One of the most exciting recent developments in the area of predictive analytics is called deep learning. Having already advanced the state of the art in many areas such as image recognition, deep learning is now poised to deliver cutting edge results for business application clients in the real estate, mortgage, and auto lending industries.
When building a statistical model, access to quality data is key. But few data sets in industry are fed directly into a machine learning algorithm. Instead, the raw fields are combined and transformed to produce features, which highlight important aspects of the data and problem being addressed to a model. Features are also a way to incorporate human business expertise into a system.
One of the challenges with features is that their creation can be time consuming, and nowadays delivering a high performing, robust system quickly is especially important. It is also difficult to be sure if a hand created feature set fully captures the data, and using more brute-force approaches to feature creation leads to a dependence on feature selection algorithms.
Enter deep learning. By training and stacking multiple layers of Restricted Boltzmann Machines or auto-encoder neural networks, a system can now construct its own complex internal representations or features automatically. And these layers can be built without the need for data tags or labels, which are frequently unavailable or incomplete in big, real world data sets. Upon the newly learned features a more typical modeling technique such as logistic regression is then used to produce the final business score.
Here at Point Predictive Inc., deep learning is one of the approaches we use to build predictive features and models more quickly. Not only faster, the transformations learned are not typically present in hand crafted feature sets, leading to more complete models. When coupled with solution delivery in the cloud, deep learning is helping to deliver strong model performance within short project timelines. Can deep learning play a role in your next analytic business application?