T-PickSeer: Visual Analysis of
Taxi Pick-up Point Selection Behavior

The study of pick-up selection of passengers is beneficial to the management of taxi operation and future urban planning, which is of great significance to the urban structure, policy making, resource allocation, and so on. These studies can be mainly divided into two parts: pattern analysis and potential pick-up exploration.

The studies of pattern analysis are mainly committed to finding behavior patterns for pick-up selection and analyzing influencing factors. Traditional methods (e.g., statistics and regression) have been proved effective and many useful conclusions have been drawn  [15]. But these suggestions are general and have trouble dealing with dynamic spatio-temporal data. More detailed information is desired. Then some research turns to the discovery and analysis of hotspots. Bi et al. [5] dug out the travel rules of different hotspots. However, further analysis of the differences between the hotspots is lacking.

Apart from these pattern analyses, some algorithms for finding and recommending potential pick-up locations are proposed. Various improved clustering algorithms are another commonly used technology. Berdeddouch et al. [4] utilized technologies of K-Means and regression to discover potential pick-up locations that are easier to find passengers. Zhang et al. [43] improved spatio-temporal clustering based on K-Means and tried to find a set of personalized pick-up locations taking drivers’ preferences into consideration by combining pick-up data and POI’s attributions. Xu et al. [38] considered the information on taxi trajectory and improved the algorithm based on HotSpotScan and Preference Trajectory Scan algorithms. These methods take into account a variety of factors to recommend pick-up points. However, a deeper analysis of the importance of different factors is necessary. Machine learning methods, such as deep learning [18], reinforcement learning [39], DeepFM [34] also play an important role in potential location discovery. However, the model’s output and access to information are limited, which confuses the user about the results. Although the above research has achieved good performance, these studies lack transparency.

In summary, existing studies are difficult for analysts who lack mathematical domain expertise to analyze. In addition, the current methods usually generate pick-up hotspots by various algorithms while further analysis of the differences between the hotspots is lacking, which also needs interpretation. It is difficult for them to intuitively display the temporal and spatial variation of the pick-up point. In this paper, we focus on the visual analytics of the taxi pick-up data and try to find different patterns of pick-up selection, which help with the setting and recommendation of the pick-up point. We designed multi-scale analysis flow and interactive operations to analyze pick-up point selection, which can help users to conduct multidimensional analysis and judgment themselves.

Thanks to the development of fruitful advanced location-sensing technologies for collocating a vast amount of traffic data, the status of the moving objects, is recorded from both spatial and temporal dimensions [12]. The analyses of these data have made many contributions to solving urban problems, such as bus route planning [35], and taxi operation [47, 20]. However, due to the complexity of urban problems and the multidimensionality of traffic data [11], some of these methods may not perform well without the involvement of domain experts. The combination of data visualization and urban computing methods enables experts to explore traffic data interactively [9].

Data query focuses on developing new visual query models to quickly query traffic data and explore traffic information. Representative works include [2, 19, 13, 31]. Filtering, sampling, aggregation, etc. combined with interaction [6] can quickly present the results the user wants. What’s more, the main task that most studies focus on is pattern mining, committed to enabling analysts to obtain patterns and insights from big data [22]. Zeng et al. [41] explored the relationship between human mobility and POIs. Deng et al. [8] studied the cascades of spatial contagions. At last, different from pattern mining, some studies have tried to take advantage of the information extracted from pattern exploration for decision making. Weng et al. [35] proposed a visual analytics system to generate optimal transit routes interactively. Liu et al. [21] solved the problem of comparing solutions rapidly for billboard placements. Visualization helps to obtain useful information and interaction is allowed to participate in the process of generating decisions, making the system perform better than the algorithms.

The success of visual analytics is inseparable from the help of three factors: the proper visual representation of data, good comparison design, and human interaction. Firstly, various proper and novel visualization designs of spatio-temporal data have been proposed to facilitate the efficient completion of data analysis tasks. Visualization of spatial properties is often map-based [13, 22], such as heat maps. Rendering and aggregation are effective solutions to visual clutter [44, 11]. For temporal properties, axis-based design is a common visual form [13]. ThemeRiver [17] and horizon map [28] can compare multiple properties over time. What’s more, calendar [27] and radial layout [45, 2] can be a good option for periodicity, which can provide a distinct comparison of different periods. Additionally, a good comparison design is essential to grasp the difference. Juxtaposition is the most commonly used method for comparison [26, 21, 35]. Superposition overlaps multiple objects to show the difference [13]. Explicit encoding presents differences visually. For instance, the grid heat map encoded by size and color in [3] showed the pattern similarity between different time periods. At last, interaction operation helps with combining human knowledge. Basic interactions like clicking and boxing can promote exploration from overview to detail, such as the flow matrix view in [35], which can be clicked for more information. What’s more, interactions such as sketch [3] can make exploration more interesting, enabling people to interact with systems.

Although these studies have been proven effective, they cannot directly used for our visual analytics tasks. For the taxi data used in this paper, relatively little research has been done on pattern analysis of passengers’ behavior for pick-up selection. It is still a challenge to explore enormous pick-up points. Problems such as visual redundancy and poor scalability are easy to appear. Therefore, we combined the multiple visualization techniques and designed a novel visual analytics system, which enables users to interactively explore pick-up selection behavior at multiple scales. In particular, a juxtaposition comparison view was designed to interact with plenty of pick-up points and gathering locations, depicting different patterns in different regions.

In this research, T-PickSeer is constructed from three types of data. i) We utilize the road network of Shenzhen and geographical data from the open street map¹¹1https://www.openstreetmap.org/. ii) We retrieve Points of Interest (POIs) data from the Baidu Map Service²²2https://api.map.baidu.com/lbsapi/getpoint/index.html. The dataset includes 190362 POI locations where each record contains the longitude, latitude, name, address, and functionality of a structure in the urban environment. iii) We use the taxi GPS data from September 1, 2019 to October 5, 2019 in Shenzhen city. The raw data records the location of each taxi with a total size of over 200 GB. Each record consists of a timestamp, taxi id, longitude, latitude, speed, travel direction, and occupancy status.

• •

  POI classification: The raw POI data is categorized into 20 industry categories and 158 detail categories. To better focus on the POI information related to our scenario, we reclassified POI data into six categories: company, education (i.e., schools and universities), entertainment (i.e., restaurants, shopping malls, and tourism), living, public service (i.e., government agency) and traffic(i.e., station and parking lot).

• •

  Trip extraction: Taxi GPS data records a sequence of key points that a taxi passed by, which is unclear for the pick-up point (Origin) and drop-off point (Destination) of a single trip. So we need to extract each OD trip from raw GPS data. It consists of three steps: i) Data cleaning. Some records that are out of the study area or occupancy status changes abnormally, are eliminated at first. ii) Pick-up and drop-off points identification. It can be extracted according to the change in occupancy status. It means that someone gets off when the taxi’s status changes from 1 to 0, the opposite means someone getting on. iii) Attribute addition. To facilitate the analysis of pick-up point selection with surrounding POIs, we attach the category attribute of the nearest POI to each OD point. On average, about two million trips are extracted each day. The metadata of each trip record is shown in Table 1.

  Field	Description
  ID	The id of the taxi
  stime	Trip start time
  slocation	Location of pick-up point
  etime	Trip end time
  elocation	Location of drop-off point
  cO	Category of the nearest POI of the O point
  cD	Category of the nearest POI of the D point

• •

  Geographical partition: In order to facilitate pattern exploration, we divide the study area into equal-sized grids. Firstly we consider the shape of the grid. Hexagon is chosen because it is recommended as a better alternative to be used as a statistical unit [25]. The main advantages are i) that hexagons are the only geometric shape for regular tessellations that shares a real border with every neighbor and not only a single point with some neighbors [1] and ii) hexagons have minimal visual ambiguity, and have a positive effect on the memory performance of object location [10]. Then we tested the grid at different resolutions (200 m, 400 m and 1 km), finding that 400-meter hexagons provide representative samples which can guarantee the details of the pick-up point distribution and is in line with the passenger’s willingness to walk the distance. After partition, we assign taxi movements to each hexagon based on their original location.

To develop a feasible and practical approach for analyzing and improving pick-up points with visual analytics, we work closely with our collaborative experts (E1, E2, E3) to derive the analytical tasks. E1 is a Ph.D. candidate specializing in urban visualization and is also one of the co-authors. E2 is an analyst in a taxi company with over 10 years of working experience and has been involved in several taxi management projects. E3 is a researcher who has long been engaged in urban computing and visual analytics. From the feedback of these experts, we summarize a set of system requirements as shown below:

• T1

  Obtain an overview of global traffic distribution in a city. The users need to grasp the spatio-temporal distribution of pick-up points over the city. A summary of daily traffic volumes provides information for time checking (T.1.1). Then the distribution of pick-ups in different regions helps to understand busy regions for further exploration (T.1.2).

• T2

  Provide the spatio-temporal patterns for regional traffic. After understanding the global distribution of pick-up points, users need regional analysis. Multiscale temporal pattern at the hourly, daily, and weekly periods is noteworthy (T.2.1). Besides, it is necessary to explore the traffic within a region (T.2.2), such as the traffic volumes of pick-ups and drop-offs, and the direction of traffic. E2 points out that drop-offs are also important information, meaning that there are potential passengers subsequently or more drivers are here. These help the driver make decisions about cruising.

• T3

  Compare the behavior of pick-up point selection with different regions. Visual comparison of situations across different regions should be supported. It is necessary to compare different locations diversely to find distribution similarities and differences, facilitating the discovery of preference patterns for pick-up point selection.

• T4

  Show detailed information for point analysis. In a focused region, users expect to explore the preference of passengers in choosing different pick-up points. A set of criteria is necessary to explore and compare the performance of different pick-up points.

The temporal view (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection BehaviorA) is designed to configure the system and provide multi-scale temporal information (T1, T2). At the top (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm A_{1}$), a date selector is provided for users to select traffic on a particular day they are interested in. Then a view selector (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm A_{2}$) is provided to configure the visibility of different views flexibly through a set of checkboxes. To obtain an overview of global traffic distribution over time for date selection (T.1.1), a calendar heatmap (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm A_{3}$) is designed to summarize the daily traffic volume of the whole city. Each rectangle represents a day and the color is encoded as the volume of the daily traffic. The darker rectangle indicates the larger traffic volume in a day.

At the bottom of the temporal view (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm A_{4}$), we design a multiscale temporal pattern chart to better reveal the temporal distribution of pick-up points in the focused region (T.2.1). It shows changes of traffic at different time granularities. Figure 2 shows the design details. The layout of blue rectangles is similar to the calendar heatmap, with each row representing a week, and each column representing the same day of the week. For each rectangle, we use length to show the daily traffic volume. Furthermore, the hourly pattern is also worthy of attention. Users are concerned about changes in hourly traffic. First, we only highlight the peak hours with higher traffic than the average per day. We split each rectangle’s length into 24 equal pieces to represent the 24 hours in a day. We encode these hours with ticks and placed them in each rectangle from left (00:00) to right (24:00). For better comparing and analyzing, the rectangles in each column are centered, with dashed gray lines indicating 12:00. Then E2 suggested that he would also like to know the details of how traffic flows change within one day, which ticks cannot show. So we provide a choice to check traffic changes hourly in a day with a line chart, which can be viewed by clicking on the rectangle. The graph is initialized to the traffic volume for the entire city and then changes with the selected region.

The map view (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection BehaviorB) provides a spatial overview of pick-ups (T.1.2) and supporting regional exploration (T.2.2).

Heatmep view. Traffic in a large-scale city typically comprises massive trips. The heatmap (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm B_{1}$) is developed to assist with mastering the global distribution of OD trips quickly and identifying regions of interest (T.1.2). It encodes the density of pick-ups at each grid. The darker, the higher number of pick-up points. From the heatmap, users can easily find crowds from a coarse-grained perspective.

Glyph view. To reveal traffic pattern in a focused region (T.2.2), we attach a glyph view (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm B_{2}$) to the heatmap. As shown in Figure 3 C, the pie chart inside represents the comparison between the pick-ups (blue) and drop-offs (green) of each grid. Two outer rings visualize the drop-offs (green) and pick-ups (blue) in the grid by the geographical directions, with thickness presenting the flow size. Users can select regions with the lasso tool (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm B_{4}$).

Design alternatives. We considered several alternative solutions during our glyph design process. We first tried the OD matrix [36] to illustrate the movement of OD trips. As shown in Figure 3A, we tried to mesh the city map and map the original grids of the entire city into every single grid. The color was used to encode the volume of each grid area to other grids. But it has several drawbacks. First, it encodes the relative position of the destination, so it is difficult to fix the position of the destination intuitively. Second, the dense grid and colors make it hard to spot patterns. Moreover, we tried the flow map [30] as shown in Figure 3B. However, it introduced a visual clutter problem when there were too many grids. So in the end we chose the glyph design (Figure 3C), which was also recognized by the experts. It can effectively display the pick-up and drop-off information and flow direction in different grids.

Point view. To assist visual linkage and smoother operations, we present a point view (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm B_{3}$) to provide a spatial context for point analysis. Different types of POIs are marked on the map with various icons and a donut graph is placed around the POI to summarize the nearby pick-ups (blue) and drop-offs (green) for taxis. If users click the icon, the detailed information for the POI will be displayed (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm B_{5}$).

To compare the behavior of pick-up point selection in different regions (T3), the comparison view (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection BehaviorC) coordinates two graphs for two regions at the same time.

Figure 4 shows the design details. For each region, a glyph on the left (Figure 4A) summarizes the information for the region. The middle pie chart displays the proportion of all-day pick-ups (pink) and drop-offs (purple) in the region. The outer arc bars represent different categories of POI with height encoding the number. Afterward, users expect to explore the preference of pick-up selection, such as the preferred period and locations. Inspired by the 1-D beeswarm graph⁶⁶6https://observablehq.com/@fil/experimental-plot-beeswarm, we design a novel 2-D visual metaphor to represent the pattern in a region (Figure 4B). As shown in Figure 4B, the x-axis represents time, and the y-axis represents the pick-up or drop-off points. The background represents the pick-up (top) and drop-off (bottom) hourly. Then we count the number of pick-ups and drop-offs by an hour and aggregate them according to their POI attributes, trying to present the time-varying flow and spatial properties of the pick-up point. These points for each hour are encoded into circles with different colors indicating different POI attributes and placed in the corresponding grids. The area of the circle indicates the quantity. Furthermore, more detailed information about the corresponding trips is needed. Experts point out that drivers are concerned about the travel duration of an order. Therefore, a circle pack (Figure 4C) is provided to summarize the travel duration by clicking on each circle. We divide trips into four categories according to the time spent on the trip: 0-10 minutes, 10-20 minutes, 20-30 minutes, and more than 30 minutes. Each circle represents a category, with the area of the circle indicating the number of trips. The darker, the longer time spent on the trip. Then force simulation is used to keep all the circles tightly connected without overlapping for readability. In this way, the beeswarm graph can provide an overview of the temporal distribution and also enable detailed local inspection. In addition, the number of circles is sometimes large, affecting the efficiency of the comparison analysis. With this in mind, we provide checkboxes on the right to support point filtering by POI attribute, facilitating focus on points of interest.

The new beeswarm graph we design can effectively display the spatio-temporal information of the region and help to interact with a large number of trips, facilitating pattern discovery. However, according to the feedback from E1, when the selected region is too large, the circles overflow occurs. So we provide the option of a stacked bar chart (Figure 4D) for users to switch. Its layout is similar to the beeswarm graph, which can more clearly show the changes over time. The original background is pushed to the top and bottom.

Design alternatives. For the beeswarm graph, we have considered aggregating data by the taxi ID to present hourly information, hoping to alleviate large-scale OD data. As shown in Figure 5, we use each circle to represent a taxi, with the area of the circle representing the number of orders and the color showing the average travel duration. However, the visual presentation is confusing due to the still large number of taxis. It is difficult for users to effectively interact with circles. So we assign the POI attribute to each pick-up point according to the nearest POI. Then we aggregate the pick-up points according to the POI attributes, and distribute the pick-up points and the drop-off points separately on the positive and negative axes of y, which effectively solves this problem. At the same time, it provides an opportunity to explore the spatial characteristics of the pick-up points.

After obtaining the patterns of the region, point-scale exploration (T4) is necessary to compare the performance of each pick-up point and explore the preference for passengers in different pick-up points. We design the rank view (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection BehaviorD) to compare and analyze pick-up points, which can help with driver cruise guidance. In reality, POIs are usually recommended as pick-up points because they are easy to find. Considering the large number of historical pick-ups, as well as the habit of passengers taking a taxi, we use POI to replace historical pick-ups for analysis. Users can also add points on the map as a supplement. They are collectively called candidate points.

Firstly, a set of criteria is required to evaluate the performance of these points. According to literature research [7, 38, 46] and expert opinions, we propose a set of criteria. Our criteria are as follows. For each candidate point i: i) AD: Accessibility is assessed by the average distance from nearby historical points within coverage to that point. $D_{ij}$ in Formula (1) means the distance from the historical point j to i. $D$ is the radius of the coverage area. $n$ is the total number of historical points. Finally, candidate points with better accessibility have higher scores. ii) AS: Average traffic speed in the vicinity represents the traffic smoothness near the candidate point. A high score indicates a safe and comfortable ride experience. iii) PL: POI level is the proportion of nearby POI categories in all, indicating the convenience of taking a taxi. iv) TF: Transfer convenience, impacting passenger source, means the proportion of transportation facilities in all POIs. v) PR and vi) DR evaluate the probability of passenger arrival and driver discovery respectively. In Formula (2), $NP_{i}$ and $ND_{i}$ mean the number of historical pick-up points and empty taxis. Then we calculate the quantity per unit length and unit time. Finally, we normalize all scores to [0, 1].

Afterward, a rank list (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm D_{1}$) with six evaluation criteria is designed for comparing candidate points. The list presents each candidate point in a row and arranges its criteria in six columns. We encode the scores by the length of bars and provide a click action on the header of each column to rank the points by different criteria. Finally a radar graph (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm D_{2}$) is placed under the list to provide a summary depiction of all candidate points. Each axis from the center to the outer edge represents a score from 0 to 1. The scores for each candidate point are connected by light orange lines. And the violin plots colored in blue are attached to each axis, showing the score distribution of all candidate points for that criterion, which emphasizes the pick-up preference. Users can click points of interest on the rank list. Then the scores of the corresponding points will be highlighted with thicker orange lines in the radar graph (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm D_{2}$).

Rich interactions are provided for users to explore pick-up point data, which are summarized as follows.

• •

  Multi-scale navigation helps users navigate effectively across different scales. In the temporal view, users can set date (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm A_{1}$) and configure the visibility of views flexibly through a set of checkboxes (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm A_{2}$). In the map view, users can select regions with the lasso tool (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm B_{4}$) or set a point by clicking on the map (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm B_{6}$).

• •

  Highlighting enables users to focus on the information of interest, which is supported in the map view and rank view. For example, the selected region and point will be highlighted in the map view (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm B_{4}$, $\rm B_{6}$) and the rank view (T-PickSeer: Visual Analysis of Taxi Pick-up Point Selection Behavior$\rm D_{2}$).

• •

  Linking connects four views in the system. For example, after setting points on the map, new points will be added at the end of the rank list. When users click on a row in the rank list, it can be visually linked to the location on the map and the score on the radar chart.

We interviewed three aforementioned experts (E1, E2, and E3) individually and collected their feedback. Each interview lasted about an hour. First, a fifteen-minute tutorial was provided to participants, which outlined the functions, visual designs, and interactions of our system. Then, participants were allowed to freely explore provided data with our system for about forty minutes. After that, we collected participants’ feedback and suggestions. The feedback from the experts was valuable based on their expertise, which is summarized as follows.

Visual designs. The experts confirmed that our system is well-designed and could be easily understood by users with different backgrounds. In particular, E1 praised the comparison view. “ It is of great help to discover the spatial property characteristics of pick-up points,” said E1. E3 favors the multiscale temporal chart. He said, “ It provides abundant temporal patterns at different granularity, which is helpful.”

Usability & Effectiveness. All three experts agreed that our proposed method is helpful and effective to analyze the pick-up points selection of passengers. E1 mentioned, “This approach supports multi-scale exploration including the city, region, and point scale. The hierarchical analysis inspires my research, and the comparison view does really help in comparing patterns of selected regions. E2 praised that T-PickSeer is helpful to make use of large amounts of GPS data collected from taxis. “T-PickSeer is an effective tool that can provide an intuitive visualization, this improves the decision interpretation. Compared to the current tools, it allows me to easily discover more detailed information to guide drivers on the cruise,” he commented. E3 said that “T-PickSeer provides visualization designs which are also easy to follow. It is friendly to the analysts lacking professional skills in statistics and machine learning methods.”

Suggestions. During the interview, three experts also provided fruitful suggestions on improving T-PickSeer. Both E1 and E2 mentioned that the region segmentation function in the map view still needs more improvement. In the current version, T-PickSeer uses a hexagonal grid to segment the urban area based on the latitude and longitude, E1 and E2 suggested that it would be better to consider the spatial contextual information during the region segmentation. For example, they hope that a building or a unit (park, campus) can be assigned to a grid, which makes sense. Instead of a unit being divided into several parts, each part is in a separate grid. E3 said that it would be more effective if T-PickSeer integrated more recommendation approaches like mainstream deep learning models. In addition, he suggested that T-PickSeer could compare these different recommendation methods and give the corresponding optimal method under different scenarios.