Skip to content

Global Explainability

Fiddler provides powerful visualizations to describe the impact of features in the model. Feature importance can be found in the Explain tab, or in the Analyze tab.

Global explanations are available in the UI for structured (tabular) and natural language (NLP) models, for both classification and regression. They are also supported via API using the Fiddler Python package. Global explanations are available for both production and dataset queries.

Tabular Models

For tabular models, Fiddler’s Global Explanation tool shows importance/impact of the features in the model.

Two global explanation methods are available:

  • Feature impact: gives the average absolute change in the model prediction when a feature is randomly ablated.
  • Feature importance: gives the average increase in loss when a feature is randomly ablated.

Feature impact/importance is displayed as percentage of all attributions.

The following is an example of feature impact for a model predicting likelihood of successful repayment of a loan:

Tabular

Language (NLP) Models

For language models, Fiddler’s Global Explanation performs ablation feature impact on a collection of text samples, determining which words have the most impact on the prediction.

For speed performance, Fiddler uses a random corpus of 200 documents from the dataset. When using the Fiddler Python package, the argument n_inputs can be changed to use a bigger corpus of texts.

Two types of visualization are available:

  • Word Cloud: displays a word cloud of top 150 words from a collection of text for this model. Fiddler provides three options:

    • Average change: it’s the average impact of a word in the corpus of documents. It takes into account the impact directionality.
    • Average absolute Feature impact: it’s the average of absolute impact of a word in the corpus of documents. It only provides the absolute impact of the word and not its directionality.
    • Occurrences: number of times the word is present in the corpus of text.
  • Bar chart: displays the top n words impact. By default, only words with at least 15 occurrences are displayed. This can be increased and will be reflected in real time in the bar chart. Fiddler provides two options:

    • Average change: it’s the average impact of a word in the corpus of documents. It takes into account the impact directionality. Since positive and negative directionality can cancel out, Fiddler provides a histogram of the individual impact by clicking on the word.
    • Average absolute Feature impact: it’s the average of absolute impact of a word in the corpus of documents. It only provides the absolute impact of the word and not its directionality.

Here is an example of word impact from a sentiment analysis model:

NLP


  1. Join our community Slack to ask any questions 

Back to top