A geospatial point data sampling method driven by sentiment feature consistency

Abstract

The present invention discloses a method for sampling geospatial point data driven by consistency of sentiment features. The method of the present invention first uses the sentiment analysis model VADER to perform sentiment analysis on the social media dataset to obtain the sentiment value expressed by the text information; then all data items are clustered using the agglomerative hierarchical method to extract the sentiment distribution of the local area, first pre-clustering is performed to obtain pre-clusters, and then bottom-up clustering is performed to obtain a clustering tree; all pre-clustered data items are sampled to maintain the global sentiment distribution and the local sentiment distribution; the sampling node method with the lowest backtracking average contribution value is used to retain the global sentiment distribution and the spatial distribution as much as possible. The present invention extracts the sentiment distribution in multi-scale areas based on the sentiment attributes and the geographical location of the data items, helping users to perform geospatial sentiment analysis on large-scale social media data with geo-tags, and is effective and practical in simplifying and exploring large social media data.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology and relates to a geospatial point data sampling method driven by sentiment feature consistency. Background Art

[0002] With the development of social media, sentiment analysis has become an increasingly valuable topic in many fields, particularly when combined with geographic visualization. In recent years, numerous methods have been developed to extract user opinions from social media, largely supporting the visual analysis of sentiment features (Aspect and Entity Extraction for Opinion Mining, Lei Zhang and Bing Liu 2014 / 1). For example, the sentiment of tweets is shaded in different colors on a map to facilitate an intuitive understanding of social reactions to large-scale events (Whisper: Tracing the Spatiotemporal Process of Information Diffusion in Real Time, Nan Cao, Yu Ru, Lin Xiaohua, Sun Lazer, Shixia Liu, and Huamin Qu 2012). Furthermore, statistics on visual variables indicate that color is the most common visual channel for conveying emotion. Therefore, using map marker colors to indicate location and encode text polarity on geographic maps can help users effectively analyze public opinion in different regions around a specific topic. However, as the scale of social media data continues to expand, overlapping visual elements of different colors interfere with users' visual perception of sentiment distribution, hindering their visual understanding of the sentiment distribution and semantics of large-scale geo-annotated social media data.

[0003] There are two main approaches to addressing this problem: filtering and sampling. Filtering methods filter data items based on their attribute characteristics (A Visual Backchannel for Large-Scale Events, Marian Dör and Daniel Gruen, 2010 / 11). However, the attributes of data items do not always correspond to their geographic distribution, so attribute-based filtering methods cannot effectively preserve the spatial distribution of simplified data items. In contrast, sampling methods can reduce visual clutter while preserving the spatial distribution.

[0004] Existing sampling techniques, such as blue noise sampling (Visual Abstraction and Exploration of Multi-class Scatterplots, Haidong Chen / Wei Chen, 2014 / 11), reduce the visual clutter of multi-class scatterplots while maintaining relative data density and class density. This method effectively preserves spatial distribution, but it can easily lose attribute characteristics in sampled visualizations. Both spatial distribution and sentiment characteristics are important for exploring large-scale geotagged social media data, posing significant challenges for the design of visual abstraction methods. Regarding sentiment analysis, researchers have proposed a model that identifies sentiment categories and categorizes them as positive, negative, and neutral. To preserve sentiment characteristics in sampled geovisualizations, it is crucial to preserve their local distribution. Furthermore, the ordering properties of sampled social media data items should be consistent with those of the original data. In the database field, a rapid sampling model (Rapid Sampling for Visualizations with Ordering Guarantees, Albert Kim / Eric Blais, 2015 / 1) proposes a method for quickly generating approximate visualizations while preserving the ordering properties. Therefore, it is an effective method for reducing dataset size while preserving the distribution of attribute values across different classes.

[0005] However, fast sampling still has the following limitations: (1) Spatial information is not considered during the sampling process, making it difficult to maintain the spatial distribution of the original data and the sampled data items. (2) The sampling rate is unstable and cannot be controlled by the user according to their own requirements, because it is difficult to obtain sampling results with consistent sorting properties in the following two situations: one is when the number of categories is large, and the other is when the three values of the attributes of the categories are similar, which makes it difficult to obtain the correct category sorting. In this work, we aim to reduce the size of large-scale social media data while preserving the local sentiment distribution and the global spatial distribution as much as possible. First, the social media data is divided into several initial clusters based on the geographical location and sentiment attributes of social media. Subsequently, starting from a large number of initial clusters, a bottom-up hierarchical clustering is iteratively merged to merge multiple clusters with similar geographical locations, and multi-tree encoding is used. Then, fast sampling is used to sample the data items of the nodes to maintain the local sentiment distribution. However, due to the randomness of the samples, the result may not effectively preserve the sentiment distribution between nodes. Summary of the Invention

[0006] The purpose of the present invention is to provide a geospatial point data sampling method driven by sentiment feature consistency.

[0007] The technical solution adopted by the present invention is:

[0008] Step (1) obtains an open source social media dataset; the attributes of each data item in the social media dataset include text information, user ID, time, and geographic coordinates.

[0009] Step (2) After stemming the text information in the social media dataset, sentiment analysis is performed using the sentiment analysis model VADER to obtain the sentiment value expressed by the text information.

[0010] Step (3) cluster all data items using the agglomerative hierarchical method to extract the sentiment distribution of the local area;

[0011] The agglomerative hierarchical method first performs pre-clustering to obtain pre-clusters, and then performs bottom-up clustering to obtain a cluster tree;

[0012] In the pre-clustering stage, data items are reasonably clustered according to kernel density and sentiment value. The specific method is:

[0013] Randomly select a data item , calculate the radius of its Poisson disk ,in is the result of kernel density estimation (KDE) calculation based on geographic coordinates, To set parameters; to use data items The geographical coordinates are the center of the circle, The data items within the radius are taken as an initial pre-cluster;

[0014] The data items in the initial pre-cluster are screened according to the sentiment value to obtain the pre-cluster;

[0015] For the data items in the initial pre-cluster Sentiment value ,like , then the data item With data items Belong to a pre-cluster, is the set emotion value threshold;

[0016] Repeat the above process until all data items belong to different pre-clusters, and each data item belongs to only one pre-cluster, and get A pre-cluster.

[0017] In the bottom-up clustering phase, we first calculate the similarity between two pre-clusters based on their geographic coordinates: With pre-cluster Average similarity , , ;in and Pre-cluster and pre-clustering The number of data items included, and Separate data items and data items The geographical coordinates of Represents the 2 norm; if the average similarity is greater than or equal to the set similarity threshold The two pre-clusters are assigned to a first-level node if the average similarity between a pre-cluster and any other pre-clusters is less than the similarity threshold , then the pre-cluster itself is used as the first-level node; using the same method, continue to cluster the first-level nodes upward to obtain the second-level nodes, and so on, to obtain a root node;

[0018] Extract the sentiment distribution of nodes at all levels and pre-clusters contained in the root node: calculate the average sentiment value of all data items in each pre-cluster , , the sentiment distribution vectors of nodes at all levels and the root node are obtained. The element of each dimension in the vector is the average sentiment value of its corresponding pre-cluster.

[0019] Step (4) sampling all pre-clustered data items to maintain the global sentiment distribution and the local sentiment distribution;

[0020] First, a data item is randomly selected from each pre-cluster as the initial sample, and the first round of sampling is performed to obtain the confidence interval of the sentiment value of each initial sample;

[0021] If the confidence intervals of one or more preclusters do not overlap with the confidence intervals of all other preclusters, then stop the subsequent sampling of these preclusters, use the initial samples as the sampling samples of these preclusters, and perform a second round of sampling on the remaining preclusters. If the confidence intervals of one or more preclusters do not overlap with the confidence intervals of all other preclusters, then stop the subsequent sampling of these preclusters, use the initial samples and the newly added samples as the sampling samples of these preclusters, and perform the next round of sampling on the remaining preclusters until the confidence intervals of all preclusters do not overlap, or there are no additional samples in the remaining preclusters, then stop sampling and obtain the sampling samples of all preclusters.

[0022] Thus, the sentiment distribution vector of the pre-cluster after sampling is obtained, that is, the sentiment distribution vector of the first-level node. The element of each dimension in the vector is the average sentiment value of the sampling samples of its corresponding pre-cluster.

[0023] Step (5) calculates the contribution of the sentiment distribution vector of each first-level node to the global sentiment distribution;

[0024] ; Indicates the Level 1 nodes The contribution value of the sentiment distribution vector to the global sentiment distribution, , is the number of first-level nodes, , ; and Pre-cluster and pre-clustering The average sentiment value of the sample; is a judgment function, but ,otherwise ; is another judgment function, but , but ;

[0025] If the sampling rate Less than the set sampling rate threshold , is the total number of data items in the social media dataset, is the number of sampled data items, then all pre-clusters contained in the second-level nodes corresponding to the first-level node with the lowest contribution value are resampled until the set sampling rate threshold is reached. .

[0026] Furthermore, the confidence interval obtained in the first round of sampling in step (4) is centered on the sentiment value of the initial sample and the interval lengths are extended up and down respectively. , the confidence probability set ;

[0027] The confidence interval obtained after the second round of sampling is centered on the average sentiment value and the length of the interval is extended up and down respectively. , Indicates the number of original data items contained in the pre-cluster, Indicates the sampling round.

[0028] Existing simplified geographic visualization technologies ignore the spatial distribution of sentiment characteristics during the sampling process, which can easily lead to misunderstandings of emotions in different places. The present invention designs an optimized hierarchical clustering method to extract the sentiment distribution in multi-scale regions based on sentiment attributes and the geographic location of data items; and adopts a sampling node method with the lowest back-test average contribution value to retain the global sentiment distribution and spatial distribution as much as possible; through some novel interactions such as the sentiment wheel, it is designed to help users conduct in-depth geospatial sentiment analysis on large-scale geotagged social media data. The present invention is used for the simplification and exploration of large-scale geotagged social media data, aiming to retain the sentiment characteristics and spatial distribution of sampled social media data items, and is effective and practical in simplifying and exploring large-scale geotagged social media data. DETAILED DESCRIPTION

[0029] A geospatial point data sampling method driven by sentiment feature consistency is as follows:

[0030] Step (1) obtains an open source social media dataset; the attributes of each data item in the social media dataset include text information, user ID, time, and geographic coordinates.

[0031] This example collects social media datasets from Yelp and Twitter. The Yelp data includes commentary information on lifestyle topics such as food and entertainment published by Phoenix between 2005 and 2015.

[0032] Step (2) After stemming the text information in the social media dataset, sentiment analysis is performed using the sentiment analysis model VADER to obtain the sentiment value expressed by the text information. Sentiment values are used to analyze and determine the sentiment polarity (such as positive, negative, neutral) and sentiment intensity of the text. The sentiment analysis model VADER is a rule-based sentiment analyzer that can handle some common sentiment intensity markers (such as "!" and "?") as well as some non-standard expressions such as abbreviations, emphasis, and repeated letters.

[0033] The sentiment analysis model VADER performs sentiment analysis by tokenizing text, breaking it down into words or phrases, and removing stop words. For each word or phrase, a sentiment lexicon is used to calculate its sentiment score. The sentiment lexicon contains a variety of sentiment terms and their corresponding sentiment intensity scores. For each text, the sentiment scores for each word or phrase are aggregated to calculate a final sentiment score. VADER categorizes the sentiment scores into three categories: positive, negative, and neutral, and provides a composite score.

[0034] Step (3) clusters all data items using the agglomerative hierarchical method to extract the sentiment distribution of local areas.

[0035] The agglomerative hierarchical method first performs pre-clustering to obtain pre-clusters, and then performs bottom-up clustering to obtain a cluster tree.

[0036] In the pre-clustering stage, data items are reasonably clustered according to kernel density and sentiment value. The specific method is:

[0037] Criterion 1: Randomly select a data item , calculate the radius of its Poisson disk ,in is the result of kernel density estimation (KDE) calculation based on geographic coordinates, To set parameters; to use data items The geographical coordinates are the center of the circle, The data items within the radius are taken as an initial pre-cluster. Compared with the definition of constant radius, the geographic spatial distribution of the pre-cluster generated by this method is more consistent with the original distribution of the data set;

[0038] Criterion 2: Filter the data items in the initial pre-cluster according to the sentiment value to obtain the pre-cluster;

[0039] For the data items in the initial pre-cluster Sentiment value ,like , then the data item With data items Belong to a pre-cluster, is the set sentiment value threshold; thereby explicitly limiting the difference in sentiment intensity between every two data items in each cluster;

[0040] Repeat the above steps until all data items belong to different pre-clusters, and each data item belongs to only one pre-cluster, and obtain A pre-cluster.

[0041] In the bottom-up clustering phase, we first calculate the similarity between two pre-clusters based on their geographic coordinates: With pre-cluster Average similarity , , ;in and Pre-cluster and pre-clustering The number of data items included, and Separate data items and data items The geographical coordinates of Represents the 2 norm; if the average similarity is greater than or equal to the set similarity threshold The two pre-clusters are assigned to a first-level node if the average similarity between a pre-cluster and any other pre-clusters is less than the similarity threshold , then the pre-cluster itself is used as the first-level node; using the same method, the first-level nodes are clustered upward to obtain the second-level nodes, and so on to obtain a root node.

[0042] Extract the sentiment distribution of nodes at all levels and pre-clusters contained in the root node: calculate the average sentiment value of all data items in each pre-cluster , , the sentiment distribution vectors of nodes at all levels and the root node are obtained. The element of each dimension in the vector is the average sentiment value of its corresponding pre-cluster.

[0043] Starting from a large number of initial pre-clusters, multiple clusters with geographical proximity are iteratively merged bottom-up. Given that each pre-cluster clearly expresses its sentiment polarity and intensity range, their average intensity is calculated to measure the average sentiment level. The average sentiment value of the pre-clusters contained in nodes at different levels is used to measure the sentiment distribution in the local area, enabling users to analyze multi-granular information such as local sentiment differences at different scales.

[0044] Step (4) samples the data items of all pre-clusters to maintain the global and local sentiment distributions. This process aims to reduce the size of the dataset while preserving the sorting properties of the average sentiment values between different pre-clusters. Correct sorting properties mean that the visualization or sorting between pre-clusters is correct, even if the actual values of the groups differ from the results of sampling the entire dataset.

[0045] First, we randomly select a data item from each pre-cluster as the initial sample, perform the first round of sampling, and obtain the confidence interval of the sentiment value of each initial sample. The confidence interval is centered on the sentiment value of the initial sample, and the interval lengths of the upper and lower extensions are , the confidence probability set , that is, the probability that the average sentiment value of a pre-cluster falls within the confidence interval of the sentiment value of this pre-cluster.

[0046] If the confidence intervals of one or more preclusters do not overlap with the confidence intervals of all other preclusters, stop sampling these preclusters, use the initial samples as the sampling samples of these preclusters, and perform a second round of sampling on the remaining preclusters. Specifically, for each of the remaining preclusters, arbitrarily select another data item as an additional sample, calculate the average sentiment value of the additional sample and the initial sample, and obtain the confidence interval of the sentiment value of each second round sample. The confidence interval is centered on the average sentiment value, and the interval lengths are extended up and down respectively. , Indicates the number of original data items contained in the pre-cluster. If the confidence intervals of one or more pre-clusters do not overlap with the confidence intervals of all other pre-clusters, then stop sampling these pre-clusters, use the initial samples and the newly added samples as the sampling samples of these pre-clusters, and perform the next round of sampling on the remaining pre-clusters until the confidence intervals of all pre-clusters do not overlap, or there are no additional samples in the remaining pre-clusters, then stop sampling, and obtain the sampling samples of all pre-clusters; the length of the upper and lower expansion intervals of each round of sampling is , Indicates the sampling round.

[0047] Thus, the sentiment distribution vector of the pre-cluster after sampling is obtained, that is, the sentiment distribution vector of the first-level node. The element of each dimension in the vector is the average sentiment value of the sampling samples of its corresponding pre-cluster.

[0048] Step (5) calculates the contribution of the sentiment distribution vector of each first-level node to the global sentiment distribution;

[0049] ; Indicates the Level 1 nodes The contribution value of the sentiment distribution vector to the global sentiment distribution, , is the number of first-level nodes, , ; and Pre-cluster and pre-clustering The average sentiment value of the sample; is a judgment function, but ,otherwise ; is another judgment function, but , but .

[0050] If the sampling rate Less than the set sampling rate threshold , is the total number of data items in the social media dataset, is the number of sampled data items, then all pre-clusters contained in the second-level nodes corresponding to the first-level node with the lowest contribution value are resampled until the set sampling rate threshold is reached. .

[0051] The method can be visualized through visualization tools to help users evaluate and optimize sampling results, such as sentiment trees and sentiment wheels. Hierarchical trees and histograms are combined into the sentiment tree to guide the sampling process and evaluate the sampling results. The sentiment wheel is used to support users in intuitively exploring and analyzing sentiment distribution across regions. A visual abstraction system integrating data modeling, visualization design, and interaction is implemented to simplify large-scale geotagged social media data. Among them, (a) the control panel displays basic information of the social media dataset, allowing users to select a dataset and set necessary parameters. (b) The word cloud view and (d) the text information view present sentiment keywords and detailed information in the area of interest. (c) The map view is designed to visualize the geographic spatial distribution and local sentiment distribution of the original and sampled datasets. (e) The sentiment tree view visualizes the stratification and sampling results of the sentiment distribution. (f) The ranking view allows users to deeply evaluate the maintenance of the sentiment distribution in local regions.

[0052] In the map view, each scatter point represents a data item, with different colors representing positive, negative, and neutral sentiment polarity. Furthermore, the location of the point is determined by the geotag attribute, and the density distribution of the scatter points is represented by a heat map. A set of interactive tools is integrated into the map view.

[0053] In the sentiment tree view, a traditional tree diagram is used to visualize the hierarchical clustering structure, in which all leaf nodes are replaced by bars of equal height. The x-coordinate of the root node of each subtree is placed in the center of the horizontal space required to layout the entire subtree. The y-coordinate depends on the depth of the root node of each subtree. In particular, some nodes are replaced by histograms to highlight the multi-scale sampling areas, and each of them represents the corresponding local sentiment distribution. The sampling rate of each leaf node is plotted in the form of a histogram in the corresponding bar chart to represent the global spatial distribution. Interactive links are supported to associate the sentiment tree with other views. For example, clicking on a node or histogram in the tree will display the sentiment distribution and geographic spatial distribution in the ranking and map views.

[0054] To facilitate user assessment of the preservation of local sentiment distribution, the pre- and post-sampling results are symmetrically displayed in a ranking view, where the average sentiment intensity of the pre-cluster is encoded using bars corresponding to the left and right numeric columns, while colors represent different sentiment polarities. Ranking is driven by the numeric columns, and sorting is based on the average. Ranking is compared by aligning the two columns horizontally, and lines in the slope graph connect pre-clusters with the same ranking, allowing users to assess changes in detail. The more horizontally aligned lines, the better the sentiment distribution is preserved. The ranking view is further linked to the map and sentiment tree views to support users in interactive geospatial sentiment analysis.

[0055] Semantic views, including word cloud and text information views, allow users to deeply understand and explore the semantic and sentiment information of areas of interest. The word cloud view displays the most frequent keywords extracted from the dataset. The scrollable text information view displays the original Twitter messages, including user ID, city, and text information. A text search function is also provided, allowing users to find the original text corresponding to words displayed in the word cloud.

[0056] In the control view, a control panel and a data overview are provided. The control panel supports interactive selection of data sets and parameter values. The user then clicks the "Run" button to start the analysis. In addition, two buttons are provided for comparing the results of other sampling methods, including random sampling and adaptive blue noise sampling. A legend of different colored dots is set to represent positive, negative, and neutral emotions. A list of statistical information can also be found in the data overview next to the map view to allow users to understand the data information more intuitively. The data overview is linked to the map and sentiment tree view through interaction. For example, after selecting a node in the sentiment tree view, the data overview will be updated with the content of the data items contained in that node.

[0057] A sentiment wheel was designed to display the sentiment distribution of a selected region before and after sampling, helping users focus on the local area of interest. The sentiment wheel consists of a circle surrounded by differently colored histograms. The histograms are arranged clockwise, with each histogram representing a precluster within that region, and its height reflecting the average sentiment value of the data items within that precluster. Lighter-colored histograms represent the average sentiment value after sampling. When interacting with the sentiment wheel, the corresponding leaf node in the sentiment tree view is highlighted, and the sorted view is updated to display the sentiment distribution of that region before and after sampling.

[0058] Users can also interactively select multiple regions and compare sentiment differences between them, allowing for deeper exploration of geospatial sentiment analysis within the map view. A sentiment scale axis is further designed to connect regions. Twenty bars are evenly distributed on either side of the axis, representing the data distribution, with bars belonging to the same category appearing symmetrically. Categories are derived by equally dividing the sentiment intensity range [−1, 1]. Bars with higher differences are highlighted, given that users may find it difficult to detect differences between regions due to small differences. Furthermore, when users select an area of interest, a text information view and document information appear, along with a word cloud view corresponding to the document data's topic keywords.

[0059] Considering the large number of pre-clusters, fully preserving the global sentiment distribution will be very difficult and will be easily affected by the huge sampling complexity. Therefore, we prioritize preserving the sentiment distribution of local regions and then optimize the selection of sampling regions according to the hierarchy to further preserve the global sentiment distribution and spatial distribution.

Claims

1. A geospatial point data sampling method driven by sentiment feature consistency, characterized by: Step (1) obtaining an open source social media dataset; the attributes of each data item in the social media dataset include text information, user ID, time, and geographic coordinates; Step (2) After extracting the stems of the text information in the social media dataset, sentiment analysis is performed using the sentiment analysis model VADER to obtain the sentiment value expressed by the text information; Step (3) cluster all data items using the agglomerative hierarchical method to extract the sentiment distribution of the local area; The agglomerative hierarchical method first performs pre-clustering to obtain pre-clusters, and then performs bottom-up clustering to obtain a cluster tree; In the pre-clustering stage, data items are reasonably clustered according to kernel density and sentiment value. The specific method is: Randomly select a data item , calculate the radius of its Poisson disk ,in is the kernel density estimation result based on geographic coordinates, To set parameters; to use data items The geographical coordinates are the center of the circle, The data items within the radius are taken as an initial pre-cluster; The data items in the initial pre-cluster are screened according to the sentiment value to obtain the pre-cluster; For the data items in the initial pre-cluster Sentiment value ,like , then the data item With data items Belong to a pre-cluster, is the set emotion value threshold; Repeat the above process until all data items belong to different pre-clusters, and each data item belongs to only one pre-cluster, and get pre-clusters; In the bottom-up clustering phase, we first calculate the similarity between two pre-clusters based on their geographic coordinates: With pre-cluster Average similarity , , ;in and Pre-cluster and pre-clustering The number of data items included, and Separate data items and data items The geographical coordinates of represents the 2-norm; If the average similarity is greater than or equal to the set similarity threshold The two pre-clusters are assigned to a first-level node if the average similarity between a pre-cluster and any other pre-clusters is less than the similarity threshold , then the pre-cluster itself is used as the first-level node; using the same method, continue to cluster the first-level nodes upward to obtain the second-level nodes, and so on, to obtain a root node; Extract the sentiment distribution of nodes at all levels and pre-clusters contained in the root node: calculate the average sentiment value of all data items in each pre-cluster , , get the sentiment distribution vector of each level of nodes and the root node. The element of each dimension in the vector is the average sentiment value of its corresponding pre-cluster; Step (4) sampling all pre-clustered data items to maintain the global sentiment distribution and the local sentiment distribution; First, a data item is randomly selected from each pre-cluster as the initial sample, and the first round of sampling is performed to obtain the confidence interval of the sentiment value of each initial sample; If the confidence intervals of one or more preclusters do not overlap with the confidence intervals of all other preclusters, then stop the subsequent sampling of these preclusters, use the initial samples as the sampling samples of these preclusters, and perform a second round of sampling on the remaining preclusters. If the confidence intervals of one or more preclusters do not overlap with the confidence intervals of all other preclusters, then stop the subsequent sampling of these preclusters, use the initial samples and the newly added samples as the sampling samples of these preclusters, and perform the next round of sampling on the remaining preclusters until the confidence intervals of all preclusters do not overlap, or there are no additional samples in the remaining preclusters, then stop sampling and obtain the sampling samples of all preclusters. Thus, the post-sampling pre-cluster sentiment distribution vector is obtained, that is, the sentiment distribution vector of the first-level node. The element of each dimension in the vector is the average sentiment value of the sampling samples of its corresponding pre-cluster; Step (5) calculates the contribution of the sentiment distribution vector of each first-level node to the global sentiment distribution; ; Indicates the Level 1 nodes The contribution value of the sentiment distribution vector to the global sentiment distribution, , is the number of first-level nodes, , ; and Pre-cluster and pre-clustering The average sentiment value of the sample; is a judgment function, but ,otherwise ; is another judgment function, but , but ; If the sampling rate Less than the set sampling rate threshold , is the total number of data items in the social media dataset, is the number of sampled data items, then all pre-clusters contained in the second-level nodes corresponding to the first-level node with the lowest contribution value are resampled until the set sampling rate threshold is reached. .

2. The sentiment feature consistency-driven geospatial point data sampling method according to claim 1, characterized in that: The confidence interval obtained from the first round of sampling in step (4) is centered on the sentiment value of the initial sample and the length of the interval extended up and down respectively , is the confidence probability set; The confidence interval obtained after the second round of sampling is centered on the average sentiment value and the length of the interval is extended up and down respectively. , Indicates the number of original data items contained in the pre-cluster, Indicates the sampling round.

3. The sentiment feature consistency-driven geospatial point data sampling method according to claim 2, characterized in that: The confidence probability .

Images