[Paper Review] Algorithmic and Statistical Challenges in Modern Large-Scale Data Analysis are the Focus of MMDS 2008

Michael W. Mahoney, Lek‐Heng Lim+1 more|ArXiv.org|Dec 19, 2008

Big Data Technologies and Applications4 citations

TL;DR

This paper summarizes the 2008 Workshop on Algorithms for Modern Massive Data Sets (MMDS 2008), which brought together researchers from computer science, statistics, mathematics, and data analysis to address algorithmic and statistical challenges in large-scale, high-dimensional, and nonlinear data. The workshop highlighted interdisciplinary approaches to modeling massive data using graphs, matrices, and statistical models, emphasizing scalable algorithms, dimensionality reduction, and kernel methods in Reproducing Kernel Hilbert Spaces (RKHS).

ABSTRACT

The 2008 Workshop on Algorithms for Modern Massive Data Sets (MMDS 2008), sponsored by the NSF, DARPA, LinkedIn, and Yahoo!, was held at Stanford University, June 25--28. The goals of MMDS 2008 were (1) to explore novel techniques for modeling and analyzing massive, high-dimensional, and nonlinearly-structured scientific and internet data sets; and (2) to bring together computer scientists, statisticians, mathematicians, and data analysis practitioners to promote cross-fertilization of ideas.

Motivation & Objective

• To explore novel algorithmic, statistical, and mathematical techniques for analyzing massive, high-dimensional, and nonlinearly structured data sets.
• To foster cross-fertilization of ideas between computer scientists, statisticians, mathematicians, and data analysis practitioners.
• To address the challenges posed by large-scale data, including sparsity, noise, and complex structure, through interdisciplinary collaboration.
• To promote the development of scalable, efficient, and statistically sound methods for modern data-intensive applications.
• To identify emerging research directions at the intersection of data mining, machine learning, and scientific computing.

Proposed method

• Organized six hour-long tutorials introducing core themes in large-scale data analysis, including graph mining, matrix algorithms, and statistical modeling.
• Utilized graph and matrix representations to model complex data, such as social networks and feature-laden datasets.
• Introduced a semiparametric formulation for Sufficient Dimensionality Reduction (SDR) using conditional independence and operators on Reproducing Kernel Hilbert Spaces (RKHS).
• Applied RKHS-based methods to optimize statistical functionals, enabling nonparametric estimation of the sufficient dimension subspace.
• Leveraged the reproducing property of RKHS to reduce function evaluation to inner products, enabling efficient computation.
• Integrated both frequentist and Bayesian perspectives, with statistical modeling emphasizing noise structure and predictive uncertainty.

Experimental results

Research questions

• RQ1 How can we develop scalable and statistically sound methods for analyzing massive, high-dimensional, and nonlinear data?
• RQ2 What are the algorithmic and statistical challenges in modeling large-scale graphs and matrices arising in real-world applications?
• RQ3 How can we reconcile the database-centric view of data as a record of events with the statistical view of data as a realization of an underlying stochastic process?
• RQ4 Can kernel methods in Reproducing Kernel Hilbert Spaces (RKHS) be used to efficiently solve complex statistical problems like Sufficient Dimensionality Reduction?
• RQ5 To what extent do existing models of large networks (e.g., Erdős-Rényi) capture the true structural properties of real-world data, such as heavy-tailed degree distributions?

Key findings

• The workshop attracted nearly 300 participants and featured 43 talks and 18 poster presentations, reflecting strong interdisciplinary interest.
• Graph and matrix models were central to understanding large-scale data, with real-world applications in social networks, web search, and scientific computing.
• The two dominant perspectives—data as a record (computer science) and data as a sample from a stochastic process (statistics)—were shown to be complementary and mutually enriching.
• Semiparametric SDR methods based on RKHS operators provided a novel, nonparametric approach to dimensionality reduction that avoids strong parametric assumptions.
• RKHS-based methods enabled efficient optimization of complex statistical functionals by reducing function evaluation to inner products.
• The workshop demonstrated that scalable algorithms, including those based on MapReduce and external memory models, are essential for handling modern data workloads.