Building Machine Learning Powered Applications by Emmanuel Ameisen

Author:Emmanuel Ameisen
Language: eng
Format: epub
Publisher: O'Reilly Media
Published: 2019-06-15T16:00:00+00:00

Dimensionality Reduction for Errors

We described vectorization and dimensionality reduction techniques for data exploration in “Vectorizing” and “Dimensionality reduction”. Let’s see how the same techniques can be used to make error analysis more efficient.

When we first covered how to use dimensionality reduction methods to visualize data, we colored each point in a dataset by its class to observe the topology of labels. When analyzing model errors, we can use different color schemes to identify errors.

To identify error trends, color each data point by whether a model’s prediction was correct or not. This will allow you to identify types of similar data points a model performs poorly on. Once you identify a region in which a model performs poorly, visualize a few data points in it. Visualizing hard examples is a great way to generate features represented in these examples to help a model fit them better.

To help surface trends in hard examples, you can also use the clustering methods from “Clustering”. After clustering data, measure model performance on each cluster and identify clusters where the model performs worst. Inspect data points in these clusters to help you generate more features.

Dimensionality reduction techniques are one way of surfacing challenging examples. To do so, we can also directly use a model’s confidence score.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.