Google and Harvard's AI Helps You Identify Unsafe Restaurants

Author


Google and Harvard have developed an AI which based on available feedbacks to identify potentially unsafe restaurants, helping you avoid food poison.

Have you ever feel anxious while going to the restaurant and not knowing the quality of food they might serve you? A new artificial intelligence (AI) has been developed by Google can give you a hand putting your anxiety away. A research conducted by the combination between the Harvard School of Public Health and the Mountain View firm depicts a model of machine learning model named FINDER, Foodborne Illness Detector in Real time. FINDER’s job is to identify possible unsafe restaurants.


Food poisoning happens frequently in America. It is estimated that thousands of Americans go to the emergency room each year due to it. However, FINDER can support local health departments and food establishments in solving these issues faster as well as preventing them from becoming a bigger health problem for the public.

According to the study's authors, FINDER receives combined and anonymous logs from users who choose to show their locations. It pinpoints search queries that indicate foodborne illness for example: “How to stop a stomachache?". Then it searches for the restaurants which were visited by the users who did the searches. Then it makes a calculation about the proportion of people who had evidence of food poisoning after visiting the restaurants via their searches.


Dealing with an ambiguous search term poses a challenge for the research team. For instance, a search for “diarrhea” on Google could possibly involve food poisoning, however, it does not indicate whole details of its symptoms. A solution for this issue was a monitored machine learning classifier that maximized those additional signals – such as results displayed in the search queries response,  total clicks on results, and what content it contains. The result is 85 percent accuracy in identifying the foodborne illness.
FINDER has been tested in two locations: Chicago and Las Vegas. Both cities health departments had a catalog of restaurants pinpointed by FINDER, which inspectors examined for health violations. For Chicago, FINDER inspected 5,880 food establishments, 71 of them were prompted. While in Las Vegas, a sum of 5,038 inspections were done, 61 of them were brought upon.

In the total restaurants identified by FINDER, 52.3 percent of them were deemed unsafe after the inspection as compared with the baseline of 24.7 percent of food establishments. Furthermore, FINDER restaurants seemingly posed a higher threat to be unsafe with all risk classification and had a larger number of violations.

All in all, FINDER performed better inspections, which is complaint inspections and routine inspections concerning the accuracy, scale, and latency (the time between people infected with foodborne illness and the outbreak).  Also, it gives the users the location of foodborne illness is more effective than the users themselves.


Researches show that people often have a tendency to blame the most recent restaurant they visited, and for that may be the reason why they choose the wrong one. Therefore, FINDER is a practical approach rather than individual complaints made by the customers as its unique methods of collecting information from various people who paid their visit to one food establishment.

The system is far from perfect. Due to the fact that food poisoning has a tendency to incubate for a long period, its symptoms could occur after a delay. FINDER only achieved the level of confidence after a considerable time had passed. Also, the odds of an unsafe restaurant identified by FINDER in a lower risk-sector posed a higher threat than its counterpart.

However, the tests illustrated an improvement of 68 percent on the category of advanced complaint systems that implemented data mining on Twitter.

The researchers further stated that their results provided substantial evidence that the tool could be used by public health departments these days for rapidly identifying and investigating where potential outbreaks might be occurring. Moreover, this model can be further expanded and developed. It should also be invested by health departments to ease the concern about food poisoning in the United States as well as the variety of other diseases spreading around the world.