A geographic point data sampling method based on user task preference learning

Abstract

The present invention discloses a geographic point data sampling method based on user task preference learning. The method applies adaptive blue noise sampling to divide the geographic space into local areas, and designs a set of visualization interfaces to present the data characteristics of these points in the local areas, so that users can intuitively select representative points according to their needs. Then, a convolutional neural network is used to model the user's preferences to further guide the sampling process of other local areas. As a result, all sampling points will retain the spatial distribution of the original data points and meet the user's preferences as much as possible. In addition, the method implements a visualization framework that integrates functions such as manual point selection, CNN training, automatic point sampling, and visualization comparison. The method of the present invention implements a geographic point data sampling based on user task preference learning, allowing users to easily obtain and evaluate geographic sampling points that meet user preferences.

Description

Technical Field

[0001] The present invention belongs to the field of information technology and relates to a geographic point data sampling method based on user task preference learning. Background Art

[0002] With the rapid development of location acquisition technology, geographic point data is widely collected and has broad application value in a variety of fields, ranging from economics and epidemiology to public policy and security. Scatter plots are a common method for visualizing geographic point data, where coordinates represent the location of geographic points, while visual elements encode attribute information, allowing users to intuitively perceive the spatial distribution and attribute relationships of geographic points. However, as the scale of geographic point data increases, the problem of overuse also arises, which seriously interferes with visual perception and hinders the exploration and analysis of geographic space. To reduce the visual clutter of large-scale geographic point data visualization, many sampling strategies have been proposed. For example, Dix and Ellis used random sampling to simplify large-scale geographic point data to enhance the visual perception of scatter plots. Density-based sampling methods aim to maintain the spatial distribution of original geographic point data; Bertini et al. proposed a non-uniform sampling strategy to reduce the visual clutter of the scatter plot while maintaining relative regional density differences; Cheng et al. adopted a stratified sampling strategy based on hash graphs to maintain outliers of the original points; Chen et al. proposed a sampling technique based on recursive subdivision, based on maintaining relative density and outliers to meet the requirements of multi-class scatter plot exploration; Zhou et al. developed an attribute-based sampling model to simplify large-scale scatter plots while retaining the spatial density and spatial autocorrelation of geographic points.

[0003] These strategies focus on preserving the original data characteristics, such as randomness, spatial distribution, and association relationships. However, users may not be interested in these predefined characteristics, and the sampling results will not be conducive to geographic data exploration in specific application scenarios. This difference will make the sampled points unable to meet the user's requirements for practical problem analysis and play the expected role in subsequent geographic data exploration. Therefore, considering user preferences and expectations for large-scale geographic point data sampling is of great significance for generating sampled geographic points that meet personalized analysis requirements. For the visual sampling of geographic point data that preserves user preferences and needs, three technical problems need to be overcome: (1) how to specify general target features that can describe various user preferences; (2) how to define a sampling model that can generate a subset of points that meet user preferences while maintaining the spatial distribution of the original geographic points; (3) how to evaluate in real time whether the characteristics of the sampled points meet user preferences and further promote geographic data exploration. Summary of the Invention

[0004] The purpose of this invention is to provide a user-driven sampling model for large-scale geographic point data visualization based on convolutional neural networks.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] Step (1) uses an adaptive blue noise sampling model to divide the geographic space into local areas, and designs a set of visual model interfaces to display the data features of the local areas, calculates the information entropy value of the local areas based on the data features of the local areas, and recommends those areas with higher entropy values ​​for users to mark points of interest based on prior knowledge and requirements, and extracts common features to construct feature vectors;

[0007] Step (2) a CNN is trained to learn the user preference according to the selected representative feature vector, and further used to guide the sampling process of other local areas; the local area determined by blue noise sampling is subjected to a user-driven sampling scheme for all local areas through the trained CNN model to generate the desired sampling results; if the general evaluation index accuracy or the accuracy obtained by dynamic time warping is less than the set threshold, the model parameters are modified and the model is trained again until the result meets the user's requirements;

[0008] The CNN model uses a preference matrix composed of one-dimensional vectors as input to the first one-dimensional CNN layer for feature learning, and the second one-dimensional CNN layer performs further feature extraction. The maximum pooling layer is used to reduce the complexity of the output, and a dropout layer is further added to alleviate the overfitting problem caused by the small amount of training data, thereby improving the processing accuracy of unseen data. In order to further learn higher-level features, two more one-dimensional CNN layers are used, followed by the maximum pooling layer and the global average pooling layer to alleviate the overfitting problem. Finally, the fully connected layer, as the last part of the network structure, plays the role of a classifier. The loss function used in the CNN training process is defined as follows: Where: N is the amount of training data selected by the user; y i Represents the label of training data point i. If the label is positive, y i =1, if it is a negative number, then y i =0; represents the probability that data point i is predicted to be the positive class and is the amount of training data selected by the user.

[0009] Step (3) designs a visualization framework that integrates manual point selection, CNN training, automatic point sampling, and visual comparison, enabling users to obtain and evaluate sampled points from the perspectives of data analysis and user preferences.

[0010] Preferably, the information entropy value of the local area is calculated, and those areas with higher entropy values ​​are recommended for users to mark points of interest based on prior knowledge and requirements. Specifically, the information entropy estimation is defined as follows: Among them, p(x i ), i = 1, 2, ···, n is the probability of a geographical point with attribute x in the interval, n represents the attribute interval specified by the user according to his requirements, when n is greater than 2, the total information entropy of the local area will be calculated by accumulating the entropy value of each attribute: Among them, m is used to construct the attributes considered when representing user preferences.

[0011] The user manually specifies the geographical points of interest in the area with high entropy value, and constructs the feature vector based on the common features extracted from the specified points of interest; the point density in the scatter plot is measured, assuming that a set of points P = {p1, p2, ...}, p m is the neighbor of point P, the density of point P: Among them, k h () is the kernel function of bandwidth h;

[0012] Construct a feature vector to represent the user's preferences: where p1, p2, p…, p i is a point selected by the user, k de i is the density value at point I, and Represents the value of attribute J of the selected point i.

[0013] Preferably, the visual comparison is as follows:

[0014] First, to quantify the attribute distribution, calculate the evaluation accuracy of the attribute: Among them, Ai is the accuracy of attribute i, n is the number of attribute i and the interval of attribute i. j ,L j and S j are the number of marking points and sampling points in interval j respectively. O ,T L and T S are the total number of original geographic points, marked points, and sampling points, respectively;

[0015] Then, the satisfaction of different user preferences within the sampling points is evaluated by using heat maps to present the spatial density of geographic points, bidirectional bar charts to show attribute distribution, and DTW charts to represent the attribute relationship between points.

[0016] The beneficial effects of the present invention are:The proposed method designs a user-driven sampling model to reduce the overdraw problem of large-scale geographic point visualization. It performs well in preserving personalized user preferences on visual abstractions of large-scale geographic point data. Adaptive spherical noise sampling and CNN models are integrated into our sampling model, which effectively preserves user preferences and spatial density of sampling points. Therefore, our sampling model not only reduces the visual clutter of large-scale geographic point data visualization, but also preserves user preferences in the sampling points. At the same time, we provide rich interactive and visual cues such as manual point selection and visual comparison (heat maps, bidirectional histograms and DTW plots), enabling users to interactively build sampling models and intuitively evaluate the sampled geographic points. Through case studies and quantitative comparisons based on real-world geographic datasets, the effectiveness of our sampling model in visual abstraction and exploration of large-scale geographic point data is further demonstrated. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0018] Figure 2 Schematic diagram of the system of the present invention. (a) is a control panel that allows users to load geographic point datasets, perform geographic space division, and list local areas according to entropy values. (b) is a model training panel that allows users to specify the parameters of the CNN model for user preference training. (c) is a bidirectional histogram, with the attribute distribution of the user-selected points on the left and the attribute distribution of sampling results based on different sampling strategies on the right. (d) is a map view that presents the geographic point dataset, which provides a rich set of visual clues such as (1) density map, (2) attribute histogram, (3) and (4) local areas of interest and (5) heat map. (e) lists basic information of the local area specified by the user. (f) is an improved global graph matrix to present different attribute relationships. (g) is a DTW graph to evaluate the preservation of attribute relationship preferences.

[0019] Figure 3 This is a description of the user preference training process based on convolutional neural networks in the present invention;

[0020] Figure 4 This is a schematic diagram of spatial density evaluation in the present invention;

[0021] Figure 5 This is a schematic diagram of attribute distribution evaluation in the present invention;

[0022] Figure 6 Schematic diagram of attribute relationship evaluation in the present invention;

[0023] Figure 7 This is a schematic diagram of the comparative evaluation of the effects of different sampling methods in the present invention. DETAILED DESCRIPTION

[0024] The present invention is described in detail below with reference to the accompanying drawings and specific implementation examples.

[0025] like Figure 1 , a user-driven sampling method for large-scale geographic point data visualization based on convolutional neural networks, the specific steps are:

[0026] Step (1) First, the geographic space is divided into local areas using an adaptive blue noise sampling model, and a set of visual model interfaces are designed to display the data features of the local areas, so that users can intuitively select local representative points based on their prior knowledge. Then a CNN is trained to learn user preferences based on the feature vectors of the selected representatives, and further used to guide the sampling process of other local areas. The local areas determined by blue noise sampling and the representatives guided by user preferences learned by CNN, the sampling points retain the spatial distribution of the original data points and meet user preferences as much as possible.

[0027] The specific steps of step (1) are:

[0028] (1-1) The user manually specifies the geographical points of interest, and constructs a feature vector based on the common features extracted from those marked points of interest. Measuring the point density in a scatter plot, assuming a set of points P = {p1, p2, ...}, p m is the neighbor of point P, the density of point P: Among them, k h () is the kernel function of bandwidth h, which determines the smoothness of the reconstructed density field.

[0029] (1-2) Construct a feature vector to represent the user's preference: where p1, p2, p…, p i is a point selected by the user, k de i is the density value of point i, and Represents the value of attribute j at selected point i.

[0030] (1-3) Design a CNN structure to extract user preference features. The loss function used in the CNN training process is defined as follows: Where: N is the amount of training data selected by the user; y i Represents the label of training data point i. If the label is positive, y i =1, if it is a negative number, then y i =0; represents the probability that data point i is predicted to be the positive class and is the amount of training data selected by the user.

[0031] (1-4) Conduct user preference training. Users use the interactive tools provided by the system to mark sampling points of interest. Based on the determined training data, the user preference model is trained online based on the CNN structure. The trained model is used to perform user-driven sampling schemes on all local areas to generate the desired sampling results.

[0032] Step (2) trains a CNN to learn user preferences based on the feature vectors of the selected representatives, and further uses it to guide the sampling process of other local areas. The local areas determined by blue noise sampling, the representatives guided by the user preferences learned by CNN, and the sampling points retain the spatial distribution of the original data points to meet the user preferences as much as possible. The specific steps are:

[0033] (2-1) Design a CNN structure to extract user preference features. The preference matrix composed of one-dimensional vectors is input into the first one-dimensional CNN layer for feature learning, and the second one-dimensional CNN layer performs further feature extraction.

[0034] (2-2) The maximum pooling layer is used to reduce the complexity of the output, and a dropout layer is further added to alleviate the overfitting problem caused by the small amount of training data, thereby improving the processing accuracy of unseen data. In order to further learn higher-level features, two more one-dimensional CNN layers are used, followed by a maximum pooling layer and a global average pooling layer to alleviate the overfitting problem. Finally, the fully connected layer, as the last part of the network structure, plays the role of a classifier. The loss function used in the CNN training process is defined as follows: Where: N is the amount of training data selected by the user; y i Represents the label of training data point i. If the label is positive, y i =1, if it is a negative number, then y i =0; represents the probability that data point i is predicted to be the positive class and is the amount of training data selected by the user.

[0035] (2-3) Conduct user preference training. Users use the interactive tools provided by the system to mark sampling points of interest. Based on the determined training data, the user preference model is trained online based on the CNN structure. Using the trained model, a user-driven sampling scheme is performed on all local areas to generate the desired sampling results. If the general evaluation indicators such as accuracy or dynamic time warping sampling points are not met, or the model parameters are modified, the model is trained again until the results meet the user's requirements. The spatial distribution of the original data points is preserved through the sampling points to meet the user's preferences as much as possible.

[0036] Step (3) designs a visualization framework that integrates manual point selection, CNN training, automatic point sampling, and visualization comparison, enabling users to obtain and evaluate sampling points from the perspective of data analysis and user preferences, such as Figure 2 The specific steps are:

[0037] (3-1) First, to quantify the attribute distribution, calculate the attribute evaluation accuracy: Among them, Ai is the accuracy of attribute i, n is the number of attribute i and the interval of attribute i. j ,L j and S j are the number of marking points and sampling points in interval j respectively. O ,T L and T S are the total number of original geographic points, marked points, and sampling points, respectively.

[0038] (3-2) Then, the satisfaction of different user preferences within the sampling points is evaluated by using heat maps to present the spatial density of geographic points, bidirectional bar charts to show attribute distribution, and DTW charts to represent the attribute relationship between points.

[0039] The effectiveness of this method is evaluated as follows:

[0040] We demonstrate the effectiveness of our proposed user-driven sampling method for large-scale geographic point data visualization using convolutional neural networks through quantitative comparison and visual evaluation. We use two datasets (D1 and D2) for evaluation, comparing our sampling method (denoted as "OURS") with two widely applicable and authoritative sampling methods: blue noise sampling (denoted as "BNS") and random sampling (denoted as "RS").

[0041] Table 1

[0042]

[0043] At three sampling rates of 1%, 5% and 10%, the above three indicators are used: as shown in Table 1, OURS and BNS have similar KDE error perspectives. RS has the smallest KDE error, indicating that its spatial distribution is quite close to the spatial distribution of the original dataset. It can be seen that RS can maintain the spatial distribution better than the other two methods, and the performance of OURS is similar to that of BNS because they are both designed based on adaptive blue noise sampling. D1 is used to evaluate the preservation of attribute distribution. As shown in the middle row of Table 1, our sampling results are better than other sampling results at all sampling rates. In particular, at a sampling rate of 5%, our sampling model completely outperforms other sampling strategies. Although there are two exceptions, the performance of OURS is also very close to the optimal method. This shows that the user-driven sampling model of the present invention is able to learn the user's preference for attribute distribution, thereby obtaining samples with similar attribute distribution through the user preference model.

[0044] like Figure 3 As shown in the figure, the user preference training steps can be divided into four steps: the user applies the interactive tools provided by the system to mark the sampling points of interest; based on the determined training data, the user preference model is trained online based on the CNN structure; using the trained model, a user-driven sampling scheme is performed on all local areas to generate the desired sampling results; if the general evaluation index accuracy or the accuracy obtained by dynamic time warping is not enough, the system will enable the user to mark more sampling points or modify the model parameters.

[0045] like Figure 4 As shown in Figure 2, the heatmap shows a comparison of the spatial density preservation in the sampling results generated by our sampling, random sampling, and adaptive blue noise sampling. We can see that the heatmaps generated by all different sampling results are very similar to the heatmap of the global original data. The heatmap generated by RS (c) is most similar to the original heatmap (a), while OURS (b) is more similar to BNS (c). This shows that the sampling method of the present invention is able to maintain the spatial density of the original data, similar to adaptive blue noise sampling, which aims to maintain spatial density, but slightly lower than random sampling, which facilitates users to perform effective geospatial analysis of the sampling results.

[0046] like Figure 5 As shown in the accuracy histogram, OURS outperforms other sampling strategies in preserving the distribution of all attributes such as PO, CLOSO and MDSO. Figure 5 (b) in the figure, it can be found that the distribution of the selected points has no obvious trend of change, and the rank of the interval distribution has not changed. For the attribute MDSO, the sampling results of OURS are also better than RS and BNS within a certain time interval, such as Figure 5 (c) in the figure. It should be noted that the attribute PO distribution obtained by OURS ( Figure 5(d)) in the interval distribution performs best, indicating that user preference may lie in attribute PO. Our sampling results can better meet the needs of personalized analysis. Similar conclusions can be drawn from the heat map, such as Figure 5 In addition, Figure 5 (a) in Fig. 3 shows that our sample points are closer to the selected points than other methods, especially in attribute PO. The above experimental results show that our user-driven sampling model has advantages in preserving attribute distribution preferences, which is of great significance for further geospatial analysis.

[0047] like Figure 6 As shown, this set of DTW charts generated by OURS, BNS and RS shows that the sampling results of the present invention present a clear trend of different attribute relationships, which are more consistent with the selected points, especially for the relationship between the combination and separation of attributes.

[0048] like Figure 7 As shown, we can intuitively see the sampling results of three different sampling methods. In general, the brown rectangle in the light-multiple matrix and the heat map show the attribution relationship between preservation and marriage. It can be concluded that compared with BNS and RS, the sampling results of the present invention better preserve the attribute relationship preference.

Claims

1. A geographic point data sampling method based on user task preference learning, characterized in that: The method is specifically: Step (1) uses an adaptive blue noise sampling model to divide the geographic space into local areas, and designs a set of visual model interfaces to display the data features of the local areas, calculates the information entropy value of the local areas based on the data features of the local areas, and recommends those areas with higher entropy values ​​for users to mark points of interest based on prior knowledge and requirements, and extracts common features to construct feature vectors; Step (2) a CNN is trained to learn the user preference according to the selected representative feature vector, and further used to guide the sampling process of other local areas; the local area determined by blue noise sampling is subjected to a user-driven sampling scheme for all local areas through the trained CNN model to generate the desired sampling results; if the general evaluation index accuracy or the accuracy obtained by dynamic time warping is less than the set threshold, the model parameters are modified and the model is trained again until the result meets the user's requirements; The CNN model uses a preference matrix consisting of one-dimensional vectors as input to the first one-dimensional CNN layer for feature learning, and the second one-dimensional CNN layer performs further feature extraction. A maximum pooling layer is used to reduce the complexity of the output, and a dropout layer is further added to alleviate the overfitting problem caused by the small amount of training data. In order to further learn higher-level features, two more one-dimensional CNN layers are used, followed by a maximum pooling layer and a global average pooling layer to alleviate the overfitting problem. Finally, the fully connected layer, as the last part of the network structure, plays the role of a classifier. The loss function used in the CNN training process is defined as follows: Where: N is the amount of training data selected by the user; y i Represents the label of training data point i. If the label is positive, y i =1, if it is a negative number, then y i =0; Represents the probability that data point i is predicted to be a positive class; Step (3) designs a visualization framework that integrates manual point selection, CNN training, automatic point sampling, and visual comparison, enabling users to obtain and evaluate sampled points from the perspectives of data analysis and user preferences.

2. A geographic point data sampling method based on user task preference learning according to claim 1, characterized in that: The information entropy value of the local area is calculated, and those areas with higher entropy values ​​are recommended for users to mark points of interest based on prior knowledge and requirements. Specifically, the information entropy estimation is defined as follows: Among them, p(x i ), i = 1, 2, ···, n is the probability of a geographical point with attribute x in the interval, and n represents the attribute interval specified by the user according to his requirements; when n is greater than 2, the total information entropy of the local area will be calculated by accumulating the entropy value of each attribute: Among them, m is used to construct the attributes considered when representing user preferences; The user manually specifies the geographical points of interest in the area with high entropy value, and constructs the feature vector based on the common features extracted from the specified points of interest; the point density in the scatter plot is measured, assuming that a set of points P = {p1, p2, ...}, p m is the neighbor of point P, the density of point P: Among them, k h () is the kernel function of bandwidth h; Construct a feature vector to represent the user's preferences: where p1, p2, p…, p i is a point selected by the user, k de i is the density value of point i, and Represents the value of attribute j at selected point i.

3. The geographic point data sampling method based on user task preference learning according to claim 1, characterized in that: The visual comparison is specifically as follows: First, to quantify the attribute distribution, calculate the evaluation accuracy of the attribute: Among them, Ai is the accuracy of attribute i, n is the number of attribute i and the interval of attribute i; j ,L j and S j are the number of marking points and sampling points in interval j respectively; T O ,T L and T S are the total number of original geographic points, marked points, and sampling points, respectively; Then, the satisfaction of different user preferences within the sampling points is evaluated by using heat maps to present the spatial density of geographic points, bidirectional bar charts to show attribute distribution, and DTW charts to represent the attribute relationship between points.

Images