To collect data Numbeo relies on user inputs and manually collected data from authoritative sources (websites of supermarkets, taxi company websites, governmental institutions, newspaper articles, other surveys, etc.). Manually collected data from established sources are entered twice per year.

They perform automatic and semi-automatic filters to filter out noise data. They utilize user behavior and previous data for the city/country to determine likelihood of a certain input whether it is considered as spam. There are more than 30 sophisticated filters which in use. The performance rate of the filter is enhanced once more inputs are included.

One of the advanced filters tries to eliminate bad training data. It digs into discarded data (spam data) and if it notices irregularities, it moves them back into the calculation. The algorithm which determines irregular spam data uses the following filter: if for a single item in a city exists a high number of classified spam data with a relatively small standard deviance from users that have more positive inputs, that means these data are misclassified and the algorithm fix it to the proper classification.

To summarize the filters, Numbeo uses heuristic technology to get the data quality. Using the existing data Numbeo periodically discards data which are most likely incorrect statistically.

Learn more here