Thanks to the cooperation between HDI and the AEGIS technical partners, the second Scenario of the Insurance Demonstrator continues to exploit at best the (updated) release of the AEGIS platform. The current results after the execution of the Scenario meet the needs of HDI, and at the same time open some interesting challenges that will be addressed with the final release of the platform and further improvements of the HDI tools.
The second Scenario of the Insurance Demonstrator is mostly related to the detection of a foreseen event of interest by the Event Detection Tool, differently from the previous one, in which the focus was on contacting the interested customers after a related event was just happened (anticipating their potential request). Through the AEGIS platform, the HDI Data Scientists can evaluate the risk exposure of the company by identifying the customers that could be affected by the events and the type of the policy/-ies held by them. Through the analysis of some features, for instance the number of accidents and the number of previous injuries, for each customer a priority value is assigned. The list of customers is then sent to the Web App for further processing by the HDI operators that will contact the customers.
The HDI Data Scientist may also send a push notification to the customers that have installed the HDI Mobile App, hence to those that had signed an AEGIS-specific terms and conditions agreement.
The notification received by the HDI customers, includes details of the event as well as information about further policy/-ies they could subscribe (type of policy, price, duration, restrictions) and contact references of the HDI agent that can support them. For this reason, every time a notification is sent to a customer, a mirroring notification is sent to the HDI Web App workspace of the responsible agent.
During the execution of the second scenario, an internal focus group has been organized at the HDI premises, involving data scientists and developers of the demonstrator, together for many iterations over the test cases, with the objective to provide a qualitative evaluation of the demonstrator.
Since the first round of evaluation, data scientists effectively run the Query Builder and Visualizer tools, correlating the features of the event with the in-house dataset regarding the customers’ policies and location, and having a fast overview of the interested customers on a map. They are now well trained in the use of the notebooks of the platform. The Web App that has been developed to allow the information exchange between the three actors involved in the process is working as expected, and the execution flow was tested by different users for each role and no issues were observed. The data scientists’ analyses within the AEGIS platform have also succeeded. The problems encountered in the first tests have been fixed in cooperation with the other partners of the consortium.
The general feedback is that the status of the demonstrator, although not yet fully integrated, is on a good direction and the different steps are clearly defined. The Event Detection Tool has been adequately trained for the second (medium) demonstrator in Italian, with the keywords ‘grandine’ and ‘grandinata’ (hailstorm), ‘polizia antisommossa’, ‘proteste’, ‘scontri’ (respectively riot police, conflict and protest). The number of tweets collected and labelled is around 5000 for each event type (including an enhanced training for the flood event, from 1000 tweets of the first phase to 5000 tweets) and neither retweets nor answers have been considered. The machine-learning algorithm of the Event Detection Tool has been improved. Developers highlighted the value of using the Event Detection Tool Python code potentially almost as-is for the streaming of tweets related to other keywords, languages and further scenarios. The use of the Web App by the data scientists has been effective and the tool has been considered clearly structured. The AEGIS platform was used for creating projects and upload datasets, and the offline Anonymiser was exploited for the management of sensitive in-house data.
The Mobile App has been used and tested, with its two basic functionalities developed ad-hoc: geolocalisation and push notifications, as well as the creation of the .csv file with customer reference name, current latitude and longitude, and the achieved results are satisfactory.
The main issue related to the second (medium) demonstrator and, in general, to the Insurance demonstrator is the privacy and security regulations. In order to respect the Italian and European Legislation about data treatment, and the Insurance specific policies, the in-house datasets stored in HDI databases are uploaded on the platform after their anonymization. The data are managed and handled only by the HDI employees that are working on the AEGIS project and only the columns of interest for the project evaluation purposes are kept. Additionally, at the end of each analysis the in-house datasets are deleted from the AEGIS platform. Finally, a potential issue in the use of the mobile app is represented by the possible reluctance of the customers to give their consent for enabling the geolocation. In general, the second (medium) demonstrator has satisfied all the HDI actors, in terms of usability of the tools and accuracy of the analysis. Toward this end, the cooperation between the HDI employees and the technical team of the project was fundamental, from the definition of the user stories (D3.1) to the test cases execution.
The complete description of the Second Insurance Scenario Execution is available in the AEGIS Deliverable D5.4 – Demonstrators Evaluation and Feedback – v2 (to be published at https://www.aegis-bigdata.eu/public-results/).
Blog post authors: HDI – GFT