Near-optimal learning of tree-structured distributions by Chow-Liu
Near-optimal learning of tree-structured distributions by Chow-Liu
复制标题
Chow-Liu 的树结构分布的近乎最优学习
DOI:
10.1145/3406325.3451066
复制
发表时间:
2021
期刊:
影响因子:
--
通讯作者:
Vinodchandran, N. V.
中科院分区:
文献类型:
--
作者:
Bhattacharyya, Arnab;Gayen, Sutanu;Price, Eric;Vinodchandran, N. V.
We provide finite sample guarantees for the classical Chow-Liu algorithm (IEEE Trans. Inform. Theory, 1968) to learn a tree-structured graphical model of a distribution. For a distributionPon Σnand a treeTonnnodes, we sayTis an ε-approximate tree forPif there is aT-structured distributionQsuch thatD(P||Q) is at most ε more than the best possible tree-structured distribution forP. We show that ifPitself is tree-structured, then the Chow-Liu algorithm with the plug-in estimator for mutual information withO(|Σ|3nε−1) i.i.d. samples outputs an ε-approximate tree forPwith constant probability. In contrast, for a generalP(which may not be tree-structured), Ω(n2ε−2) samples are necessary to find an ε-approximate tree. Our upper bound is based on a new conditional independence tester that addresses an open problem posed by Canonne, Diakonikolas, Kane, and Stewart (STOC, 2018): we prove that for three random variablesX,Y,Zeach over Σ, testing ifI(X;Y∣Z) is 0 or ≥ ε is possible withO(|Σ|3/ε) samples. Finally, we show that for a specific treeT, withO(|Σ|2nε−1) samples from a distributionPover Σn, one can efficiently learn the closestT-structured distribution in KL divergence by applying the add-1 estimator at each node.
登录
查看更多内容
DOI:
--
发表时间:
2018
期刊:
and Automata
影响因子:
--
作者:
Diakonikolas, Ilias;Gouleakis, Themis;Peebles, John;Price, Eric
通讯作者:
Price, Eric
DOI:
--
发表时间:
2016
期刊:
Annual Conference Computational Learning Theory
影响因子:
--
作者:
C. Daskalakis;Qinxuan Pan
通讯作者:
Qinxuan Pan
DOI:
10.1145/3391403.3399541
发表时间:
2020
期刊:
21st ACM Conference on Economics and Computation
影响因子:
--
作者:
Brustle, Johannes;Cai, Yang;Daskalakis, Constantinos
通讯作者:
Daskalakis, Constantinos
DOI:
--
发表时间:
2004
期刊:
Conference on Uncertainty in Artificial Intelligence
影响因子:
--
作者:
Mukund Narasimhan;J. Bilmes
通讯作者:
J. Bilmes
影响因子:
2.5
作者:
C. Canonne;Ilias Diakonikolas;D. Kane;Alistair Stewart
通讯作者:
Alistair Stewart