An evaluation system, method, and program product for multi-view evaluation

Abstract

An evaluation system, method and program product for multi-view evaluation. The system acquires multi-view original data of the same evaluation object, aligns the data according to object identification, time identification and task identification, and forms a time sequence multi-view data set; generates a quality vector, an uncertainty vector and a view representation based on the data set, constructs a cross-time multi-view evidence graph, and calculates a credible weight; performs consistency propagation and conflict suppression on the view representation based on the evidence graph and the uncertainty vector to obtain a modified representation; generates an intervention variable and an intervention response by taking the modified representation as a variable node of a structural causal model; calculates a fusion evaluation score based on the credible weight, the modified representation and the intervention response, and calculates a conflict contribution value according to the fusion evaluation score and the intervention response; when the conflict contribution value exceeds a threshold, the modified representation, the credible weight and the edge weight of the evidence graph are updated jointly, and the updated fusion evaluation score and the explainable information are output. The present application can improve the accuracy, robustness and explainability of the evaluation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and data fusion technology, and in particular to an evaluation system, method and program product for multi-perspective evaluation, which is applicable to intelligent evaluation scenarios that perform quality modeling, uncertainty modeling, evidence graph fusion, causal intervention analysis, conflict attribution and backtracking correction on multi-source heterogeneous data of the same evaluation object. Background Technology

[0002] With the development of sensor networks, business information systems, intelligent algorithm models, and multimodal data acquisition technologies, different types of data can often be obtained from multiple perspectives for the same assessment object. For example, in scenarios such as equipment status assessment, risk target identification, business event analysis, behavioral object evaluation, or sample quality judgment, the same object may simultaneously correspond to sensor data, text data, image data, business record data, model prediction results, and historical status data. Multi-perspective assessment techniques, by fusing data from different sources, with different structures, and at different time granularities, help improve the accuracy and stability of assessment results.

[0003] Existing multi-view evaluation methods typically include feature concatenation, weighted fusion, attention fusion, graph model propagation, or multi-model ensemble. While these methods can utilize complementary information from different perspectives to some extent, they still suffer from the following problems in complex, multi-source, heterogeneous environments:

[0004] First, existing methods typically fuse data based only on perspective features or model outputs, without fully considering quality factors such as data integrity, timeliness, noise level, source stability, and task relevance. This can lead to low-quality perspectives potentially causing significant interference with the final evaluation results.

[0005] Second, existing methods are insufficient in modeling uncertainty, especially in distinguishing between cognitive uncertainty caused by insufficient model understanding and data uncertainty caused by observation noise and data fluctuations, making it difficult to accurately quantify the reliability of different perspectives.

[0006] Third, existing graph model fusion methods mostly focus on modeling the relationship between different perspectives at the same time, lacking the modeling of cross-time state evolution relationships, and are difficult to reflect the changing patterns of the same or related perspectives at consecutive times.

[0007] Fourth, when there is information conflict between different perspectives, existing methods often deal with it through simple averaging, weight suppression or consistency constraints, which makes it difficult to judge the source of conflict from a causal perspective, and also makes it difficult to quantify the degree of contribution of a certain perspective to the deviation or inconsistency of the final evaluation results.

[0008] Fifth, the existing evaluation process is usually a one-way calculation process, that is, the evaluation result is directly output after feature fusion is completed. It lacks a closed-loop mechanism to jointly backtrack and correct the perspective representation, fusion weight and graph structure relationship based on the conflict attribution results, which leads to insufficient robustness of the evaluation system in scenarios with data drift, abnormal noise, missing perspective or relationship distortion.

[0009] Therefore, there is an urgent need for an assessment system, method, and procedure that can comprehensively consider multi-perspective data quality, uncertainty, cross-temporal evidentiary relationships, and causal intervention responses, and can retrospectively correct based on conflict contribution results, in order to improve the accuracy, robustness, and interpretability of assessment results. Summary of the Invention

[0010] The purpose of this invention is to provide an evaluation system, method, and program product for multi-perspective evaluation, in order to solve the problems existing in the current multi-perspective evaluation technology, such as insufficient utilization of data quality, insufficient modeling of uncertainty, insufficient expression of cross-time evidence relationships, difficulty in causally attributing conflict sources, and lack of closed-loop correction mechanisms.

[0011] This invention aligns object, time, and task identifiers on multi-perspective raw data of the same evaluation object to form a time-series multi-perspective dataset. It further generates quality vectors, uncertainty vectors, and perspective representations to construct a cross-time multi-perspective evidence graph, and calculates credible weights based on these vectors. The cross-time multi-perspective evidence graph and uncertainty vectors are then used to perform consistency propagation and conflict suppression on the perspective representations, resulting in a revised representation. The revised representation is then used as variable nodes in a structural causal model to generate intervention variables and intervention responses. A fusion evaluation score is calculated based on the credible weights, revised representation, and intervention response, and a conflict contribution value is calculated based on the fusion evaluation score and intervention response. When the conflict contribution value exceeds a threshold, the revised representation, credible weights, and edge weights of the cross-time multi-perspective evidence graph are jointly updated, thus forming a closed-loop processing mechanism of "evaluation—causal attribution—backtracking correction—re-evaluation," improving the accuracy, robustness, and interpretability of multi-perspective evaluation.

[0012] To achieve the above objectives, the present invention provides a multi-perspective evaluation system, comprising:

[0013] The data alignment module acquires multi-view raw data of the same evaluation object and aligns it according to object identifier, time identifier, and task identifier to form a time-series multi-view dataset. The quality and uncertainty modeling module generates quality vectors, uncertainty vectors, and view representations for each view at each time point based on the time-series multi-view dataset. The evidence graph construction module constructs a cross-time multi-view evidence graph based on the view representations. The credibility weight calculation module calculates credibility weights based on the quality vectors, uncertainty vectors, and the cross-time multi-view evidence graph. The consistency propagation and conflict suppression module propagates the view representations based on the cross-time multi-view evidence graph and the uncertainty vectors. The system includes: a revised representation; a causal intervention modeling module, which uses the revised representation as variable nodes in a structural causal model to apply intervention to the target perspective to generate intervention variables, and calculates the intervention response by combining the revised representations of other perspectives; a fusion evaluation module, which calculates a fusion evaluation score based on the credible weights, the revised representation, and the intervention response; a conflict causal attribution module, which calculates a conflict contribution value based on the fusion evaluation score and the intervention response; a backtracking correction module, which jointly updates the revised representation, the credible weights, and the edge weights of the cross-time multi-perspective evidence graph when the conflict contribution value exceeds a threshold; and a result output module, which outputs the updated fusion evaluation score and interpretable information.

[0014] This invention also provides a multi-perspective evaluation method, comprising:

[0015] The process involves: acquiring raw data from multiple perspectives of the same evaluation object and aligning it by object identifier, time identifier, and task identifier to form a time-series multi-perspective dataset; generating quality vectors, uncertainty vectors, and perspective representations for each perspective at each time step based on the time-series multi-perspective dataset; constructing a cross-time multi-perspective evidence graph based on the perspective representations; calculating confidence weights based on the quality vectors, uncertainty vectors, and the cross-time multi-perspective evidence graph; performing consistency propagation and conflict suppression on the perspective representations based on the cross-time multi-perspective evidence graph and uncertainty vectors to obtain a corrected representation; using the corrected representations as variable nodes in a structural causal model, applying intervention to the target perspective to generate intervention variables, and calculating the intervention response in conjunction with the corrected representations of other perspectives; calculating a fusion evaluation score based on the confidence weights, the corrected representations, and the intervention response; calculating a conflict contribution value based on the fusion evaluation score and the intervention response; jointly updating the edge weights of the corrected representations, the confidence weights, and the cross-time multi-perspective evidence graph when the conflict contribution value exceeds a threshold; and outputting the updated fusion evaluation score and interpretable information.

[0016] The present invention also provides a computer program product, including a computer-readable storage medium on which a computer program is stored, wherein when the computer program is executed by a processor, the evaluation method of the above-mentioned multi-view evaluation is implemented.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects:

[0018] 1. Improve the accuracy of multi-perspective assessment. This invention does not simply splice or weight multiple perspectives, but first aligns the original multi-perspective data by object identification, time identification, and task identification to form a time-series multi-perspective dataset. Then, it performs a fusion assessment based on quality vectors, uncertainty vectors, cross-time multi-perspective evidence graphs, and credibility weights. This allows the fusion assessment score to comprehensively reflect data quality, perspective reliability, inter-perspective relationships, and temporal evolution, thereby improving the accuracy of the assessment results.

[0019] 2. Enhanced ability to suppress low-quality and high-uncertainty data. This invention generates quality and uncertainty vectors through a quality and uncertainty modeling module, and uses them for trusted weight calculation, consistency propagation, and conflict suppression. This allows the influence of perspectives with low integrity, poor timeliness, high noise, unstable sources, or high uncertainty in the fusion evaluation to be reasonably suppressed, reducing the interference of abnormal perspectives on the evaluation results.

[0020] 3. Achieving cross-temporal evidence relationship modeling. This invention constructs a cross-temporal multi-view evidence graph, which not only expresses the relationship between different perspectives at the same moment, but also the cross-temporal relationship between the same perspective or related perspectives at adjacent moments. This enables the evaluation process to utilize historical state changes, continuous moment correlations, and temporal evolution patterns, thereby improving its adaptability to dynamic objects or temporal scenarios.

[0021] 4. Improve the reliability of conflict source identification. This invention uses modified representations as variable nodes in a structural causal model, applies intervention to the target perspective and calculates the intervention response, and then calculates the conflict contribution value based on the fused evaluation score and intervention response. This makes conflict identification no longer rely solely on surface differences or correlation biases, but can quantify the degree of contribution of the target perspective to the conflict in the evaluation results from the perspective of causal intervention response.

[0022] 5. Establishing a closed-loop backtracking and correction mechanism. When the conflict contribution value exceeds the threshold, this invention does not only adjust the final evaluation result, but also jointly updates and corrects the representation, credible weights, and edge weights of the cross-time multi-view evidence graph. This allows the representation layer, weight layer, and graph structure layer to be corrected synchronously, forming a closed-loop processing mechanism of evaluation, attribution, correction, and re-evaluation. This improves the robustness of the system in scenarios with noise anomalies, data drift, missing perspectives, and distorted relationships.

[0023] 6. Improve the interpretability of evaluation results. This invention outputs an updated fusion evaluation score and interpretable information. The interpretable information may include the credible weights of each perspective, the conflict contribution value, the perspective identifier that triggered the backtracking correction, the change in edge weights before and after correction, the change in credible weights before and after correction, and the change in the fusion evaluation score before and after correction, enabling users to understand the basis for the formation of the final evaluation result, the source of conflict, and the correction process.

[0024] 7. Facilitates software deployment and engineering applications. This invention protects the system, method, and program products, which can be deployed in a modular software form on servers, edge computing nodes, terminal devices, or cloud platforms. It is suitable for scenarios such as device status assessment, risk target assessment, business event analysis, behavioral object evaluation, multi-model output fusion, and intelligent analysis of multi-source heterogeneous data. Attached Figure Description

[0025] Figure 1 This is a block diagram of the multi-view evaluation system provided by the present invention.

[0026] Figure 2 This is a flowchart of the multi-perspective evaluation method provided by the present invention. Detailed Implementation

[0027] The following detailed description, in conjunction with specific embodiments, further illustrates the multi-perspective assessment system, method, and program product provided by this invention. This embodiment is applicable to scenarios involving the fusion assessment of multi-source heterogeneous data on the same assessment object. The assessment object can be equipment status, business events, risk targets, behavioral objects, sample objects, or entities to be analyzed. The multi-perspective raw data can originate from different sensors, different business systems, different algorithm models, different data modalities, or different time segments. This embodiment addresses the technical problems of inconsistent multi-perspective data quality, significant uncertainties, conflicts between perspectives, the easy suppression of real anomalies by ordinary smoothing fusion, and the need for interpretable output of assessment results. It sequentially executes data alignment, quality and uncertainty modeling, cross-time multi-perspective evidence graph construction, credible weight calculation, consistency propagation and conflict suppression, causal intervention modeling, fusion assessment, conflict causal attribution, backtracking correction, and result output, forming a closed-loop multi-perspective assessment processing chain.

[0028] In this embodiment, the same letters represent the same meaning, and different letters represent different meanings; the same terms represent the same technical features, and different terms represent different technical features. In the following formulas, superscripts... Indicates time index, subscript Indicates the first One perspective, subscript Indicates the first From multiple perspectives; when the same letter has different subscripts and superscripts, its overall symbol is interpreted as different parameters. In this implementation, the modules are coupled sequentially according to the data flow order of "raw data - quality and uncertainty - graph structure - credible weight - corrected representation - causal intervention - fusion evaluation - conflict attribution - retrospective correction - interpretable output", and form a feedback loop when the conflict contribution value exceeds the threshold.

[0029] First Embodiment

[0030] like Figure 1 As shown, the first embodiment of the present invention provides a multi-perspective evaluation system, including a data alignment module, a quality and uncertainty modeling module, an evidence graph construction module, a credible weight calculation module, a consistency propagation and conflict suppression module, a causal intervention modeling module, a fusion evaluation module, a conflict causal attribution module, a backtracking correction module, and a result output module.

[0031] The data coupling relationship between the modules is as follows: the data alignment module outputs a time-series multi-perspective dataset; the quality and uncertainty modeling module generates a quality vector, an uncertainty vector, and a perspective representation based on the time-series multi-perspective dataset; the evidence graph construction module constructs a cross-time multi-perspective evidence graph based on the perspective representation; the credibility weight calculation module calculates credibility weights based on the quality vector, the uncertainty vector, and the cross-time multi-perspective evidence graph; the consistency propagation and conflict suppression module propagates and updates the perspective representation based on the cross-time multi-perspective evidence graph and the uncertainty vector to obtain a corrected representation; the causal intervention modeling module generates intervention variables and intervention responses based on the corrected representation; the fusion evaluation module calculates a fusion evaluation score based on the credibility weights, the corrected representation, and the intervention response; the conflict causal attribution module calculates a conflict contribution value based on the fusion evaluation score and the intervention response; the backtracking correction module jointly updates the corrected representation, the credibility weights, and the edge weights of the cross-time multi-perspective evidence graph when the conflict contribution value exceeds a threshold; and the result output module outputs the updated fusion evaluation score and interpretable information.

[0032] In a preferred embodiment, the system further includes conflict type discrimination logic, cross-time edge freezing logic, and true anomaly retention logic. The conflict type discrimination logic distinguishes between low-quality noise conflicts, information-missing conflicts, graph relationship distortion conflicts, and true anomaly conflicts before backtracking correction. The cross-time edge freezing logic temporarily freezes the corresponding cross-time edge weights when stable support relationships exist at consecutive times, preventing one-time anomalies from erroneously weakening historical stable relationships. The true anomaly retention logic retains key corrected representations from the target perspective when the task relevance and source stability of the target perspective are both high and the intervention response changes significantly, preventing true anomalies from being misjudged as noise and suppressing them.

[0033] The three logics mentioned above are all embedded between the backtracking correction module and the conflict causation attribution module, so that when the system performs backtracking correction, it does not simply reduce the weight of the conflict perspective, but selects different correction paths according to the type of conflict source, thereby improving the ability of multi-perspective evaluation to identify complex and abnormal scenarios.

[0034] Data alignment module

[0035] The data alignment module is used to acquire multi-view raw data of the same evaluation object and align them according to object identifier, time identifier and task identifier to form a time-series multi-view dataset.

[0036] In one specific embodiment, the data alignment module reads raw data records from multiple data sources, which may include sensor acquisition terminals, business databases, model inference interfaces, log systems, image acquisition devices, text recording systems, or historical sample libraries. The data alignment module first determines whether data from different sources corresponds to the same evaluation object based on object identifiers; then, it maps data with different sampling frequencies, recording periods, and time granularities to a unified time axis based on time identifiers; and finally, it filters data fields, data fragments, or model output results related to the current evaluation task based on task identifiers.

[0037] A time-series multi-view dataset can be represented as:

[0038]

[0039] in, Represents a time-series multi-view dataset. Indicates the first A perspective at any moment Multi-perspective raw data, Indicates the total number of viewpoints. Indicates the total time. Indicates the viewpoint number. Indicates a time index.

[0040] When the time granularity of different perspectives is inconsistent, the data alignment module can perform window aggregation on high-frequency data and nearest neighbor mapping or time interpolation on low-frequency data. When the data structures of different perspectives are inconsistent, the data alignment module can perform field standardization, unit normalization, missing value marking, and encoding format unification. After processing, the same evaluation object forms a multi-perspective original data set that can be jointly evaluated at the same time.

[0041] In a preferred embodiment, the data alignment module also generates an alignment confidence identifier. , used to indicate the first A perspective at any moment The system determines whether the original data was obtained through direct acquisition, window aggregation, nearest neighbor mapping, or time interpolation. The alignment confidence flag is input to the quality and uncertainty modeling module and used to correct for timeliness indicators and data uncertainty. Therefore, when data from a certain perspective is obtained through interpolation or nearest neighbor mapping, its credibility in subsequent fusion evaluation can be appropriately reduced to avoid biases introduced by the alignment method itself being mistaken for true observation results.

[0042] Alignment confidence flags can be represented as:

[0043]

[0044] in, Indicates the first A perspective at any moment Alignment confidence flags, Indicates direct collection of identifiers. Indicates the window aggregation identifier, Indicates the nearest neighbor mapping identifier. Indicates the time interpolation identifier.

[0045] This module provides a unified data input foundation for subsequent modules. It avoids fusion biases caused by object mismatch, time misalignment, task inconsistency, or unclear alignment methods, thereby improving the reliability of subsequent quality modeling, evidence graph construction, and causal attribution.

[0046] Quality and Uncertainty Modeling Module

[0047] The quality and uncertainty modeling module is used to generate quality vectors, uncertainty vectors, and viewpoint representations for each viewpoint at each time step based on the time-series multi-view dataset.

[0048] 1. Mass Vector Generation

[0049] The mass vector can be represented as:

[0050]

[0051] in, Indicates the first A perspective at any moment The mass vector, Indicates the first A perspective at any moment Integrity indicators Indicates the first A perspective at any moment Timeliness indicators Indicates the first A perspective at any moment Noise index, Indicates the first A perspective at any moment Source stability indicators, Indicates the first A perspective at any moment Task relevance metrics.

[0052] In one specific embodiment, the integrity index is used to represent the completeness of valid fields, valid samples, or valid observations in the original data; the timeliness index is used to represent the time proximity between the original data and the target evaluation time; the noise index is used to represent the degree of outliers, the intensity of disturbances, or the degree of abnormal fluctuations in the original data; the source stability index is used to represent the stability of the output state of the corresponding data source within the historical time window; and the task relevance index is used to represent the degree of correlation between the corresponding perspective and the current evaluation task objective.

[0053] In a preferred embodiment, the quality vector is also associated with an alignment confidence identifier. Coupling. When When directly collecting identifiers, the timeliness index retains its original value calculated from the time difference; when When representing nearest neighbor mapping identifiers or time interpolation identifiers, the timeliness and completeness indicators are adjusted according to preset reduction factors. In this way, the system can distinguish between true observations and estimated values ​​obtained by time completion.

[0054] 2. Generation of Uncertainty Vectors

[0055] The uncertainty vector can be represented as:

[0056]

[0057] in, Indicates the first A perspective at any moment The uncertainty vector, Indicates the first A perspective at any moment Cognitive uncertainty, Indicates the first A perspective at any moment Data uncertainty.

[0058] Cognitive uncertainty is used to characterize the degree of uncertainty caused by insufficient model knowledge coverage, uncertain model parameters, or out-of-sample distribution; data uncertainty is used to characterize the degree of uncertainty caused by observation noise, acquisition error, data fluctuation, or unavoidable random disturbances.

[0059] In one specific embodiment, cognitive uncertainty can be estimated through model ensemble, Bayesian approximate inference, or Monte Carlo dropout; data uncertainty can be estimated through residual variance, reconstruction error, observation noise, or differences in repeated observations within the same time window.

[0060] In a preferred embodiment, data uncertainty Also with alignment confidence flags Coupling, when the first A perspective at any moment When data is obtained through time interpolation or nearest neighbor mapping, data uncertainty increases; when data is directly collected and the source has high stability, data uncertainty decreases. In this way, the system can pass on the estimation error introduced by the data alignment process to subsequent reliable weight calculation and propagation adjustment.

[0061] 3. Generation of viewpoint representation

[0062] The perspective can be represented as:

[0063]

[0064] in, Indicates the first A perspective at any moment From the perspective of Indicates the first Feature extraction models corresponding to each perspective Indicates the first The feature extraction model parameters corresponding to each viewpoint Indicates the first A perspective at any moment Multi-perspective raw data.

[0065] For structured data, the feature extraction model can be a multi-layer perceptron; for time-series data, the feature extraction model can be a temporal convolutional network, a gated recurrent unit, or a long short-term memory network; for text data, the feature extraction model can be an encoder model; and for image data, the feature extraction model can be a convolutional neural network.

[0066] 4. Joint calibration of quality and uncertainty

[0067] In a preferred embodiment, the quality and uncertainty modeling module also generates joint calibration coefficients. This is used to uniformly represent the first digit in subsequent trusted weight calculation and propagation adjustment. A perspective at any moment The degree of reliable participation.

[0068]

[0069] in, Indicates the first A perspective at any moment The joint calibration coefficients, Represents the normalized mapping function. This represents the calibration coefficient corresponding to the quality score. This represents the calibration coefficient corresponding to the uncertainty penalty term. Indicates by the first A perspective at any moment The quality score is obtained from the quality vector. Indicates by the first A perspective at any moment The uncertainty penalty term is obtained from the uncertainty vector.

[0070] The joint calibration coefficient can be input into the trusted weight calculation module and the consistency propagation and conflict suppression module to enable high-quality, low-uncertainty perspectives to have stronger fusion participation capabilities, while low-quality, high-uncertainty perspectives are restricted in their propagation impact.

[0071] 5. Model Training Methods

[0072] The training samples for the feature extraction model consist of historical multi-view raw data, historical evaluation labels, and auxiliary reconstruction targets or cross-view comparison sample pairs. During training, supervised learning, self-supervised learning, or semi-supervised learning methods can be used. If supervised learning is used, historical evaluation results are used as supervision labels to minimize the error between the predicted evaluation results and historical evaluation results. If self-supervised learning is used, the model is trained using reconstruction loss, temporal consistency loss, or cross-view comparison loss. If semi-supervised learning is used, the task loss of labeled samples and the consistency loss of unlabeled samples are used together as optimization objectives.

[0073] In a preferred embodiment, the feature extraction model also incorporates a cross-viewpoint contrastive learning loss, which brings high-quality viewpoint representations of the same evaluation object closer together at the same time, while keeping representations of different evaluation objects or low-quality conflicting viewpoints further apart. The resulting viewpoint representations can more accurately reflect the support and conflict relationships between viewpoints in subsequent evidence map construction.

[0074] The output of the quality and uncertainty modeling module includes , , and optional .in, and Input to the trusted weight calculation module, and Input consistency propagation and conflict suppression module, Input evidence graph construction module, Input the trusted weight calculation module and the consistency propagation and conflict suppression module.

[0075] This module transforms raw multi-view data into fusionable, measurable, and interpretable intermediate variables. It models data quality, uncertainty, alignment confidence, and viewpoint representation separately, enabling subsequent weight calculations and conflict suppression to simultaneously consider data reliability, model reliability, alignment reliability, and viewpoint semantic representation.

[0076] Evidence diagram construction module

[0077] The evidence graph construction module is used to construct a multi-view evidence graph across time based on the viewpoint representation.

[0078] The evidence diagram spanning multiple time perspectives can be represented as:

[0079]

[0080] in, Indicates time Evidence images spanning multiple time periods and perspectives. Indicates time The set of nodes formed by various perspectives Indicates time The set of edges between nodes from different perspectives. Indicates time The set of edge weights between nodes from different perspectives. This represents the set of cross-time edges connecting nodes at adjacent moments of the same or related viewpoints.

[0081] The set of nodes can be represented as:

[0082]

[0083] in, Indicates time The set of nodes, Indicates the first A perspective at any moment From the perspective of Indicates the first A perspective at any moment From the perspective of Indicates the first A perspective at any moment This is expressed from the perspective of [the subject / perspective].

[0084] The edge weight can be represented as:

[0085]

[0086] in, Indicates the first The first perspective and the first A perspective at any moment The right to the border, This represents the edge weight generating function. Indicates the first A perspective at any moment From the perspective of Indicates the first A perspective at any moment From the perspective of Indicates the first A perspective at any moment The mass vector, Indicates the first A perspective at any moment The mass vector, Indicates the first A perspective at any moment The change in state relative to adjacent time points. Indicates the first A perspective at any moment The change in state relative to adjacent time points.

[0087] in, Can be derived from the first A perspective at any moment The perspective of the first The differences between the viewpoint representations at adjacent time points are determined; Can be derived from the first A perspective at any moment The perspective of the first The differences between the viewpoint representations of each viewpoint at adjacent times are determined.

[0088] In a preferred embodiment, the evidence graph construction module also generates edge type identifiers. This is used to distinguish between supporting edges, conflicting edges, and edges to be verified.

[0089]

[0090] in, Indicates the first The first perspective and the first A perspective at any moment Edge type identifier, This indicates support for edge identifiers. Indicates conflict edge identifiers. This indicates an edge to be verified. If two perspectives have high similarity, high task relevance, and consistent state change direction over time, the edge type is identified as a supporting edge. If two perspectives differ significantly and at least one perspective has high quality and low uncertainty, the edge type is identified as an edge to be verified. If two perspectives differ significantly and one perspective has low quality or high uncertainty, the edge type is identified as a conflicting edge.

[0091] In a further embodiment, the evidence graph construction module also considers the cross-time edge set. Set a freeze indicator , used to indicate the first Whether the cross-time edge between adjacent moments of a viewpoint enters a frozen state.

[0092]

[0093] in, Indicates the first A perspective at any moment The corresponding freeze flag for cross-time edges, This indicates freezing edges across time. This indicates that time-crossing edges are not frozen. When the first... When a perspective exhibits high source stability, low data uncertainty, and state change below a preset stability threshold across multiple consecutive time points, the corresponding cross-time edge enters a frozen state; when high conflict contribution values ​​appear in subsequent consecutive time points and causal intervention responses support the occurrence of real changes, the corresponding cross-time edge is unfrozen.

[0094] By setting edge type identifiers and cross-time edge freeze identifiers, the system can distinguish between "conflicts caused by occasional noise" and "conflicts caused by real state changes," avoiding the erroneous destruction of stable historical relationships due to single-point anomalies and also avoiding the smoothing out of real mutation errors.

[0095] Evidence graph construction module receives As the core input, and can be combined , , and generation of state changes over time , , and The aforementioned The output is sent to the trusted weight calculation module, the consistency propagation and conflict suppression module, and the backtracking correction module.

[0096] This module aims to unify the horizontal and longitudinal relationships between multiple perspectives into a graph structure. Its beneficial effect is that it enables the evaluation process to not only rely on a single perspective but also utilize the mutual support, conflict, and verification relationships between perspectives, as well as the laws governing temporal evolution, thereby improving stability and anomaly detection capabilities in dynamic evaluation scenarios.

[0097] Trustworthy weight calculation module

[0098] The credible weight calculation module is used to calculate credible weights based on the quality vector, the uncertainty vector, and the cross-time multi-view evidence graph.

[0099] The credibility weight can be expressed as:

[0100]

[0101] in, Indicates the first A perspective at any moment Confidential weights, Indicates the first A perspective at any moment The unnormalized weight score is obtained by combining the quality vector, the uncertainty vector, and the neighborhood support strength of the cross-time multi-view evidence graph. Indicates the first A perspective at any moment Unnormalized weighted scores Indicates the total number of viewpoints. Indicates the viewpoint number involved in the normalization calculation.

[0102] The unnormalized weighted score can be expressed as:

[0103]

[0104] in, Indicates the first A perspective at any moment Unnormalized weighted scores The weighting coefficients representing the quality score. The weighting coefficients represent the uncertainty penalty terms. The weighting coefficients represent the strength of neighborhood support. The weighting coefficients represent the joint calibration coefficients. Indicates by the first A perspective at any moment The quality score is obtained from the quality vector. Indicates by the first A perspective at any moment The uncertainty penalty term is obtained from the uncertainty vector. Indicates the first A perspective at any moment In terms of neighborhood support strength in multi-view evidence maps across time, Indicates the first A perspective at any moment The joint calibration coefficients.

[0105] The neighborhood support strength can be expressed as:

[0106]

[0107] in, Indicates the first A perspective at any moment neighborhood support strength, Indicates the first The first perspective and the first A perspective at any moment The right to the border, Indicates the first The first perspective and the first A perspective at any moment Edge type identifier, This indicates support for edge identifiers. Indicates conflict edge identifiers. Indicates an indicator function, Indicates the first The perspective is different from the first. One perspective.

[0108] In one specific embodiment, if the first If a data source has high data integrity, good timeliness, low noise, high source stability, and strong task relevance, and exhibits high support edge neighborhood strength in the cross-time multi-view evidence graph with other high-quality perspectives, then its unnormalized weight score is... Higher, corresponding to a higher credibility weight Larger; if the first If a viewpoint has high uncertainty or many conflicting edges, its credibility weight is reduced.

[0109] In a preferred embodiment, the trusted weight calculation module also retains the trusted weights before correction. This is used by the subsequent result output module to calculate the change in confidence weights, and by the backtracking correction module to determine whether oversuppression has occurred. When the confidence weight of a certain highly task-relevant perspective decreases by more than the weight change threshold in a single correction, the system enters a review state for that perspective, and the causal intervention response determines whether it is a noise conflict or a real anomaly.

[0110] This module receives , , , and Output The aforementioned Input fusion evaluation module and backtracking correction module.

[0111] This module quantifies the credibility of different perspectives participating in the fusion evaluation at the current moment. It avoids low-quality, high-uncertainty, or graph-structure-lacking perspectives from having an excessive impact on the final evaluation results, and improves the ability to distinguish between supportive and conflicting relationships by identifying edge types, thereby enhancing the ability to discriminate credibility weights.

[0112] Consistency Propagation and Conflict Suppression Module

[0113] The consistency propagation and conflict suppression module is used to propagate and update the perspective representation based on the cross-time multi-view evidence graph and the uncertainty vector to obtain a corrected representation.

[0114] The correction can be represented as:

[0115]

[0116] in, Indicates the first A perspective at any moment The correction indicates that Indicates the first A perspective at any moment From the perspective of Indicates the first A perspective at any moment From the perspective of Indicates the first The first perspective and the first A perspective at any moment The right to the border, Indicates the first A perspective at any moment The uncertainty vector, Indicates the first A perspective at any moment The uncertainty vector, Indicates by the first A perspective at any moment The uncertainty vector and the first A perspective at any moment The propagation regulation function is jointly determined by the uncertainty vectors. Indicates the first The first perspective and the first A perspective at any moment Edge type identifier, Represents the propagation symbol function determined by the edge type identifier.

[0117] The propagation regulation function can be expressed as:

[0118]

[0119] in, Represents the propagation regulation function. Indicates the first A perspective at any moment The uncertainty vector, Indicates the first A perspective at any moment The uncertainty vector, Indicates by the first A perspective at any moment The uncertainty penalty term is obtained from the uncertainty vector. Indicates by the first A perspective at any moment The uncertainty penalty term is obtained from the uncertainty vector.

[0120] The propagation symbol function can be expressed as:

[0121]

[0122] in, Indicates by the first The first perspective and the first A perspective at any moment The propagation symbol function is determined by the edge type identifier. Indicates the edge type identifier. This indicates support for edge identifiers. Indicates the edge identifier to be verified. Indicates conflict edge identifiers. This represents the conflict suppression coefficient.

[0123] In one specific embodiment, when the first The first perspective and the first When the edges between the three perspectives are supporting edges and the uncertainty between them is low, the first... The representation of the first perspective on the second The correction effect of each perspective is strong; when the edge is an edge to be verified, the edge will not participate in the propagation correction for the time being; when the edge is a conflict edge, the information in the conflict direction is suppressed by the conflict suppression coefficient to prevent the spread of highly uncertain conflict information.

[0124] In a preferred embodiment, if the time-bound edge is frozen... Then the first The revised representation from each perspective retains a high proportion of historically stable representations during propagation and updating; if Then the first Each perspective can be fully updated based on neighborhood information at the current moment. Therefore, the system can suppress occasional noise in a steady state and allow for rapid response in a dynamically changing state.

[0125] Consistency Propagation and Collision Suppression Module Receive , , , and Output The aforementioned Input the causal intervention modeling module, the fusion evaluation module, and the retrospective correction module.

[0126] This module incorporates neighborhood support information from the graph structure into the viewpoint representation, while suppressing the negative propagation of highly uncertain or conflicting viewpoints. This improves the consistency, stability, and noise resistance of the viewpoint representation, and prevents genuine anomalies from being masked by simple averaging processes.

[0127] Causal Intervention Modeling Module

[0128] The causal intervention modeling module is used to apply intervention to the target perspective to generate intervention variables by using the modified representation as the variable node of the structural causal model, and to calculate the intervention response by combining the modified representations of other perspectives.

[0129] The intervention variable can be represented as:

[0130]

[0131] in, Indicates the first A perspective at any moment Intervention variables, Indicates the first A perspective at any moment The correction indicates that Indicates the application to the first A perspective at any moment Intervention disturbance variables.

[0132] The intervention response can be expressed as:

[0133]

[0134] in, Indicates the first A perspective at any moment Intervention response after intervention is applied This refers to the structural causal model. Indicates the first A perspective at any moment Intervention variables, Indicates except the first Other perspectives besides the one mentioned at any given moment The modified representation set.

[0135] in, It can be represented as:

[0136]

[0137] in, Indicates except the first Other perspectives besides the one mentioned at any given moment The modified representation set, Indicates the first A perspective at any moment The correction indicates that Indicates the total number of viewpoints. Indicates the first The perspective is different from the first. One perspective.

[0138] In one specific embodiment, the intervention perturbation variable These can be null perturbations, shielding perturbations, enhancing perturbations, substitution perturbations, or controlled offset perturbations, used to simulate changes in the evaluation response after the target perspective is altered, weakened, or replaced. By comparing the difference between the intervention response and the current fusion evaluation score, the degree of causal influence of the target perspective on the current evaluation result can be determined.

[0139] In a preferred embodiment, the causal intervention modeling module also generates counterfactual consistency identifiers. , used to indicate the first A perspective at any moment Whether the intervention response is consistent with counterfactual responses in similar historical scenarios.

[0140]

[0141] in, Indicates the first A perspective at any moment Counterfactual consistency markers, Indicates the first A perspective at any moment Intervention response after intervention is applied Indicates the first of similar historical scenarios The average intervention response from each perspective at the corresponding time. Indicates the counterfactual consistency threshold. This indicates an indicator function.

[0142] when When this occurs, it indicates that the current intervention response is consistent with similar counterfactual responses in the past; when If the current intervention response deviates from the historical counterfactual response of the same type, it is necessary to further combine the quality vector, edge type and conflict contribution value to determine whether it belongs to a real anomaly or anomalous noise.

[0143] Structural causal model training methods

[0144] The training samples for the structural causal model include a set of historical modified representations, historical intervention samples, and historical evaluation results. If a known causal prior relationship exists, it can be added to the model training as a structural mask or constraint edge. If no explicit causal prior relationship exists, causal structure learning can be performed first based on the historical modified representation sequence to obtain candidate causal structures, and then the structural causal model can be trained based on the candidate causal structures.

[0145] Training objectives may include prediction loss, intervention consistency loss, and counterfactual consistency loss. Prediction loss aims to make the structural causal model output approximate historical true evaluation results; intervention consistency loss aims to ensure the structural causal model maintains stable and interpretable response changes under intervention variable inputs; counterfactual consistency loss aims to constrain counterfactual responses to remain reasonably consistent across similar scenarios and preserve response biases in truly anomalous scenarios. The parameters of the structural causal model can be updated using gradient descent, stochastic gradient descent, or adaptive moment estimation optimization algorithms.

[0146] Causal intervention modeling module receives Output , and optional .in, Input fusion evaluation module and conflict causal attribution module, Input conflict causal attribution module and backtracking correction module.

[0147] This module characterizes the impact of changes in the target perspective on the evaluation results from a causal perspective. It enables the system not only to identify differences in correlation between perspectives but also to quantify the causal perturbation effect of the target perspective on the evaluation results, and to help distinguish between ordinary noise and genuine anomalies through counterfactual consistency indicators.

[0148] Fusion Assessment Module

[0149] The fusion assessment module is used to calculate the fusion assessment score based on the credible weights, the modified representation, and the intervention response.

[0150] The integration assessment score can be expressed as:

[0151]

[0152] in, Indicates time The integration assessment score This represents an evaluation function that integrates the weighted modified representation with the intervention response. Indicates the first A perspective at any moment Confidential weights, Indicates the first A perspective at any moment The correction indicates that Indicates the first A perspective at any moment Intervention response, Indicates from the first From one perspective to the first A perspective at any moment The set of intervention responses Indicates the total number of viewpoints.

[0153] In one specific embodiment, the evaluation function It can be implemented using a multilayer perceptron, regression model, classification model, scoring model, or rule mapping function. Weighted correction representation. The intervention response set is used to reflect the overall perspective fusion result based on credible weights and modified representations. This module is used to reflect the responses of different perspectives to the evaluation results under causal intervention. The fusion evaluation module inputs both types of information into the evaluation function to obtain the fusion evaluation score at the current moment.

[0154] In a preferred embodiment, the fusion evaluation module also bases its evaluation on counterfactual consistency identifiers. Generate causal reliable gating values The causal reliability gating value is then used to modulate the impact of the intervention response on the fusion assessment score. This indicates that the intervention response from this perspective conforms to historical counterfactual patterns, and the intervention response can serve as a strong constraint; if Furthermore, if the perspective is of high quality and highly relevant to the task, the intervention response is retained as a true anomaly candidate; if Furthermore, if the quality of the perspective is low or the uncertainty is high, the intervention response will be regarded as a noisy intervention response and weakened.

[0155] The causal reliability gate value can be expressed as:

[0156]

[0157] in, Indicates the first A perspective at any moment The causal reliability gate value, Represents the normalized mapping function. This represents the gating coefficient corresponding to the counterfactual consistency flag. This represents the gating coefficient corresponding to the quality score. This represents the gating coefficient corresponding to the uncertainty penalty term. Indicates the first A perspective at any moment Counterfactual consistency markers, Indicates by the first A perspective at any moment The quality score is obtained from the quality vector. Indicates by the first A perspective at any moment The uncertainty penalty term is obtained from the uncertainty vector.

[0158] Evaluation function training method

[0159] Evaluation function The training samples include historical credible weights, historical corrected representations, historical intervention responses, and historical assessment labels. During training, the error between the predicted fusion assessment score and the historical assessment labels can be used as the primary loss, while the consistency constraint between the intervention response and the assessment results can be used as the auxiliary loss, so that the fusion assessment score conforms to the historical assessment results without deviating from the perspective influence relationship revealed by the causal intervention response.

[0160] In a preferred embodiment, the evaluation function also incorporates an anomaly retention loss. For samples labeled as true anomalies in historical data, the anomaly retention loss constraint system must not completely weaken the influence of a few key anomaly perspectives due to the consistency of the majority of common perspectives; for samples labeled as noisy anomalies in historical data, the anomaly retention loss constraint system reduces the impact of noisy perspectives on the fusion evaluation score. Thus, the fusion evaluation module can simultaneously achieve noise suppression and true anomaly retention.

[0161] Fusion evaluation module receives , , and optional Output The aforementioned Input conflict causal attribution module and result output module.

[0162] This module unifies credible weights, corrected representations, and intervention responses into the final evaluation result. This allows the fused evaluation score to simultaneously reflect data quality, graph structure consistency, and causal response characteristics, improving the accuracy and interpretability of the evaluation results.

[0163] Conflict Causal Attribution Module

[0164] The conflict causal attribution module is used to calculate the conflict contribution value based on the fusion assessment score and the intervention response.

[0165] The conflict contribution value can be expressed as:

[0166]

[0167] in, Indicates the first A perspective at any moment Conflict contribution value, Represents norm operations, Indicates time The integration assessment score Indicates the first A perspective at any moment Intervention response.

[0168] In one specific embodiment, when the first Intervention response obtained after intervention from a different perspective Compared with the current integration assessment score When the difference is large, it indicates that the first Each perspective has a strong perturbation or conflict with the current assessment results, and its conflict contribution value Higher; when the difference between the two is small, it indicates that the first Individual perspectives that are consistent with the overall assessment results or whose causal influence is weak have a lower contribution value to the conflict.

[0169] In a preferred embodiment, the conflict causation attribution module also calculates a conflict type identifier. , used to indicate the first A perspective at any moment The types of conflicts.

[0170]

[0171] in, Indicates the first A perspective at any moment Conflict type identifier, This represents a conflict type discrimination function. Indicates the first A perspective at any moment Conflict contribution value, Indicates the first A perspective at any moment The mass vector, Indicates the first A perspective at any moment The uncertainty vector, Indicates the first The first perspective and the first A perspective at any moment Edge type identifier, Indicates the first A perspective at any moment Counterfactual consistency markers.

[0172] Conflict type identifier This can include low-quality noise conflicts, information-missing conflicts, graph relation distortion conflicts, and real anomaly conflicts. If the... If the conflict contribution value of the first perspective is high, the noise index is high, and the source stability or uncertainty is low, then the conflict type is low-quality noise conflict; if the first perspective has high conflict contribution value, high noise index, low source stability or high uncertainty, then the conflict type is low-quality noise conflict; If the integrity index of the first perspective is low and the historical state is stable across time edges, then the conflict type is an information missing conflict; if the first perspective is low and the historical state is stable across time edges, then the conflict type is an information missing conflict. If a viewpoint has high quality but its edge types with multiple viewpoints are consistently abnormal, then the conflict type is graph relationship distortion conflict; if the first viewpoint has high quality but its edge types with multiple viewpoints are consistently abnormal, then the conflict type is graph relationship distortion conflict. If a conflict is characterized by high relevance to the task from a given perspective, high stability of the source, high contribution value of the conflict, and counterfactual consistency indicators showing deviation from historical similar responses, then the conflict type is a genuine anomalous conflict.

[0173] Conflict causality attribution module receives , , , , and Output and The aforementioned and Input backtracking correction module and result output module.

[0174] This module pinpoints inter-perspective conflicts to specific perspectives, moments, and conflict types. This enables the system to quantify conflict sources from a causal intervention and response perspective and to perform differentiated backtracking corrections based on conflict type, avoiding the application of the same suppression strategy to all conflict perspectives.

[0175] Backtracking Correction Module

[0176] The backtracking correction module is used to jointly update the corrected representation, the trusted weight, and the edge weights of the cross-time multi-view evidence graph when the conflict contribution value exceeds the threshold.

[0177] The joint update process can be represented as:

[0178]

[0179] in, This represents the backtracking correction function. Indicates the first A perspective at any moment Conflict contribution value, Indicates the first A perspective at any moment The mass vector, Indicates the first A perspective at any moment The uncertainty vector, Indicates the first A perspective at any moment The correction indicates that Indicates the first A perspective at any moment Confidential weights, Indicates the first The first perspective and the first A perspective at any moment The right to the side.

[0180] In one specific embodiment, the system pre-sets a conflict threshold. When the following conditions are met:

[0181]

[0182] At that time, a backtracking correction is triggered. Among them, Indicates the first A perspective at any moment Conflict contribution value, This indicates the conflict threshold.

[0183] In a preferred embodiment, the backtracking correction function further introduces a conflict type identifier. Cross-time edge freeze mark and counterfactual consistency mark The joint update process is represented as:

[0184]

[0185] in, Indicates the first A perspective at any moment Conflict type identifier, Indicates the first A perspective at any moment The corresponding freeze flag for cross-time edges, Indicates the first A perspective at any moment The counterfactual consistency marker is used; the meanings of the other symbols are consistent with those in the preceding text.

[0186] When the conflict type is a low-quality noise conflict, the backtracking correction module reduces the credibility weight of the target viewpoint and weakens the conflict edge weights between the target viewpoint and its neighboring nodes. When the conflict type is an information-missing conflict, the backtracking correction module compensates for the target viewpoint's corrected representation by using cross-time edges and supporting neighboring nodes. When the conflict type is a graph relationship distortion conflict, the backtracking correction module re-estimates the relevant edge weights and updates the edge type identifiers. When the conflict type is a real anomaly conflict, the backtracking correction module retains the key corrected representation of that viewpoint, limits the decrease in credibility weight, and adjusts the weight allocation of other viewpoints to prevent real anomalies from being suppressed by most ordinary viewpoints.

[0187] In a further embodiment, the backtracking correction module sets a correction confidence threshold value. This is used to determine the magnitude of the correction in this round.

[0188]

[0189] in, Indicates the first A perspective at any moment The corrected confidence gate value, Represents the normalized mapping function. This represents the gating coefficient corresponding to the conflict contribution value. This represents the gating coefficients corresponding to the joint calibration coefficients. This represents the gating coefficient corresponding to the uncertainty penalty term. Indicates the first A perspective at any moment Conflict contribution value, Indicates the first A perspective at any moment The joint calibration coefficients, Indicates by the first A perspective at any moment The uncertainty penalty term is obtained from the uncertainty vector.

[0190] A higher correction confidence threshold indicates greater confidence in the current correction direction; a lower threshold indicates a reduction in the correction magnitude to avoid overcorrection. This approach allows the backtracking correction module to avoid excessive adjustments to viewpoint weights or edge weights in a single conflict determination, thus improving the stability of the closed-loop iteration.

[0191] Updated , and The data is resubmitted to the fusion evaluation module and the conflict causation attribution module to obtain an updated fusion evaluation score and an updated conflict contribution value. If the updated conflict contribution value still exceeds the threshold, backtracking correction continues until the conflict contribution value does not exceed the threshold or the preset number of iterations is reached.

[0192] This module feeds back the conflict causal attribution results to the representation layer, weight layer, and graph structure layer. This forms a closed loop of evaluation, attribution, correction, and re-evaluation, improving the system's robustness in scenarios involving noise anomalies, data drift, missing perspectives, distorted relationships, and coexisting real anomalies.

[0193] Result Output Module

[0194] The results output module is used to output the updated fusion evaluation score and interpretable information.

[0195] In one specific embodiment, the interpretable information output by the result output module includes the credible weights of each viewpoint, the conflict contribution value, the viewpoint identifier that triggered the backtracking correction, the change in edge weights before and after correction, the change in credible weights before and after correction, and the change in fusion evaluation score before and after correction.

[0196] The change in confidence weights before and after the correction can be expressed as:

[0197]

[0198] in, Indicates the first A perspective at any moment The change in the credible weight, Indicates the first A perspective at any moment Corrected credibility weights Indicates the first A perspective at any moment The credibility weight before correction.

[0199] The change in edge weights before and after the correction can be expressed as:

[0200]

[0201] in, Indicates the first The first perspective and the first A perspective at any moment The change in edge weight, Indicates the first The first perspective and the first A perspective at any moment The revised edge weights Indicates the first The first perspective and the first A perspective at any moment Edge weights before modification.

[0202] The change in the fusion assessment score before and after the correction can be expressed as:

[0203]

[0204] in, Indicates time The change in the fusion assessment score Indicates time Revised fusion assessment score Indicates time The original fusion assessment score.

[0205] In a preferred embodiment, the result output module also outputs a conflict type identifier. Edge type identifier Cross-time edge freeze mark Counterfactual consistency markers and corrected confidence gate value Through the above output, users can identify which perspectives support the current evaluation results, which perspectives conflict, whether the conflicts are noise or real anomalies, which cross-time relationships are preserved, and the credibility of this round of corrections.

[0206] The result output module receives the updated... , , , , , , , , The system backtracks and corrects the results, outputting evaluation results and interpretable information.

[0207] This module transforms the system's internal evaluation process into outputs that users can understand, verify, and use. It enhances the transparency of multi-perspective evaluation results, making it easier for users to identify key perspectives, sources of conflict, the basis for corrections, and the basis for retaining genuine anomalies.

[0208] Second Embodiment

[0209] The second embodiment provides an evaluation method for multi-view evaluation performed by the multi-view evaluation system described in the first embodiment, based on the multi-view evaluation system described in the first embodiment; the system modules, variable meanings and data coupling relationships described in the first embodiment are applicable to this embodiment.

[0210] like Figure 2 As shown, the second embodiment of the present invention provides a multi-view evaluation method, which can be executed by the above-mentioned multi-view evaluation system, and includes the following steps.

[0211] Step S1: Acquire and align the raw data from multiple perspectives

[0212] Obtain raw data from multiple perspectives for the same evaluation object, and align them according to object identifier, time identifier, and task identifier to form a time-series multi-perspective dataset. The output of this step Enter step S2.

[0213] In a preferred embodiment, step S1 also generates an alignment confidence identifier. The alignment confidence flag is then input into step S2 to correct the quality vector and uncertainty vector.

[0214] The purpose of this step is to unify data from different sources, at different time granularities, and with different structures. Its beneficial effect is that it provides a unified input for subsequent quality modeling, evidence graph construction, and fusion evaluation, and explicitly propagates the estimation errors caused by alignment methods to subsequent steps.

[0215] Step S2: Generate the quality vector, uncertainty vector, and viewpoint representation.

[0216] Based on time-series multi-perspective datasets , generate the first A perspective at any moment mass vector Uncertainty vector and perspective representation .in, and Enter step S4, Enter steps S3 and S5.

[0217] In a preferred embodiment, step S2 further generates joint calibration coefficients. Then input it into steps S4 and S5.

[0218] The feature extraction model involved in this step is trained according to the aforementioned training method. The purpose of this step is to transform the raw data into computable quality, uncertainty, and representation variables. Its beneficial effect is to provide a reliable data foundation for subsequent weight calculations and conflict suppression.

[0219] Step S3: Construct a multi-perspective evidence map across time.

[0220] Based on perspective representation Constructing evidence maps across time and multiple perspectives The multi-view evidence graph spanning time includes edges between different viewpoints at the same moment and edges spanning time. This step outputs... Enter steps S4, S5, and S9.

[0221] In a preferred embodiment, step S3 also generates an edge type identifier. and time-bound edge freeze marker The edge type identifier and the cross-time edge freeze identifier are then input into steps S4, S5, S8, and S9, respectively.

[0222] The purpose of this step is to establish relationships between perspectives and across time evolution. Its beneficial effect is that it enables the system to be evaluated using lateral perspective support, conflict relationships, and longitudinal temporal continuity.

[0223] Step S4: Calculate the confidence weight

[0224] Based on mass vector Uncertainty vector Joint calibration coefficients Edge type identifier and evidence diagrams from multiple perspectives across time Calculate the confidence weight The output of this step Enter steps S7 and S9.

[0225] The purpose of this step is to determine the credibility of each perspective in the fusion evaluation. Its beneficial effect is to suppress low-quality, high-uncertainty, and graph-structure-lacking perspectives, while enhancing the contribution of high-quality supporting perspectives.

[0226] Step S5: Perform consensus propagation and conflict suppression

[0227] Based on evidence images from multiple perspectives across time Uncertainty vector Edge type identifier and time-bound edge freeze marker , representation of perspective The update is propagated and a corrected representation is obtained. The output of this step Enter steps S6, S7, and S9.

[0228] This step aims to refine the perspective representation using neighborhood evidence while suppressing the propagation of conflicts with high uncertainty. Its beneficial effects include improving the consistency and noise resistance of the perspective representation and enabling differentiated processing between stable and truly changing states.

[0229] Step S6: Perform causal intervention modeling

[0230] To correct As variable nodes in a structural causal model, intervention is applied to the target perspective, generating intervention variables. And combined with the modified representation set from other perspectives Calculate the intervention response The output of this step Enter steps S7 and S8.

[0231] In a preferred embodiment, step S6 also generates a counterfactual consistency identifier. The counterfactual consistency identifier is then input into steps S7, S8, and S9.

[0232] The structural causal model involved in this step is trained according to the aforementioned structural causal model training method. The purpose of this step is to obtain the causal response of changes in the target perspective to the evaluation results. Its beneficial effect is to provide causal evidence for the identification of conflict sources.

[0233] Step S7: Calculate the fusion assessment score

[0234] Based on credibility weight Correction representation Intervention response and counterfactual consistency mark Calculate the fusion assessment score The output of this step Enter steps S8 and S10.

[0235] The evaluation function involved in this step is trained according to the aforementioned evaluation function training method. The purpose of this step is to generate a fusion evaluation result specific to the evaluation object. Its beneficial effect is that the evaluation result simultaneously reflects credibility, graph structure consistency, causal response, and the mechanism for preserving true anomalies.

[0236] Step S8: Calculate the conflict contribution value and conflict type identifier.

[0237] Based on fusion evaluation score and intervention response Calculate the conflict contribution value And based on the conflict contribution value Mass vector Uncertainty vector Edge type identifier and counterfactual consistency mark Calculate the conflict type identifier The output of this step and Enter steps S9 and S10.

[0238] This step aims to quantify the contribution of the target perspective to the current assessment results and identify the type of conflict. Its beneficial effect lies in enabling causal attribution of conflict sources and preparation for differentiated corrections.

[0239] Step S9: Perform backtracking correction

[0240] When the conflict contribution value Exceeding the conflict threshold At that time, based on the conflict type identifier Cross-time edge freeze mark Counterfactual consistency markers Mass vector and uncertainty vector Joint update and revision indicate Credible weight And the edge weights of evidence graphs across time and multiple perspectives The updated results are then re-entered into steps S7 and S8 to form a closed-loop iteration.

[0241] In a preferred embodiment, step S9 further adjusts the confidence threshold value. Adjust the magnitude of this round of corrections to avoid over-correction.

[0242] This step serves to feed back the conflict attribution results to the representation layer, weight layer, and graph structure layer. Its beneficial effect is to improve the system's robustness in complex and abnormal scenarios and to implement different correction strategies for low-quality noise, missing information, distorted graph relationships, and genuine anomalies.

[0243] Step S10: Output the result

[0244] Output the updated fusion evaluation score and interpretable information. The interpretable information includes the credible weights of each perspective, the conflict contribution value, the perspective identifier that triggered the backtracking correction, the change in edge weights before and after the correction, the change in credible weights before and after the correction, the change in the fusion evaluation score before and after the correction, the conflict type identifier, the edge type identifier, the cross-time edge freeze identifier, the counterfactual consistency identifier, and the corrected confidence gate value.

[0245] The purpose of this step is to provide users with the final evaluation results and the basis for those results. Its benefit lies in improving the interpretability and usability of the evaluation results.

[0246] Third Embodiment

[0247] A third embodiment provides a computer program product, including a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements the aforementioned multi-perspective evaluation method.

[0248] In one specific embodiment, the computer-readable storage medium may be a read-only memory, random access memory, solid-state drive, magnetic disk, optical disk, flash memory, or other tangible storage medium capable of storing program code. The computer program may be divided into multiple program units, each corresponding to processing logic such as data alignment, quality and uncertainty modeling, evidence graph construction, credible weight calculation, consistency propagation and conflict suppression, causal intervention modeling, fusion evaluation, conflict causal attribution, backtracking correction, and result output.

[0249] When the processor calls and executes the computer program, it reads the original data of the same evaluation object from multiple perspectives, generates a quality vector, an uncertainty vector, a perspective representation, and joint calibration coefficients, constructs a cross-time multi-perspective evidence graph, an edge type identifier, and a cross-time edge freeze identifier, calculates the confidence weight, generates a modified representation, performs causal intervention to obtain intervention variables, intervention responses, and counterfactual consistency identifiers, calculates the fusion evaluation score, conflict contribution value, and conflict type identifier, and performs joint backtracking correction based on the modified confidence gating value when the conflict contribution value exceeds the threshold, and finally outputs the updated fusion evaluation score and interpretable information.

[0250] In summary, this implementation aligns the original data of the same evaluation object from multiple perspectives using object identifiers, time identifiers, and task identifiers, generating quality vectors, uncertainty vectors, and perspective representations. It constructs a cross-temporal multi-perspective evidence graph, calculates credible weights based on quality, uncertainty, and graph structure, and obtains a corrected representation through consistency propagation and conflict suppression. Furthermore, it utilizes a structural causal model to intervene in the target perspective, generating an intervention response, and calculates the conflict contribution value based on the fused evaluation score and the intervention response. When the conflict contribution value exceeds a threshold, it jointly updates the corrected representation, credible weights, and edge weights of the cross-temporal multi-perspective evidence graph, thus forming a closed-loop multi-perspective evaluation mechanism consisting of data alignment, quality modeling, graph structure modeling, causal intervention, conflict attribution, and retrospective correction. By further introducing alignment confidence identifiers, edge type identifiers, cross-temporal edge freezing identifiers, counterfactual consistency identifiers, conflict type identifiers, and corrected confidence gate values, it can more accurately distinguish between low-quality noise, missing information, graph relationship distortion, and true anomalies, avoiding erroneous suppression of true anomalies, while reducing the impact of noise and data drift on the fused evaluation results, thereby improving the accuracy, robustness, and interpretability of the multi-perspective evaluation results.

Claims

1. A multi-perspective evaluation system, characterized in that, include: The data alignment module is used to acquire multi-view raw data of the same evaluation object and align them according to object identifier, time identifier and task identifier to form a time-series multi-view dataset. The quality and uncertainty modeling module is used to generate quality vectors, uncertainty vectors, and perspective representations for each perspective at each time point based on the time-series multi-perspective dataset. The evidence graph construction module is used to construct a multi-view evidence graph across time based on the viewpoint representation; The credible weight calculation module is used to calculate the credible weight based on the quality vector, the uncertainty vector, and the cross-time multi-view evidence graph. The consistency propagation and conflict suppression module is used to propagate and update the perspective representation based on the cross-time multi-perspective evidence graph and the uncertainty vector to obtain a corrected representation; the causal intervention modeling module is used to apply intervention to the target perspective using the corrected representation as the variable node of the structural causal model to generate intervention variables, and calculate the intervention response by combining the corrected representations of the other perspectives. The fusion assessment module is used to calculate the fusion assessment score based on the credible weights, the corrected representation, and the intervention response; The conflict causation attribution module is used to calculate the conflict contribution value based on the fusion assessment score and the intervention response; The backtracking correction module is used to jointly update the corrected representation, the trusted weight, and the edge weights of the cross-time multi-view evidence graph when the conflict contribution value exceeds the threshold. The results output module is used to output the updated fusion evaluation score and interpretable information.

2. The multi-view evaluation system according to claim 1, characterized in that, The time-series multi-view dataset generated by the data alignment module satisfies: in, Represents a time-series multi-view dataset. Indicates the first A perspective at any moment Multi-perspective raw data, Indicates the total number of viewpoints. Indicates the total time. Indicates the viewpoint number. The time index is represented; the data alignment module performs window aggregation, nearest neighbor mapping or time interpolation on multi-view raw data with different sampling frequencies, and performs field standardization, dimension normalization, missing value marking and encoding format unification on multi-view raw data with different structures.

3. The multi-view evaluation system according to claim 1, characterized in that, The quality vector and uncertainty vector generated by the quality and uncertainty modeling module satisfy the following: in, Indicates the first A perspective at any moment The mass vector, Indicates the first A perspective at any moment Integrity indicators Indicates the first A perspective at any moment Timeliness indicators Indicates the first A perspective at any moment Noise index, Indicates the first A perspective at any moment Source stability indicators, Indicates the first A perspective at any moment Task relevance metrics; in, Indicates the first A perspective at any moment The uncertainty vector, Indicates the first A perspective at any moment Cognitive uncertainty, Indicates the first A perspective at any moment Data uncertainty; the perspective represents the following: in, Indicates the first A perspective at any moment From the perspective of Indicates the first Feature extraction models corresponding to each perspective Indicates the first Feature extraction model parameters corresponding to each viewpoint Indicates the first A perspective at any moment Multi-perspective raw data.

4. The multi-view evaluation system according to claim 1, characterized in that, The evidence graph constructed by the evidence graph construction module across time and multiple perspectives satisfies the following: in, Indicates time Evidence images spanning multiple time periods and perspectives. Indicates time The set of nodes formed by various perspectives Indicates time The set of edges between nodes from different perspectives. Indicates time The set of edge weights between nodes from different perspectives. This represents the set of cross-time edges connecting nodes at adjacent time points of the same or related viewpoints; the edge weights in the set of edge weights satisfy: in, Indicates the first The first perspective and the first A perspective at any moment The right to the border, This represents the edge weight generating function. Indicates the first A perspective at any moment From the perspective of Indicates the first A perspective at any moment From the perspective of Indicates the first A perspective at any moment The mass vector, Indicates the first A perspective at any moment The mass vector, Indicates the first A perspective at any moment The change in state relative to adjacent time points. Indicates the first A perspective at any moment The change in state relative to adjacent time points.

5. The multi-view evaluation system according to claim 1, characterized in that, The trusted weights calculated by the trusted weight calculation module satisfy the following: in, Indicates the first A perspective at any moment Credibility weight, Indicates the first A perspective at any moment Unnormalized weighted scores Indicates the first A perspective at any moment Unnormalized weighted scores Indicates the total number of viewpoints. Indicates the viewpoint number participating in the normalization calculation; the unnormalized weight score satisfies: in, The weighting coefficients representing the quality score. The weighting coefficients represent the uncertainty penalty terms. The weighting coefficients represent the strength of neighborhood support. The weighting coefficients represent the joint calibration coefficients. Indicates by the first A perspective at any moment The quality score is obtained from the quality vector. Indicates by the first A perspective at any moment The uncertainty penalty term is obtained from the uncertainty vector. Indicates the first A perspective at any moment In terms of neighborhood support strength in the cross-time multi-view evidence map, Indicates the first A perspective at any moment The joint calibration coefficients; the neighborhood support strength satisfies: in, Indicates the first The first perspective and the first A perspective at any moment The right to the border, Indicates the first The perspective is different from the first. One perspective.

6. The multi-view evaluation system according to claim 1, characterized in that, The corrected representation obtained by the consensus propagation and conflict suppression module satisfies: in, Indicates the first A perspective at any moment The correction indicates that Indicates the first A perspective at any moment From the perspective of Indicates the first A perspective at any moment From the perspective of Indicates the first The first perspective and the first A perspective at any moment The right to the border, Indicates the first A perspective at any moment The uncertainty vector, Indicates the first A perspective at any moment The uncertainty vector, Indicates by the first A perspective at any moment The uncertainty vector and the first A perspective at any moment The propagation regulation function is jointly determined by the uncertainty vectors. Indicates the first The first perspective and the first A perspective at any moment Edge type identifier, The propagation sign function is defined by the edge type identifier; the propagation adjustment function satisfies: in, Indicates by the first A perspective at any moment The uncertainty penalty term is obtained from the uncertainty vector. Indicates by the first A perspective at any moment The uncertainty penalty term is obtained from the uncertainty vector.

7. The multi-view evaluation system according to claim 1, characterized in that, The intervention variables and intervention responses generated by the causal intervention modeling module satisfy the following: in, Indicates the first A perspective at any moment Intervention variables, Indicates the first A perspective at any moment The correction indicates that Indicates the application to the first A perspective at any moment Intervention disturbance variables; in, Indicates the first A perspective at any moment Intervention response after intervention is applied Represents a structural causal model. Indicates except the first Other perspectives besides the one mentioned at any given moment The set of modified representations for the remaining perspectives; the set of modified representations for the remaining perspectives satisfies: in, Indicates the first A perspective at any moment The correction indicates that Indicates the total number of viewpoints. Indicates the first The perspective is different from the first. One perspective.

8. The multi-view evaluation system according to claim 1, characterized in that, The fusion evaluation score calculated by the fusion evaluation module and the conflict contribution value calculated by the conflict causal attribution module satisfy the following: in, Indicates time The integration assessment score This represents an evaluation function that integrates the weighted modified representation with the intervention response. Indicates the first A perspective at any moment Confidential weights, Indicates the first A perspective at any moment The correction indicates that Indicates the first A perspective at any moment Intervention response, Indicates from the first From one perspective to the first A perspective at any moment The set of intervention responses Represents the total number of viewpoints; the conflict contribution value satisfies: in, Indicates the first A perspective at any moment Conflict contribution value, Represents norm operations, Indicates time The integration assessment score Indicates the first A perspective at any moment Intervention response; the backtracking correction module, when satisfying At that time, according to: The corrected representation, the trusted weights, and the edge weights of the cross-time multi-view evidence graph are jointly updated, wherein, Indicates the conflict threshold. This represents the backtracking correction function. Indicates the first A perspective at any moment The mass vector, Indicates the first A perspective at any moment The uncertainty vector.

9. A multi-perspective evaluation method, characterized in that, include: Obtain raw data from multiple perspectives of the same evaluation object, and align them according to object identifier, time identifier, and task identifier to form a time-series multi-perspective dataset; Based on the aforementioned time-series multi-view dataset, generate quality vectors, uncertainty vectors, and view representations for each view at each time step; construct a cross-time multi-view evidence graph based on the view representations; and calculate confidence weights based on the quality vectors, uncertainty vectors, and the cross-time multi-view evidence graph. Based on the cross-time multi-view evidence graph and the uncertainty vector, consistency propagation and conflict suppression are performed on the viewpoint representation to obtain a modified representation; the modified representation is used as the variable node of the structural causal model to apply intervention to the target viewpoint to generate intervention variables, and the intervention response is calculated by combining the modified representations of the other viewpoints. A fusion evaluation score is calculated based on the credible weights, the corrected representation, and the intervention response. The conflict contribution value is calculated based on the fusion assessment score and the intervention response; When the conflict contribution value exceeds the threshold, the corrected representation, the credible weight, and the edge weights of the cross-time multi-view evidence graph are jointly updated; the updated fusion evaluation score and interpretable information are output.

10. A computer program product, characterized in that, It includes a computer-readable storage medium on which a computer program is stored, which, when executed by a processor, implements the multi-view evaluation method as described in claim 9.

Images