according to the time points covered by the window. The abnormal behavior time point is determined from the abnormal score of the time point according to the preset quantile threshold, and the corresponding abnormal dimension information is obtained by combining the abnormal score of the dimension level.
7. The method for supervising an electronic bidding and tendering transaction platform according to claim 1, characterized in that, The process of generating a set of potential behavioral paths based on state transition relationships and solving for the minimum energy path from the normal state to the abnormal state includes: Obtain the time point of abnormal behavior and the corresponding abnormal dimension information, obtain the behavior topology phase information and the deformation result of the regulatory space; generate a state set by discretizing the bid entity identifier, timestamp, behavior type and resource identifier, and establish the state reachability relationship according to the time sequence to form an initial state transition set. A potential path space for bidding behavior is constructed, which consists of an event layer and a constraint layer. The event layer carries the temporal directed structure of states and transitions. The constraint layer generates gating rules based on the behavioral topology phase information and the deformation results of the regulatory space. A gating mark is set on each transition to give the result of passing, restricting or eliminating, thus obtaining a set of restricted state transitions. For each migration in the restricted state migration set, migration energy is configured. The migration energy is obtained by linear combination of time interval deviation component, resource sharing intensity component, role coupling intensity component, file change magnitude component, fund path change amount component, and operation rhythm mutation component according to preset weights. For each time point of abnormal behavior, two levels of pruning are performed in the restricted state transition set based on the corresponding abnormal dimension information. Time coverage pruning only retains the transitions to reachable time points, and dimension matching pruning only retains the transitions that have a positive correlation with the abnormal dimension, generating a set of candidate potential paths with time points as the target. Dynamic programming is used to determine the minimum energy potential path for the candidate potential path set. The key bidding entities, key resources and key behavior sequences involved in the potential path are extracted, and the abnormal behavior time points, corresponding abnormal dimension information, key bidding entities, key resources and key behavior sequences are output.
8. The method for supervising an electronic bidding and tendering transaction platform according to claim 1, characterized in that, The generation of causal chains for anomalous behavior based on minimum energy paths includes: Obtain the minimum energy potential path determined for each abnormal behavior time point. The minimum energy potential path includes a state transition sequence arranged in chronological order, as well as the key bidding entities, key resources, and key behavior sequence corresponding to the state transition sequence. Based on the state transition order in the minimum energy potential path, the original operation data is time-aligned and rearranged to generate an event chain that corresponds one-to-one with the minimum energy potential path. Each event in the event chain includes a bidder identifier, behavior type identifier, timestamp, resource identifier, and corresponding original operation data index. An abnormal behavior causal chain is constructed based on an event chain. The abnormal behavior causal chain consists of multiple causal nodes and causal edges connecting the causal nodes. The causal nodes are generated by events in the event chain, and the causal edges are generated by the state transition relationship between adjacent events. Each causal edge is associated with and recorded with the corresponding migration energy, behavior topology phase information identifier, and regulatory space deformation result identifier. An evidence chain package is generated based on the causal chain of the abnormal behavior. The evidence chain package includes the time point of the abnormal behavior, the corresponding abnormal dimension information, the list of key bidding entities, the list of key resources, the sequence of key behaviors, the details of the event chain, and the original operation data index and timestamp information corresponding to the event chain. The abnormal behavior causal chain and evidence chain package are output as the regulatory result of the electronic bidding and tendering transaction platform, and the abnormal behavior causal chain and evidence chain package are stored in the regulatory database.
9. An electronic bidding and tendering transaction platform supervision system, implementing the electronic bidding and tendering transaction platform supervision method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition and sequence construction module is used to collect operational data throughout the entire bidding process and construct a multivariate time series of bidding behavior in chronological order. The behavior topology construction module is used to construct a behavior topology structure based on the relationship between behavior events and resource usage, and to calculate topological invariants to generate behavior topology phase information; The multidimensional regulatory space analysis module is used to construct a multidimensional regulatory space based on regulatory indicators and perform curvature, density, twist rate and volume change analysis to obtain the deformation results of the regulatory space. Anomaly detection module is used to detect anomalies in multivariate bidding behavior time series using the improved TranAD model, and output the time points of abnormal behavior and the corresponding anomaly dimension information. The latent path and minimum energy solution module is used to construct the latent path space by combining the time points of abnormal behavior, the topological phase information of behavior, and the deformation results of the regulatory space, and to solve for the minimum energy path. The causal chain output module is used to generate causal chains of abnormal behavior based on the minimum energy path and output them for regulatory review and verification.