Horizontal Decision Tree Learning with Parallel Conflict Resolution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing solutions for high-rate data stream processing, such as Streaming Parallel Decision Tree (SPDT) and Scalable Advanced Massive Online Analysis (SAMOA), face scalability issues due to centralized model update computations, which restrict their ability to handle extremely high data arrival rates effectively.
Innovation Solution
The method involves distributing data record instances to multiple model update processing items for parallel candidate leaf splitting actions, resolving conflicts, and applying tree structure changes in a distributed manner, enabling horizontal parallelism in decision tree learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If centralized model update computation is used (as in SPDT and SAMOA), then model consistency is maintained, but scalability is limited due to processing bottlenecks
Solution Approach 1:
The patent segments the centralized model update computation into distributed model update operations across multiple processing instances. Each instance maintains and updates its own copy of the decision tree model independently, eliminating the single-point bottleneck while maintaining model consistency through conflict resolution mechanisms.
Solution Approach 2:
The patent introduces a conflict resolution mechanism as an intermediary that coordinates updates across distributed instances. When multiple instances attempt to update the same model parameters, the conflict resolution mechanism mediates these updates to ensure consistency, replacing the traditional centralized update approach.
2Measurement precision
If data is processed sequentially (as in MOA), then model update accuracy is maintained, but processing speed cannot handle high data arrival rates
Solution Approach 1:
The patent segments the sequential processing pipeline into parallel processing streams where multiple instances simultaneously process different portions of the data stream and perform model updates independently, then reconcile results through conflict resolution.
Solution Approach 2:
The patent implements dynamic parallel processing where the degree of parallelism can adapt to data arrival rates. Multiple processing instances dynamically process data in parallel and coordinate updates, allowing the system to scale processing speed while maintaining accuracy through distributed model updates.
3Productivity
If distributed model updates are implemented without conflict resolution, then processing parallelism increases, but model consistency deteriorates
Solution Approach 1:
The patent introduces a conflict resolution mechanism as an intermediary that coordinates updates across distributed instances. When multiple instances attempt to update the same model parameters, the conflict resolution mechanism mediates these updates to ensure consistency, replacing the traditional centralized update approach.
Solution Approach 2:
The patent implements feedback mechanisms where processing instances monitor and communicate model state changes to each other. This feedback loop allows distributed instances to coordinate their updates and resolve conflicts, maintaining model consistency while preserving processing parallelism.
Data Source
AI summary
A mechanism is provided in a data processing system for distributed tree learning. A source processing instance distributes data record instances to a plurality of model update processing items. The plurality of model update processing items determine candidate leaf splitting actions in a decision tree in parallel based on the data record instances. The plurality of model update processing items send the candidate leaf splitting actions to a plurality of conflict resolve processing items. The plurality of conflict resolve processing items identifies conflict leaf splitting actions. The plurality of conflict resolve processing items applies tree structure changes to the decision tree in the plurality of model update processing items.