SmartAdP: Visual Analytics of Large-scale Taxi Trajectories for Selecting Billboard Locations

论文
视频

作者

香港科技大学

Dongyu Liu
Huamin Qu

浙江大学CAD&CG国家重点实验室

Di Weng
Yingcai Wu

澳门大学

Yuhong Li

微软亚洲研究院

Jie Bao
Yu Zheng

摘要

由于缺乏有效的工具来进行深入分析以做出明智的决策，所以很长一段时间以来，广告策划人员一直难以立即制定解决方案并快速进行比较。在这项研究中，我们尝试使用可视化分析，结合最先进的挖掘和可视化技术，利用大规模的全球定位系统轨迹数据来解决这个问题。特别是，我们推出了SmartAdP，这是一个交互式可视分析系统，可应对两大挑战，包括在巨大的解决方案空间中找到好的解决方案，以及以直观的方式对解决方案进行比较。一个集成了新颖的可视化驱动的数据挖掘模型的交互式框架使广告策划者能够有效地制定好的候选解决方案。此外，我们提出了一组耦合的可视化:一个基于隐喻的glyph的解决方案视图，用于可视化不同解决方案之间的相关性；以紧凑方式显示广告牌位置的位置视图；和呈现解决方案的多类型排序的排序视图。该系统已经通过实际数据集和领域专家访谈的案例研究进行了演示。我们的方法可以适用于其他位置选择问题，例如使用轨迹数据选择零售店或餐馆的位置。

Introduction

Billboard advertising works because it is seen by everyone on the roads.
Appropriate locations

挑战：

How to acquire such comprehensive data in a cheap way
- Conventional approaches are expensive and time-consuming
How to formulate a good solution quickly?
- Large solution space
- Different opinions on the optimal
How to visually compare several candidate solutions to find the optimal one
- Which advertising strategy is better…
- What if I spend more $50000 to buy three more billboards…

贡献：

利用出租车轨迹数据对广告牌位置选择问题进行系统描述，并对设计要求和空间进行深入讨论和总结。
一个交互式框架，通过新颖的可视化驱动的数据挖掘模型和定制的特定于应用程序的数据索引机制来生成广告牌解决方案。
一套新的可视化技术，使最终用户能够探索多种解决方案的主要特征，并比较它们之间的共性和差异。

Related Work

Trajectory query processing
- New Query: finding points within user specified regions that can cover maximum trajectories (trajectories may have different weight)
- We propose a visualization-driven data mining model
Multicriteria decision making in a spatial context
- Integrations of computational methods with interactive visualization tools by multiple coordinated views have been suggested to solve MCDM problem.
- We first apply this mechanism on billboard location selection problem with taxi trajectory data.
Taxi trajectory visualization & Visual comparison
- We develop a visual analytics system tailored to the billboard location selection problem.

Background

Dataset

The Tianjin (China) dataset
- Road network data, 133726 road segments and 99007 vertices (Intersections)
- GPS trajectory data, 4 million trajectories
- POI data, 154633 points

Performance Indicators:

Traffic volume
Traffic speed
Traffic OD
Environment

Reach: The percentage of covered target trajectories among all target trajectories.
Opportunities to see (OTS)/Frequency
Gross rating points (GRP)

Task Analysis

Spatio-temporal distribution: How are the target trajectories distributed across the city? What is the difference between weekday and weekend?
Location recommendation: How many billboards should be placed in the target areas? Where are the optimal locations?
Location assessment: How good is a billboard location? Why is it selected for a billboard?
Solution assessment: How effective is a billboard solution? How does it satisfy customers’ requirements?
Solution comparison: What are the differences and similarities among multiple candidate solutions?
Solution classification: How many groups of candidate solutions exist? How these groups differentiate from each other?
Solution ranking: What is the ranking of multiple solutions? Which ones are optimal?

System Architecture

Model

Indexes for speeding up the mining algorithm
A GPS trajectory is a sequence of time-oriented spatial points

Mining Algorithm
Problem: identifying a set of locations with the maximum trajectory coverage value (the sum of the weights of all the covered trajectories)

Each trajectory has a weight
Each location has a cost
K-Location Query
Selection & Updating
t-Budget Constraint Query
Utilization ratio = current coverage value / cost

Visual Design

Solution Generator
- Dashboard View
- Map View
- Solution Preview
Solution Explorer
- Solution View
- Location View
- Ranking View
Interactions
- Details-on-demand
- Filtering and highlighting
- Linking

Evaluation

Conclusion

在本文中，我们系统地研究了利用大量轨迹数据来确定最佳广告牌位置的问题。通过与最终用户的密切合作，我们能够从广告牌广告策划人员面临的两大挑战中得出结论，即以一种即时且准确的方式创建和比较多种解决方案。因此，我们提出了SmartAdP，这是一个交互式可视分析系统，它结合了一个新的应用驱动的挖掘模型和几个设计良好的可视化和交互技术。我们进行案例研究和专家访谈来演示该系统。积极的反馈和深入的见解显示了我们系统的有用性和有效性。

Future Work

Improving the system
- Improve the mining algorithm
- Integrate additional types of data
Extending the application scenarios
- apply the approach to other resource allocation-related applications