multivariate dyadic regression trees for sparse learning problems xi chen machine learning...

Multivariate Dyadic Regression Trees for Sparse Learning Problems

Xi ChenMachine Learning Department

Carnegie Mellon University(joint work with Han Liu)

Content

Experimental Results

Statistical Property

Multivariate Regression and Dyadic Regression Tree

Tree Learning Algorithm

Multivariate Dyadic Regression Tree for Sparse Learning

Multivariate Regression Model

Multivariate Regression Model

Predictors Responses

Estimate : Minimize the L2-risk

Empirical Risk Minimization

Tree Based Method

Estimation using tree based methodsWhy trees? Simplicity of Design Good Interpretability Easy Implementation Good Practical Performance

Tree Based Method

CART (Classification and Regression Tree)[Breiman 1984]

No. of terminal nodesHard to be theoretically analyzed!

Dyadic Decision/Regression Tree

Dyadic Split[Scott 2004]

Sparse Model

Lower Minimax Rate of Convergence of the risk

Slow

Fast

Sparse Model

Regression Tree

Piecewise Constant

Piecewise Linear

Piecewise Polynomial

Gamma-Ray Burst 845

Multivariate Dyadic Regression Tree (MDRT)

Active Set

Rule 1

Rule 2

Multivariate Dyadic Regression Tree (MDRT) Variable Selection

Multivariate Dyadic Regression Tree

Regularization Parameter

Fine partitionSparse Model

Lower degree poly

Statistical Property

Assumption 1:

Assumption 2:

Convergence Rate

Minimax Rate

Tree Learning Algorithm

Loss:

Minimize the cost

Tree Learning Algorithm

Tree-growing stage

Pruning-back stage

Randomized

Greedy

Experimental Results

Methods Compared

Methods

Greedy MDRT with M=1 MDRT(G, M=1)

Randomized MDRT with M=1 MDRT(R, M=1)

Greedy MDRT with M=0 MDRT(G, M=0)

Randomized MDRT with M=0 MDRT(R, M=0)

Classification and Regression Tree CART

Piecewise LinearPiecewise Constant

Generalized Nonlinear Model

Experimental Results

Synthetic Data

Linear Model

Additive Model

Experimental Results

Experimental Results

Real Data (MSE)

10 artificial variables from Unif(0,1)

15 artificial variables from Unif(0,1)

Never selected in 20 runs for M=1

Conclusion

Multivariate Regression Tree Model Dyadic Split A novel penalization term Theoretically, achieve nearly optimal minimax

rate for (α,C) smooth function Empirically, conduct variable selection for sparse

models Efficient computation tree learning algorithm

Extensions Classification Trees Forest Extensions