PhD Defense of Jesper – Data Driven Mobility

Four years ago, Jesper joined the Datalab to combine data science at Pon with PhD research at the Vrije Universiteit of Amsterdam. Through close collaboration with various Pon companies, he was able to develop novel methodology and implement part of it at Pon. The research was published in multiple papers, combined into a PhD thesis, and successfully defended in the aula of the VU in Amsterdam.

The thesis is titled “Data Driven Mobility” and revolves around extracting value from data through various analysis. These analyses start at so called descriptive analysis, which aim to describe what has happened in the past. The next level is predictive analysis, which aim to predict what is going to happen in the future. The final and perhaps holy grail are prescriptive analysis, which aim to prescribe what to do, as to anticipate future events.

The research was executed in close collaboration with Pon. Four research tracks with four Pon companies were set up: with Shuttel, PAH, Gazelle, and FOCUS.

The collaboration with Shuttel revolved around mobility choices. We analyzed what type of mobility is most relevant in which area in the Netherlands and predicted which travel mode (car or public transport) is most relevant for any trip between two coordinates. The results are promising, we can predict the correct mobility type with 97% accuracy.

The second analysis involved predicting the impact of changes in the automotive dealer network. The research was executed in close collaboration with the PAH network management team. The methodology developed can compute the impact of various network scenarios, like adding or removing a dealer in certain areas.

The research track at Gazelle started by analyzing the point of sales inventory data of dealers. This data is noisy, as the automated systems do not share the data as consistently as desired. Through a novel outlier detection methodology, we can extract reliable data from the raw data.

The final research track focused on sales forecasting for Gazelle and FOCUS. Existing methodology did not yield desirable results, thus, we investigated whether new machine learning methodology can help. The model developed combines the data of all models of which historic data is available, and generates predictions that are ±10% more accurate than existing methodology.

  • Jesper Slik

    Senior AI Specialist