03. Video: How Do We Know Our Recommendations Are Good?

Training and Testing Data For Recommendations

In the last lesson, you were making recommendations by providing a list of popular items, or a list of items that the user hadn't observed but that someone with similar tastes had observed. However, understanding if these recommendations are good in practice means that you have to deploy these recommendations to users and see how it impacts your metrics (sales, higher engagement, clicks, conversions, etc.).

You may not want your recommendations to go live to understand how well they work. In these cases, you will want to split your data into training and testing portions. In these cases, you can train your recommendation engine on a subset of the data, then you can test how well your recommendation engine performs on a test set of data before deploying your model to the world.

However, the cases you saw in the last lesson, where just a list of recommendations was provided, don't actually lend themselves very well to training and testing methods of evaluation. In the next upcoming pages, you will be introduced to matrix factorization, which actually does work quite well for these situations.