PROJECTS
T-patterns
Discovering Temporal Patterns
​​
In the insurance industry, timely and effective interaction with customers are at the core of everyday operations and processes that are key for a satisfactory customer experience. These interactions often result in sequences of data derived from events that occur over time. Such recurrent patterns can provide valuable information that can be used in a variety of ways to improve customer related work-flows. This project proposed an algorithm to uncover such time patterns that takes into account the time between events to form such patterns. We use temporal customer data generated in multiple different use cases (e.g. customer satisfaction and fraud) to show that this algorithm successfully detects patterns that occur in the insurance context.
Papers related to project:
-
T-Patterns in Business (https://ssrn.com/abstract=3066839)
-
Discovering Temporal Patterns from Insurance Interaction Data (IAAI 2019, link coming soon)
​​
Collaborators: University of Wisconsin - Madison
Pedestrian
detection
Pedestrian Detection in drone images
​
Use of drones for home inspection is a common application in insurance industry. Pilots fly drones over insured houses and take pictures to access condition of house and roof. This brings some serious privacy and regulatory concern. We want to gather pictures from all perspectives of the house but at the same time respect privacy of neighborhood by not including people in the picture. To solve the problem, we build a pedestrian detection system that scans incoming images and removes people out of the picture. We build a semantic segmentation model that masks people at pixel level while retaining all other information in the picture.
​
Papers related to project:
-
Learning View Invariant Semantic Segmentation for UAV Video Sequences (SDM 2018)
Collaborators: University of Wisconsin - Madison
Mileage
Extraction
Mileage Extraction from odometer images
​
Odometer mileage is an important piece of information in auto insurance both for quote generation and claims processing. It is used alongside License Plate number and Vehicle Identification Number(VIN). Adjusters and customers need to fill out these three fields and upload supporting pictures for each. In an effort to make this process more user friendly, we plan to auto fill these numbers form the uploaded images. This will not only save manual work for customers/adjustors but also remove possible human errors during the process.
Odometer mileage extraction is a challenging problem because of the poor quality of uploaded images, often due to poor camera quality, non-uniform lighting and incorrect camera orientation. We design a robust odometer mileage extraction model by cascading two object detectors. The system is trained to detect odometer display in the picture and recognize characters inside odometer display. Finally, a post processing algorithm gathers the detected characters and extracts mileage number.
Storm
Loss
Storm loss Modeling​
​
Catastrophic events (CAT) such as large hail storms, hurricanes, wildfire, and tornadoes can occur with limited forewarning and can result in significant losses for insurance companies in a very short period of time. In addition, when such CAT events do occur, insurance companies are hard pressed to mobilize claims adjusters quickly to service significantly increased customer claim volume. Accurate prediction of the volume and location of expected claims resulting from CAT events can greatly help with resource planning and deployment so that affected customers are serviced in a timely manner.
​
We're developing models to predict volume and location of claims due to large hail storms. Using copulas and graph neural networks, the models account for the nature of the event and specifics of the exposures in the affected area, in addition to non-IID correlations and complications due to neighborhood clustering.
​
Collaborators: University of Wisconsin - Madison
Chatbot
Knowledge Representation for Conversational Agents
​
American Family has a few conversational agents (chatbots) for answering FAQs and executing workflows using natural language. We have built our own chatbot platform that is able to interactively refine and disambiguate incoming queries based on structured descriptions of the data. Currently, we are building the foundation for the next generation of American Family chatbots that can do automated reasoning, smart search and insurance-domain question answering using an enterprise knowledge graph.