Example of binary classifier with numerical features using deep learning

Question

I would like to get more understanding of deep learning. Browsing the web I find applications in speech recognition and hand-written digits. However I would be interested to get some guidance on how to apply this in the classical setting:

• binary classifier
• numerical features (each sample is a numerical vector of $K$ entries, no 2D pixels or such).

I am doing my own experiments choosing learning rates, number of hidden neurons and so on, but I would be happy to see an application by somebody more experienced.

The software that I use offers weight initialziation using Restricted Boltzmann Machines (RBMs). I wonder whether this is useful in this context and whether the other special techniques that one encounters in the literature (convolutional NN) are useful here to.

Could anybody share a blog post, a paper or personal experience?

Answer 1

I used Binary classification for sentiment analysis of texts. I converted sentences into vectors by taking appropriate vectorizer and classified using OneVsRest classifier.

On another approach, my words were converted into vectors and there, I used a CNN based approach to classify. Both when tested on my datasets were giving comparable results as of now.

If you have vectors, there are already really good approaches available for binary classification which you can try. On Binary Classification with Single–Layer Convolutional Neural Networks is a good read for you for classification using CNNs for starters.

This is one of the first blogs I read to gain more knowledge about this and doesn't require much of pre-requisites to understand(I am assuming you know the basics about convolution and Neural Networks).

Answer 2

Deep learning is a fancy thing now in ML since it has been outperforming other ML algorithms in many respects. Convolutional neural networks is one of the methods to implement Deep learning and it is highly applicable to different data types such as images, signals (time series) and text.

I mainly use CNN for images and signals. In my application I have done binary and multiclass classification. The theoretical background to both is the same. Depending on the problem and data at hand you may want to break your multiclass classification to simple binary ones and the combine them at a later stage. These are things that comprise model selection.

Applying deep (neural) networks to a feature dataset can be thought in two aspects. One is applying to text data after vectorizing as explained by @HimaVarsha. I have not worked with text data, but my applications has mainly coming from images ,signals or catalogues (astronomy catalogues). I have tried using catalogue data which is basically numerical features (2nd aspect) in deep neural networks but did not get significant improvement in results as made the network more deep -- it can be either the problem with my design or maybe the features itself. (Also this method is not often considered a way of learning in deep learning, because deep learning is more of learning directly from the data. But it's another philosophical question) With convolutional neural networks I have got nice results with deeper layers.

In addition to the blog link in the first answer I would also suggest this. Keras blog also recommended if you would like to look into CNNs.