MagPunk
EN
Understand MagPunkmagpunk.com / en / understand
InteractiveDocument
Contents
MagPunk answersHow It Reads Content2.3.6

“Why This Label?” — Explainability

It is the transparency screen showing exactly with the contribution of which word the machine learning model on your device gave that label to a news item.

When the system assigns a label, there is no black box behind it. When you press the ‘Why this label?’ button, you clearly see from which word the model received how many points while making that decision. As a requirement of the transparency invariant; the sum of the mathematical contributions of the words and the base value of the system (Σ(word contribution) + base = score) is always equal to the final score of the label.

Example: Seeing transparently on the screen that the word ‘earthquake’ in a headline provided a direct contribution of +0.42 points to the model.

Go deeper

Most modern deep learning systems hide their decisions behind inexplicable massive matrices. The lightweight (≈7 MB) classical classifier in MagPunk, on the other hand, is wholly explainable due to its linear features. This feature allows the user to question the algorithm’s decisions. Being able to see how the algorithm thinks (or how it errs) is the most effective way to keep in mind that they are statistical prediction tools, instead of blindly trusting artificial systems.

Up next2.3.7 Rule Engine: Fallback System if the Model Fails to Load

2.3.6 M6: Why This Label?

M6

See how words contribute to the verdict score.

The words push toward the label, yet the verdict is no.
The headlines here were written as examples. The verdicts and contributions are the real output of the model published on the Transparency page. In the app you see the same explanation for any item via the "?" button, including the full list and the model's base tendency. The sum of word contributions alone does not decide the label: the threshold is set separately for each language.
From the app
The calculation panel behind the question mark next to a label: base, word contributions, decision score and threshold. Silent recording — tap to play.