[SYSTEMDS-3149] Decision Tree cleanups and Builtin

• Type: Task
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: SystemDS 3.2
• Epic Link: New Built-in functions
• Assignee: Matthias Boehm
• Reporter: Shafaq Siddiqi
• Created: 04/Oct/21 11:49
• Updated: 05/May/23 18:58
• Resolved: 25/Mar/23 15:34

Description

Cleanup of existing decision tree algorithm in SystemDS "scripts/algorithms/decision-tree.dml" and convert it into dml builtin function.

As a preparation for a holistic cleanup of decision tree and random forest scripts (in scripts/builtin and scripts/algorithms), we should first introduce primitives for information gain, entropy, and gini (SYSTEMDS-3184), and devise vectorized prediction scripts.

In a first step, this task should introduce a new builtin/decisionTreePredict that implements the different strategies of the Hummingbird paper [1]. Initial tests can hard-code the vectorized decision tree representation and focus on testing the prediction procedure. The builtin function signature might expose a 'method' attribute to select the different strategies.

[1] Supun Nakandala, Karla Saur, Gyeong-In Yu, Konstantinos Karanasos, Carlo Curino, Markus Weimer, Matteo Interlandi:
A Tensor Compiler for Unified Machine Learning Prediction Serving. OSDI 2020: 899-917, https://www.usenix.org/system/files/osdi20-nakandala.pdf

Issue Links

• is blocked by: SYSTEMDS-3184 Builtin for computing information gain using entropy and gini
• links to: GitHub Pull Request #103
• links to: GitHub Pull Request #1506

Sub-Tasks

• Decision Tree Predict: Resolved
• Cleanup decisionTree builtin: Closed
• Cleanup randomForest builtin: Closed
• Additional decisionTree/randomForest inference methods: Closed
• Optional Data Compaction in decisionTree/randomForest: Closed