CN111783832A - An Interactive Selection Method for Spatio-temporal Data Prediction Models - Google Patents

Abstract

An interactive selection method based on a spatiotemporal data prediction model, which filters, cleans, deletes outliers, and completes missing values from the original data; then combines the Canopy and K-Means clustering algorithms to cluster the above data to obtain K clusters and extract the first K1 cluster areas to calculate the boundaries; after the data of each area is differentially processed, the Random Forest algorithm, SVM algorithm, ARIMA algorithm, and LSTM algorithm are used to train and establish models, and then each area is based on the model. Make predictions; display the obtained prediction result data with visual glyphs in each area on the map; after completing the above steps, optimize the layout of the area glyphs and area connections. The interactive exploration component provided by the system helps users to effectively distinguish the differences of various models in an intuitive way. The font design and geographic layout design of the present invention enables the user to intuitively conduct in-depth analysis of the prediction output.

Description

An Interactive Selection Method for Spatio-temporal Data Prediction Models

technical field

The invention relates to an interactive selection method of a spatiotemporal data prediction model.

Background technique

With the development of society, a large amount of spatiotemporal data is generated in various fields, such as: tracking industrial production, conducting financial transactions and monitoring the environment. Analysis of spatiotemporal data typically includes statistical analysis, model recommendation, and optimal prediction, which has important implications for many aspects of society. Making reasonable predictions by analyzing spatiotemporal data can help researchers grasp social/technological trends and make better decisions.

The core role of spatiotemporal prediction is to establish a prediction model. The prediction methods for spatiotemporal data can be divided into two categories: traditional statistical methods and machine learning methods. Traditional statistical methods are usually based on user observations and require the following steps: data sampling, data plotting, and curve fitting/parameter estimation, including moving average models (MA), autoregressive moving average models (ARMA), and autoregressive synthesis Models such as moving average (ARIMA). Machine learning methods are mainly based on data classification and require the following steps: data decomposition, model training, and model prediction, including models such as support vector machines (SVM), random forests, and nonparametric regression models. With the development, many advanced models, such as artificial neural network (ANN) and long short-term memory network (LSTM), have emerged to help people gain insight into spatiotemporal data in different application domains.

Since different types of models are suitable for different application scenarios and may have their specific limitations, researchers have conducted a series of studies on model recommendation to help people have better choices in model selection. However, there are some challenges in their work. One of the difficulties is that their research focus is data-driven. Although a lot of interesting information can be found through data transformation analysis, there is also a lot of application scenario information to consider. At the same time, this can be difficult due to various levels of uncertainty. Another challenge is that they do not have a proper explanation for the predictive performance of the models. Users may wonder why this model works well while others do not in specific situations, and the advantages of different models and their scope of use.

At this stage, in the visual analysis of spatiotemporal prediction models, visual analysis systems combine multiple machine learning models with interactive visualizations, allowing users to examine models and visually compare the prediction performance between different models. Many visualization researchers perform analysis by linking model performance metrics such as accuracy and precision to visualization components such as line charts, bar charts, and heatmaps. Through an interactive exploration process, users can learn about the multiple possibilities of a predictive model and gain insight into how multiple models differ. However, existing work for multi-model steering and selection aims to detect performance metrics without comprehensively considering input data and model outputs, and insufficient understanding of the relationship between spatiotemporal data and model outputs may lead to incorrect model selection.

Based on the above questions, we carefully design a visualization framework combining time series/geospatial data and the performance of various prediction models, which is beneficial for machine learning beginners and non-experts to choose and understand models. The present invention proposes a visual analysis system that enables users to interactively check the similarity of the predicted outputs of spatiotemporal data models. In contrast to other work, we aim to map geographic information to the output of a predictive model, with simultaneous correlation of multiple models. Clustering methods have been applied to similarity analysis, combined with the correlation matrix view, users can intuitively understand the similarity of different models. At the same time, we design novel glyphs that allow users to intuitively perform in-depth analysis of the prediction output. Furthermore, based on mouse scaling, we design multi-layered layouts that help eliminate geospatial overlap when the user closely observes the glyphs linked to the model's predicted output. At each level, a force-directed graph algorithm has been applied to the glyphs to avoid collisions. In addition, a well-designed timeline view can help users quickly understand the original time series.

In various fields, data forecasting is a common data analysis method. However, it is difficult for beginners and non-experts to choose a suitable forecasting model and understand the forecasting performance and usage scenarios of different models. Existing model recommendation systems recommend a most suitable prediction model to users based on data-driven, and do not give a proper explanation for the prediction performance of the prediction model, so it is difficult for users to find interesting points from it. Users may wonder why this model works well while others do not in specific situations, and the advantages of different models and their scope of use.

SUMMARY OF THE INVENTION

Existing work for multi-model steering and selection aims to detect performance metrics rather than comprehensively considering input data and model outputs, and insufficient understanding of the relationship between spatiotemporal data and model outputs may lead to incorrect model selection. In order to overcome the deficiencies of the prior art, the present invention provides an interactive selection method of a spatiotemporal data prediction model.

In order to solve the above-mentioned technical problems, the present invention provides the following technical solutions:

An interactive selection method for a spatiotemporal data prediction model, which visually interprets and contrasts the spatiotemporal prediction model based on a map area, and the method includes the following steps:

(1) Obtain the shared bicycle data, delete the bicycle data that is not in our analysis area and the riding time is abnormal, and judge whether the data of each bicycle at all time points on all dates in this area is empty, if it is empty, use 0 Fill in and make a data set;

(2) Combine the Canopy clustering algorithm and the K-means clustering algorithm to cluster the data set obtained above, and take the first K1 clusters as the prediction area; then, construct the traffic data of the K1 prediction areas. Model and predict, the test results include predicted data, real data, MAE value of predicted data, RMSE value of predicted data, R-Squared value and 1-R-Squared value of predicted data;

(3) The prediction results of the four prediction algorithms for each region are presented in a visualized glyph. The analysis and display steps are as follows:

(3-1) Use the multi-layer radar map as the first layer of the area map: take the (1-R-Squared) value, MAE value and RMSE value obtained above as the vertices of the radar map in anti-clockwise order. The vertices display the parameter names and numerical values. The four algorithms are displayed by stacking four layers of radar images. Each layer of radar images is represented by a different texture. The backslashes represent the Random Forest algorithm, and the vertical bars represent the LSTM algorithm. , the horizontal bar represents the ARIMA algorithm, and the slash system represents the SVM algorithm;

(3-2) Use the histogram as the second layer of the area map: draw the R-Squared value, the MAE value and the RMSE value obtained above in a counterclockwise order and distribute the histogram on the second layer. The height of each column is Indicates the numerical value, and each column displays the parameter name; different textures are used to represent different prediction algorithms to obtain prediction results, where the backslashes represent the Random Forest algorithm, the vertical bars represent the LSTM algorithm, and the horizontal bars represent the is the ARIMA algorithm, and the slashes represent the SVM algorithm; the different gray levels of the histogram of the same texture represent the prediction results obtained by different training models of the same algorithm;

(4) Layout algorithm based on map prediction area glyph; according to the center position of the area, place the original position of the glyph at the center point of the area, and use the improved force-oriented layout algorithm to re-layout the overlapping glyphs. The process of the layout algorithm is as follows: :

(4-1) Input K1 initial node positions, the radius of the node is r, and calculate the distance between the nodes, if the node distance is less than 2r, the two nodes coincide;

(4-2) Calculate the relative position between each node to obtain the relative matrix M. For the nodes {a1, a2....} (i=K1), the calculated matrix M is:

Among them, for each node, define its upper right label as 0, its upper left label as 1, its lower left label as 2, and its lower right label as 3, and then calculate the relative relationship between each original node and all other nodes position, get the K1 order matrix;

(4-4) Calculate the displacement generated by the force action between the two nodes, and the displacement calculation formula of the two points is:

Where x represents the displacement generated by the force action between two points, Δx represents the difference between the abscissas of the two points, Δy represents the difference between the ordinates of the two points, and k is the force action coefficient;

(4-5) For each node, calculate the displacement generated by the force action between it and all other nodes, and accumulate to obtain the unit displacement of each node;

(4-6) According to the unit displacement of each node, update the coordinates of K1 nodes in turn;

(4-7) Calculate the updated relative position between each node to obtain the relative matrix M1, and compare M and M1, that is, compare the relative position between the points and the relative position between the updated points; For the two nodes P ₁ and P ₂ whose positions have changed, the maximum displacement of the two points is calculated according to the original relative angle and the original coordinates of the two points, and the coordinates of P ₂ are updated;

(4-8) Repeat steps (4-4) to (4-7), and iterate n times until all nodes do not overlap;

(5) Use lines to connect the glyphs of similar areas; calculate the attributes of the nodes of the area, the attributes are composed of various indicators of the prediction algorithm, and then the clustering algorithm classifies similar points into one category, and finally uses line segments to connect the corresponding areas of these areas. Similar glyphs, and create Bezier curves for beautification. The algorithm process of connecting glyphs with line segments is as follows:

(5-1) Input a series of coordinate points S={P ₁ , P ₂ ···P _t };

(5-2) Select the first point P ₁ as the initial point p ₀ ;

(5-3) Calculate the Euclidean distance between p ₀ and other points, select the closest point p _k , save the line segment [p ₀ , p _k ] and the closest distance dis, and delete p ₀ from S;

(5-4) Let p ₀ =p _k , repeat step (5-3);

(5-5) Iterate (t-1) times, that is, stop the iteration when there is only one element in S, and save the connection line as lines0, and save the line length distance0;

(5-6) Select P ₂ ...P _t as the initial point p ₀ in turn, and repeat steps (5-2) to (5-5) to obtain (t-1) types of connection lines and their line lengths, and Screen out the connection lines whose line segments do not intersect with each other (the common endpoints are not considered to intersect), and then select the connection line with the shortest line length as the line segment connection scheme;

(6) A line layout algorithm is proposed to detect the collision between line segments and nodes and re-plan the line paths. The implementation steps are as follows:

(6-1) The radius of the node is r, and the distance d from each node to the line segment is detected. If r>d, it is detected as a collision between the node and the line segment; otherwise, continue to judge the next node;

(6-2) When the node collides with the line segment, judge the relative position of the node and the line segment, if the node is on the left side of the line segment, select the stagnation point on the right side of the node, otherwise, select the stagnation point on the left side of the node;

(6-3) Connect the stagnation point and the node center, obtain two intersection points on the circumference of the dissimilar nodes, find the intersection point closest to the line segment, and then generate a virtual node within the threshold θ of this point, and connect the virtual node with the similar node.

The technical idea of the present invention is to design an interactive visual analysis system for interactive analysis, understanding and visual comparison of the prediction model and its performance. Firstly, the spatial data is analyzed, then the prediction model is established through the data set, and the prediction is completed. Finally, a set of visual glyphs are designed to better display the prediction results of the model. At the same time, a novel layout algorithm is proposed to solve the collision between glyphs and glyphs. The problem and the glyph-line-segment collision problem demonstrate the differences and correlations between prediction models. It helps users to better understand the impact of the forecasting model and its parameters on the forecasting results, as well as the differences between the forecasting results and different models.

Beneficial effects of the present invention: through visual analysis and comparison of prediction models, combined with the performance of time series/geospatial data and various prediction models, comprehensively considering the relationship between spatiotemporal data and model output, an interactive visual analysis system is designed, allowing users to Interactively explore the similarity of model prediction outputs on spatiotemporal data to gain a deeper understanding of model performance. Innovative glyph design and geographic layout design allow users to intuitively perform in-depth analysis of the forecast output.

Description of drawings

FIG. 1 is a flow chart of the present invention.

FIG. 2 is a visual glyph diagram of the present invention.

FIG. 3 is a font layout diagram of the present invention, wherein (a) represents the regional font, and (b) represents the regional font after layout.

FIG. 4 is a layout diagram of the connection lines of the present invention, wherein (a) represents the connection lines of the regional fonts, and (b) represents the connection lines of the regional fonts after the layout.

specific implementation

The present invention will be described in detail below according to the accompanying drawings and preferred embodiments, and the purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

Referring to Fig. 1 to Fig. 4, an interactive selection method of a spatiotemporal data prediction model includes the following steps:

(1) Obtain the shared bicycle data, delete the bicycle data that is not in our analysis area and the riding time is abnormal, and judge whether the data of each bicycle at all time points on all dates in this area is empty, if it is empty, use 0 Fill in and make a data set;

(2) Combine the Canopy clustering algorithm and the K-means clustering algorithm to cluster the data set obtained above, and take the first K1 clusters as the prediction area; then, construct the traffic data of the K1 prediction areas. Model and predict, the test results include predicted data, real data, MAE value of predicted data, RMSE value of predicted data, R-Squared value and 1-R-Squared value of predicted data;

(3) Propose a visual font to display the prediction results of the four prediction algorithms in each region (as shown in Figure 2). The analysis and display steps are as follows:

(3-1) Use the multi-layer radar map as the first layer of the area map: take the (1-R-Squared) value, MAE value and RMSE value obtained above as the vertices of the radar map in anti-clockwise order. The vertices display the parameter names and numerical values. The four algorithms are displayed by stacking four layers of radar images. Each layer of radar images is represented by a different texture. The backslashes represent the Random Forest algorithm, and the vertical bars represent the LSTM algorithm. , the horizontal bar represents the ARIMA algorithm, and the slash system represents the SVM algorithm;

(3-2) Use the histogram as the second layer of the area map: draw the R-Squared value, the MAE value and the RMSE value obtained above in a counterclockwise order and distribute the histogram on the second layer. The height of each column is Represents the numerical value, and each column displays the parameter name; different textures are used to represent different prediction algorithms to obtain prediction results, where the backslashes represent the RandomForest algorithm, the vertical bars represent the LSTM algorithm, and the horizontal bars represent the ARIMA algorithm, the slashes represent the SVM algorithm; the different gray levels of the histogram of the same texture represent the prediction results obtained by different training models of the same algorithm;

(4) Layout algorithm based on map prediction area glyph; according to the center position of the area, place the original position of the glyph at the center point of the area, and use the improved force-oriented layout algorithm to re-layout the overlapping glyphs (as shown in Figure 3), The process of the layout algorithm is:

(4-1) Input K1 initial node positions, the radius of the node is r, and calculate the distance between the nodes, if the node distance is less than 2r, the two nodes coincide;

(4-2) Calculate the relative position between each node to obtain the relative matrix M. For the nodes {a1, a2....} (i=K1), the calculated matrix M is:

Among them, for each node, define its upper right label as 0, its upper left label as 1, its lower left label as 2, and its lower right label as 3, and then calculate the relative relationship between each original node and all other nodes position, get the K1 order matrix;

(4-4) Calculate the displacement generated by the force action between the two nodes, and the displacement calculation formula of the two points is:

Where x represents the displacement generated by the force action between two points, Δx represents the difference between the abscissas of the two points, Δy represents the difference between the ordinates of the two points, and k is the force action coefficient;

(4-5) For each node, calculate the displacement generated by the force action between it and all other nodes, and accumulate to obtain the unit displacement of each node;

(4-6) According to the unit displacement of each node, update the coordinates of K1 nodes in turn;

(4-7) Calculate the updated relative position between each node to obtain the relative matrix M1, and compare M and M1, that is, compare the relative position between the points and the relative position between the updated points; For the two nodes P ₁ and P ₂ whose positions have changed, the maximum displacement of the two points is calculated according to the original relative angle and the original coordinates of the two points, and the coordinates of P ₂ are updated;

(4-8) Repeat steps (4-4) to (4-7), and iterate n times until all nodes do not overlap.

(5) Use lines to connect the glyphs of similar areas (as shown in Figure 4(a)); calculate the attributes of the nodes in the area. Finally, line segments are used to connect similar glyphs corresponding to these areas, and Bezier curves are created for beautification. The algorithm process of connecting glyphs with line segments is as follows:

(5-1) Input a series of coordinate points S={P ₁ , P ₂ ···P _t };

(5-2) Select the first point P ₁ as the initial point p ₀ ;

(5-3) Calculate the Euclidean distance between p ₀ and other points, select the closest point p _k , save the line segment [p ₀ , p _k ] and the closest distance dis, and delete p ₀ from S;

(5-4) Let p ₀ =p _k , repeat step (5-3);

(5-5) Iterate (t-1) times, that is, stop the iteration when there is only one element in S, and save the connection line as lines0, and save the line length distance0;

(5-6) Select P ₂ ...P _t as the initial point p ₀ in turn, and repeat steps (5-2) to (5-5) to obtain (t-1) types of connection lines and their line lengths, and Filter out the connection lines whose line segments do not intersect with each other (the common endpoints are not considered to intersect), and then select the connection line with the shortest line length as the line segment connection scheme 5

(6) Referring to FIG. 4, in a preferred embodiment, a line layout algorithm is proposed to detect the collision between line segments and nodes and re-plan the line paths. The implementation steps are as follows:

(6-1) The radius of the node is r, and the distance d from each node to the line segment is detected. If r>d, it is detected as a collision between the node and the line segment; otherwise, continue to judge the next node;

(6-2) When the node collides with the line segment, judge the relative position of the node and the line segment, if the node is on the left side of the line segment, select the stagnation point on the right side of the node, otherwise, select the stagnation point on the left side of the node;

(6-3) Connect the stagnation point and the node center, obtain two intersection points on the circumference of the dissimilar nodes, find the intersection point closest to the line segment, and then generate a virtual node within the threshold θ of this point, and connect the virtual node with the similar node.

This embodiment designs an interactive visual analysis system, which is used for interactive analysis, understanding and visual comparison of the prediction model and its performance. Firstly, the spatial data is analyzed, then the prediction model is established through the data set, and the prediction is completed. Finally, a set of visual glyphs are designed to better display the prediction results of the model. At the same time, a novel layout algorithm is proposed to solve the collision between glyphs and glyphs. The problem and the glyph-line-segment collision problem demonstrate the differences and correlations between prediction models. It helps users to better understand the impact of the forecasting model and its parameters on the forecasting results, as well as the differences between the forecasting results and different models.

Those of ordinary skill in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still understand the Modifications are made to the technical solutions described in the foregoing examples, or equivalent replacements are made to some of the technical features. All modifications and equivalent replacements made within the spirit and principle of the invention shall be included within the protection scope of the invention.

Claims (1)

1. the interactive selection method of a spatiotemporal data prediction model, is characterized in that, described method comprises the following steps: (1) Obtain the shared bicycle data, delete the bicycle data that is not in our analysis area and the riding time is abnormal, and judge whether the data of each bicycle at all time points on all dates in this area is empty, if it is empty, use 0 Fill in and make a data set; (2) Combine the Canopy clustering algorithm and the K-means clustering algorithm to cluster the data set obtained above, and take the first K1 clusters as the prediction area; then, construct the traffic data of the K1 prediction areas. Model and predict, the test results include predicted data, real data, MAE value of predicted data, RMSE value of predicted data, R-Squared value and 1-R-Squared value of predicted data; (3) The prediction results of the four prediction algorithms for each region are presented in a visualized glyph. The analysis and display steps are as follows: (3-1) Use the multi-layer radar map as the first layer of the area map: take the (1-R-Squared) value, MAE value and RMSE value obtained above as the vertices of the radar map in anti-clockwise order. The vertices display the parameter names and numerical values. The four algorithms are displayed by stacking four layers of radar images. Each layer of radar images is represented by a different texture. The backslashes represent the Random Forest algorithm, and the vertical bars represent the LSTM algorithm. , the horizontal bar represents the ARIMA algorithm, and the slash system represents the SVM algorithm; (3-2) Use the histogram as the second layer of the area map: draw the R-Squared value, the MAE value and the RMSE value obtained above in a counterclockwise order and distribute the histogram on the second layer. The height of each column is Indicates the numerical value, and each column displays the parameter name; different textures are used to represent different prediction algorithms to obtain prediction results, where the backslashes represent the Random Forest algorithm, the vertical bars represent the LSTM algorithm, and the horizontal bars represent the is the ARIMA algorithm, and the slashes represent the SVM algorithm; the different gray levels of the histogram of the same texture represent the prediction results obtained by different training models of the same algorithm; (4) Layout algorithm based on map prediction area glyph; according to the center position of the area, place the original position of the glyph at the center point of the area, and use the improved force-oriented layout algorithm to re-layout the overlapping glyphs. The process of the layout algorithm is as follows: : (4-1) Input K1 initial node positions, the radius of the node is r, and calculate the distance between the nodes, if the node distance is less than 2r, the two nodes coincide; (4-2) Calculate the relative position between each node to obtain the relative matrix M. For the nodes {a1, a2....} (i=K1), the calculated matrix M is:

Among them, for each node, define its upper right label as 0, its upper left label as 1, its lower left label as 2, and its lower right label as 3, and then calculate the relative relationship between each original node and all other nodes position, get the K1 order matrix; (4-4) Calculate the displacement generated by the force action between the two nodes, and the displacement calculation formula of the two points is:

Where x represents the displacement generated by the force action between two points, Δx represents the difference between the abscissas of the two points, Δy represents the difference between the ordinates of the two points, and k is the force action coefficient; (4-5) For each node, calculate the displacement generated by the force action between it and all other nodes, and accumulate to obtain the unit displacement of each node; (4-6) According to the unit displacement of each node, update the coordinates of K1 nodes in turn; (4-7) Calculate the updated relative position between each node to obtain the relative matrix M1, and compare M and M1, that is, compare the relative position between the points and the relative position between the updated points; For the two nodes P ₁ and P ₂ whose positions have changed, the maximum displacement of the two points is calculated according to the original relative angle and the original coordinates of the two points, and the coordinates of P ₂ are updated; (4-8) Repeat steps (4-4) to (4-7), and iterate n times until all nodes do not overlap; (5) Use lines to connect the glyphs of similar areas; calculate the attributes of the nodes of the area, the attributes are composed of various indicators of the prediction algorithm, and then the clustering algorithm classifies similar points into one category, and finally uses line segments to connect the corresponding areas of these areas. Similar glyphs, and create Bezier curves for beautification. The algorithm process of connecting glyphs with line segments is as follows: (5-1) Input a series of coordinate points S={P ₁ , P ₂ …P _t }; (5-2) Select the first point P ₁ as the initial point p ₀ ; (5-3) Calculate the Euclidean distance between p ₀ and other points, select the closest point p _k , save the line segment [p ₀ , p _k ] and the closest distance dis, and delete p ₀ from S; (5-4) Let p ₀ =p _k , repeat step (5-3); (5-5) Iterate (t-1) times, that is, stop the iteration when there is only one element in S, and save the connection line as lines0, and save the line length distance0; (5-6) Select P ₂ ... P _t as the initial point p ₀ in turn, repeat steps (5-2) to (5-5) to obtain (t-1) types of connection lines and their line lengths, and filter out Connecting lines whose line segments do not intersect with each other, and then select a connecting line with the shortest line length as the line segment connection scheme; (6) A line layout algorithm is proposed to detect the collision between line segments and nodes and re-plan the line paths. The implementation steps are as follows: (6-1) The radius of the node is r, and the distance d from each node to the line segment is detected. If r>d, it is detected as a collision between the node and the line segment; otherwise, continue to judge the next node; (6-2) When the node collides with the line segment, judge the relative position of the node and the line segment, if the node is on the left side of the line segment, select the stagnation point on the right side of the node, otherwise, select the stagnation point on the left side of the node; (6-3) Connect the stagnation point and the node center, obtain two intersection points on the circumference of the dissimilar nodes, find the intersection point closest to the line segment, and then generate a virtual node within the threshold θ of this point, and connect the virtual node with the similar node.

Images