A logistics claim risk labeling method and system based on relative difference hierarchical clustering
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-21
Smart Images
Figure CN122434261A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of logistics claims risk identification and intelligent risk control technology, specifically involving a logistics claims risk labeling method and system based on relative difference hierarchical clustering. Background Technology
[0002] As the focus of competition in the logistics industry gradually shifts from scale expansion to improving service quality and customer experience, claims processing efficiency and claims risk control have become crucial management aspects for logistics companies. Companies typically need to classify customer claims based on the actual compensation amounts, claim amounts, and discrepancies in historical waybills, and accordingly adopt differentiated processing strategies such as expedited claims processing, enhanced review, or focused verification.
[0003] In existing methods for identifying claims risks, a common approach is to classify claims based on the amount difference or a pre-set fixed threshold. However, the amount difference itself is easily affected by the magnitude of the claim: for shipments with high payouts, a large absolute difference does not necessarily indicate an anomaly; for shipments with low payouts, a small absolute difference may already reflect a significant overpayment. Therefore, ignoring the crucial factor of the actual payout amount and relying solely on a uniform threshold or a single difference indicator for judgment often leads to inconsistent risk standards between samples from different amount ranges, affecting the accuracy of risk labeling and the interpretability of the business.
[0004] Furthermore, claims data often suffers from outlier samples, significant differences in distribution across different amount levels, and unstable risk category boundaries. Existing technologies lack a systematic solution for an integrated approach of "amount stratification—density clustering—distribution fitting—risk label mapping" in logistics claims scenarios, making it difficult to balance labeling accuracy, rule adaptability, and subsequent business implementation needs. Summary of the Invention
[0005] To address the problems in existing technologies, such as distorted risk assessment due to differences in monetary amounts, poor adaptability of uniform thresholds, and inconsistent risk standards across different compensation ranges, this invention proposes a logistics claims risk labeling method based on hierarchical clustering of relative differences. This method includes: acquiring historical waybill claims data, which at least includes the actual compensation amount and the claim amount; calculating the claim difference based on the actual compensation amount and the claim amount, and constructing a relative difference index; identifying and labeling outliers in the relative difference; dividing the sample into multiple monetary layers based on the quantiles of the actual compensation amount; performing cluster analysis on the relative difference samples using the DBSCAN clustering algorithm to obtain intra-layer clustering results; performing Gaussian mixture model fitting on the intra-layer clustering results to obtain the risk range boundaries for each monetary layer; labeling the waybill claims based on the risk range boundaries, and outputting the risk labeling results.
[0006] A logistics claims risk labeling system based on relative difference hierarchical clustering, the system includes: a data acquisition module, a difference construction module, an anomaly identification module, an amount hierarchical module, a hierarchical clustering module, a distribution fitting module, and a risk output module;
[0007] The data acquisition module is used to acquire historical waybill claims data; the historical waybill claims data includes at least the actual compensation amount and the claim amount.
[0008] The difference construction module calculates the claim difference based on the actual compensation amount and the claim amount, and constructs a relative difference index;
[0009] The anomaly identification module is used to identify and mark outliers in the relative difference.
[0010] The amount stratification module divides the sample into multiple amount strata based on the quantile of the actual compensation amount;
[0011] The hierarchical clustering module is used to perform cluster analysis on the relative difference samples to obtain intra-layer clustering results;
[0012] The distribution fitting module is used to perform Gaussian mixture model fitting on the intra-layer clustering results to obtain the risk interval boundaries of each amount layer;
[0013] The risk output module marks the waybill claim based on the risk range boundary and outputs the risk marking results.
[0014] The beneficial effects of this invention are:
[0015] This invention constructs a relative difference index instead of a simple absolute difference as the core variable for risk identification, which can significantly reduce the judgment bias caused by differences in monetary magnitude and improve the comparability between different waybill samples. This invention uses a stratified mechanism based on actual compensation amounts to form adaptive risk boundaries within different monetary strata, avoiding misjudgments caused by uniform thresholds and improving the precision of risk labeling. This invention combines DBSCAN density clustering with Gaussian mixture model fitting, which can identify the local density structure of sample distribution and form stable and interpretable risk interval boundaries, enhancing the robustness and business interpretability of the model. The risk labeling results output by this invention can be directly used for the diversion management of logistics claims processes, which is beneficial to improving processing efficiency, reducing the risk of excessive compensation, and providing data support for subsequent pre-sales risk assessment and insurance product recommendations. Attached Figure Description
[0016] Figure 1 This is an overall flowchart of the present invention;
[0017] Figure 2 This is a block diagram of the logistics claims risk labeling system of the present invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] A logistics claims risk labeling method based on hierarchical clustering of relative differences, such as Figure 1 and Figure 2 As shown, the method includes: acquiring historical waybill claims data, which at least includes the actual compensation amount and the claim amount; calculating the claim difference based on the actual compensation amount and the claim amount, and constructing a relative difference index; identifying and marking outliers in the relative difference; dividing the sample into multiple amount layers based on the quantiles of the actual compensation amount; performing cluster analysis on the relative difference samples using the DBSCAN clustering algorithm to obtain intra-layer clustering results; performing Gaussian mixture model fitting on the intra-layer clustering results to obtain the risk interval boundaries of each amount layer; labeling the waybill claims based on the risk interval boundaries, and outputting the risk labeling results.
[0020] In this embodiment, historical waybill claims data is first obtained, including at least the actual compensation amount and the claim amount. The claim difference D is calculated based on these two fields, and a relative difference R is further constructed to transform the difference indicator into a proportional quantity comparable to the amount, reducing dimensional interference between samples of different compensation scales.
[0021] Preferably, the relative difference is constructed in the form of R = D / (actual compensation amount + ε), where ε is a preset minimum amount.
[0022] After obtaining the relative differences, outlier identification and labeling are performed on the samples. Preferably, at least one of the interquartile range method and the standard deviation method is used to determine the outlier boundary. Samples exceeding the outlier boundary are not directly deleted, but are marked and retained to balance the stability of subsequent modeling and the needs of risk control auditing.
[0023] Specifically, determining the outlier boundary of relative differences using the interquartile range (IQR) and / or standard deviation method includes: obtaining a relative difference sequence; calculating the first quartile Q1 and the third quartile Q3 of the relative difference sequence, and obtaining the interquartile range according to IQR = Q3 - Q1, where the first outlier boundary obtained based on the IQR method is [Q1 - k1·IQR, Q3 + k1·IQR], and k1 is a preset strict coefficient; calculating the mean μ and standard deviation σ of the relative difference sequence, and obtaining the second outlier boundary based on the standard deviation method is [μ - k2·σ, μ + k2·σ], where k2 is a preset standard deviation coefficient; when both the IQR and standard deviation methods are used simultaneously, the intersection interval of the first outlier boundary and the second outlier boundary is taken as the final outlier boundary, or the stricter upper and lower bounds of the two methods are selected as the final outlier boundary; samples exceeding the final outlier boundary are marked as outliers and retained in the original sample set for subsequent risk identification and manual review.
[0024] In this embodiment, the samples are stratified according to the actual compensation amount.
[0025] Preferably, the samples are divided into low-amount, low-to-medium-amount, medium-to-high-amount, and high-amount strata according to the 25th, 50th, and 75th percentiles. This stratification makes the monetary magnitudes of the samples in each stratum more similar, thus establishing a unified comparative basis for subsequent risk boundary identification.
[0026] In this embodiment, within each monetary level, DBSCAN density clustering analysis is performed using the relative difference as the core input variable. The neighborhood radius can be adaptively determined based on the k-distance distribution of the samples in the current level, and the minimum number of samples can be set as a certain proportion of the sample size in the current level, preferably 5% and at least 3. Through DBSCAN clustering, the initial cluster structure within each level can be obtained based on the sample density and proximity relationships in the data space.
[0027] Specifically, the DBSCAN clustering algorithm is used to perform clustering analysis on the relative difference samples, including: constructing a one-dimensional feature sample set for the relative difference samples in each amount layer; determining the neighborhood radius eps of DBSCAN clustering based on the k-distance distribution of the samples in the current amount layer, preferably by obtaining the distance value corresponding to the point with the largest gradient change by elbow detection on the k-distance sorting curve as the neighborhood radius eps; determining the minimum number of samples min_samples based on the sample size n of the current amount layer, wherein the minimum number of samples satisfies min_samples=max(3,α·n), where α is a preset proportional parameter, preferably 0.05; performing DBSCAN clustering on the relative difference samples in the current amount layer using the neighborhood radius eps and the minimum number of samples min_samples to obtain the cluster label of each sample; dividing the relative difference samples into core samples, boundary samples, and noise samples according to the density connectivity and proximity relationships of the samples in the feature space, and generating an initial cluster structure within the layer for subsequent distribution fitting.
[0028] After obtaining the initial cluster structure within each layer, a Gaussian mixture model is further used to fit the sample distribution of each amount layer to obtain a smoother and more interpretable risk interval boundary. Preferably, the means of each component obtained from the fitting are sorted from largest to smallest and mapped sequentially to reasonable claims, excessive claims, and serious over-limit claims; samples with absolute values exceeding a preset serious threshold can be directly classified into the serious over-limit claim category to enhance the ability to identify high-risk samples.
[0029] Finally, based on the risk range boundaries obtained from each amount level, the waybills to be labeled are classified into three categories: reasonable claims, excessive claims, and serious over-limit claims, and the risk label results are output. These label results can be used for scenarios such as rapid claims processing, manual review, financial risk control, and subsequent model training for logistics companies.
[0030] Specifically, the Gaussian mixture model fitting of the intra-layer clustering results includes: performing Gaussian mixture model fitting on the relative difference samples corresponding to the initial cluster structure obtained by DBSCAN clustering within each amount layer to obtain the mean, variance, and mixture weight of a preset number of Gaussian components; determining the component affiliation of each relative difference sample based on the posterior probability of the Gaussian components to obtain Gaussian mixture model component labels; sorting the mean of each Gaussian component according to its numerical value, mapping components with larger mean values to reasonable claims, components with medium mean values to excessive claims, and components with smaller mean values to severe over-limit claims; determining the risk interval boundaries of each amount layer based on the boundary points of adjacent Gaussian components, and forming reasonable claim intervals, excessive claim intervals, and severe over-limit claim intervals accordingly; when the absolute value of the relative difference exceeds a preset severe threshold, the corresponding sample is directly labeled as severely over-limit.
[0031] In this embodiment, the determination of the mean value can be set to the following preferred range: if the mean value μ of the Gaussian component is within the range of [-3.801, 0), it is determined to be a larger mean value; if the mean value μ of the Gaussian component is within the range of [-11.210, -1.227), it is determined to be a middle mean value; if the mean value μ of the Gaussian component is less than -11.210, it is determined to be a smaller mean value. Wherein, [-3.801, 0) is the preferred range formed by the lower and upper bounds of the reasonable demand interval for each amount level, [-11.210, -1.227) is the preferred range for the excessive demand interval for each amount level, and less than -11.210 is the preferred range for the severely excessive demand interval; in different amount levels, the mean value range can be adaptively fine-tuned according to the fitting result corresponding to that amount level.
[0032] Example 1: Basic Risk Labeling Example
[0033] Obtain historical waybill claims data, extract the fields of actual compensation amount and claim amount, and calculate the claim difference D = actual compensation amount - claim amount; further construct the relative difference according to R = D / (actual compensation amount + ε), where ε is a preset minimum amount to avoid division by zero anomalies when the actual compensation amount is zero.
[0034] Outlier identification is performed on samples with relative differences. The interquartile range method is used to determine the initial outlier boundary, which can be combined with the standard deviation method for verification. Samples that exceed the boundary are marked as outliers, but are not directly deleted from the original dataset.
[0035] The sample was divided into four monetary strata based on the 25%, 50%, and 75% percentiles of the actual compensation amount. DBSCAN cluster analysis was performed on the relative differences within each monetary stratum to obtain the cluster label results within the stratum.
[0036] Gaussian mixture model was applied to the clustering results of each amount layer, and the fitted components were mapped to reasonable claims, excessive claims, and serious over-limit claims according to the mean size, so as to obtain the risk classification criteria for each amount layer.
[0037] Based on the aforementioned risk classification criteria, each historical waybill is labeled and a risk label corresponding to each waybill is output.
[0038] Example 2: Parameter Adaptation Example
[0039] Based on Example 1, the neighborhood radius of DBSCAN is automatically determined by the elbow position of the k-distance sorting curve to adapt to the density changes of samples in different monetary levels; the minimum number of samples is set to 5% of the current monetary level sample size and at least 3.
[0040] During the Gaussian mixture model fitting stage, a severe threshold rule can be set for the severe excess category. Samples with relative differences whose absolute values exceed the severe threshold can be directly classified into the severe excess category to improve the ability to identify extreme risk samples.
[0041] In one specific implementation, the identification results of each layer can be output in the same order as the original waybill, and the sample quantity, category distribution, threshold file and chart file of each amount layer can be recorded simultaneously to facilitate subsequent review and reproduction.
[0042] Example 3: Business Application Example
[0043] The risk label results output by this invention are applied to the logistics claims process: For waybills with reasonable claims, they are given priority to enter the fast processing or automatic compensation process; for waybills with excessive claims, they enter the enhanced review process; for waybills with serious over-limit claims, they trigger the key review or risk control interception mechanism.
[0044] By implementing the aforementioned differentiated distribution, the risk of excessive compensation can be controlled while ensuring service efficiency. Furthermore, historical claims risk tags can also be used in reverse for user profiling analysis, pre-sales protection recommendations, and high-risk waybill identification, forming a closed-loop management mechanism that extends from post-claims settlement to pre-claims prevention.
[0045] A logistics claims risk labeling system based on relative difference hierarchical clustering, the system includes: a data acquisition module, a difference construction module, an anomaly identification module, an amount hierarchical module, a hierarchical clustering module, a distribution fitting module, and a risk output module;
[0046] The data acquisition module is used to acquire historical waybill claims data; the historical waybill claims data includes at least the actual compensation amount and the claim amount.
[0047] The difference construction module calculates the claim difference based on the actual compensation amount and the claim amount, and constructs a relative difference index;
[0048] The anomaly identification module is used to identify and mark outliers in the relative difference.
[0049] The amount stratification module divides the sample into multiple amount strata based on the quantile of the actual compensation amount;
[0050] The hierarchical clustering module is used to perform cluster analysis on the relative difference samples to obtain intra-layer clustering results;
[0051] The distribution fitting module is used to perform Gaussian mixture model fitting on the intra-layer clustering results to obtain the risk interval boundaries of each amount layer;
[0052] The risk output module marks the waybill claim based on the risk range boundary and outputs the risk marking results.
[0053] The system implementation and method implementation of the present invention are the same.
[0054] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A logistics claims risk labeling method based on hierarchical clustering of relative differences, characterized in that, include: Obtain historical waybill claims data, which includes at least the actual compensation amount and the claim amount; The claim difference is calculated based on the actual compensation amount and the claim amount, and a relative difference index is constructed; outlier identification and marking are performed on the relative difference. The sample is divided into multiple amount layers based on the quantiles of the actual compensation amount; the DBSCAN clustering algorithm is used to perform cluster analysis on the relative difference samples to obtain the intra-layer clustering results; a Gaussian mixture model is applied to the intra-layer clustering results to obtain the risk interval boundaries of each amount layer; the waybill claims are marked according to the risk interval boundaries, and the risk marking results are output.
2. The logistics claims risk labeling method based on relative difference hierarchical clustering according to claim 1, characterized in that, The claim difference is the difference between the actual compensation amount and the claim amount; the relative difference R is R = D / (actual compensation amount + ε), where ε is a preset minimum amount used to avoid division by zero anomalies when the actual compensation amount is zero.
3. The logistics claims risk labeling method based on relative difference hierarchical clustering according to claim 1, characterized in that, Outlier identification and labeling of relative differences are performed as follows: the interquartile range method and / or standard deviation method are used to determine the outlier boundary of the relative difference; for samples that exceed the outlier boundary, a label retention method is used for processing.
4. The logistics claims risk labeling method based on relative difference hierarchical clustering according to claim 3, characterized in that, Determining the outlier boundary of relative differences using the interquartile range (IQR) and / or standard deviation method includes: obtaining a relative difference sequence; calculating the first quartile Q1 and the third quartile Q3 of the relative difference sequence, and obtaining the interquartile range IQR based on the third quartile Q3 and the first quartile Q1, wherein the first outlier boundary obtained based on the IQR method is [Q1-k1·IQR,Q3+k1·IQR], where k1 is a preset strict coefficient; calculating the mean μ and standard deviation σ of the relative difference sequence, and obtaining the second outlier boundary based on the standard deviation method as [μ-k2·σ,μ+k2·σ], where k2 is a preset standard deviation coefficient; when both the IQR and standard deviation methods are used simultaneously, the intersection of the first and second outlier boundaries is taken as the final outlier boundary, or the stricter upper and lower bounds of the two methods are selected as the final outlier boundary; samples exceeding the final outlier boundary are marked as outliers and retained in the original sample set for subsequent risk identification and review.
5. The logistics claims risk labeling method based on relative difference hierarchical clustering according to claim 1, characterized in that, Dividing the sample into multiple monetary tiers includes: dividing the sample into low-amount, low-to-medium-amount, medium-to-high-amount, and high-amount tiers based on the 25%, 50%, 75%, and 100% quantiles of the actual compensation amount, so that the samples within the same monetary tier remain similar in terms of monetary magnitude.
6. The logistics claims risk labeling method based on relative difference hierarchical clustering according to claim 1, characterized in that, The DBSCAN clustering algorithm is used to perform clustering analysis on relative difference samples, which includes: constructing a one-dimensional feature sample set for relative difference samples in each amount layer; determining the neighborhood radius eps of DBSCAN clustering based on the k-distance distribution of samples in the current amount layer; determining the minimum number of samples min_samples based on the sample size n of the current amount layer; performing DBSCAN clustering on the relative difference samples in the current amount layer using the neighborhood radius eps and the minimum number of samples min_samples to obtain the cluster label of each sample; and dividing the relative difference samples into core samples, boundary samples, and noise samples based on the density connectivity and proximity relationships of the samples in the feature space, and generating an initial cluster structure within the layer for subsequent distribution fitting.
7. A logistics claims risk labeling method based on relative difference hierarchical clustering according to claim 6, characterized in that, The neighborhood radius eps is obtained by performing elbow detection on the k-distance sorted curves to obtain the distance value corresponding to the point with the largest gradient change as the neighborhood radius eps.
8. A logistics claims risk labeling method based on relative difference hierarchical clustering according to claim 6, characterized in that, The minimum number of samples, min_samples, is: min_samples=max(3,α·n); Where α is a preset ratio parameter and n is the sample size.
9. The logistics claims risk labeling method based on relative difference hierarchical clustering according to claim 1, characterized in that, The Gaussian mixture model fitting process for the intra-layer clustering results includes: fitting a Gaussian mixture model to the relative difference samples corresponding to the initial cluster structure obtained by DBSCAN clustering within each amount layer, obtaining the mean, variance, and mixture weight of a preset number of Gaussian components; determining the component affiliation of each relative difference sample based on the posterior probability of the Gaussian components, obtaining Gaussian mixture model component labels; sorting the mean of each Gaussian component according to its numerical value, mapping components with larger mean values to reasonable claims, components with medium mean values to excessive claims, and components with smaller mean values to severe over-limit claims; determining the risk interval boundaries of each amount layer based on the boundary points of adjacent Gaussian components, and forming reasonable claim intervals, excessive claim intervals, and severe over-limit claim intervals accordingly; when the absolute value of the relative difference exceeds a preset severe threshold, the corresponding sample is directly labeled as severely over-limit.
10. A logistics claims risk labeling system based on relative difference hierarchical clustering, the system being used to execute the logistics claims risk labeling method based on relative difference hierarchical clustering as described in any one of claims 1 to 9, characterized in that, The system includes: a data acquisition module, a difference construction module, an anomaly identification module, a monetary stratification module, a stratified clustering module, a distribution fitting module, and a risk output module; The data acquisition module is used to acquire historical waybill claims data; the historical waybill claims data includes at least the actual compensation amount and the claim amount. The difference construction module calculates the claim difference based on the actual compensation amount and the claim amount, and constructs a relative difference index; The anomaly identification module is used to identify and mark outliers in the relative difference. The amount stratification module divides the sample into multiple amount strata based on the quantile of the actual compensation amount; The hierarchical clustering module is used to perform cluster analysis on the relative difference samples to obtain intra-layer clustering results; The distribution fitting module is used to perform Gaussian mixture model fitting on the intra-layer clustering results to obtain the risk interval boundaries of each amount layer; The risk output module marks the waybill claim based on the risk range boundary and outputs the risk marking results.