Summary: combining the data of insurance companies with those of streetview found a correlation between the house where you live and how many accidents you can (the study does not disclose the results of house types)
"Tell me what it looks like the house where you live and I'll tell you how many chances you have of making a car accident." It sounds like a meaningless phrase, and it has its logic, to heed a new studio - pretty amazing! - ukasz Kidziski of Stanford University (USA) and Kinga Wojciechowska-Kita of the University of Warsaw (Poland).
Their research is based on observation of Street View images, and had work to find a possible correlation between the type of home in which a person lives and the likelihood that the same person has to be involved in a car accident .
The theorem Street View. The whole party by analyzing data from 20,000 people who had taken out insurance car in Poland between 2013 and 2015: the researchers and their team have tried, for each insured person, the address in Google Street View and, after having downloaded an image, they classified according to type (condo, detached house, townhouse etc.), the age and aesthetic conditions. Finally they cross these data with those supplied dallassicurazione to extrapolate a possible correlation with the likelihood that an insured presented a complaint.
An image comes from the study that has identified a link between type of dwelling and probability of incurring a traffic accident. | University of Warsaw
Surprise effect. The results? Surprising. Kidziski and Kita-Wojciechowska found that the type of house they live in an insured a good predictor of the likelihood it will present a complaint allassicurazione: The visible characteristics of a picture of a house can be predictive (but the study does not reveal in detail how, ed) the risk of motor vehicle accidents, regardless of the classically used variables such as age or zip code. Even know the status dellabitazione could rise by 10% the probability that unassicuratore able to identify a customer likely to file claims. And according to the researchers, the current study only a "test": its accuracy could be improved by using more and using more precise analysis of the data amount.
And privacy? Using data such as dellabitazione faade where you live - with current laws - it goes far beyond the limit of what can indulge unassicurazione against those who choose it. "Consent given by customers to the company to store their addresses do not necessarily mean a consent for storing information on the appearance of their homes, say the researchers themselves. And the correlation between home and accidents could open a Pandora's box: other firms could take advantage, with good peace of privacy "the insurance industry could be rapidly followed by banks, since there is a proven correlation between insurance risk models and credit risk score.
Vote for Street! Not the first time that Google's Street View data is used to steal sensitive information and / or confidential. Two years ago, another researcher at Stanford, Timnit Gebru, used the Google Street View images to see how they voted some cities of the United States, starting from photos of parked cars. Thanks to an algorithm, the Gebru team failed to grasp the correlation between the types of vehicles and the data from the US Census and the mode of voting of the presidential elections in 35 cities, in each district examined.
The question we wanted to answer was: given the pattern of the vehicles in an area, the algorithm could accurately predict demographic data registered in the US census data and the presidential vote? The answer was s: Looking motor vehicles classified in every neighborhood, we can infer a wide range of demographics, socioeconomic attributes and its residents, political preferences, "the researchers explained. In this case, unlike the study that showcases relationship homes with customer profiles of car insurance, had been provided some more details more: for example, the saloons were more closely associated with the Democrats, while the pickup is "abbinavano" of pi to the owners who voted Republican. "We found that walking a city for 15 minutes counting sedans and pickup trucks, can reliably determine if the city voted Democrat or Republican, "says Gebru.
Results in both cases that raise questions about the way of sets of seemingly innocent data can filter out personal information and how organizations should be able to use them. And most importantly, what you can find us using a simple web app or a social network?
From Focus