Calibrating Predictions to Decisions: A Novel Approach to Multi-Class Calibration

When facing uncertainty, decision-makers want predictions they can trust. A machine learning provider can convey confidence to decision-makers by guaranteeing their predictions are distribution calibrated -- amongst the inputs that receive a predicted class probabilities vector $q$, the actual distr...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Zhao, Shengjia, Kim, Michael P, Sahoo, Roshni, Ma, Tengyu, Ermon, Stefano
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title
container_volume
creator Zhao, Shengjia
Kim, Michael P
Sahoo, Roshni
Ma, Tengyu
Ermon, Stefano
description When facing uncertainty, decision-makers want predictions they can trust. A machine learning provider can convey confidence to decision-makers by guaranteeing their predictions are distribution calibrated -- amongst the inputs that receive a predicted class probabilities vector $q$, the actual distribution over classes is $q$. For multi-class prediction problems, however, achieving distribution calibration tends to be infeasible, requiring sample complexity exponential in the number of classes $C$. In this work, we introduce a new notion -- \emph{decision calibration} -- that requires the predicted distribution and true distribution to be ``indistinguishable'' to a set of downstream decision-makers. When all possible decision makers are under consideration, decision calibration is the same as distribution calibration. However, when we only consider decision makers choosing between a bounded number of actions (e.g. polynomial in $C$), our main result shows that decisions calibration becomes feasible -- we design a recalibration algorithm that requires sample complexity polynomial in the number of actions and the number of classes. We validate our recalibration algorithm empirically: compared to existing methods, decision calibration improves decision-making on skin lesion and ImageNet classification with modern neural network predictors.
doi_str_mv 10.48550/arxiv.2107.05719
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2107_05719</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2107_05719</sourcerecordid><originalsourceid>FETCH-LOGICAL-a679-fbf8b207c3a0939b14c78d37dfef6c39350a0abef887f122a5951d336ff35ae3</originalsourceid><addsrcrecordid>eNo9j81KxDAURrNxIaMP4GryAq1J76RJ3JX6C6OO6L7cpIkGYlPSOujbS2fE1cfhgwOHkAvOyo0Sgl1i_g77suJMlkxIrk_JS4sxmIxzGN7pLrs-2DmkYaJzotfOhmmBK9rQp7R3kTbjmBPaj-V-_IpzKNqI00T_LWk4Iyce4-TO_3ZFXm9v3tr7Yvt899A22wJrqQtvvDIVkxaQadCGb6xUPcjeO19b0CAYMjTOKyU9ryoUWvAeoPYeBDpYkfXRekjqxhw-Mf90S1p3SINfNrRJvg</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>Calibrating Predictions to Decisions: A Novel Approach to Multi-Class Calibration</title><source>arXiv.org</source><creator>Zhao, Shengjia ; Kim, Michael P ; Sahoo, Roshni ; Ma, Tengyu ; Ermon, Stefano</creator><creatorcontrib>Zhao, Shengjia ; Kim, Michael P ; Sahoo, Roshni ; Ma, Tengyu ; Ermon, Stefano</creatorcontrib><description>When facing uncertainty, decision-makers want predictions they can trust. A machine learning provider can convey confidence to decision-makers by guaranteeing their predictions are distribution calibrated -- amongst the inputs that receive a predicted class probabilities vector $q$, the actual distribution over classes is $q$. For multi-class prediction problems, however, achieving distribution calibration tends to be infeasible, requiring sample complexity exponential in the number of classes $C$. In this work, we introduce a new notion -- \emph{decision calibration} -- that requires the predicted distribution and true distribution to be ``indistinguishable'' to a set of downstream decision-makers. When all possible decision makers are under consideration, decision calibration is the same as distribution calibration. However, when we only consider decision makers choosing between a bounded number of actions (e.g. polynomial in $C$), our main result shows that decisions calibration becomes feasible -- we design a recalibration algorithm that requires sample complexity polynomial in the number of actions and the number of classes. We validate our recalibration algorithm empirically: compared to existing methods, decision calibration improves decision-making on skin lesion and ImageNet classification with modern neural network predictors.</description><identifier>DOI: 10.48550/arxiv.2107.05719</identifier><language>eng</language><subject>Computer Science - Learning ; Statistics - Machine Learning ; Statistics - Methodology</subject><creationdate>2021-07</creationdate><rights>http://arxiv.org/licenses/nonexclusive-distrib/1.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,780,885</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2107.05719$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2107.05719$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Zhao, Shengjia</creatorcontrib><creatorcontrib>Kim, Michael P</creatorcontrib><creatorcontrib>Sahoo, Roshni</creatorcontrib><creatorcontrib>Ma, Tengyu</creatorcontrib><creatorcontrib>Ermon, Stefano</creatorcontrib><title>Calibrating Predictions to Decisions: A Novel Approach to Multi-Class Calibration</title><description>When facing uncertainty, decision-makers want predictions they can trust. A machine learning provider can convey confidence to decision-makers by guaranteeing their predictions are distribution calibrated -- amongst the inputs that receive a predicted class probabilities vector $q$, the actual distribution over classes is $q$. For multi-class prediction problems, however, achieving distribution calibration tends to be infeasible, requiring sample complexity exponential in the number of classes $C$. In this work, we introduce a new notion -- \emph{decision calibration} -- that requires the predicted distribution and true distribution to be ``indistinguishable'' to a set of downstream decision-makers. When all possible decision makers are under consideration, decision calibration is the same as distribution calibration. However, when we only consider decision makers choosing between a bounded number of actions (e.g. polynomial in $C$), our main result shows that decisions calibration becomes feasible -- we design a recalibration algorithm that requires sample complexity polynomial in the number of actions and the number of classes. We validate our recalibration algorithm empirically: compared to existing methods, decision calibration improves decision-making on skin lesion and ImageNet classification with modern neural network predictors.</description><subject>Computer Science - Learning</subject><subject>Statistics - Machine Learning</subject><subject>Statistics - Methodology</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2021</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNo9j81KxDAURrNxIaMP4GryAq1J76RJ3JX6C6OO6L7cpIkGYlPSOujbS2fE1cfhgwOHkAvOyo0Sgl1i_g77suJMlkxIrk_JS4sxmIxzGN7pLrs-2DmkYaJzotfOhmmBK9rQp7R3kTbjmBPaj-V-_IpzKNqI00T_LWk4Iyce4-TO_3ZFXm9v3tr7Yvt899A22wJrqQtvvDIVkxaQadCGb6xUPcjeO19b0CAYMjTOKyU9ryoUWvAeoPYeBDpYkfXRekjqxhw-Mf90S1p3SINfNrRJvg</recordid><startdate>20210712</startdate><enddate>20210712</enddate><creator>Zhao, Shengjia</creator><creator>Kim, Michael P</creator><creator>Sahoo, Roshni</creator><creator>Ma, Tengyu</creator><creator>Ermon, Stefano</creator><scope>AKY</scope><scope>EPD</scope><scope>GOX</scope></search><sort><creationdate>20210712</creationdate><title>Calibrating Predictions to Decisions: A Novel Approach to Multi-Class Calibration</title><author>Zhao, Shengjia ; Kim, Michael P ; Sahoo, Roshni ; Ma, Tengyu ; Ermon, Stefano</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-a679-fbf8b207c3a0939b14c78d37dfef6c39350a0abef887f122a5951d336ff35ae3</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2021</creationdate><topic>Computer Science - Learning</topic><topic>Statistics - Machine Learning</topic><topic>Statistics - Methodology</topic><toplevel>online_resources</toplevel><creatorcontrib>Zhao, Shengjia</creatorcontrib><creatorcontrib>Kim, Michael P</creatorcontrib><creatorcontrib>Sahoo, Roshni</creatorcontrib><creatorcontrib>Ma, Tengyu</creatorcontrib><creatorcontrib>Ermon, Stefano</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv Statistics</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Zhao, Shengjia</au><au>Kim, Michael P</au><au>Sahoo, Roshni</au><au>Ma, Tengyu</au><au>Ermon, Stefano</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>Calibrating Predictions to Decisions: A Novel Approach to Multi-Class Calibration</atitle><date>2021-07-12</date><risdate>2021</risdate><abstract>When facing uncertainty, decision-makers want predictions they can trust. A machine learning provider can convey confidence to decision-makers by guaranteeing their predictions are distribution calibrated -- amongst the inputs that receive a predicted class probabilities vector $q$, the actual distribution over classes is $q$. For multi-class prediction problems, however, achieving distribution calibration tends to be infeasible, requiring sample complexity exponential in the number of classes $C$. In this work, we introduce a new notion -- \emph{decision calibration} -- that requires the predicted distribution and true distribution to be ``indistinguishable'' to a set of downstream decision-makers. When all possible decision makers are under consideration, decision calibration is the same as distribution calibration. However, when we only consider decision makers choosing between a bounded number of actions (e.g. polynomial in $C$), our main result shows that decisions calibration becomes feasible -- we design a recalibration algorithm that requires sample complexity polynomial in the number of actions and the number of classes. We validate our recalibration algorithm empirically: compared to existing methods, decision calibration improves decision-making on skin lesion and ImageNet classification with modern neural network predictors.</abstract><doi>10.48550/arxiv.2107.05719</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2107.05719
ispartof
issn
language eng
recordid cdi_arxiv_primary_2107_05719
source arXiv.org
subjects Computer Science - Learning
Statistics - Machine Learning
Statistics - Methodology
title Calibrating Predictions to Decisions: A Novel Approach to Multi-Class Calibration
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-10T14%3A02%3A49IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=Calibrating%20Predictions%20to%20Decisions:%20A%20Novel%20Approach%20to%20Multi-Class%20Calibration&rft.au=Zhao,%20Shengjia&rft.date=2021-07-12&rft_id=info:doi/10.48550/arxiv.2107.05719&rft_dat=%3Carxiv_GOX%3E2107_05719%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true