Watershed clustering method for regional health big data

Through the watershed clustering method, the mixed k-nearest neighbor map and water collection basin merging technology is used to solve the problem of pre-specifying the number of classes and processing complex structures in regional health big data, and efficient data clustering and accurate identification are achieved, which improves clustering and clustering. Class performance and accuracy of results.

CN118940060AInactive Publication Date: 2024-11-12HUAIHUA UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410940107.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

When processing regional health big data, existing clustering methods have defects, oversegment and overabsorbing of pre-specified classes, which makes it difficult to effectively deal with multiple complex data structures, such as multi-density, irregular shapes and overlapping class boundaries, resulting in clustering. Class performance and accuracy are not high.

Method used

The watershed clustering method is adopted to construct a mixed k-nearest neighbor map, identify core points and boundary points, and use altitude and nearest neighbor information to discover the water collection basin, combine the stability factor and the similarity of the water collection basin, and use the boundary factor and The Laida law aggregates unclustered data points to achieve automatic clustering.

Benefits of technology

Overcome the limitations of traditional clustering methods, realize automatic clustering and accurate identification of the number of classes, improve the clustering performance of multi-dimensional and multi-modal data, avoid the generation of invalid connections and noise data, and improve the clustering results. accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940060A_ABST
    Figure CN118940060A_ABST
Patent Text Reader

Abstract

The invention discloses a watershed clustering method for regional health big data, and belongs to the field of big data processing and pattern recognition, and the method comprises the following steps: constructing a mixed k-nearest neighbor graph based on an input data set, the mixed k-nearest neighbor graph comprising an edge weight set and a vertex attribute set, respectively identifying a core point and a boundary point based on the roughness factor and the boundary factor of the vertex attribute set; discovering a water collection basin set and a water collection basin connection set based on the altitude and neighbor information of the core point and neighbor members of the core point; identifying effective water collecting basin connections from the water collecting basin connection set based on a stable factor, and combining the water collecting basins by utilizing water collecting basin level similarity to obtain combined water collecting basins; and aggregating the data points which are not clustered to a water collecting basin by utilizing the neighbor information and the boundary points to obtain a final data cluster. According to the method, the limitation that the existing watershed clustering method and the traditional density clustering method cannot process various complex structures is overcome, and the clustering performance and quality are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of big data processing and pattern recognition, and in particular relates to a watershed clustering method for regional health big data. Background Art

[0002] The rapid growth of population and the informatization of medical data have led to the emergence of massive, diversified, high-dimensional and structurally liberalized health data in different regions. The scale of data patterns (classes, clusters and groups) of these large-scale health data also increases with the growth of time span and sample size. Detecting and mining different health data patterns from regional health big data, such as healthy people, sub-healthy people, people with different diseases, and the pattern information reflected by different people, can not only promote the development of new medical treatments such as accurate treatment based on health big data, the concept and technology of "preventive treatment by traditional Chinese medicine", and gene targeted therapy, but also promote the development of the health care industry, thereby reducing the waste of precious medical and health data resources. However, in the process of structural analysis of data space (sample space), we found that the complexity of data is manifested in many forms, such as: (1) diversity of class shapes; (2) large differences in scale and boundary overlap between classes; (3) weak connectivity within a class is close to the distance between classes; (4) the dimensionality curse of data leads to empty space and distance concentration.

[0003] Existing clustering methods have mitigated the impact of complex data structures on data clustering to a certain extent. For example, the center clustering method (K-Means) solves the detection of Gaussian classes that can better detect boundary overlap. The density clustering algorithm method (DBSCAN) can identify irregular data patterns. The density peak clustering method (DPC) can solve the overlap of class boundaries and detect irregular data classes based on the specified number of classes, density peak points and density diffusion strategies. However, due to the large scale of regional health big data, the mixture of multiple complex data structures, and the inclusion of many data patterns, it is difficult to obtain prior knowledge of data patterns. In addition, the influence of outlier data points and abnormal data classes makes clustering methods based on pre-specified class data such as K-Means and DPC perform poorly. Although the DBSCAN method does not require the pre-specified number of clusters and runs fast, it and its variant methods are still affected by the unified threshold, making it difficult to solve the density differences of different classes and the influence of high overlap of class boundaries, resulting in over-segmentation and over-absorption. In addition, the existing watershed clustering method has achieved good performance, but its secondary terrain topology and angle-based outlier point calculation cost is large, resulting in high time complexity. This is also the deficiency and limitation of existing density clustering, center clustering and other methods. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a watershed clustering method for regional health big data to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above object, the present invention provides a watershed clustering method for regional health big data, comprising:

[0006] Constructing a hybrid k-nearest neighbor graph based on an input data set, wherein the hybrid k-nearest neighbor graph includes an edge weight set and a vertex attribute set, and identifying core points and boundary points based on a roughness factor and a boundary factor of the vertex attribute set, respectively;

[0007] Finding a water basin set and a water basin connection set based on the altitude and neighbor information of the core point and its neighbor members;

[0008] identifying valid water collection basin connections from the water collection basin connection set based on a stability factor, and merging the water collection basins using water collection basin-level similarity to obtain a merged water collection basin;

[0009] The neighbor information and boundary points are used to aggregate the unclustered data points into the catchment basin to obtain the final data clustering.

[0010] Optionally, the process of constructing a hybrid k-nearest neighbor graph includes:

[0011] Using Euclidean distance to calculate the k-nearest neighbor information of the data point on the input data set to obtain the edge weight set of the mixed k-nearest neighbor graph; wherein the edge weight set includes: k-nearest neighbors, net k-nearest neighbors, mutual k-nearest neighbors and shared k-nearest neighbors;

[0012] A vertex attribute set of a hybrid k-nearest neighbor graph is calculated based on the k-nearest neighbor information of the data point; wherein the vertex attribute set includes altitude, roughening factor, boundary factor and angle factor.

[0013] Optionally, the elevation of a data point is calculated as:

[0014]

[0015] In the formula, α(x i ) represents the data point x i The altitude, w_dist(x i ,x j ) represents x i To the nearest neighbor x j Distance and x i The ratio coefficient of the distance to its net t nearest neighbors, β(x i ,x j ) represents x i With neighbor x j The average distance between them based on shared neighbors, pN t (x i ) represents xi The net t neighbors of , t≤k.

[0016] Optionally, the roughening factor of a data point is calculated as:

[0017]

[0018] In the formula, θ(x i ) represents the roughening factor, CFI(x i ,x j ) represents an indicator function.

[0019] Optionally, the boundary factor of a data point is calculated as:

[0020]

[0021] In the formula, π(x i ) represents the boundary factor, E(x i ) represents the equidistant points in the opposite direction, M(x i ) represents the local center point, BFI(x j ,M(x i ),E(x i )) represents the angle indicator function.

[0022] Optionally, the angle factor of a data point is calculated as:

[0023]

[0024] In the formula, represents the angle factor, Represents x i The radius of the normalized t-nearest sphere, ζ(x i ) represents x i The normalized center offset of the t-nearest-neighbor sphere, Represents x i The distance to the last member of the net t-neighborhood, SRPC(pN t (x i ))={PG1,…,PG j ,…,PG m}, PG i Represents point x i About the point x of t's nearest neighbor j The point set of shared neighbor groups, PGS(SRPC(pN t (x i ))) represents the set of all binary subsets of grouping points, γ(x i ) represents x i The angular density formed with the grouped net t neighbors.

[0025] Optionally, the process of discovering the collection basin set and the collection basin connection set includes:

[0026] S1. Arranging all data points in the hybrid k-nearest neighbor graph in ascending order of altitude in a vertex attribute set to obtain an initial data point queue, and initializing a water collection basin set and a water collection basin connection set;

[0027] S2, sequentially extracting the head data point of the initial data point queue, and when the head data point is a core point and has not been aggregated, adding the current head data point to the new water collection basin set;

[0028] S3, sequentially extracting neighbor members from the k-nearest neighbor queue of the current head data point, and performing aggregation when the altitude and neighbor information of the current head data point and the neighbor members meet the shared neighbor rule, Laida rule and voting method;

[0029] S4. Execute S2-S3 in a loop until the data point is empty to obtain a water collection basin set and a water collection basin connection set.

[0030] Optionally, the process of merging catchment basins using catchment basin-level similarity includes:

[0031] S5, arranging the connection pairs in ascending order based on the altitudes of the connection pairs in the water collection basin connection set to obtain a connection pair queue;

[0032] S6, sequentially extract the head connection pair of the connection pair queue, calculate the stability factor of the members in the connection pair, and execute S7 when the stability factor is greater than 0.2, otherwise execute S6;

[0033] S7. When at least one of the members in the connection pair is a basin point of a water collection basin, the water collection basins are merged; when none of the members in the connection pair is a basin point and at least one of the members in the connection pair is located at an altitude below half of the water collection basin to which it belongs, the water collection basins are merged;

[0034] S8. Execute S6-S7 in a loop. When the connection pair queue is empty, output the merged water collection basin.

[0035] Optionally, the calculation formula of the stability factor is:

[0036]

[0037] In the formula, χ(x i ) represents the stability factor, x i represents any member in the catchment basin connection (i, j); θ(x i ) represents x i The roughening factor of Represents x i Angle factor; max() represents the maximum value operation; Represents x iThe water collection basin; Represents x i The set of competing points for the catchment basin to which it belongs.

[0038] Optionally, the process of aggregating unclustered data points into catchment basins using neighbor information and boundary points includes:

[0039] S9, obtaining non-core points based on the unclustered data points, and sorting the non-core points in ascending order according to altitude to obtain a non-core point queue;

[0040] S10, sequentially extract the head non-core point of the non-core point queue, and sequentially obtain the nearest and clustered data points from the k nearest neighbors of the current head non-core point. When the local area elevation and neighbor information of the current head non-core point and the clustered data points in the k nearest neighbors meet the shared nearest neighbor rule and the Laida rule, the current head non-core point is aggregated to the water collection basin to which the clustered data points in the k nearest neighbors belong, otherwise execute S10;

[0041] S11, when the current team head non-core point is a boundary point, according to the shared nearest neighbor rule, the current team head non-core point is aggregated into the water collection basin to which the clustered data points in the k nearest neighbors belong;

[0042] S12, execute S10-S11 in a loop, and when the non-core point queue is empty, output the final data clustering.

[0043] Compared with the prior art, the present invention has the following advantages and technical effects:

[0044] The present invention utilizes the idea of ​​watershed algorithm to realize automatic clustering and obtain the number of classes, overcoming the defect that the traditional clustering method requires the pre-specification of the number of classes; combining the terrain topology method and the overlapping Gaussian distribution method, a new watershed clustering method for multi-dimensional and multi-modal regional health big data is proposed.

[0045] This paper combines the mutual k-nearest neighbor rule, Laida's rule and the quartile-based variability measurement method to propose a roughening factor to measure the coreness of data points. Compared with the core point measurement method based on intercept and density threshold of density clustering, it integrates graph theory, terrain topology, local information and data variability measurement, and further improves the effect of identifying core points.

[0046] Compared with the density clustering method that simply considers density information, the present invention combines the boundary properties and roughness factors of data points and proposes a stability factor for detecting effective water collection basin connections, thereby achieving accurate identification of effective connections between water collection basins, avoiding aggregation of invalid connections between subclasses of different classes, and further improving clustering accuracy.

[0047] The present invention overcomes the limitations of existing watershed clustering methods and traditional density clustering methods that fail to handle a variety of complex structures, including complex structures such as multiple densities, irregular shapes, weak intra-class connectivity, and overlapping class boundaries, and improves clustering performance and results. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0049] Figure 1 A flow chart of a watershed clustering method for regional health big data according to an embodiment of the present invention;

[0050] Figure 2 A flowchart of establishing a hybrid k-nearest neighbor graph according to an embodiment of the present invention;

[0051] Figure 3 A flow chart of finding a water collection basin according to an embodiment of the present invention;

[0052] Figure 4 A flow chart of merging water collection basins according to an embodiment of the present invention;

[0053] Figure 5 The flowchart of aggregating non-core points according to an embodiment of the present invention is shown in FIG. DETAILED DESCRIPTION

[0054] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0055] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0056] Embodiment 1

[0057] Aiming at the pattern detection problem of regional health big data and overcoming the shortcomings of existing clustering methods, and under the premise of lack of prior knowledge of the number of classes, a watershed clustering method for regional health big data is provided, which overcomes the shortcomings and limitations of traditional density clustering, center clustering, graph clustering and other methods.

[0058] The present invention discloses a watershed clustering method for regional health big data. Based on the idea of ​​a watershed algorithm, the weight set of edges (k nearest neighbors, net k nearest neighbors, mutual k nearest neighbors and shared nearest neighbors) and the attribute set of vertices (elevation, roughness factor, angle factor and boundary factor) of a mixed k nearest neighbor graph are calculated using sample dissimilarity, and the core points and boundary points are respectively identified using the roughness factor and the boundary factor; according to the idea that the altitudes of adjacent areas are similar, the nearest neighbor propagation method, the Gaussian overlap method and the shared nearest neighbor method are used to identify the watershed basins composed of the core points, and then the watershed basins are merged according to the stability factor and the watershed basin level similarity; finally, the boundary factor, the Laida rule and the shared nearest neighbor rule are used to aggregate non-core points to achieve accurate recognition of data patterns.

[0059] like Figure 1 As shown, in this embodiment, a watershed clustering method for regional health big data is provided, which mainly realizes clustering of data patterns based on a hybrid k-nearest neighbor graph and a roughening factor, and includes the following steps: constructing a hybrid k-nearest neighbor graph based on the input data set, the hybrid k-nearest neighbor graph including an edge weight set and a vertex attribute set, identifying core points and boundary points based on the vertex attribute set; identifying a watershed set and a watershed connection set based on the altitude and neighbor information of the core point and the neighbor members of the core point; based on the watershed connection set, calculating the stability factor and the watershed level similarity to merge the watersheds to obtain a merged watershed; obtaining non-core points that are not clustered into the watershed based on the data set and the merged watershed; and aggregating the non-core points to obtain the final data clustering.

[0060] like Figure 2 As shown, the process of establishing a hybrid k-nearest neighbor graph G includes: Step 1, according to the input data point and the Euclidean distance, calculate the k-nearest neighbors (including itself) and net k-nearest neighbors (excluding itself) of each data point, calculate the mutual k-nearest neighbors and shared k-nearest neighbors based on the k-nearest neighbors, and obtain the edge weight set of the hybrid k-nearest neighbor graph. Step 2, according to the geometric information of the hybrid k-nearest neighbor graph and the data space, calculate the vertex attribute set of the hybrid k-nearest neighbor graph, including altitude (topography), roughening factor (identifying core points), boundary factor, angle factor, and then use the roughening factor and boundary factor to identify the core points and boundary points respectively.

[0061] The elevation calculation process of the vertex attribute set of the hybrid k-nearest neighbor graph (G) includes:

[0062] (1) Using t nearest neighbor information and shared nearest neighbors, we can obtain the point pair (x i ,x j ) based on the distance and distance of shared t neighbors, dist_snn(x i ,x j ), the formula is as follows:

[0063]

[0064] In the formula, dist(x i ,x q ) represents x based on Euclidean distance i and x q The difference between SNN t (x i ,x j ) represents x i The t nearest neighbors of x j The intersection of t nearest neighbors.

[0065] (2) Based on the mutual t-nearest neighbor and shared t-nearest neighbor methods, we can obtain x i With neighbor x j The mean distance between the shared neighbors, β(x i ,x j ), the formula is as follows:

[0066]

[0067] Where, pN t (x i ) represents x i The net t neighbors (excluding itself x i ), t≤k; RNN t (x i ,x j ) represents the point pair (x i ,x j ) whether they satisfy mutual t-nearest neighbors, if they do, the value is logically true, otherwise it is logically false; |SNN t (x i ,x j )| represents the modulus of the shared t-neighbor set.

[0068] (3) Using the net t nearest neighbors, we can obtain x i To the nearest neighbor x j Distance and x i The ratio coefficient of the distance to its net t neighbors, w_dist(x i ,x j ), the formula is as follows:

[0069]

[0070] Using the distance mean and ratio coefficient based on shared neighbors, we get the data point x i The altitude is as follows:

[0071]

[0072] The calculation process of the roughening factor of the vertex attribute set of the mixed k-nearest neighbor graph includes:

[0073] (1) Using data point x j The altitudes of the t nearest neighbors are obtained to obtain the mean and variance of the altitude set, and the weakened left-hand Laida law and the weakened right-hand Laida law are established. The formula is as follows:

[0074]

[0075] In the formula, α(N t (x j )) represents the data point x in the data set D j The elevation of all members of the t-nearest neighbor; μ() represents the mean operation; σ() represents the variance operation; True and False represent logical true and logical false respectively; FPCW(α(x i ),α(x j )) represents the data point x i The altitude satisfies x j The Zolaida law of the 2σ law of the terrain in the t-neighboring area is weakened; RPCW(α(x j ),α(x i )) represents the data point x i The terrain of the t neighboring area satisfies x j The right Laida law of the 2σ law at altitude is weakened.

[0076] Based on the weakening of the left Laida rule and the right Laida rule, as well as the mutual t-nearest neighbors, the weakening of the Laida rule is obtained, and the formula is as follows:

[0077]

[0078] In the formula, PCW(x i ,x j ) is True, indicating that x i The topography of the t neighboring area is related to x j The terrain topology within the t-nearest region appears as a Gaussian terrain anomaly.

[0079] (2) Using data point x i The altitude of the t nearest neighbors is obtained, and the quartile distance of the altitude set is obtained. The weakened left quartile rule and the weakened right quartile rule are established. The formula is as follows:

[0080]

[0081] In the formula, α(N t (x i )) represents the data point x in the data set D i The elevations of all members of t's nearest neighbors; Q L () indicates the lower quartile operation; Q U Indicates the upper quartile operation; IQR(x i) represents the nearest neighbor N t (x i )The distance between the upper quartile and the lower quartile of the altitude set; True and False represent logical true and logical false respectively; RIQRW t (x i ,x j ) represents the data point x i The altitude satisfies x i The lower quartile rule of the topography in the neighboring area of ​​t is weakened; RIQRW t (x j ,x i ) represents the data point x i The altitude satisfies x j The upper quartile rule of the terrain in the neighboring area of ​​t is weakened.

[0082] Based on the weakening of the lower quartile rule and the weakening of the upper quartile rule, as well as the mutual t-nearest neighbors, the weakening of the quartile rule is obtained, and the formula is as follows:

[0083]

[0084] In the formula, IQRW(x i ,x j ) is True, indicating that x i The topography of the t neighboring area is related to x j The terrain topology within the t-neighboring area appears as interval terrain anomaly.

[0085] (3) Using the weakening of the Laida rule and the quartile rule, the roughening factor of the core point identification is obtained. The formula is as follows:

[0086]

[0087] In the formula, not RNN t (x i ,x j ) represents the point pair (x i ,x j ) does not satisfy the mutual nearest neighbor rule and is logically false; CFI(x i ,x j ) is an indicator function.

[0088] The calculation process of the boundary factor of the vertex attribute set of the mixed k-nearest neighbor graph includes:

[0089] (1) Based on data point x i t nearest neighbors, calculate x i The local center point M(x i ), the equidistant point in the opposite direction E(x i ), the formula is as follows:

[0090] E(x i )=x i M(x i )×2+M(x i )

[0091] In the formula, E(x i ) is the vector M(x i )x i Points in the direction satisfying |x i M(x i )|=|x i E(x i )|.

[0092] (2) Based on the local center point M(x i ) and the equidistant point E(x i ), and obtain the angle indication function, the formula is as follows:

[0093]

[0094] In the formula, the symbol · represents the inner product operation of the vector; E(x i )M(x i ) and E(x i )x j Represent two different vectors respectively.

[0095] (3) Based on the t nearest neighbor and angle indicator function, the boundary factor is obtained, and the formula is as follows:

[0096]

[0097] When π(x i ) is equal to 0, or x i The net t nearest neighbor member x j Satisfies ||E(x i )x j ||≤||E(x i )x i ||, data point x i It is a boundary point.

[0098] (4) The formula of the boundary point set BPS is as follows:

[0099] BPS

[0100] ={x i |π(x i )==0or(π(x j )==0and||E(x i )x j ||≤||E(x i )x i ||),xj ∈pN t (x i )}

[0101] In the formula, |||| represents the modular operation of the vector.

[0102] The calculation process of the angle factor of the vertex attribute set of the mixed k-nearest neighbor graph is:

[0103] (1) Using data point x i and x j t nearest neighbors, and calculate their center points M(x i ) and M(x j ), get x i The radius and center offset of the normalized t-nearest sphere (point x i To point M(x i ) are expressed as and ζ(x i ), the formula is as follows:

[0104]

[0105] In the formula, and Respectively represent x i and x j The distance to the last member of the net t-neighborhood; dist() represents the distance operation.

[0106] (2) Calculate point x i The angle density formed with the grouped net t neighbors is as follows:

[0107]

[0108] In the formula, SRPC(pN t (x i ))={PG1,…,PG j ,…,PG m}, PG i Represents point x i About the point x of t's nearest neighbor j The set of points of shared neighbor groups, namely PG j The adjacent members of the same group satisfy the requirement of sharing no less than 50% of their neighbors; represents the set of all binary subsets of grouped points, where BS() represents the binary subset operation of the set.

[0109] (3) The formula for the angle factor is as follows:

[0110]

[0111] In the formula, || represents the modulus of the set.

[0112] The process of discovering the water collection basin includes: arranging the data points in ascending order according to their elevation, sequentially taking data points from the head of the queue, and using nearest neighbor propagation, overlapping Gaussian method and shared nearest neighbor method to simulate the immersion of the water collection basin to achieve the discovery of the water collection basin.

[0113] The specific implementation process includes: S1, arranging all data points in ascending order of the altitude in the vertex attribute set for the mixed k-nearest neighbor graph to obtain an initial data point queue, and initializing the water collection basin set and the water collection basin connection set; S2, sequentially extracting the head data point of the initial data point queue, and when the head data point is a core point and has not been aggregated, adding a new water collection basin to the water collection basin set for the current head data point; S3, sequentially extracting neighbor members from the k-nearest neighbor queue of the current head data point, and when the altitude and neighbor information of the current head data point and the neighbor members meet the shared neighbor rule, Laida rule and voting method, aggregation is performed; S4, looping S2-S3 until the data point is empty to obtain the water collection basin set and the water collection basin connection set.

[0114] That is, according to the input mixed k-nearest neighbor graph G, firstly sort all data points in ascending order according to altitude and store them in queue L. At the same time, initialize the catchment set B and the catchment connection set BL to be empty. Then, take data point i from the head of queue L in order, and use the coreness of the data point, neighbor information, Laida's rule, shared neighbor method, etc. to find the catchment and catchment connection, and save them in B and BL respectively.

[0115] like Figure 3 As shown, the specific steps include:

[0116] For the input mixed k-nearest neighbor graph G, all data points are sorted in ascending order using the vertex attribute elevation value and saved in the queue L; the water basin set B and the water basin connection set BL are initialized as empty sets respectively.

[0117] Sequentially take data point i from the head of queue L. If i is not aggregated by any catchment basin, take bi as the catchment basin to which i belongs, and add bi to B; then, sequentially take neighbor member j from i's k-nearest neighbor queue, where data point j is the core point. When the elevation and neighbor information of the k-nearest neighbor areas of i and j satisfy the shared neighbor rule, Laida rule and voting method, aggregation is performed in two ways: ① When j does not belong to any catchment basin, aggregate i and j and add j to bi; ② When j has been aggregated to other catchments, i and j are core points, add the catchment connection pair (i, j) to BL, and add i to the catchment competition point set Bcp to which j belongs, and add j to the catchment competition point set Bcp to which i belongs. ③ Until L is empty, the catchment set B and the catchment connection set BL are obtained.

[0118] The process of merging water collection basins includes: S5, arranging the connection pairs in ascending order based on the altitudes of the connection pairs of the water collection basin connection set to obtain a connection pair queue; S6, sequentially extracting the head connection pairs of the connection pair queue, calculating the stability factors of the members in the connection pair, and executing S7 when the stability factor is greater than 0.2, otherwise executing S6; S7, merging the water collection basins when at least one of the members in the connection pair is a basin point of the water collection basin; merging the water collection basins when none of the members in the connection pair are basin points and at least one of the members in the connection pair is located at an altitude below half of the water collection basin to which it belongs; S8, looping through S6-S7, and outputting the merged water collection basin when the connection pair queue is empty.

[0119] That is, the stability factor is used to identify effective catchment basin connections, and the catchment basins are merged in combination with the catchment basin-level similarity to obtain data clustering.

[0120] The specific implementation process includes:

[0121] The elevations of the connection pairs of the catchment basin connection set BL are used to sort them in ascending order into the queue LS. Then, the connection pair (i, j) is taken from LS. If the stability factor threshold of (i, j) is 0.2, the catchment basin is merged. When LS is empty, a cluster consisting of core points is obtained.

[0122] like Figure 4 As shown, the specific steps include:

[0123] For the input catchment basin connection set BL, all connection pairs (i, j) in all BL are arranged in ascending order of the elevation of the data point with the smaller elevation value in each connection pair (i, j) in LS, and saved in the queue LS.

[0124] Sequentially take out the basin connection (i, j) from the head of the LS team and calculate the stability factor of the members of the connection pair. The formula is as follows:

[0125]

[0126] In the formula, x i represents any member in the catchment basin connection (i, j), or a data point in the data set D; θ(x i ) represents x i The roughening factor of Represents x i Angle factor; max() represents the maximum value operation; Represents x i The water collection basin; Represents x i The set of competing points for the catchment basin to which it belongs.

[0127] When the stability of any point in the connection relationship between the water collection basins is x(x i)>0.2, the catchment connection relationship can be used to merge catchments. When the catchment connection satisfies the catchment level similarity, there are two strategies for merging catchments: (1) When i or j is a basin point of the catchment (the lowest elevation point in the catchment), the catchment where j is located is merged into the catchment where i belongs; (2) When i and j are not basin points, and the altitude of i or j is at or below half the altitude of the catchment where it belongs, the catchment where j belongs is merged into the catchment where i belongs; When LS is empty, a cluster consisting of core points is obtained.

[0128] The process of aggregating non-core points includes: S9, sorting the non-core points in ascending order according to the altitude to obtain a non-core point queue; S10, sequentially extracting the head non-core point of the non-core point queue, and sequentially obtaining the nearest and clustered data points from the k nearest neighbors of the current head non-core point. When the local area altitude and neighbor information of the current head non-core point and the clustered data points in the k nearest neighbors meet the shared nearest neighbor rule and the Laida rule, the current head non-core point is aggregated to the water collection basin to which the clustered data points in the k nearest neighbors belong, otherwise S11 is executed; S11, when the current head non-core point is a boundary point, according to the shared nearest neighbor rule, the current head non-core point is aggregated to the water collection basin to which the clustered data points in the k nearest neighbors belong; S12, looping through S10-S11, and when the non-core point queue is empty, outputting the final data clustering,

[0129] That is, according to k nearest neighbor, boundary factor, shared nearest neighbor method and Laida's rule, the unassigned non-core points are aggregated. First, the non-core points with non-boundary properties are aggregated using shared nearest neighbor and Laida's rule; then the non-core points with boundary properties are aggregated to the clusters to which their k nearest neighbor members that have been clustered belong, and the final clustering result is obtained. The non-core points that are not aggregated are noise data.

[0130] The specific implementation process includes:

[0131] For the input data set D, the hybrid k-nearest neighbor graph G, the shared nearest neighbor method and the Laida rule are used to aggregate the unclustered non-core points into the catchment basin B to obtain the final clustering.

[0132] like Figure 5 As shown, the specific steps include:

[0133] According to the catchment basin and data set D, non-core points that are not clustered into the catchment basin B are obtained, sorted in ascending order of altitude, and saved in the queue UL.

[0134] Sequentially take data point i from the head of the queue UL, then sequentially select the nearest and clustered data point j from i's k nearest neighbors, and then perform aggregation according to the following two-step strategy: (1) If the local area elevation and neighbor information of i and j satisfy the shared nearest neighbor rule and the Laida rule, i is aggregated to the catchment basin to which j belongs; otherwise, execute (2) i is a boundary point, and according to the shared nearest neighbor rule, i is clustered to the catchment basin where j belongs. When UL is empty, the aggregation of non-core points ends, and the final clustering result B is obtained. Non-core points that are not aggregated are noise data.

[0135] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A watershed clustering method for regional health big data, characterized in that: include: Constructing a hybrid k-nearest neighbor graph based on an input data set, wherein the hybrid k-nearest neighbor graph includes an edge weight set and a vertex attribute set, and identifying core points and boundary points based on a roughness factor and a boundary factor of the vertex attribute set, respectively; Finding a water basin set and a water basin connection set based on the altitude and neighbor information of the core point and its neighbor members; identifying valid water collection basin connections from the water collection basin connection set based on a stability factor, and merging the water collection basins using water collection basin-level similarity to obtain a merged water collection basin; The neighbor information and boundary points are used to aggregate the unclustered data points into the catchment basin to obtain the final data clustering.

2. The watershed clustering method for regional health big data according to claim 1, characterized in that: The process of constructing a hybrid k-nearest neighbor graph includes: Using Euclidean distance to calculate the k-nearest neighbor information of the data point on the input data set to obtain the edge weight set of the mixed k-nearest neighbor graph; wherein the edge weight set includes: k-nearest neighbors, net k-nearest neighbors, mutual k-nearest neighbors and shared k-nearest neighbors; A vertex attribute set of a hybrid k-nearest neighbor graph is calculated based on the k-nearest neighbor information of the data point; wherein the vertex attribute set includes altitude, roughening factor, boundary factor and angle factor.

3. The watershed clustering method for regional health big data according to claim 2, characterized in that: The formula for calculating the elevation of a data point is: In the formula, α(x i ) represents the data point x i The altitude, w_dist(x i ,x j ) represents x i To the nearest neighbor x j Distance and x i The ratio coefficient of the distance to its net t nearest neighbors, β(x i ,x j ) represents x i With neighbor x j The average distance between them based on shared neighbors, pN t (x i ) represents x i The net t neighbors of , t≤k.

4. The watershed clustering method for regional health big data according to claim 3, characterized in that: The roughening factor calculation formula of the data point is: In the formula, θ(x i ) represents the roughening factor, CFI(x i ,x j ) represents an indicator function.

5. The watershed clustering method for regional health big data according to claim 4, characterized in that: The boundary factor calculation formula of a data point is: In the formula, π(x i ) represents the boundary factor, E(x i ) represents the equidistant points in the opposite direction, M(x i ) represents the local center point, BFI(x j ,M(x i ),E(x i )) represents the angle indicator function.

6. The watershed clustering method for regional health big data according to claim 5, characterized in that: The angle factor calculation formula of the data point is: In the formula, represents the angle factor, Represents x i The radius of the normalized t-nearest sphere, ζ(x i ) represents x i The normalized center offset of the t-nearest-neighbor sphere, Represents x i The distance to the last member of the net t-neighborhood, SRPC(pN t (x i ))={PG1,…,PG j ,...,PG m }, PG i Represents point x i About the point x of t's nearest neighbor j The point set of shared neighbor groups, PGS(SRPC(pN t (x i ))) represents the set of all binary subsets of grouping points, γ(x i ) represents x i The angular density formed with the grouped net t neighbors.

7. The watershed clustering method for regional health big data according to claim 2, characterized in that: The process of discovering the collection basin sets and collection basin connection sets includes: S1. Arranging all data points in the hybrid k-nearest neighbor graph in ascending order of altitude in a vertex attribute set to obtain an initial data point queue, and initializing a water collection basin set and a water collection basin connection set; S2, sequentially extracting the head data point of the initial data point queue, and when the head data point is a core point and has not been aggregated, adding the current head data point to the new water collection basin set; S3, sequentially extracting neighbor members from the k-nearest neighbor queue of the current head data point, and performing aggregation when the altitude and neighbor information of the current head data point and the neighbor members meet the shared neighbor rule, Laida rule and voting method; S4. Execute S2-S3 in a loop until the data point is empty to obtain a water collection basin set and a water collection basin connection set.

8. The watershed clustering method for regional health big data according to claim 7, characterized in that: The process of merging catchment basins using catchment basin-level similarity includes: S5, arranging the connection pairs in ascending order based on the altitudes of the connection pairs in the water collection basin connection set to obtain a connection pair queue; S6, sequentially extract the head connection pair of the connection pair queue, calculate the stability factor of the members in the connection pair, and execute S7 when the stability factor is greater than 0.2, otherwise execute S6; S7. When at least one of the members in the connection pair is a basin point of a water collection basin, the water collection basins are merged; when none of the members in the connection pair is a basin point and at least one of the members in the connection pair is located at an altitude below half of the water collection basin to which it belongs, the water collection basins are merged; S8. Execute S6-S7 in a loop. When the connection pair queue is empty, output the merged water collection basin.

9. The watershed clustering method for regional health big data according to claim 8, characterized in that: The calculation formula of the stability factor is: In the formula, χ(x i ) represents the stability factor, x i represents any member in the catchment basin connection (i, j); θ(x i ) represents x i The roughening factor of Represents x i Angle factor; max() represents the maximum value operation; Represents x i The water collection basin; Represents x i The set of competing points for the catchment basin to which it belongs.

10. The watershed clustering method for regional health big data according to claim 8, characterized in that: The process of aggregating unclustered data points into catchment basins using neighbor information and boundary points includes: S9, obtaining non-core points based on the unclustered data points, and sorting the non-core points in ascending order according to altitude to obtain a non-core point queue; S10, sequentially extract the head non-core point of the non-core point queue, and sequentially obtain the nearest and clustered data points from the k nearest neighbors of the current head non-core point. When the local area elevation and neighbor information of the current head non-core point and the clustered data points in the k nearest neighbors meet the shared nearest neighbor rule and the Laida rule, the current head non-core point is aggregated to the water collection basin to which the clustered data points in the k nearest neighbors belong, otherwise execute S10; S11, when the current team head non-core point is a boundary point, according to the shared nearest neighbor rule, the current team head non-core point is aggregated into the water collection basin to which the clustered data points in the k nearest neighbors belong; S12, execute S10-S11 in a loop, and when the non-core point queue is empty, output the final data clustering.

Citation Information

Cited By

  • Abnormal data detection method and system for regional health big data

    CN119648646A