Fast and memory ecient AUC-ROC approximation in Stream Learning
| Autor: | Albakour, Subhy, Manine, Alain-Pierre, Alphonse, Erick |
|---|---|
| Přispěvatelé: | Institut Polytechnique de Paris (IP Paris), BREUIL, Florent |
| Jazyk: | angličtina |
| Rok vydání: | 2021 |
| Předmět: | |
| Zdroj: | Actes CNIA 2021 CNIA 2021 : Conférence Nationale en Intelligence Artificielle CNIA 2021 : Conférence Nationale en Intelligence Artificielle, Jun 2021, Bordeaux (en ligne), France. pp 68-75 |
| Popis: | International audience; Machine Learning applied to never-ending data streams has unique resource-consumption constraints that require processing one data-point at a time without storing. Much research has been devoted to develop algorithms that learn from data streams, and most algorithms are available inlibraries such as MOA1 and River2. However, less is done on the evaluation of the models generated by these algorithms. Area under the ROC Curve (AUC-ROC) has more discrimination power, as an evaluation metric, than accuracy and all other confusion-matrix based metrics. Nonetheless, computing the AUC-ROC of a model is expensive, and violates streaming constraints in practical applications, as it requires the entire history of the model’s predictions when computed. In this paper, we show how we can extend sketching algorithms to summarize the prediction history of a model, and then produce a high-quality approximation of its AUC-ROC in bounded memory and constant update time. This renders AUC-ROC practical for evaluating and selecting Stream Learning models. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|