Insurance recommendation method based on big data

Abstract

The invention relates to the technical field of insurance recommendation based on big data, and discloses an insurance recommendation method based on big data. Collecting user geographical behaviors and time information, mapping to a unified plane coordinate system, and dividing grids according to an equal area principle; counting the multi-period access times of each grid, calculating access density and logarithmic smooth heat, and constructing a dynamic thermodynamic model; extracting a median and a median absolute deviation in a historical period to generate a heat baseline, and judging an abnormal hotspot according to a deviation degree; the discrete hotspots are combined into a continuous region by adopting four-connection clustering, a composite score is generated by combining the average anomaly degree and the area of the region, candidate insurance products are mapped through a preset risk interval, weighted sorting is performed according to the score, and finally a personalized recommendation list is output. According to the scheme, short-time burst and large-range slight risks are taken into consideration, the fine mapping capacity of unification of multi-source data, high-robustness anomaly detection and scale perception is achieved, and the insurance recommendation accuracy and the user satisfaction degree are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data-based insurance recommendation, in particular to a big data-based insurance recommendation method. BACKGROUND

[0002] With the wide popularity of mobile terminals and the Internet of Things, user geographic location information collection has become the basis for various intelligent services. The existing technology usually calls third-party map services or is based on high-precision GIS platforms to project latitude and longitude data into a plane coordinate system to support path planning, navigation and regional analysis. However, this approach has high dependence on external services, high computational cost and difficulty in deployment in lightweight or edge devices. In addition, multi-source positioning data formats are not unified, often requiring additional data cleaning and format conversion, resulting in complex and inefficient backend processing procedures.

[0003] In terms of spatial discretization, traditional heat maps and geographic information analysis often use administrative divisions, business districts, or irregular polygon partitions. These units have large area differences, making it difficult to compare the access intensity and risk level of each region fairly. Some methods use tile maps for slicing statistics. However, since the map slicing level is fixed, it cannot adapt to different granularity requirements and it is difficult to balance overall performance and local accuracy, resulting in rough capture of hotspots or excessive computational load. For the construction of spatio-temporal access distribution, existing techniques often perform statistics based on fixed hours or days, ignoring the timeliness of short-term fluctuations and the smoothness of long-term trends. At the same time, simply using the number of visits as the heat weight is easily disturbed by traffic peaks or temporary activities, resulting in severe heat map noise and frequent false positives, making it difficult to effectively identify real risk areas. In terms of heat and anomaly determination, traditional methods often rely on mean and standard deviation for threshold judgment, or set a fixed threshold based on experience. These two methods perform poorly when the data distribution is asymmetric or contains extreme values. The mean / standard deviation is sensitive to outliers, and a slight burst of traffic can distort the baseline. A fixed threshold cannot adapt to regional differences, which may miss potential risks of moderate intensity or frequently misreport routine fluctuations. In addition, existing spatial hotspot detection methods often stop at the single-point level and do not aggregate adjacent hotspots, making it difficult to distinguish between isolated hotspots and noise. Moreover, there is a lack of quantitative evaluation of both the size and intensity of hotspots, making it difficult to provide accurate basis for subsequent business decisions. If an insurance recommendation system relies solely on static user profiles or collaborative filtering models based on interest preferences, it ignores the dynamic changes in geographic scenarios and spatio-temporal risks, causing the recommendation results to deviate from the actual needs of users. Finally, in terms of insurance product recommendation, current methods often use rules or simple classification models based on user age, occupation, and preferences for sorting, lacking deep matching of regional risks and product protection capabilities. Single-dimensional priority setting leads to a lack of effective linkage between the recommendation list and the risk map, making it difficult for users to obtain personalized insurance solutions that align with their behavior patterns and current safety situation.

[0004] To this end, the present invention aims to provide an insurance recommendation method based on big data, which performs fine-grained grid-based statistics, logarithmic smoothing, and robust baseline construction on the hotspot access frequency of users in a two-dimensional plane, enabling anomaly detection in sudden or continuously high-frequency activity areas. Then, it integrates discrete hotspots into continuous regions using spatial connected clustering, and accurately maps regions of different risk levels to pre-set insurance product intervals by combining area and anomaly degree in a composite score. Finally, it outputs a personalized recommendation list in weighted order. This method can balance real-time, robustness, and interpretability, and is suitable for dynamic insurance demand matching in scenarios that require spatio-temporal monitoring of crowd gathering and travel concentration. SUMMARY

[0005] The application provides an insurance recommendation method based on big data, which promotes the solution to the problems mentioned in the background art.

[0006] The application provides the following technical scheme: an insurance recommendation method based on big data, comprising: Collecting geographic behavior data and establishing a two-dimensional space coordinate system, wherein the geographic behavior data comprises user position information and corresponding time; Grid division is performed on the research area according to the equal-area principle to construct a regular space unit; The number of visits of each grid in multiple time periods is counted to form a space-time visit distribution; The visit density and heat index of each grid are calculated to construct a dynamic heat model; The heat statistical characteristics of each grid in the historical period are extracted to generate a heat baseline index; Based on the deviation degree of the heat baseline and the current heat, it is determined whether the grid is an abnormal hot spot; The hot spot grid determined as abnormal is spatially clustered to construct a hot spot area, and the hot spot area is matched with an insurance product set; The candidate insurance products obtained through matching are weighted and sorted to finally generate a personalized insurance recommendation list.

[0007] Optionally, the collecting of geographic behavior data and the establishment of a two-dimensional space coordinate system, wherein the geographic behavior data comprises user position information and corresponding time, specifically comprises: Obtaining records from a user terminal or a background log, wherein the records comprise ; Among them, is the latitude of the th record, and the unit is degree; is the longitude of the th record, and the unit is degree; is the time of the th record, and the unit is second; In the range of the research area, the latitude and longitude of the southwest corner point are selected as the plane coordinate origin; a unified reference point is determined to facilitate the consistency of subsequent coordinate conversion; A two-dimensional space rectangular coordinate system is established, wherein the positive east direction corresponding to the increasing direction of the longitude is selected as the axis; the positive north direction corresponding to the increasing direction of the latitude is selected as the axis; the plane direction is defined to ensure that the meaning of the mapped coordinates is clear; The average radius of the earth is obtained, denoted as ; each record is mapped to a plane coordinate by approximating a small area plane, specifically as follows: , ; wherein, is the horizontal coordinate of the i-th record; is the horizontal coordinate of the i-th record; is the vertical coordinate of the i-th record. is the vertical coordinate of the i-th record.

[0008] Optionally, the study area is divided into a grid according to the equal-area principle to construct regular spatial units, specifically including: In the plane coordinate system, the grid side length is selected as , the area is divided into a plurality of square grid units, specifically: ; wherein, the area of each grid is ; a unique index is assigned to each grid, and the center coordinates are recorded; wherein, is the center point coordinates of the grid .

[0009] Optionally, the number of visits of each grid in a plurality of time periods is counted to form a space-time visit distribution, specifically including: Set the time statistical granularity as , the current time is divided into a plurality of time periods ; For each grid and the selected current time period , the number of records falling into the grid is counted, specifically: ; wherein, is the number of visits of the grid in the time period; is a counting function to calculate the number of elements in a set.

[0010] Optionally, the visit density and heat index of each grid are calculated to construct a dynamic heat model, specifically including: Set the grid visit density as: ; Set the predefined reference density as ; Calculate the grid heat index, specifically: .

[0011] Optionally, the heat statistical features of each grid in the historical period are extracted to generate a heat baseline index, specifically including: For each grid , the heat sequence of the previous time periods is collected; The median heat is calculated as: ; Where, The median of the sequence is calculated. The median absolute deviation is calculated as: ; Where, The absolute value of the difference between and is calculated.

[0012] Optionally, the degree of deviation of the heat baseline from the current heat is used to determine whether the grid is an abnormal hot spot, specifically including: For each grid The abnormality is calculated as: ; Where, is a very small constant to prevent division by zero; The abnormality threshold is set as ; If , mark the grid as an abnormal hot spot: ; If , do not mark.

[0013] Optionally, the hot spot grids determined to be abnormal are spatially clustered to construct a hot spot area, and matched with a set of insurance products, specifically including: Define adjacency as any two grids and , if they satisfy , they are adjacent; merge spatially adjacent hot spots into continuous areas; Initialize all middle grids as unvisited; Iterate through each unvisited grid , perform the following steps: S1, take any and unvisited, set the queue ; S2, pop the head of the queue , and its four adjacencies If they are in and unvisited, add them to and mark them as visited; S3, repeat step S2 until is empty, and the resulting set is a region ; S4, repeat steps S1 to S3 for the remaining unvisited hot spots until is generated; Get the number of region grids, denoted as ; The total area of the calculation region is calculated as ; The average abnormality of the calculation region is calculated as ; The total area of the research region is calculated as ; The composite risk score of the region is set as ; The set of all insurance products is set as , and each product The pre-defined risk application interval is set as ; The mapping function from the region to the candidate insurance product set is set as: ; Wherein, is the candidate insurance product set corresponding to the region.

[0014] Optionally, the matched candidate insurance products are weighted and sorted to generate a personalized insurance recommendation list, specifically including: Combining the candidate sets of all regions: ; wherein, is the set of all candidate insurance products; For each product , the recommendation weight is set as: ; wherein, is the product recommendation weight; In descending order of , the final recommendation sequence is generated.

[0015] The present application has the following beneficial effects: 1. Directly using the southwest corner as the custom origin point and small-scale plane projection avoids the high computational load and error accumulation of complex projection systems, realizing parallel preprocessing of large-scale user data. Lightweight latitude and longitude to coordinate conversion can adapt to any non-global positioning scenario; reducing spatial distortion caused by different projections in subsequent calculations. Eliminate the ambiguity of latitude and longitude units in multi-source logs to provide a unified and accurate spatial reference for subsequent gridding.

[0016] 2. Unified square grid design with equal area, and save static coordinates at the center of the grid to simplify cross-grid adjacency calculation. Equal-area division allows subsequent dimensionless processing of heat and area ratio comparison to be directly performed; center coordinate pre-storage reduces real-time calculation overhead. Avoid density calculation inconsistencies caused by different grid shapes or sizes; support fast spatial indexing under massive data.

[0017] 3. The time period length As an adjustable parameter, it allows for a fine-grained trade-off between real-time alerts and smoothing analysis. The timeframe can be flexibly set from 5 to 60 minutes or several hours depending on the scenario, balancing response speed and robustness in anomaly detection. Fixed long-period statistics are insufficient to capture sudden events; this method can arbitrarily switch between timeframes from seconds to hours.

[0018] 4. First, normalize the number of visits to a density that is consistent with area and time, then generate a dimensionless heat index using a logarithmic function. Logarithmic smoothing weakens the influence of extreme high-density values ​​while preserving low-density differences, ensuring that hotspot identification neither overemphasizes regular high-traffic areas nor ignores hidden risk points. This overcomes the shortcomings of the original density normalization or Z-score methods, which are sensitive to extreme values ​​and require mean and standard deviation, thus improving robustness in long-tailed visit distribution scenarios.

[0019] 5. The median and median absolute deviation are used to construct the baseline heat distribution, replacing the traditional mean and standard deviation. The median and median absolute deviation are more robust to outliers and less susceptible to distortion by extreme access spikes or missing data, ensuring the stability and reliability of the baseline characteristics. This avoids serious interference of sudden events or abnormal accesses in historical data on the baseline parameters, enhancing the accuracy of subsequent outlier measurements.

[0020] 6. Based on This standardized deviation metric directly outputs a dimensionless anomaly score. The metric is adaptive to baseline fluctuations; a larger deviation is required to identify anomalies in grids with high inherent fluctuations, while subtle anomalies are more easily captured in areas with low fluctuations. It resolves the dilemma of choosing between absolute and relative differences, enabling accurate anomaly identification under varying background noise levels.

[0021] 7. Using the Manhattan adjacency rule of the grid index, automatically generate non-overlapping spatial anomaly regions; simultaneously consider the average anomaly degree and region proportion to form a single dimensionless score. The risk intervals correspond to the scoring points, dynamically selecting the most suitable product set. Clustering takes into account spatial continuity, avoiding misclassifying isolated points as major risk areas; risk scoring considers both intensity and scale, reasonably balancing large-scale minor anomalies with small-scale severe anomalies; the interval mapping method is more flexible than a single threshold, naturally dividing risk segments and mapping them to different product tiers. This overcomes the shortcomings of using only connectivity components, which cannot quantify size and intensity or accurately match insurance tiers, seamlessly integrating spatial and anomaly information to form a business-driven product mapping.

[0022] 8. The candidate set mapped by all regions is globally weighted and accumulated, and ranked and output accordingly. When the user has risks in multiple regions, the weights of the candidate products will naturally accumulate, and products that can cover multiple risk scenarios will be preferentially recommended; at the same time, when the risk in a single region is high, the products mapped by this region will also be highlighted. The contradiction between regional isolated recommendation and global consistency is solved, so that the recommended results reflect both local features and overall risk layout. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 The flowchart of the present application is shown. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0025] Embodiments, with reference to Figure 1 A big data-based insurance recommendation method, comprising: Collecting geographic behavior data and establishing a two-dimensional spatial coordinate system, wherein the geographic behavior data includes user location information and corresponding time; Grid division is performed on the research area according to the equal-area principle to construct a regular spatial unit; The number of visits of each grid in multiple time periods is counted to form a spatio-temporal visit distribution; The visit density and heat index of each grid are calculated to construct a dynamic heat model; The heat statistical features of each grid in the historical period are extracted to generate a heat baseline index; Based on the deviation degree of the heat baseline and the current heat, it is determined whether the grid is an abnormal hot spot; The hot spot grids determined to be abnormal are spatially clustered to construct a hot spot region, and matched with a set of insurance products; The candidate insurance products obtained by matching are weighted and ranked to finally generate a personalized insurance recommendation list.

[0026] Through data collection and gridding processing, the scheme realizes the fine splitting and spatial discretization of user movement trajectory. By collecting geographic behavior data and establishing a two-dimensional coordinate system, the problem of non-uniform format of multi-source positioning data that cannot be directly used for spatial analysis is solved, ensuring that all subsequent calculations are completed under the same metric plane coordinate, efficiently compatible with various terminal data, and reducing coordinate conversion errors. By dividing the study area into grids according to the equal area principle, the problem of uneven density comparison caused by inconsistent area units is solved, and the comparability and consistency of statistical results are ensured. In the process of forming the spatio-temporal access distribution and heat model, by counting the number of visits to each grid in multiple time periods, the defect of single sampling being easily disturbed by temporary fluctuations is overcome, forming a more robust spatio-temporal behavior portrait; by calculating the access density and constructing the heat index, the difficulty of distinguishing high-frequency access areas from actual abnormal hotspots is solved, and potential risk hotspots can still be detected after logarithmic smoothing of high-flow areas, improving the robustness of the model to noise. In baseline extraction and anomaly determination, by extracting the median and median absolute deviation of historical periods to construct the heat baseline index, the problem of mean and standard deviation being sensitive to extreme values and easily disturbed by abnormal access is solved, improving the stability of the baseline and making the subsequent anomaly determination more reliable; by comparing the deviation of the current heat from the baseline, the adaptive identification of hotspots under different background noise is realized, taking into account both sudden small-scale high-density and large-scale slight deviation risk scenarios. In spatial clustering and product mapping, by merging single-point hotspots into continuous areas through four-connected clustering algorithm, the error of isolated anomalies being misjudged as large-area risk is solved, highlighting the real risk clusters; by combining regional average anomaly degree and relative area to generate a composite score and map it to a segmented product interval, the problem of traditional single-threshold mapping being rough and difficult to reflect risk size differences is solved, realizing fine product matching considering both risk intensity and size. Finally, in step eight, by weighting and sorting all regional candidate products according to the composite score, the problem of balancing multiple regional and product recommendation results is solved, highlighting both the preferred products covering multiple high-risk areas and the single regional high-value demand, ultimately generating an insurance recommendation list that meets the user's individual risk portrait. The overall structure of the scheme is clear and self-consistent, which can effectively improve the accuracy of insurance recommendation and user satisfaction.

[0027] The geographic behavior data includes user location information and its corresponding time, specifically including: From the user terminal or background log records, including ; Wherein, is the latitude of the th record, in degrees; For the first The longitude of each record, in degrees; For the first The time of each record is in seconds; collect user location data at different points in time to support subsequent thermal calculations; Within the study area, the latitude and longitude of the southwesternmost point were selected. As the origin of the plane coordinate system, a unified reference point is established to facilitate consistency in subsequent coordinate transformations. Establish a two-dimensional rectangular coordinate system, where the due east direction is selected, corresponding to the direction of increasing longitude. Axis; select true north as the direction corresponding to increasing latitude. Axis; defines the plane direction to ensure that the meaning of the coordinates is clear after mapping; Obtain the Earth's average radius, denoted as . By approximating a small region as a plane, each record is mapped to planar coordinates, specifically as follows: , ;in, For the first The x-coordinate of each record; For the first The vertical coordinate of each record; convert latitude and longitude in degrees to planar coordinates in meters for easier gridding.

[0028] By acquiring location information and time data from user terminals or backend logs, the problem of unified access to multi-source heterogeneous positioning data is solved. This step centrally collects records from different devices and in different formats and stores them in the form of "latitude-longitude-time" triplets. The advantage is that it provides a unified and standardized input for all subsequent spatial analyses, avoiding data loss and misinterpretation caused by inconsistent data formats. By selecting the latitude and longitude of the southwesternmost point of the study area as the plane reference origin, the analysis bias caused by different coordinate benchmarks used by different platforms or data sources is solved. This method locks a fixed and easily understood reference point globally, ensuring that all subsequent coordinate transformations are based on the same standard, guaranteeing comparability and consistency between different batches or real-time data. By establishing a reference point with due east as the reference point... The axis, with due north as the direction. The two-dimensional Cartesian coordinate system of the axes solves the problem of ambiguous direction definition when mapping geographic latitude and longitude to a plane. This step clarifies the direction of the coordinate axes, making the mapped location points readable and visually appealing, which helps in the accurate execution of subsequent gridding, heatmap visualization, and spatial clustering. This is achieved by using the Earth's average radius. And small area plane approximation, latitude and longitude conversion to plane coordinates in meters, solve the problem of inconsistent dimension and subsequent statistical calculation difficulty caused by direct use of "degree" unit. The mapping method not only unifies the coordinate unit to meters, but also only involves multiplication, cosine and constant conversion in the calculation process, which is simple and controllable in error, significantly reduces the dependence on high-precision geographic information system (GIS) system, and also improves the running efficiency of edge device or lightweight server.

[0029] Through the coupling application of the above series of steps, a closed loop process from data acquisition to coordinate conversion is formed, which ensures the end-to-end consistency of the entire spatial analysis link. Without frequent calling of external map interface or third party service, efficient and stable geographic data preprocessing can be completed in local environment, which lays a solid foundation for subsequent grid division, heat calculation and risk detection, and reduces the operation cost and external dependence risk.

[0030] The research area is divided into grids according to the equal area principle, and regular spatial units are constructed, which specifically include: In the plane coordinate system, the grid side length is selected as The area is divided into a plurality of square grid units, which are specifically: Discrete continuous space into manageable units for statistics and analysis; Wherein, the area of each grid is ; A unique index is allocated to each grid , and the center coordinates are recorded ; wherein, is the center point coordinates of the grid ; maintain the addressability and uniqueness of the grid position for subsequent statistics.

[0031] By selecting the grid side length L in the plane coordinate system and dividing the area into a plurality of square grid units, the problem that continuous geographic space cannot be directly analyzed is solved. After discretizing the area into equal-area square units, the originally difficult-to-quantify spatial distribution becomes regular and controllable, and any access, heat or risk index in a grid can be directly accumulated and compared, thereby greatly simplifying the subsequent spatio-temporal statistical calculation. By calculating and defining the area A = L 2 of each grid, the problem of index incomparability caused by different unit areas is solved. After unifying the area, the numerical meaning of the access frequency, access density and heat index of each grid is completely consistent, and the index from different grids can be compared vertically or horizontally, avoiding misjudgment caused by area difference, thereby improving the accuracy and fairness of risk assessment. By allocating a unique index to each grid and recording the center coordinates , solves the problem of difficult to trace and locate the spatial unit. The number and center point coordinates of each grid provide a clear "anchor point" for subsequent data mounting, querying and visualization. Whether in heat map, alarm system or product mapping link, it can quickly locate to specific grid and efficiently perform data aggregation and display, reducing the complexity of spatial correlation operation.

[0032] The regular grid structure formed by the above steps fundamentally solves the problems of "how to divide the unit most reasonably" and "how to ensure the comparability of indicators" in geographic spatial statistics. The benefits are that spatial statistics and analysis have universality and can be applied to various scenarios such as urban transportation, public safety and environmental monitoring; the preset grid size L can be flexibly adjusted according to application requirements, supporting coarse-grained trend monitoring or fine-grained anomaly capture; the entire discretization process only relies on basic geometric operations and does not require external map dependence, making it easy to deploy in lightweight platforms or edge computing environments.

[0033] The number of visits of each grid in multiple time periods is counted to form a spatio-temporal visit distribution, specifically including: Let the time statistical granularity be , the current time is divided into several time periods ; define the statistical window size to control the heat response speed and smoothness; The value of the time statistical granularity directly determines the "observation window" width of the user's geographic heat change, thereby affecting the response speed, smoothness and calculation overhead of the statistical result. The specific effects are as follows: 1. Response speed and smoothness effect: Smaller, such as a few minutes to tens of minutes, can capture heat mutations in short periods and discover short-term hotspots in time; the corresponding heat sequence is more jittered and noisier, which may misjudge a one-time discrete visit as an anomaly; it is suitable for scenarios that are sensitive to second to minute-level travel fluctuations and require real-time early warning. Larger, such as several hours to a whole day, smooths out short-term access peaks and valleys, avoiding false alarms caused by short-term peaks or valleys; the detection of abnormal bursts will be delayed and cannot respond quickly; it is suitable for overall traffic trend analysis and business requirements with high tolerance to sudden events. 2. Robustness of anomaly determination: smaller , the number of visits in each time period is less, , the estimated variance is large, and the baseline statistics are more sensitive to single outliers; larger , the count is more concentrated, and the baseline is more stable, but if is too large, the true abnormal short-term fluctuations may be smoothed out and is not easy to trigger. According to the trade-off between the response length of the early warning and the false alarm rate, the time statistical granularity is generally selected as 15 minutes to 1 hour; for ultra-short-time burst scenarios (such as traffic accidents, outdoor activities), it can be appropriately reduced to 5 to 10 minutes; for daily traffic monitoring, it can be increased to 1-2 hours to reduce noise and computing overhead.

[0034] For each grid and the selected current time period , the number of records falling into the grid is counted, specifically: ; wherein, is the number of visits of the grid in the time period; is the counting function, which counts the number of elements in the set; quantifies the access intensity of the grid in unit time as the input for density calculation.

[0035] By setting an adjustable time statistical granularity and dividing the current time into several time periods, the problem of difficult balance between response speed and smoothing degree of space-time heat map is solved. The advantage is that when the business needs a second-level or minute-level fast warning, the can be set to 5-10 minutes to capture short-time hotspots in time; and in the scene of focusing on macro trends or reducing noise false alarms, the can be adjusted to 1-2 hours to avoid frequent misjudgments caused by peak-valley fluctuations, thereby providing a highly customizable monitoring experience. By analyzing the influence of different values on the robustness of abnormality determination, the problem of single outlier access or "short-time burst fluctuation being easily amplified or completely ignored under traditional fixed time window is solved. The advantage is that when is smaller, although the algorithm is more sensitive, it can be combined with a higher abnormal threshold to reduce the false alarm rate; when is larger, the algorithm has stronger recognition ability for regions with continuous high density, while avoiding the neglect of short-term abnormalities, achieving a balance between accurate warning and tolerance of fluctuations for different business sides. By counting the number of visits in each grid and the selected time period, solves the problem of lack of quantitative indicators for spatial units and difficulty in density conversion. The advantage is that the formed space-time access distribution can accurately reflect the regional heat trend, providing a highly reliable data basis for the calculation of density and heat index; at the same time, this statistical method is simple and easy to understand, and can be efficiently implemented in batch processing or stream processing environment. By explicitly writing the statistical window and counting function, the problem of fuzzy statistical process, algorithm non-reproducible or different execution between platforms is solved. The advantage is that when the team compares tests in different environments or versions, they only need to reproduce the results according to the fixed formula and parameters, ensuring the maintainability and scalability of the system, and also facilitating subsequent scientific tuning of , window length and threshold.

[0036] Through the above-mentioned comprehensive design of the space-time access distribution construction process, the dual problems of the traditional heat map, i.e., the lack of consideration of the space-time correlation, the easy missing of short-time high-risk or the slow response to long-term trends, are solved. The advantage is that the system can not only acutely capture instantaneous abnormal hotspots, but also maintain data smoothing and stability on a large time scale, thereby providing solid and flexible space-time data support for accurate risk assessment and insurance recommendation.

[0037] The method further comprises the following steps of: The grid access density is set as: ; the number of times is converted into a density index unified in area and time, facilitating cross-grid comparison; The pre-defined reference density is set as ; The original density ratio is dimensionless, and the compression strength of high-density data is controlled through logarithmic scaling; The smaller the ratio , the larger the corresponding , that is, the more sensitive to hotspots, and the smaller the density of low-density grids, may exceed 1, so that non-real abnormality is also amplified, and if , a large number of false hotspots will be generated, and the false positive rate will increase; The larger the ratio , the smaller the corresponding , and the overall compression of the value is reduced, and the difference between the real high-density area and the ordinary area is reduced, and if , the actual high-density grid is difficult to be highlighted, the sensitivity is reduced, and the false negative rate increases; the average density or the median density of the whole region or the typical business period can be used as the initial , which is fine-tuned to the best balance point through historical backtesting, so as to balance the false negative rate and the false positive rate; The grid heat index is calculated, specifically as: ; the density is logarithmically smoothed to weaken the influence of extreme values, and a dimensionless heat value is generated.

[0038] By standardizing the number of visits into a density index unified in area and time, in the present scheme, the calculation is performed for each grid, solving the problem that the number of visits between grids of different sizes or different statistical time lengths cannot be directly compared. This step converts the discrete visit quantity originally in "times" into a continuous density in the form of "times m -2 s -1 ", so that the access intensity of each grid can be compared fairly across space and time period, and the advantage is that the comparability and accuracy of the overall heat model are significantly improved, laying a solid foundation for subsequent abnormal identification. By introducing the reference density And the density ratio is dimensionless. The pre-defined reference density is set in this scheme , and as a dimensionless ratio, and then scaled by the logarithmic function , which solves the deviation problem of high-density areas dominating the thermal results due to excessive absolute values. The advantages are: log compression can weaken the influence of extremely high access density, while ensuring that small differences in low-density areas are not completely hidden, achieving dual sensitivity to hot spots and cold areas; at the same time, dimensionless processing reduces the risk of overflow in numerical operations, improving system stability. Through historical backtesting and fine-tuning of the reference density , the scheme recommends using the average density or median of the entire region or typical business period as the initial , combined with historical data backtesting, dynamically fine-tuned to the best balance point of false positive rate and false negative rate, solving the problem that a single threshold setting is difficult to balance sensitivity and robustness. The advantages are: the operation team can flexibly adjust the parameters according to different monitoring scenarios (such as traffic peak, holiday travel), so that the model can quickly respond to real heat anomalies, and can also suppress pseudo-hot spots caused by short-term fluctuations, improving the practicality and maintainability of the recommendation system in real business environment. Through log smoothing to generate a dimensionless heat index, the final dynamic heat model represents the grid heat with , solving the defect that direct addition or comparison of access density in different time periods and different regions is easily disturbed by extreme values. The advantages are: the dimensionless heat index not only has a unified dimension, and the log smoothing allows the algorithm to effectively distinguish between continuous growth and sudden jumps, making it more intuitive to reflect risk intensity in visual display and threshold determination, improving the accuracy of risk identification and business interpretability. At the same time, the calculation of the heat index only relies on simple logarithmic operations, which is easy to efficiently parallelize in large-scale distributed platforms.

[0039] The heat statistical features of each grid in the historical period are extracted to generate a heat baseline index, which specifically includes: For each grid , collect the heat sequence of the previous periods; obtain historical behavior to construct a baseline distribution; Calculate the median heat as: ; obtain a robust representative value of heat, resistant to outliers; where is the median of the sequence; The median absolute deviation is calculated as: ; quantify the heat dispersion degree, which is used as the denominator for anomaly calculation; where is the absolute value of the difference between and .

[0040] By collecting the heat sequence of the previous T time periods for each grid, the problem that the single time period heat cannot reflect the long-term behavior characteristics of the region is solved. This step continuously aggregates historical heat data to form a complete heat change curve, which has the advantage of accurately capturing the fluctuation range of the region in the normal period, providing sufficient data support for subsequent baseline construction, and avoiding misjudgment caused by abnormal access or missing data in isolated periods. By calculating the median of the historical heat sequence as the baseline representative value, the problem of mean being too sensitive to extreme values and being easily distorted by sudden large-scale access is solved. The median is naturally resistant to outliers, and even if a small amount of "brush screen" or temporary heavy traffic occurs at a certain time, it will not have a significant impact on the baseline, thereby ensuring a robust description of the heat baseline and making subsequent anomaly measurement more reliable and stable. By calculating the median absolute deviation of the difference between the historical heat sequence and the median, the problem of standard deviation being inaccurate in the case of skewed distribution or a small number of outliers is solved. MAD (Median Absolute Deviation) as a measure of heat dispersion is also robust to outlier access, and has the advantage of being able to objectively quantify the overall fluctuation amplitude while ensuring that it is not pulled up by extreme peaks, providing a reliable and stable scale for the denominator of the anomaly measure. By combining the median and the median absolute deviation to construct the heat baseline index, the deficiency of a single statistical quantity that cannot consider both "central tendency" and "dispersion" is solved. The combined use of and can accurately locate the "normal level" of the regional hot spot and clearly reflect the fluctuation intensity of the hot spot region, which has the advantage of providing "double insurance" for anomaly detection: only when the current heat is both above the central tendency and far from the normal fluctuation range will it be marked as a true anomaly, thereby significantly reducing the false alarm rate and the false negative rate. Through the closed-loop design of the above historical baseline construction process, the problem of disconnection between real-time heat calculation and historical behavior reference is solved. The real-time system can continuously call the baseline indicators and for comparison, without the need to recalculate the complete historical sequence each time, which has the advantage of significantly reducing the online calculation pressure while ensuring accuracy, allowing the big data platform to complete anomaly monitoring and alarm at a frequency of seconds or even higher, meeting the needs of high concurrency and large-scale scenarios.

[0041] The degree of deviation of the heat baseline from the current heat is used to determine whether the grid is an abnormal hot spot, specifically including: For each grid , the anomaly degree is calculated as: , which quantifies the deviation of the current heat from the baseline; where is a very small constant to prevent division by zero, such as ; Set the anomaly threshold as ; anomaly threshold The smaller, the easier to meet , so more grids are marked as abnormal, high sensitivity, can capture subtle but important fluctuations, normal jitter as an abnormal, false positives increase, affect product recommendations targeting; anomaly threshold The larger, only when the baseline is extremely significant deviation is determined to be abnormal, false negatives decrease, recommendations focus on the most significant hot spots, not sensitive to moderate intensity but potential risk areas, false negatives increase; first with 1.0 as a neutral starting point, combined with business risk tolerance and recommendation capacity requirements A / B testing or backtracking analysis, adjust to the optimal; If , mark the grid as an abnormal hotspot: ; filter out hot spots that deviate significantly from the baseline; If , do not mark.

[0042] By calculating the standardized anomaly degree of the grid, the problem of "different grid background fluctuation amplitude is not the same, it is difficult to unify the determination of abnormal" is solved. This step carries out non-dimensionalization processing on the deviation of the current heat and the baseline, so that in the grid with large baseline fluctuation, only the truly significant heat change can produce high anomaly degree, and in the grid with stable background, even small amplitude heat jump can be accurately captured. The advantage is that this "adaptive" deviation measurement can identify abnormalities in a variety of spatial backgrounds, without the need to set thresholds for each grid, greatly simplifying the parameter tuning workload, while significantly improving the accuracy and consistency of anomaly detection. By introducing a small constant to prevent the denominator from being zero, the problem of "zero or near-zero historical heat fluctuation completely, division by zero error or anomaly degree value out of control" is solved. This small constant ensures the stability of mathematical operations and does not substantially interfere with normal denominator values. The advantage is that the algorithm can also calculate reasonable anomaly degrees when it faces new online areas or just initializes the grid, providing seamless support for risk monitoring in new areas, avoiding system crashes or ignoring risks due to insufficient historical data. Through the adjustable anomaly threshold, the problem of "global unified threshold difficult to balance high sensitivity and low false positive rate" is solved. When the threshold is small, more subtle deviations can be captured, which helps to discover potential risks early; when the threshold is large, regular fluctuations can be filtered out, focusing on extreme hotspots and reducing false alarms. The advantage is that the operation team can quickly determine the optimal δ value according to the business needs for "early warning frequency" and "recommendation list size" through A / B testing or historical backtracking analysis, so as to neither miss key risks nor let users be tired of too many "pseudo-hotspots", thus achieving the best balance between sensitivity and stability. Through the adjustable anomaly threshold , mark the cell as an abnormal hotspot, otherwise do not mark, solve the "how to extract the actual hotspot from the numerical deviation" landing problem. The binary decision logic is simple and easy to implement in a high-efficiency streaming or batch processing pipeline. The advantage is that the whole system only needs one comparison operation to complete the hotspot screening of ten thousand grids, supports high concurrency and low delay online monitoring, and provides timely and accurate spatial location basis for real-time risk early warning and insurance product recommendation.

[0043] Through the overall coupling application of the above steps, the problem of "lack of clear decision path from heat change to product recommendation" is solved. The determination of abnormal hotspots directly triggers subsequent spatial clustering and product mapping, ensuring that only the real risk points that have passed the quantitative deviation test will enter the recommendation process. The recommendation system eliminates the cumbersome manual screening link, realizes the full-automatic closed loop from raw geographic data to the final insurance scheme, improves the operation efficiency, and also enhances the credibility and explainability of the recommendation results, ultimately improves the user satisfaction and conversion rate.

[0044] The abnormal hotspot grid determined is spatially clustered to construct a hotspot area, and matched with an insurance product set, specifically including: Define the adjacency as any two grids and , if , then it is adjacent; merge the spatially adjacent hotspots into a continuous area; Initialize all grid in the area to unvisited; Traverse each unvisited grid , execute the following steps: S1, take any and unvisited, set the queue ; S2, pop the head of the queue , and its four adjacent if in and unvisited, then add and mark it as visited; S3, repeat step S2 until is empty, and the resulting set is a region ; S4, repeat steps S1 to S3 for the remaining unvisited hotspots until is generated; Get the number of region grids, denoted as ; Calculate the total area of the region as ; Calculate the average abnormality of the region as ; Calculate the total area of the study area as ; Provide area and anomaly intensity indicators for each region; Set the regional composite risk score as follows: A single risk score is generated by combining the size of the region and the intensity of the anomaly. Let the complete set of insurance products be Each product The predefined risk application range is Threshold and The selection of [a specific factor] directly determines the risk score for each region. The specific impact of mapping to different product categories is as follows: 1. The breadth of the mapping range; the wider the threshold range, the more each product... The applicable risk range covers more possibilities. To ensure that there are at least one candidate product across all risk levels and that the mapping is comprehensive, avoiding the problem of too many products for each region, which leads to a lengthy result list and reduced recommendation accuracy and differentiation, a narrower threshold range means that each product only covers smaller fluctuations. The mapping results are more refined, and the recommended products are highly matched with regional risks; if the threshold is set too narrowly, some... 1. Falling into the gap region results in no product mapping. 2. Threshold is too high or too low, lower limit. Low, product Opening up to low-risk areas will increase product recall rates in areas with inflated risk levels, but may also allow users in areas with already low risk to see overly insured products; lower limit The risk level is too high. The product is only available to medium- and high-risk areas, which will reduce recommendations to low-risk areas and optimize filtering capabilities. However, this may result in some medium- and low-risk areas being excluded. Not enough products could be matched. Maximum The upper limit is too high, making the product suitable for higher-risk areas and easier to cover extreme risk scenarios, but it may also lead to over-recommendation of the product in less extreme areas; The risk level is too low, meaning the product is only suitable for low to medium risk areas. This avoids recommending insufficiently protected products in high-risk areas, but it also presents the problem of having no products available in high-risk areas. 3. Practical Recommendations: Quantitative division based on risk score distribution; first, analyze all historical or real-time calculated regional scores... Perform quantile analysis, such as 20%, 50%, and 80%, to divide the risk into several intervals from low to high; for example, a low-risk zone. medium-risk area High-risk areas Then, different product sets are assigned to different risk ranges, and these ranges are mapped to the respective products. ;This ensures that no risk segment is missed; each risk segment corresponds to a reasonable number of products; as new data is constantly updated, the interval boundaries , The mapping function can be recalculated regularly to keep the dynamic fit of the mapping function.

[0045] The mapping function from the region to the candidate risk set is set as: ; accurately map the region score to the product set matching the risk; wherein, is the set of candidate insurance products corresponding to the region.

[0046] By defining four-connected adjacency and performing breadth-first clustering, spatially adjacent abnormal hotspots are automatically merged into continuous hotspot regions, solving the problem of "single-point anomalies cannot reflect the true geographical distribution of risks." This step can fuse locally discrete high-abnormality grids into regions with spatial continuity, avoiding the misjudgment of isolated noise points as large-scale risks. The advantage is that the regions formed after clustering are closer to the actual event impact range, making subsequent risk assessment and product matching more focused, reducing the fragmentation of the recommendation results, and improving the accuracy of the recommendation coverage. By initializing all hotspots as "unvisited" and performing queue enqueue and adjacency expansion for each unvisited grid, the algorithm engineering problem of "how to efficiently traverse and construct non-overlapping clusters" is solved. This clustering process only relies on simple queue operations and does not require complex data structures or third-party libraries, ensuring linear time complexity and efficient execution in large-scale grid scenarios. The advantage is that even with tens of thousands of hotspots online simultaneously, this method can complete clustering in a short time, providing reliable spatial partitioning for real-time or near-real-time insurance recommendations. By calculating the number of grids, total area, and average abnormality of each clustered region, the problem of "unable to quantify the risk intensity and scale of clustered regions" is solved. This step not only obtains the geometric range of the clusters but also obtains the statistical description of the region's abnormality, providing multi-dimensional input for subsequent comprehensive scoring. The advantage is that in risk assessment, both "the larger the area, the wider the impact" and "the higher the average deviation, the more intense the risk" are considered, making the final score truly reflect the overall risk situation of the region. By multiplying the region size ratio and the average abnormality to generate a composite risk score, the defect of "single dimension cannot consider both scale and intensity" is solved. This score combines the region size and heat deviation, ensuring that it neither overestimates the risk due to small-scale extreme deviation nor ignores potential risks due to large-scale slight deviation. The advantage is that the single dimensionless score generated for each region can be directly used for product mapping and sorting, simplifying subsequent logic and enhancing the interpretability and business controllability of the score. By dynamically dividing risk intervals based on quantile analysis and assigning each product to a risk interval, the problem of "how to accurately map products to risk regions" is solved. This step not only ensures that each product is assigned to a risk region that matches its risk level but also guarantees that the product set in each risk region is reasonable and diverse, making the recommendation results more accurate and comprehensive. By precisely mapping regional scores, the problem of "fixed threshold mapping failing to accommodate different risk levels and product tiers" is solved. Using the 20%, 50%, and 80% quantiles of historical or real-time regional scores as risk segment boundaries, low, medium, and high-risk zones can be adaptively divided, and corresponding product sets can be configured for different zones. The advantages are: it ensures product coverage for all risk segments (no omissions in the mapping), while also allowing for appropriately refined recommendations, concentrating high-value products in high-risk zones, medium-value products in medium-risk zones, and entry-level products in low-risk zones, significantly improving the matching accuracy of recommendation results and user satisfaction. This is achieved through a mapping function. This function directly outputs a set of candidate insurance products, solving the technical challenge of "how to translate complex scoring logic into executable recommendations." It uses simple range judgments to filter products, making it intuitive and easy to manage. The advantages include: supporting rapid iteration and parameter tuning, allowing insurance operations teams to adjust settings at any time. The boundaries are fed back to the recommendation system in real time, improving system flexibility and business responsiveness.

[0047] In summary, this step, through a complete process design from fine-grained clustering to multi-dimensional scoring and then to flexible mapping, thoroughly solves the core challenges of spatial risk identification and accurate product matching. It outperforms traditional static threshold or single-point mapping methods in terms of efficiency, reliability, and interpretability, providing a solid foundation and tangible business value for large-scale, dynamic insurance recommendations.

[0048] The process of weighting and sorting the matched candidate insurance products to generate a personalized insurance recommendation list includes: Merge the candidate sets of all regions: ;in, For all candidate insurance products; integrate candidate products from various regions; For each product Set the recommendation weight as follows: ;in, Assign weights to products; measure the suitability of each product across all regions; according to Sort in descending order to generate the final recommendation sequence. .

[0049] By merging the candidate product sets of all regions, the problem of isolated candidate sets of each abnormal region being difficult to manage uniformly is solved, and the beneficial effect is that all candidate products from different geographical hotspots are integrated into a set C, avoiding omission and repetition in the cross-regional recommendation process, so that subsequent sorting and display can be based on a complete and consistent product pool, greatly improving the overall control ability of the system for multi-region risks. By calculating the weighted recommendation weight for each product, the problem of "some products being suitable for multiple high-risk regions but being difficult to compare with products suitable for only a single region" is solved, and the beneficial effect is that the composite risk scores of each region are organically accumulated, so that products that can truly cover multiple risk hotspots or have outstanding performance in high-risk regions naturally have higher weights. This not only quantifies the cross-region applicability of products, but also makes the recommendation results more in line with the user's multi-scenario protection needs, improving the accuracy of product matching and the degree of personalization of recommendations. By arranging the final recommendation sequence in descending order of recommendation weight, the problem of not being able to clearly display the "top priority" recommendation in a large number of candidate products is solved, and the beneficial effect is that the product that best matches the user's overall risk profile is placed at the front of the list based on weight-based sorting, ensuring that the user can see the product that best meets their multi-region risk needs at the first time, and also reducing the user's selection cost through ordered display. This approach balances system automation decision-making and user experience, significantly improving the readability and conversion rate of the recommendation list.

[0050] In summary of the above steps, the present scheme forms a closed-loop and interpretable recommendation path through "set merging-weight accumulation-priority sorting", which not only takes into account the comprehensive measurement of cross-regional risks, but also realizes the fine quantification of the value contribution of each product, significantly improving the accuracy, stability and user satisfaction of the insurance recommendation system in multi-dimensional scenarios.

[0051] It should be noted that, in this document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0052] The above description is only the preferred embodiment of the present application, and it should be noted that for those skilled in the art, without departing from the technical principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should also be considered as the protection scope of the present application.

Claims

1. A big data-based insurance recommendation method, characterized in that, include: Collect geographic behavior data and establish a two-dimensional spatial coordinate system, wherein the geographic behavior data includes user location information and its corresponding time; The study area was divided into grids according to the principle of equal area to construct regular spatial units; The number of visits to each grid within multiple time periods is counted to form a spatiotemporal access distribution; Calculate the access density and heat index of each grid and construct a dynamic thermal model; Extract the heat statistics of each grid in historical time periods to generate heat baseline indicators; Based on the degree of deviation between the heat baseline and the current heat, determine whether the grid is an abnormal hotspot; Spatial clustering is performed on hotspot grids identified as anomalous to construct hotspot regions, which are then matched with a set of insurance products. The matched candidate insurance products are weighted and sorted to generate a personalized insurance recommendation list.

2. The insurance recommendation method based on big data according to claim 1, characterized in that, The process of collecting geographic behavior data and establishing a two-dimensional spatial coordinate system includes user location information and its corresponding time, specifically: Obtain from user terminal or background logs 1 record, including ; in, For the first The latitude of each record, in degrees; For the first The longitude of each record, in degrees; For the first The time of each record, in seconds; Within the study area, the latitude and longitude of the southwesternmost point were selected. As the origin of the plane coordinate system, a unified reference point is established to facilitate consistency in subsequent coordinate transformations. Establish a two-dimensional rectangular coordinate system, where the due east direction is selected, corresponding to the direction of increasing longitude. Axis; select true north as the direction corresponding to increasing latitude. Axis; defines the plane direction to ensure that the meaning of the coordinates is clear after mapping; Obtain the Earth's average radius, denoted as . By approximating a small region as a plane, each record is mapped to planar coordinates, specifically as follows: , ;in, For the first The x-coordinate of each record; For the first The vertical axis of each record.

3. The insurance recommendation method based on big data according to claim 2, characterized in that, The process of dividing the study area into grids according to the principle of equal area and constructing regular spatial units specifically includes: In the planar coordinate system, select a grid side length of... The region is divided into several square grid units, specifically: ； The area of ​​each cell ; Assign a unique index to each grid And record its center coordinates. ;in, For grid The coordinates of the center point.

4. The insurance recommendation method based on big data according to claim 3, characterized in that, The method of statistically analyzing the number of visits to each grid within multiple time periods to form a spatiotemporal access distribution specifically includes: Let the granularity of time statistics be... , the current moment Divide forward into several time periods ; For each grid and the selected current time period The number of records falling into this grid is counted, specifically: ;in, For grid Number of visits within a time period; Let be a counting function, and let be the number of elements in a set.

5. The insurance recommendation method based on big data according to claim 4, characterized in that, The calculation of the access density and heat index of each grid, and the construction of a dynamic thermal model, specifically includes: Set the grid access density to: ; Let the predefined reference density be ; The grid heat index is calculated as follows: .

6. The insurance recommendation method based on big data according to claim 5, characterized in that, The step of extracting the heat statistics of each grid over historical periods to generate a heat baseline index specifically includes: For each cell Before collecting Heat sequence for each time period ; The median heat is calculated as follows: ; in, Calculate the median of the sequence; The median absolute deviation is calculated as follows: ; in, For calculation and The absolute value of the difference.

7. The insurance recommendation method based on big data according to claim 6, characterized in that, The determination of whether a grid is an abnormal hotspot based on the deviation between the heat baseline and the current heat specifically includes: For each cell The outlier is calculated as follows: ; in, To prevent division by zero for extremely small constants; Set the abnormal threshold as ;like If so, mark the cell as an abnormal hotspot: ； like If it is, then no mark is made.

8. The insurance recommendation method based on big data according to claim 7, characterized in that, The process of spatially clustering the hotspot grids identified as anomalous to construct hotspot regions and matching them with a set of insurance products specifically includes: Adjacency is defined as any two grids and If satisfied If they are adjacent, then they are adjacent; All The middle grid is initialized to unvisited; Traverse each unvisited grid Perform the following steps: S1, take any And not visited, set up a queue ; S2, pop up the queue head Connect its four neighbors If in If not visited, then join. And mark the access; S3. Repeat step S2 until... If empty, the resulting set is a region. ; S4. Repeat steps S1 to S3 for the remaining unvisited hotspots until a result is generated. ; Get the number of grid cells in the region, denoted as ; The total area of ​​the calculation region is ; The average anomaly of the calculated region is ; Calculate the total area of ​​the study region as follows: ; Set the regional composite risk score as follows: ; Let the complete set of insurance products be Each product The predefined risk application range is ; The mapping function from the region to the candidate insurance product set is set as follows: ； in, This is the set of candidate insurance products corresponding to this region.

9. The insurance recommendation method based on big data according to claim 8, characterized in that, The process of weighting and sorting the matched candidate insurance products to generate a personalized insurance recommendation list includes: Merge the candidate sets of all regions: ;in, For all candidate insurance types; For each product Set the recommendation weight as follows: ;in, Weighting for product recommendations; according to Sort in descending order to generate the final recommendation sequence. .

Images