Multi-directional ensembles of decision trees, or MERCS models, are:
- fast
- versatile
- robust to missing values at prediction time
Motivation
- You can cope with missing values
- You can switch targets
- You can switch the amount of targets
Advantages
Given
what's the value of ?
X
Y
A versatile ML model can solve this.
versatile
Motivation
A typical ML model can solve this
Given
what's the value of ?
X
Y
- Idealized scenario: if you do not know the exact task on beforehand, this paradigm breaks down.
- E.g.: anomaly detection (e.g. on sensor data), prediction in spreadsheets, ...
Disadvantage
Versatile models comparison
PGM
kNN
NN
Nominal + Numeric data
Interpretable
Scalable
MERCS
Problem
MERCS should handle any prediction task
- At training time: prediction task unknown
- At prediction time: you do not have the 'perfect' tree at your disposal
- Prediction algorithms overcome this discrepancy
MERCS MODEL
f_1
f_2
f_3
QUERIES
q_1
q_2
q_3
Chaining
Bottom-Up Chaining
Query:
\{A_1, A_2\} \rightarrow \{A_3\}
Use most appropriate models, given
\{A_1, A_2\}
Use most appropriate models, given
\{A_1, A_2, A_4\}
MERCS-models
A unidirectional ML
model can solve this
unidirectional
A multi-directional ML model can solve this, and other things too!
multi-directional
Given what is ?
X
Y
Versatile
MERCS should handle any prediction task
MERCS MODEL
= what you have
QUERIES FROM USER
= what you need
f_1
f_2
f_3
q_1
q_2
q_3
Robust to missing data
Build query-specific compositions
QUERIES FROM USER
= what you need
q_1
q_2
q_3
MISSING DATA
SOLUTION
MERCS-spotlight
By eliavw
MERCS-spotlight
SML Spotlight Presentation
