Data-Driven Mobility Services

As part of the cooperation between ISMLL and DnA the following data-driven case studies are being investigated and the cooperation has led to the establishment of Data Analytics Research Center (DARC)

Ongoing Work Packages

Automatic Damage Assessment for Cars

Machine Learning as well as sophisticated object detection and image processing techniques can help in improving the mobile customer experience by knowing and avoiding turn-in penalties beforehand. This will be achieved by providing a smart digital assistant on the mobile phone device using its internal camera only. The idea is to streamline the lease-end process. Undoubtedly, there is more the customer can do besides washing the car and cleaning from the inside before the hand-over to minimize costly penalties. There are supposed to be almost identical, recurring damage patterns when returning leased vehicles which can be used to make predictions on current images based on labeled past ones (supervised learning). The initial model generation will be based on a large set of historic, labeled and priced pictures and inspection reports.

Asset Return Forecast

An important decision at the end of the leasing period of a vehicle is to identify ‘the next best use’ of the returned vehicle. Typically, this process begins after the return of the vehicle to the manufacturer. The ability to anticipate the return condition of these leased vehicles beforehand would reduce the delay in reusing the vehicle for the next use case. The return date, return mileage and return damage are key targets required by VWFS to determine the future use of a leased vehicle. Our work, predicts these key targets using machine learning algorithms (ML) from the leasing contract data.

Next Best Offer

In the Next Best Offer Workpackage we aim to research and develop recommender systems that can generate personalized vehicle recommendations for existing VFWS private and business customers. An increasingly large number of vehicles are offered from many brands under the Volkswagen organization with many vehicle and corresponding financing specific configuration options, necessitating the use of advanced Machine Learning based Recommender Systems to generate personalized recommendations for customers. Besides enriching the customer experience, there also additional business implications of up-selling and cross-selling across brands that can realize an increase in business value. From a research prespective, there are complex problems that require dedicated solutions since existing recommender systems solutions are either limited in the modeling capacity or are inapplicable in light of challenges such as using sequential information, modeling multiple vehicle specific configuration attributes as targets, providing robust predictions for new users and items with few samples and generating diverse yet applicable recommendations.

Used Car Recommendation Engine

Selling used cars is usually done after the end of the leasing contracts through online business-to-business auctions where dealers are presented with various options to choose from and bid on.

In order to enhance the dealers’ experience and improve the number of sales, a state-of-the-art attribute-aware multi-relational sequential model is developed by extending and adapting the recent SASRec model for auction settings. We trained the recommender model using the two available historical relational interactions between user and items, namely the purchase and bidding interactions. Using both relations provides the model with richer data to learn dealers’ preferences as both interactions are highly correlated.

The following steps in this project aim to improve the current system further and add more features to the model, which will lead to a better dealer's experience and achieve the business target in increasing the number of sales.

Completed Work Packages

Predictive Parking

By applying advanced Machine Learning techniques, it is possible to predict free parking spots and enhance with them FS parking apps. Hence, the integration enables a more convenient and seamless user experience for drivers. This in turn will increase the usage of the parking apps and possibly enable new pricing-options for the new service. The idea is to identify customer behavior, seasonality or peaks of parking behavior to predict available parking spots. This is done by using historical and real-time on-street parking data and transaction capabilities. In addition, existing data can be enriched by using publicly available open data, i.e. weather forecasts, city infrastructure or events, to improve the model.

Residual Value Forecast for Cars

Given the car’s configuration settings, list price and contract start date we want to estimate the car’s residual value at a future date X with estimated mileage per-year Y

Partner: Volkswagen Financial Services

Contact:
Lars Schmidt-Thieme,
Mofassir ul Islam Arif
Kiran Madhusudhanan
Shereen Elsayed
Shayan Jawed

• Mofassir ul Islam Arif, Felix Wieland, Chiara Bianchin, Andre Hintsches, Katrin Lange, Mohsan Jameel, Lars Schmidt-Thieme (2022):
  Object Regression: Multi-Modal Data Enhanced Object Detection for Leasing Vehicle Return Assessment, in The International Conference on Digital Image Computing: Techniques and Applications (DICTA).
• Shayan Jawed, Mofassir ul Islam Arif, Ahmed Rashed, Kiran Madhusudhanan, Shereen Elsayed, Mohsan Jameel, Alexei Volk, Andre Hintsches, Marlies Kornfeld, Katrin Lange, Lars Schmidt-Thieme (2022):
  AI and Data-Driven Mobility at Volkswagen Financial Services AG, in arXiv.
• Mofassir ul Islam Arif, Mohsan Jameel, Josif Grabocka , Lars Schmidt-Thieme (2020):
  Phantom Embeddings: Using Embeddings Space for Model Regularization in Deep Neural Networks , in LWDA.
• Ahmed Rashed, Shayan Jawed, Lars Schmidt-Thieme (2020):
  MultiRec: A Multi-Relational Approach for Unique Item Recommendation in Auction Systems, in 14th ACM Recommender Systems Conference (RecSys 2020).
• Mohsan Jameel, Mofassir ul Islam Arif, Andre Hintsches, Lars Schmidt-Thieme (2020):
  Automation of Leasing Vehicle Return Assessment Using Deep Learning Models, in European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD 2020).
• Shayan Jawed, Ahmed Rashed, Lars Schmidt-Thieme (2019):
  Multi-step Forecasting via Multi-task Learning, in Proceedings of the 2019 IEEE International Conference on Big Data (IEEE BigData 2019) In The IEEE Big Data Conference 2019.
• Mofassir ul Islam Arif, Mohsan Jameel, Lars Schmidt-Thieme (2019):
  Directly Optimizing IoU for Bounding Box Localization, in 5th Asian Conference on Pattern Recognition.
• Ahmed Rashed, Shayan Jawed, Jens Rehberg, Josif Grabocka, Lars Schmidt-Thieme, Andre Hintsches (2019):
  A Deep Multi-Task Approach for Residual Value Forecasting, in European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML PKDD 2019).
• Hadi S. Jomaa, Josif Grabocka, Lars Schmidt-Thieme (2019):
  A Hybrid Convolutional Approach for Parking Availability Prediction, in International Conference on Convolutional Neural Networks (IJCNN2019) .