0
$\begingroup$

First than all, I am not sure if this questions is more about Machine Learning, or if its Artificial Intelligence, if not, just let me know I will delete it.

At my company we need to create a solution for banks, where a client comes in and they want to to open a bank account.

They need to know if that person is a politician or political exposed person, maybe they work in the european comission, or they are family from a pep for example.

The business users has lots of data sources where to get these people, for example: http://www.europarl.europa.eu/meps/en/full-list/all

They want to train a Model (Machine Learning), where the end user can enter the name: Bill Clinton for example, and then the system has to return the percentage of a person being political or not.

Obviosly some persons are 100% politicial and the percentage will be 100%.

But if they enter a name that is not in any of their data sources, how would I train a model to decide if its pep or not?

quite confused

thanks

$\endgroup$
  • 2
    $\begingroup$ From just the name, even "Bill Clinton" is not going to be 100% political. In fact, probably less than 20%. Most names are repeated. According to Google, I am a well-known jazz musician in the USA, the author of a fishing book in Australia, a Microsoft developer in the UK. $\endgroup$ – Neil Slater Jul 11 '19 at 11:57
  • $\begingroup$ good point indeed $\endgroup$ – Luis Valencia Jul 11 '19 at 11:59
  • 1
    $\begingroup$ In a certain sense, this is like asking "given just the last digit, write an AI to determine if a number is prime". One, you do not have enough information to start with, and two, you skipped the step of determining whether an AI is in fact the right tool for the job. $\endgroup$ – MSalters Jul 12 '19 at 15:21
2
$\begingroup$

But if they enter a name that is not in any of their data sources, how would I train a model to decide if its pep or not?

Based on just a person's name and nothing else, the accuracy of this model is going to be very low. Consider that most first name, surname combinations in Europe are going to be repeated across the population.

However, the accuracy might still be slightly better than guessing. Some families and social classes could be more likely to be involved in political work, and a statistical model would pick up on that.

To train the model, take your positive names, and combine with a random selection of names of people that are known to be "not political" or similar enough. It doesn't matter if some of the names are the same provided you are confident in your data. Probably you could just take a phone directory or the electoral register or some other list of general names. Provided your "political" people are a small fraction of all people, this will work well enough even if you have some of them in the negative class.

Ideally you mix those name groups in the rough proportion that the bank expects to see "political" and "non-political" customers, so that your data set is a good representation of the target population.

Then you train a classifier on the names. As this is text and sequence data, you will need a solution for that. Possibly LSTM would be suitable architecture, but so might some feature selection from the names in a more simple ML model.

Remember to hold back some data (both positive and negative cases) for cross-validating and testing the model.

Expect the accuracy of your model when testing to be low. Very low. I would not at all be surprised to find the end result unusable by itself.

If this is seriously to be part of some bank's account setup process, there needs to be additional data used later in the process. A gate for additional checks based purely on someone's name will perform very poorly in my opinion.

| improve this answer | |
$\endgroup$
  • $\begingroup$ thanks for the answer, thats why I asked, I am not an expert, but guessing something with just the name seems very difficult if not impossible to have a usable result. $\endgroup$ – Luis Valencia Jul 11 '19 at 12:17
  • $\begingroup$ @LuisValencia: Yes, unless you pick a target that is strongly correlated with the name, such as what page of a telephone directory they will appear on, or whether two people are related (and this second one is never something you really want to guess with no other data) $\endgroup$ – Neil Slater Jul 11 '19 at 12:20
  • $\begingroup$ I'm not sure why this is upvoted. For a bank, a solution like this is criminally bad. Banks have a legal "Know Your Customer" obligation. Using something like this would be a reason for the supervising authority to summon the CTO for a chewing out. Seriously. $\endgroup$ – MSalters Jul 12 '19 at 15:19
  • $\begingroup$ @MSalters: I agree the situation is bad if this was used for real. That is not what the question is about however, and out of scope for AI Stack Exchange. I think the caveat in the last paragraph covers all that is necessary here. There is no evidence that anyone is considering doing this for real other than a thought experiment, and even if there were I think the correct place to handle those concerns would be Workplace, Security or other Stack Exchanges. All this site can realistically offer is "statistically the result will be very poor". $\endgroup$ – Neil Slater Jul 12 '19 at 15:40
  • $\begingroup$ @MSalters: If you can offer an improvement to the answer which covers concerns, but doesn't jump on the OP just for asking a "can AI really do this" kind of question (answer, as here: yes, sort of, but it would not be usable, or possibly: no, not really, but you might see the needle move) then I'd be happy to incorporate it provided it doesn't turn the whole answer into: "don't do this" $\endgroup$ – Neil Slater Jul 12 '19 at 15:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.