Computer Science Master's Degree Thesis

Title: Leveraging Spatio-Temporal Traffic Patterns to Enhance Travel Time Estimation

Institution: University of Bonn, Institute for Informatics 3

Summary

This thesis focuses on improving travel time estimation (TTE), a key component in smart city mobility applications. The work introduces a pattern-aware ensemble approach that combines state-of-the-art TTE methods (TEMP, LightGBM, DeepTTE) through a novel weight calculation mechanism informed by spatio-temporal traffic patterns. A custom clustering algorithm was developed using a distance matrix that captures the interplay of multiple traffic-related features. Experiments on the Porto Taxi Dataset show that the proposed method achieves a mean absolute error of 51.39 seconds, outperforming the conventional ensemble baseline as well as two state-of-the-art models. While the approach does not consistently surpass the best single method across all metrics, results highlight the potential of incorporating traffic pattern analysis into ensemble TTE systems and point toward promising directions for future research.

The main goal of this thesis is to enhance travel time estimation by integrating an ensemble-averaging approach with pattern extraction. More precisely, we aim to extract patterns in a given trip data by clustering, and use this information to build targeted predictive models in an ensemble. Throughout the thesis, we investigate the following questions:

1. How can we learn (spatio-temporal) patterns from trip data?

2. How can we incorporate the information from learned patterns into an ensemble approach to improve travel time estimation?

3. Can the pattern-based ensemble approach enhance travel time estimation?

4. How does changing the group of features for extracting patterns change the overall performance of the approach?

The thesis, LaTeX source code and defense slides

Here is a link to the Pdf file to the thesis. Here you can find the LaTeX source code for the thesis. The presentation for my thesis defense can be found here.

The model architecture

The architecture of the proposed model consists of four main building blocks, as shown in the following figure: data preprocessing, prediction generation, pattern extraction and pattern-based weight computation.

1. The preprocessing block is responsible for data cleaning, feature engineering, feature transformation, feature selection and partitioning.

2. In the prediction generation block, $n$ state-of-the-art methods for travel time estimation are implemented and trained on the preprocessed data independently from each other. Each predictor returns a prediction $p_i$ for $i ∈ [1, n]$.

3. The pattern extraction block employs a clustering algorithm to extract patterns from the dataset, and divides it into qualitative classes. The clustering ensures that similar trips are grouped together within the same pattern class.

4. The resulting class labels and predictions are combined in the pattern-based weight computation, which decides on how strongly or weakly a predictor should affect the final prediction of a trip based on its performance in the pattern class it belongs to.

Pattern Extraction

We propose the following distance measure: $D_{i,j}^{\text{all}} = \alpha_{\text{traj}} D_{i,j}^{\text{traj}} + \alpha_{\text{temp}} D_{i,j}^{\text{temp}} + \alpha_{\text{categ}} D_{i,j}^{\text{categ}}$ where $D_{i,j}^{\text{traj}}$ is the distance between trip $i$ and $j$ based on their trajectory, $D_{i,j}^{\text{temp}}$ based on their temporal features, $D_{i,j}^{\text{categ}}$ based on their categorical features. For more information on the distance metric and it's motivation, please have a look at Chapter 4 of the thesis.