RankMap: A Framework for Distributed Learning From Dense Data Sets.

Area of Science:

• Computer Science
• Machine Learning
• Data Science

Background:

• Iterative learning algorithms are crucial for many data science applications.
• Processing massive and dense datasets presents significant computational challenges.
• Existing frameworks often struggle with scalability and efficiency for large-scale iterative tasks.

Purpose of the Study:

• Introduce RankMap, a novel framework for efficient execution of iterative learning algorithms.
• Address the challenges of processing massive and dense datasets in a distributed computing environment.
• Provide a scalable and efficient solution for contemporary machine learning applications.

Main Methods:

• Developed a platform-aware, end-to-end framework named RankMap.
• Exploited data structure factorization into lower rank subspaces.
• Created sparse, low-dimensional data representations for efficient mapping and scheduling.
• Designed matrix-based and graph-based APIs for framework adoption.
• Evaluated on real-world datasets using distributed computing infrastructure.

Main Results:

• Achieved up to two orders of magnitude improvement in memory usage, execution speed, and bandwidth.
• Demonstrated significant performance gains on sparse recovery and power iteration problems.
• Maintained learning accuracy comparable to prior state-of-the-art methods.
• Successfully processed datasets with up to 1.8 billion nonzeros on up to 244 cores.

Conclusions:

• RankMap offers a highly efficient and scalable solution for iterative learning on massive datasets.
• The framework's data factorization and sparse representation techniques are key to its performance improvements.
• RankMap facilitates the practical application of advanced learning algorithms in distributed environments.