Multitask Learning and Bandits via Robust Statistics DOI
Kan Xu, Hamsa Bastani

SSRN Electronic Journal, Journal Year: 2024, Volume and Issue: unknown

Published: Jan. 1, 2024

Decision-makers often simultaneously face many related but heterogeneous learning problems. For instance, a large retailer may wish to learn product demand at different stores solve pricing or inventory problems, making it desirable jointly for serving similar customers; alternatively, hospital network patient risk providers allocate personalized interventions, hospitals populations. Motivated by real datasets, we study natural setting where the unknown parameter in each instance can be decomposed into shared global plus sparse instance-specific term. We propose novel two-stage multitask estimator that exploits this structure sample-efficient way, using unique combination of robust statistics (to across instances) and LASSO regression debias results). Our yields improved sample complexity bounds feature dimension d relative commonly-employed estimators; improvement is exponential "data-poor" instances, which benefit most from learning. illustrate utility these results online embedding our within simultaneous contextual bandit algorithms. specify dynamic calibration appropriately balance bias-variance tradeoff over time, improving resulting regret context d. Finally, value approach on synthetic datasets.

Language: Английский

Multitask Learning and Bandits via Robust Statistics DOI
Xu Kan, Hamsa Bastani

Management Science, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 6, 2025

Decision makers often simultaneously face many related but heterogeneous learning problems. For instance, a large retailer may wish to learn product demand at different stores solve pricing or inventory problems, making it desirable jointly for serving similar customers; alternatively, hospital network patient risk providers allocate personalized interventions, hospitals populations. Motivated by real data sets, we study natural setting where the unknown parameter in each instance can be decomposed into shared global plus sparse instance-specific term. We propose novel two-stage multitask estimator that exploits this structure sample-efficient way, using unique combination of robust statistics (to across instances) and LASSO regression debias results). Our yields improved sample complexity bounds feature dimension d relative commonly employed estimators; improvement is exponential “data-poor” instances, which benefit most from learning. illustrate utility these results online embedding our within simultaneous contextual bandit algorithms. specify dynamic calibration appropriately balance bias-variance trade-off over time, improving resulting regret context d. Finally, value approach on synthetic sets. This paper was accepted J. George Shanthikumar, science. Supplemental Material: The appendix files are available https://doi.org/10.1287/mnsc.2022.00490 .

Language: Английский

Citations

2
Using Contingency Management with a Deposit Contract to Increase Toothbrushing Accuracy with College Students DOI

Briar N. Moroney,

Sharon A. Reeve, Kenneth F. Reeve

et al.

Journal of Behavioral Education, Journal Year: 2025, Volume and Issue: unknown

Published: Jan. 28, 2025

Language: Английский

Citations

0
Exploring parental opinions on oral hygiene behavior and knowledge of their young children in Lithuania: a cross-sectional survey study DOI Creative Commons
Yvonne A.B. Buunk‐Werkhoven,

Rasa Tamulienė,

Daiva Mačiulienė

et al.

Frontiers in Oral Health, Journal Year: 2025, Volume and Issue: 6

Published: April 29, 2025

Background An appropriately formulated oral health education program carefully based on research, can increase knowledge, change behavior in a positive direction and improve self-confidence. This study aimed to examine parental opinions their children's hygiene (OHB) knowledge (OHK) among pre- primary school children Kaunas, Lithuania. Methods In this cross-sectional study, an online 33-question survey was conducted before after World Oral Health Day March 20 assess the skills, eating habits, demographics of 5–12 year children. A total 532 parents participated, with data from 420 parents, mainly married mothers (average age 37.3 years) being analyzed. Most participants had higher education, lived one three children, average 7 years for oldest child. Results used manual toothbrush. The adapted OHB index showed that most generally good control over tooth brushing many twice daily using fluoride toothpaste. One-third always re-brushed child's teeth child brushed independently. Parents demonstrated strong care, as reflected high scores OHK index. correlation found between ( r = 0.14, p 0.05). Younger were more frequently, linked frequent re-brushing, particularly less than 10 years, better but did not demonstrate OHB. Conclusions insights gained into help implement evidence-based preventive approach practices.

Language: Английский

Citations

0
Improving Treatment Responses through Limited Nudges: A Data-Driven Learning and Optimization Approach DOI
Esmaeil Keyvanshokooh, Kyra Gan, Yongyi Guo

et al.

Published: Jan. 1, 2024

Language: Английский

Citations

0
Evaluate Closed-Loop, Mindless Intervention in-the-Wild: A Micro-Randomized Trial on Offset Heart Rate Biofeedback DOI
Yiran Zhao, Tanzeem Choudhury

Published: Sept. 22, 2024

Language: Английский

Citations

0
Multitask Learning and Bandits via Robust Statistics DOI
Kan Xu, Hamsa Bastani

SSRN Electronic Journal, Journal Year: 2024, Volume and Issue: unknown

Published: Jan. 1, 2024

Decision-makers often simultaneously face many related but heterogeneous learning problems. For instance, a large retailer may wish to learn product demand at different stores solve pricing or inventory problems, making it desirable jointly for serving similar customers; alternatively, hospital network patient risk providers allocate personalized interventions, hospitals populations. Motivated by real datasets, we study natural setting where the unknown parameter in each instance can be decomposed into shared global plus sparse instance-specific term. We propose novel two-stage multitask estimator that exploits this structure sample-efficient way, using unique combination of robust statistics (to across instances) and LASSO regression debias results). Our yields improved sample complexity bounds feature dimension d relative commonly-employed estimators; improvement is exponential "data-poor" instances, which benefit most from learning. illustrate utility these results online embedding our within simultaneous contextual bandit algorithms. specify dynamic calibration appropriately balance bias-variance tradeoff over time, improving resulting regret context d. Finally, value approach on synthetic datasets.

Language: Английский

Citations

0