AI - Best of both worlds

I love the power of an AI model. That power comes from the extent of the data and the problem to solve. There are plenty of structured data problems in business that don’t require a deep learning model, but when unstructured data presents itself, it can add value to predictive models. What type of business problems have unstructured data? Everything along the customer lifecycle, because email text is typically available and is a good place to start.

If you’re maintaining a production model, you may periodically test new input variables without requiring a large refactoring effort. That’s ideal because performance may improve at minimal resource cost. The concept in this case is similar, with little disruption to the production model but a bit of work up front on the deep learning model. Luckily, it’s pretty easy to incorporate AI into existing models in production. The best of both worlds, one might say. The following is an example process incorporating email data.

• START WITH PRE-TRAINED EMBEDDINGS DATASET

• PREPARE EMAIL TEXT DATA

• ADD LAYERS TO EMBEDDINGS TO TRAIN RELEVANT EMAIL TEXT

• NAME THE INTERMEDIATE LAYER BEFORE OUTPUT LAYER

• FIT MODEL WITH ADDITIONAL LAYERS

• DEFINE NEW MODEL THAT RETRIEVES NAMED INTERMEDIATE LAYER

• PREDICT INTERMEDIATE OUTPUT ON EMAIL TEXT
  

The predicted intermediate output of the final step is a set of numbers in structured form that represents the trained text data. What does that give you? New variables to add to your current predictive model. If you’re able to handle a full refactor, there could be more powerful models to test. But if you’re looking for a quick win utilizing untapped text data, this could be an exciting step into gaining AI insights.