Software Engineering Stack Exchange is a question and answer site for professionals, academics, and students working within the systems development life cycle. It only takes a minute to sign up.
Sign up to join this community
I have 10,000 addresses from a city, which all have a region field in the database. When a new address is entered, then I want my software to automatically detect the region of the address .
I think it should be implemented with some sort of machine learning algorithm. How can I do this?
And with every newly inserted address, the machine should learn to detect the region of the new address. Is there any library for machine-Learning algorithms (like aforge.net for neural networks)?
I think it should be implemented with some sort of machine learning algorithm.
Nope
How can I do this?
Use a shapefile with polylines of the regions (they are more or less files full of coordinate pairs with a bit of metadata associated). Use something like the Google Maps Geocoding API to geocode the address (you send an address, and it sends back a coordinate pair). Write a simple algrithm* to determine which polygon from the shapefile the geocoded coordinates lie within. You can find shapefiles all over the web, especially from government agencies such as NOAA. The USGS has a decent collection too. I believe this solves the problem without breaking any of the laws of robotics, so I would not even bother with an AI-oriented solution. :)
*I would start here for a good reference to get you started. Also, do not forget that the earth is curved, so distance calculations work a bit different than on flat plane (think radians).
I think there're clear rules on how cities are partitioned into quarters or regions. You should ask your local administration on where they draw the borders. Then you could, for example, retrieve the location data of the address (latitude and longitude might work) and simply check in which region's boundaries this address is in. There's no need for a learning algorithm for this problem.
However, if you can't acquire the boundary data for the partitions then I'd try to find the nearest known region, probably by distance. Again, I see no sense in an evolving algorithm or some sort of AI here. Either you can determine the region deterministically by known boundaries or you can try to find the nearest known region. AI is imho an overkill for such a task. You'd have to constantly recalculate initially guessed region-boundaries and evaluate them and then update existing addresses of which the region is known to be uncertain. Also, you'd have to feed the system constantly with addresses of which the region is known to verify uncertain regions.
But as regions are very unlikely to change their borders, I'd just try to obtain the boundaries, like stated above, from the local administration.
If your addresses are poorly formed, like this extreme example:
John Q. Public
Lives on the coner of west and main
The city 20 miles east of New York.. Sally is the mayor
I would suggest looking into a natural language processing toolkit like OpenNLP. Then, you could build up a corpus of these poorly formed addresses and train the algorithm on the corpus.
But in most cases, Falcon is right, and there's no need to involve AI.
You're trying to classify the addresses, and associate the classification groups with regions.
You could pour your 10,000 addresses + regions into a random forest. Or build several for an ensemble. The trick would be how to build the inputs: you might have to use a "word bag" approach, with a boolean for each street name and a few fields for the discreet values like street address. That would be a big input, but that's OK; sometimes the features of a training set can run into the thousands (or more).
Split your data up into training/testing sets, though. Pour 9,000 of the addresses into the random forest, then use the other 1,000 to test it and see what % accuracy you get. There are fancier ways to split it up, but that's a good start.
In Python, scikit-learn is always a good choice. sci-kit learn will have other classification schemes that might be even better than random forest for this task.
