What would you recommend if I wanted to start learning about Deep Learning to become proficient enough to be able to build a fraud detection system?<p>As it is right now, the NGO I work for has a manual screening process to find possible fraudulent profiles. I would like to be able to use a Deep Learning based system to help them with that task.
The first question you need to ask is, why do you think you want to use Deep Learning? DL is hardly the only machine learning technique you might apply in this domain, and it might not be the best. If you don't have a background in Machine Learning, I'd suggest starting with a broad based overview of some sort, to give you some context for choosing a technique, instead of just going with what seems to be trendy.<p>All of that said, Deep Learning can mean different things, but to the extent that people apply the term in association with neural networks, a "deep" neural network is just a neural network with multiple hidden layers. So if you learn about basic neural networks with back-propagation, you are implicitly learning about deep neural networks. To that end, the Andrew Ng and/or Geoffrey Hinton classes on Coursera are pretty good for getting your feet wet (and the Andrew Ng class is also something of a "broad based overview" as mentioned earlier).<p>There is also a DL specific class using TensorFlow, running on Udacity.<p>Anyway, all of <i>that</i> said... the one thing about neural networks is that they are a supervised learning approach, which means you need labeled training data to train on. Now that in itself might not be a big deal... you can probably find some old transactions that were flagged as fraudulent in your system, and train against those. BUT... this means you are pretty much always vulnerable to <i>new</i> fraud techniques, and you have to always create new training data and re-train the model, etc. which may involve a lot of human effort. That is, this approach has little ability to adapt on its own.<p>There are other algorithms which are usually lumped into the broad category of "anomaly detection" that might be a better fit for doing fraud detection (or you may actually want to use both), where the algorithms don't need to be pre-trained using labeled data.
A couple of thoughts:<p>1. If the process is being done manually, there might not be an adequate corpus to train a deep learning tool.<p>2. The business goal is not to detect fraud with deep learning. The business goal is to detect fraud.<p>3. Computational techniques that assist humans in detecting fraud are a way to achieve the business goal. Deep learning is just one computational technique.<p>4. Quick access to records of previous fraud for comparison by humans might improve detection. It's also a step in the direction of developing a training corpus because it might identify relevant features.<p>Good luck.
If you're over step 0 and are pretty sure deep learning is right for the job, I suggest steps 1: download Torch, and 2: get comfortable with putting your data into Tensors. As a software developer I found Torch much easier to learn than alternatives like Tensorflow, probably due to the comfortable imperative style.