A stroke-based interaction method for clustering guidance and refinement

Abstract

This invention provides a clustering guidance and refinement method based on stroke interaction, belonging to the field of data processing technology. Specifically, it includes: using adjusted parameters as adjustment thresholds, performing clustering analysis on a scatter plot based on the MST clustering algorithm to divide the nodes of the scatter plot into different cluster groups; obtaining the number of discrete nodes between different adjacent cluster groups; further, based on the distance between discrete nodes and adjacent nodes of adjacent cluster groups, automatically generating different types of connection lines between different adjacent cluster groups; determining the type of connection lines between different adjacent cluster groups based on the automatically generated connection lines and the modification results of manually interacted connection lines; and re-clustering the scatter plot according to the type of connection lines between different adjacent cluster groups, thereby improving the accuracy of scatter plot clustering.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to a clustering guidance and refinement method based on stroke interaction. Background Technology

[0002] In recent years, with the development of vectorization technology and nonlinear dimensionality reduction algorithms, vectorizing data into dense high-dimensional vectors and then reducing the dimensionality to a two-dimensional plane to form a scatter plot has become one of the standard technical paths for exploring large-scale data. Specifically, in the invention patent CN202410114975.0 "A Method for Optimizing and Classifying User Data in Multi-Energy Distribution Networks", several clusters are obtained by clustering the scatter plot based on the final centroid. The electricity consumption sequences of all users are then classified and stored according to the clusters. However, ideally, the clustering results, dimensionality reduction results, and user perception should be aligned, that is, visually significant clusters should be the categories obtained by clustering and conform to people's perception of the data. However, this is often difficult to achieve in practice.

[0003] To address the aforementioned technical problems, this invention provides a clustering guidance and refinement method based on stroke interaction. Summary of the Invention

[0004] The purpose of this invention is to provide a clustering guidance and refinement method based on stroke interaction.

[0005] To address the aforementioned technical problems, this invention provides a clustering guidance and refinement method based on stroke interaction, characterized by specifically including:

[0006] Obtain the number of nodes in the scatter plot and determine the adjustment parameters of the clustering algorithm based on the distance between different nodes;

[0007] Using the adjustment parameters as the adjustment threshold, cluster analysis is performed on the scatter plot based on the MST clustering algorithm to divide the nodes of the scatter plot into different cluster groups;

[0008] The number of discrete nodes between different adjacent cluster groups is obtained, and further, based on the distance between the discrete nodes and the adjacent nodes of the adjacent cluster groups, different types of connection lines are automatically generated between different adjacent cluster groups.

[0009] Based on the modification results of automatically generated connection lines and manually interacted connection lines, the correction type of connection lines for different adjacent cluster groups is determined, and the scatter plot is re-clustered according to the correction type of connection lines for different adjacent cluster groups.

[0010] A further technical solution is that the distance between the different nodes is determined according to the Euclidean distance function.

[0011] A further technical solution is that the method for determining the adjustment parameters of the clustering algorithm is as follows:

[0012] The number of nearby nodes of different nodes and the distance between different nearby nodes are determined by the distance between different nodes and different adjacent nodes. The aggregation degree of different nodes is determined by combining the average distance between the node and different adjacent nodes.

[0013] The number of nodes of different node types is determined based on different clustering degree evaluation values, and the adjustment parameters of the clustering algorithm are determined based on the number of nodes of different node types and the clustering degree evaluation values.

[0014] A further technical solution is that when the distance between the adjacent node and the node is within a preset distance, the adjacent node is determined to be a close-range node.

[0015] A further technical solution involves re-clustering the scatter plot based on the type of connection lines between different adjacent cluster groups, specifically including:

[0016] When the type of the connection line between adjacent cluster groups is a refined connection line, the adjacent cluster groups are classified into the same cluster group.

[0017] When the connecting line between adjacent clusters is a refined dividing line, the adjacent clusters are no longer further divided.

[0018] The beneficial effects of this invention are as follows:

[0019] 1. High interaction efficiency, applicable to data of millions: The proposed method only involves a large amount of computation in VP-tree construction and MST construction, with a time complexity of O(nlogn). However, these can be completed in the pre-computation stage, independent of the interaction process, and only need to be calculated once. The four strokes involved in the stroke interaction process are all local adjustments, and the time complexities of the nearest neighbor query and MST tree splitting and merging are O(logn), O(n), and O(1), respectively. Therefore, the interaction is completed in near real-time, even when facing data points of millions.

[0020] 2. It combines ease of use and reliability. The four strokes proposed in the method have clear and direct purposes, which align with the user's basic intentions during cluster optimization. Figure 1 Therefore, it is easy to understand and master, even for ordinary users. Secondly, because all four strokes only require outlining the target area without precisely drawing the closed boundary, it is simpler and more convenient to operate than the direct allocation method. Finally, the interaction results of the four strokes are predictable, which is more advantageous than the human-in-the-loop method in human-computer collaboration. Figure 1 It is more reliable in terms of consistency and is suitable for non-professional users.

[0021] Other features and advantages will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0023] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0024] Figure 1 This is a flowchart of a multimodal knowledge graph integration method.

[0025] Figure 2 This is a schematic diagram of a guiding curve.

[0026] Figure 3 This is a schematic diagram of a refined curve.

[0027] Figure 4 This is a flowchart illustrating the method for determining the adjustment parameters of a clustering algorithm.

[0028] Figure 5 This is an example diagram illustrating the use of the method of this invention to perform cluster refinement and cluster guidance. Detailed Implementation

[0029] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the embodiments set forth herein; rather, they are provided so that the invention will be thorough and complete, and the concept of the exemplary embodiments will be fully conveyed to those skilled in the art. The same reference numerals in the drawings denote the same or similar structures, and therefore their detailed description will be omitted.

[0030] The terms “a,” “one,” “the,” and “the” are used to indicate the existence of one or more elements / components / etc.; the terms “include” and “have” are used to indicate an open-ended meaning of inclusion and that there may be other elements / components / etc. in addition to the listed elements / components / etc.

[0031] In the era of big data, visualization has become an important analytical tool, helping people better understand and interpret data. First, visualization presents data graphically, making complex information, such as relationships between data points, clusters, and outliers, intuitive and easy to discover. Second, visualization also facilitates exploratory data analysis; through interaction, users can freely manipulate and explore data, raise new questions, and find answers. Finally, visualization can help decision-makers better understand the current situation, evaluate the merits of different options, and support data-driven decision-making processes. Overall, visualization has irreplaceable value in data analysis, insight discovery, and decision-making, providing an intuitive, interactive, and effective way to understand and utilize data.

[0032] In recent years, with the development of vectorization technology and nonlinear dimensionality reduction algorithms, vectorizing data into dense high-dimensional vectors and then reducing the dimensionality to a two-dimensional plane to form a scatter plot has become one of the standard technical paths for exploring large-scale data. Leveraging the excellent visual expressiveness of scatter plots, this approach can help us quickly peek into patterns in the data, such as clusters and outliers, and is applicable to almost all data types, including multi-attribute data, text, audio, and video data. Typically, after dimensionality reduction, clustering is performed, and data points are colored according to the clustering results to provide category information and enhance the visual distinction of clusters. Ideally, the clustering results, dimensionality reduction results, and user perception should be aligned; that is, visually significant clusters represent the categories obtained from clustering and conform to people's perception of the data. However, this is often difficult to achieve in practice. First, dimensionality reduction inevitably leads to information loss, potentially resulting in "false tearing" of the dimensionality reduction results, where the same type of data is split into several visual clusters, or even clusters that are far apart from each other. Second, dimensionality reduction and clustering are usually two independent processes, and they offer different perspectives on data insight. Finally, different people may have different data classification benchmarks, which may also differ from the algorithms used.

[0033] To bridge the differences among the three, researchers developed several interactive clustering optimization methods, mainly including two categories: direct assignment and human-in-the-loop method.

[0034] Direct assignment refers to allowing users to select a subset of data points and reassign categories after clustering using a lasso or similar tool. This method is simple and straightforward, with the decision-making power entirely resting with the user. However, the edges of clusters are often rugged and closely spaced, making it difficult to select the complete edges of clusters. This requires not only meticulous operation but also real-time user decision-making. This method lacks automation and is more suitable for simple scenarios where cluster separation is obvious in the dimensionality reduction results.

[0035] The human-in-the-loop approach refers to a method that aligns human, machine, and data through a human-in-the-loop concept. The process involves humans interacting with the machine to convey their cognitive intentions. The machine then modifies model parameters and methods, updating clustering and dimensionality reduction results. Humans then review the results and interact further, repeating this process multiple times. This method fully integrates human and machine intelligence, with decision-making power shared between them. However, this approach has a high learning cost, and the direct manipulation of data through selection and labeling—translating user intentions into model parameter modifications—can lead to ambiguity and misunderstanding, causing convergence failures across multiple iterations. Furthermore, this method alters the layout results, which is unacceptable in certain scenarios.

[0036] To overcome the shortcomings of previous methods, this invention develops a stroke-based clustering optimization method deeply integrated with the MST (Minimum Spanning Tree) clustering algorithm. Optimization is divided into two categories: guidance and clustering, corresponding to the stages before and after automatic clustering, respectively. Specifically, cluster guidance allows users to explicitly express their knowledge through strokes, guiding the clustering algorithm to provide initial clustering results; cluster refinement allows users to pick clusters through strokes and refine the clustering results by merging and splitting clusters. Unlike direct assignment methods, this invention's method does not require selecting complete cluster edges, but only needs to simply outline their distribution area, making it more efficient and convenient. Unlike human-in-the-loop methods, this invention's method does not change the scatter distribution obtained from dimensionality reduction, has an intuitive and simple interaction, low learning cost, and produces predictable deterministic results.

[0037] Example 1

[0038] To solve the above technical problems, such as Figure 1 As shown, this invention provides a clustering guidance and refinement method based on stroke interaction, characterized in that it specifically includes:

[0039] Obtain the number of nodes in the scatter plot and determine the adjustment parameters of the clustering algorithm based on the distance between different nodes;

[0040] Using the adjustment parameters as the adjustment threshold, cluster analysis is performed on the scatter plot based on the MST clustering algorithm to divide the nodes of the scatter plot into different cluster groups;

[0041] MST-based clustering

[0042] An ideal clustering algorithm for scatter plots should meet the following five requirements:

[0043] R1: Able to identify visually salient clusters

[0044] R2: Able to remove outliers that act as background noise.

[0045] R3: No need to specify the number of clusters in advance.

[0046] R4: Suitable for clusters of any shape, whether convex or concave.

[0047] R5: Has the potential to support interactive clustering refinement

[0048] This invention ultimately selected a clustering algorithm based on MST (Minimum Spanning Tree). R2, R3, and R4 excluded most classic algorithms, including KMeans (R2, R3, R4), spectral clustering (R2, R3), and hierarchical clustering (R2). Even with carefully tuned parameters, the Meanshift algorithm failed to provide satisfactory outlier identification results. DBSCAN effectively supports the first four requirements, but its parameters are not intuitive, and parameter tuning requires a time-consuming and laborious trial-and-error process. The MST-based clustering algorithm, in particular, satisfies all five requirements, especially R5, which inherently supports a stroke-based clustering tuning interaction.

[0049] The MST clustering algorithm has two core steps. The first step is to construct a global minimum spanning tree based on the Euclidean distance between data points. The second step is pruning, cutting edges longer than a threshold α, so that each independent subtree constitutes a cluster (R1). Increasing the threshold α tends to merge adjacent clusters; conversely, decreasing it will create more small clusters. Therefore, α controls the global clustering granularity. Smaller clusters can be filtered out using another parameter, size, which represents the cluster size (R2). Clearly, the MST clustering algorithm does not require specifying the number of clusters (R3) and is adaptable to clusters of various shapes (R4). Furthermore, it has two other significant advantages:

[0050] The interaction of adjusting the parameter α to generate a new set of clusters is real-time. This interaction only requires traversing all edges and comparing the edge length with α, with a time complexity of O(n), thus enabling real-time interaction even with millions of nodes. The first step of building the MST is relatively time-consuming, with a time complexity of O(nlogn), but it can be performed in the preprocessing step, independent of the pruning process, and only needs to be executed once.

[0051] The algorithm's two parameters—α and size—both have clear physical meanings. With the support of real-time interaction, users can quickly understand their impact on clustering results and master how to adjust them.

[0052] It is particularly important to emphasize that the essence of the MST clustering algorithm is dataset partitioning induced by reachability relationships. The following concepts are defined: if the distance from point a to point b is less than α, then a is said to be directly reachable from b; if there is a path connecting point a and point b by a direct reachability relationship, then a is said to be reachable from b; if there is a reachable path between point set A and point set B, then point set A is said to be reachable from point set B. These three relationships are symmetric. Therefore, the clusters obtained by the MST algorithm have the following properties: any pair of clusters is unreachable from each other, and any two points within each cluster are reachable from each other. This is quite similar to the mechanism by which the human visual system recognizes clusters, i.e., there are significant gaps between clusters, while data points within a cluster are adjacent to each other.

[0053] Automatic generation of different types of connection lines between different adjacent clusters is achieved by considering the number of discrete nodes between different adjacent clusters and the distance between different discrete nodes and the adjacent nodes of adjacent clusters.

[0054] Based on the automatically generated connection lines and the modification results of the connection lines obtained through manual interaction, the types of connection lines for different adjacent cluster groups are determined, and the scatter plot is re-clustered according to the types of connection lines for different adjacent cluster groups.

[0055] There are two strategies for aligning dimensionality reduction results, clustering results, and user cognition. One is a pre-processing strategy, which involves passing prior human knowledge of clustering to the automatic clustering algorithm before execution, thereby influencing the automatic clustering results. The other is a post-processing strategy, which involves manually splitting and merging clusters after obtaining the automatic clustering results to correct them. This invention develops a stroke-based interaction system based on MST, allowing users to guide the generation and refinement of clustering results by drawing simple curves, thus realizing both strategies.

[0056] like Figure 2 As shown, curves are divided into two types: splitting curves and connecting curves. The former splits a point set by creating artificial intervals between two sets to disrupt reachability; the latter merges points by introducing virtual points to actively construct reachable paths. The stroke interactions for cluster guidance and cluster refinement are identical in terms of the interaction actions, but the algorithms differ in their specific operations on the data points. All interaction implementations are deeply integrated with the MST clustering algorithm; each type of curve will be described in detail below.

[0057] Guided Connectivity: The goal is to guide the clustering algorithm to identify multiple visual clusters as the same cluster. The user only needs to draw a curve connecting them. First, the algorithm samples the curve at equal intervals with a density less than the parameter α. Then, the sampled points, i.e., the red points in the figure, participate as virtual points in the subsequent MST clustering algorithm. They form reachable paths between clusters, so connected clusters are identified as the same cluster. Finally, the sampled points are removed from the subtree (cluster).

[0058] Guided dividing lines: The goal is to prevent clustering algorithms from classifying a set of points into the same cluster; the user only needs to draw a dividing line. First, the algorithm identifies data points on both sides of the curve whose distance from the curve is less than or equal to α / 2 (the points outlined in black in the graph). These points are then excluded from the subsequent MST clustering algorithm. The two sets of points remaining on both sides of the curve are identified as two separate clusters because the distance between them is greater than α and there is no reachable path between them. Finally, the excluded points are assigned to clusters based on the proximity principle.

[0059] like Figure 3 As shown, the refined connector algorithm aims to merge multiple identified clusters into a single cluster, requiring only one curve to connect them. First, the algorithm identifies data points whose distance from the curve is less than α. Then, it merges the subtrees containing these points, resulting in the merged cluster.

[0060] Refined dividing line: The goal is to split an identified cluster into two clusters, requiring only one dividing line to be drawn. First, the algorithm identifies points on both sides of the curve whose distance from the curve is less than α / 2 (the points outlined in black in the graph) and removes them from their respective subtrees. Next, the MST clustering algorithm is re-executed on this cluster. Because the two sets of remaining data points on both sides of the curve are unreachable, they are identified as two separate clusters. Finally, the deleted nodes are added to their corresponding subtrees (clusters) based on proximity.

[0061] Whether in the guided or refined phase, users can perform a clustering identification after drawing multiple dividing and connecting lines consecutively. The influence of each curve is considered sequentially, which helps users fully express prior knowledge and significantly improves the efficiency of clustering adjustments. For guided interactions, the process ends after clicking the confirmation button for automatic clustering; for refined interactions, each interaction is accompanied by an update of the tree structure, thus supporting cumulative corrections under continuous interaction. All interactions can be completed in near real-time, even with millions of data points. The main computations involved in the interaction phase are finding points within a given range from the curve and merging and splitting the minimum spanning tree. The former can be transformed into a problem of querying points within a given distance from a point in the scatter plot by discretizing the curve. Based on the pre-built VP-tree data structure, the time complexity of this query is O(logn), where n is the number of data points. Constructing the VP-tree structure and performing MST clustering are independent of the interaction process, can be completed in a pre-computation step, and only need to be executed once. For the latter, splitting the minimum spanning tree requires traversing all edges to identify the longest connected subgraph, so the time complexity is O(n). Merging the minimum spanning tree only requires a set merging operation, so the time complexity is O(1).

[0062] The stroke-based interactive method proposed in this invention fully integrates the advantages of direct assignment and human-in-the-loop methods while overcoming their disadvantages. Based on the fundamental data structure provided by MST clustering, stroke-based interaction supports two stages: pre-clustering guidance and post-clustering refinement. Stroke-based interaction effectively combines machine intelligence and user prior knowledge, and utilizes the inherent distribution structure of the data. On the one hand, users only need to draw rough strokes to optimize clustering, avoiding the time-consuming and laborious process of carefully outlining target data points required by direct assignment. On the other hand, the positions of the scattered points remain unchanged before and after adjustment, and the adjustment results are predictable, avoiding the high uncertainty and low interpretability of the human-in-the-loop method.

[0063] Specifically, the distance between the different nodes is determined according to the Euclidean distance function.

[0064] Specifically, such as Figure 2 As shown, the method for determining the adjustment parameters of the clustering algorithm is as follows:

[0065] The number of nearby nodes of different nodes and the distance between different nearby nodes are determined by the distance between different nodes and different adjacent nodes. The aggregation degree of different nodes is determined by combining the average distance between the node and different adjacent nodes.

[0066] The number of nodes of different node types is determined based on different clustering degree evaluation values, and the adjustment parameters of the clustering algorithm are determined based on the number of nodes of different node types and the clustering degree evaluation values.

[0067] Furthermore, when the distance between the adjacent node and the node is within a preset distance, the adjacent node is determined to be a close-range node.

[0068] In another embodiment, the method for determining the adjustment parameters of the clustering algorithm is as follows:

[0069] Obtain the number of nodes, determine whether the number of nodes is greater than the preset number of nodes; if so, determine the adjustment parameters of the clustering algorithm through the preset parameters; if not, proceed to the next step.

[0070] The number of nearby nodes of different nodes is determined by the distance between different nodes and different adjacent nodes. The number of filtering nodes is determined based on the number of nearby nodes of a node. It is then determined whether the number of filtering nodes is greater than the preset number of filtering nodes. If so, the adjustment parameters of the clustering algorithm are determined by the preset parameters. If not, proceed to the next step.

[0071] The number of nearby nodes of different nodes and the distance between different nearby nodes are determined by the distance between different nodes and different adjacent nodes. The clustering degree evaluation value of different nodes is determined by combining the average distance between the node and different adjacent nodes. It is then determined whether the number of nodes whose clustering degree evaluation value meets the requirements is greater than the preset node number limit. If so, the adjustment parameters of the clustering algorithm are determined by the preset parameters. If not, proceed to the next step.

[0072] The number of nodes of different node types is determined based on different clustering degree evaluation values, and the adjustment parameters of the clustering algorithm are determined based on the number of nodes of different node types and the clustering degree evaluation values.

[0073] In another embodiment, the method for determining the adjustment parameters of the clustering algorithm is as follows:

[0074] The number of nearby nodes of different nodes and the distance between different nearby nodes are determined by the distance between different nodes and different adjacent nodes. The aggregation degree of different nodes is determined by combining the average distance between the node and different adjacent nodes.

[0075] Determine whether the sum of the clustering degree evaluation values ​​of all nodes meets the requirements. If yes, proceed to the next step; otherwise, determine the adjustment parameters of the clustering algorithm through preset parameters.

[0076] The nodes are divided into clustered nodes and scattered nodes according to different clustering degree evaluation values. It is determined whether the number of clustered nodes meets the requirements. If yes, proceed to the next step. If not, the adjustment parameters of the clustering algorithm are determined by preset parameters.

[0077] The comprehensive clustering evaluation value of the clustering node is determined by the number of clustering nodes and the clustering evaluation value of different clustering nodes. It is then determined whether the comprehensive clustering evaluation value of the clustering node meets the requirements. If yes, proceed to the next step; otherwise, the adjustment parameters of the clustering algorithm are determined by the preset parameters.

[0078] The overall clustering degree of the dispersed nodes is determined by the number of dispersed nodes and the clustering degree evaluation of different dispersed nodes, and the adjustment parameters of the clustering algorithm are determined by combining the overall clustering degree evaluation of the clustered nodes.

[0079] Furthermore, the nodes of the scatter plot are divided into different cluster groups, specifically including:

[0080] Construct a global minimum spanning tree based on the Euclidean distance between nodes;

[0081] Edges whose length in the global minimum spanning tree is greater than the adjustment parameter are pruned to form independent subtrees, and the nodes of the scatter plot are divided into different cluster groups based on the independent subtrees.

[0082] Specifically, such as Figure 4 As shown, the automatic generation of different types of connecting cables is performed, specifically including:

[0083] Obtain discrete nodes between adjacent cluster groups, and take the nodes of adjacent cluster groups whose distance from the discrete node is less than a preset distance as the adjacent group nodes of the discrete node;

[0084] The correlation coefficient between a discrete node and its neighboring clusters is determined by the number of adjacent cluster nodes in different adjacent clusters and the distance between the discrete node and the adjacent cluster nodes in different adjacent clusters. Based on the correlation coefficient, the discrete node is divided into associated nodes and unassociated nodes.

[0085] The group association coefficient between adjacent clusters is determined by the number of associated nodes between adjacent clusters and the association coefficient of different associated nodes, the number of unassociated nodes and the association coefficient of different unassociated nodes, and the matching type connection line between adjacent clusters is automatically generated based on the group association coefficient.

[0086] Furthermore, the types of connecting lines include guided connecting lines and guided dividing lines.

[0087] In another embodiment, the automatic generation of different types of connecting lines specifically includes:

[0088] Obtain discrete nodes between adjacent cluster groups. Nodes of adjacent cluster groups whose distance to the discrete node is less than a preset distance are taken as adjacent group nodes of the discrete node. Determine whether the number of adjacent group nodes in different cluster groups meets the requirements. If yes, proceed to the next step. If no, automatically generate connection lines between the adjacent cluster groups based on guided dividing lines.

[0089] The node correlation between different cluster groups and the discrete nodes is determined based on the distance between adjacent group nodes and different discrete nodes in different cluster groups and the number of adjacent group nodes. It is then determined whether there are cluster groups whose node correlation does not meet the requirements. If so, the connection lines between the adjacent cluster groups are automatically generated based on the guided dividing line. If not, proceed to the next step.

[0090] The correlation coefficient between the discrete node and the adjacent clusters is determined by the number of adjacent group nodes in different adjacent clusters and the distance between the discrete node and the adjacent group nodes in different adjacent clusters. It is then determined whether the sum of the correlation coefficients of the discrete node meets the requirements. If yes, proceed to the next step; otherwise, automatically generate the connection line between the adjacent clusters based on the guided dividing line.

[0091] The discrete nodes are divided into associated nodes and unassociated nodes based on the correlation coefficient. The group correlation coefficient between adjacent clusters is determined by the number of associated nodes between adjacent clusters and the correlation coefficient of different associated nodes, the number of unassociated nodes and the correlation coefficient of different unassociated nodes, and the automatic generation of matching connection lines between adjacent clusters based on the group correlation coefficient.

[0092] In another embodiment, the automatic generation of different types of connecting lines specifically includes:

[0093] Obtain discrete nodes between adjacent cluster groups, determine whether the number of discrete nodes between adjacent cluster groups meets the requirements. If yes, proceed to the next step; otherwise, automatically generate connection lines between adjacent cluster groups based on guided dividing lines.

[0094] The nodes of the adjacent clusters that are less than a preset distance from the discrete node are taken as the adjacent group nodes of the discrete node. It is determined whether there are discrete nodes whose number of adjacent group nodes meets the requirements. If yes, proceed to the next step. If no, the connection line between the adjacent clusters is automatically generated based on the guided dividing line.

[0095] The correlation coefficient between a discrete node and its neighboring clusters is determined by the number of adjacent cluster nodes in different adjacent clusters and the distance between the discrete node and its neighboring cluster nodes in different adjacent clusters. Based on the correlation coefficient, the discrete node is divided into associated nodes and unassociated nodes. It is then determined whether the number of associated nodes meets the requirements. If yes, proceed to the next step; otherwise, automatically generate the connection line between the adjacent clusters based on the guided dividing line.

[0096] The group association coefficient between adjacent clusters is determined by the number of associated nodes between adjacent clusters and the association coefficient of different associated nodes, the number of unassociated nodes and the association coefficient of different unassociated nodes, and the matching type connection line between adjacent clusters is automatically generated based on the group association coefficient.

[0097] It should be noted that the scatter plot is re-clustered based on the type of connection lines between different adjacent cluster groups, specifically including:

[0098] When the type of the connection line between adjacent cluster groups is a refined connection line, the adjacent cluster groups are classified into the same cluster group.

[0099] When the connecting line between adjacent clusters is a refined dividing line, the adjacent clusters are no longer further divided.

[0100] Figure 5 Example diagrams illustrating cluster refinement and cluster guidance using the method of this invention are shown. In both examples, users can merge and split clusters according to their preferences with only a few strokes, thereby aligning dimensionality reduction results, clustering results, and user perception.

[0101] Based on the above embodiments, the following technical effects are expected to be achieved:

[0102] 1. High interaction efficiency, applicable to data of millions: The proposed method only involves a large amount of computation in VP-tree construction and MST construction, with a time complexity of O(nlogn). However, these can be completed in the pre-computation stage, independent of the interaction process, and only need to be calculated once. The four strokes involved in the stroke interaction process are all local adjustments, and the time complexities of the nearest neighbor query and MST tree splitting and merging are O(logn), O(n), and O(1), respectively. Therefore, the interaction is completed in near real-time, even when facing data points of millions.

[0103] 2. It combines ease of use and reliability. The four strokes proposed in the method have clear and direct purposes, which align with the user's basic intentions during cluster optimization. Figure 1 Therefore, it is easy to understand and master, even for ordinary users. Secondly, because all four strokes only require outlining the target area without precisely drawing the closed boundary, it is simpler and more convenient to operate than the direct allocation method. Finally, the interaction results of the four strokes are predictable, which is more advantageous than the human-in-the-loop method in human-computer collaboration. Figure 1 It is more reliable in terms of consistency and is suitable for non-professional users.

[0104] In the description of this specification, the terms "an embodiment," "a preferred embodiment," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0105] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. For those skilled in the art, the embodiments of the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the embodiments of the present invention should be included within the protection scope of the embodiments of the present invention.

Claims

1. A clustering guidance and refinement method based on stroke interaction, characterized in that, Specifically, it includes: The number of nodes in the scatter plot is obtained, and the adjustment parameters of the clustering algorithm are determined by combining the distance between different nodes; the data types processed by the scatter plot include text, audio, and video data; Using the adjustment parameters as the adjustment threshold, cluster analysis is performed on the scatter plot based on the MST clustering algorithm to divide the nodes of the scatter plot into different cluster groups; The number of discrete nodes between different adjacent cluster groups is obtained, and further, based on the distance between the discrete nodes and the adjacent nodes of the adjacent cluster groups, different types of connection lines are automatically generated between the different adjacent cluster groups; wherein, the types of connection lines include guided connection lines and guided splitting lines; specifically: The nodes of the adjacent clusters that are less than a preset distance from the discrete node are taken as the adjacent group nodes of the discrete node. It is determined whether the number of adjacent group nodes in different clusters meets the requirements. If yes, proceed to the next step. If no, automatically generate the connection line between the adjacent clusters based on the guided dividing line. The node correlation between different cluster groups and the discrete nodes is determined based on the distance between adjacent group nodes and different discrete nodes in different cluster groups and the number of adjacent group nodes. It is then determined whether there are cluster groups whose node correlation does not meet the requirements. If so, the connection lines between the adjacent cluster groups are automatically generated based on the guided dividing line. If not, proceed to the next step. The correlation coefficient between the discrete node and the adjacent clusters is determined by the number of adjacent group nodes in different adjacent clusters and the distance between the discrete node and the adjacent group nodes in different adjacent clusters. It is then determined whether the sum of the correlation coefficients of the discrete node meets the requirements. If yes, proceed to the next step; otherwise, automatically generate the connection line between the adjacent clusters based on the guided dividing line. The discrete nodes are divided into associated nodes and unassociated nodes according to the correlation coefficient. The group correlation coefficient between adjacent clusters is determined by the number of associated nodes between adjacent clusters and the correlation coefficient of different associated nodes, the number of unassociated nodes and the correlation coefficient of different unassociated nodes, and guided connection lines between adjacent clusters are automatically generated based on the group correlation coefficient. Based on the automatically generated connectors and the modification results of the connectors obtained through manual interaction, the types of connectors for different adjacent cluster groups are determined, and the scatter plot is re-clustered according to the types of connectors for different adjacent cluster groups. Specifically, this includes: When the type of the connection line between adjacent cluster groups is a guided connection line, the adjacent cluster groups are classified into the same cluster group. When the connecting line between adjacent clusters is a guided dividing line, the adjacent clusters will not be further divided.

2. The clustering guidance and refinement method based on stroke interaction as described in claim 1, characterized in that, The distance between the different nodes is determined according to the Euclidean distance function.

3. The clustering guidance and refinement method based on stroke interaction as described in claim 1, characterized in that, The method for determining the adjustment parameters of the clustering algorithm is as follows: The number of nearby nodes of different nodes and the distance between different nearby nodes are determined by the distance between different nodes and different adjacent nodes. The aggregation degree of different nodes is determined by combining the average distance between the node and different adjacent nodes. The number of nodes of different node types is determined based on different clustering degree evaluation values, and the adjustment parameters of the clustering algorithm are determined based on the number of nodes of different node types and the clustering degree evaluation values.

4. The clustering guidance and refinement method based on stroke interaction as described in claim 3, characterized in that, When the distance between the adjacent node and the node is within a preset distance, the adjacent node is determined to be a close-range node.

5. The clustering guidance and refinement method based on stroke interaction as described in claim 1, characterized in that, The nodes of the scatter plot are divided into different cluster groups, specifically including: Construct a global minimum spanning tree based on the Euclidean distance between nodes; Edges whose length in the global minimum spanning tree is greater than the adjustment parameter are pruned to form independent subtrees, and the nodes of the scatter plot are divided into different cluster groups based on the independent subtrees.

Images