scaling multivariate statistics to massive data algorithmic problems and approaches alexander gray...

Scaling Multivariate Statistics to Massive Data

Algorithmic problems and approaches

Alexander GrayGeorgia Institute of Technology

www.fast-lab.org

Core methods ofstatistics / machine learning / mining

1. Querying: spherical range-search O(N), orthogonal range-search O(N), spatial join O(N2), nearest-neighbor O(N), all-nearest-neighbors O(N2)

2. Density estimation: mixture of Gaussians, kernel density estimation O(N2), kernel conditional density estimation O(N3)

3. Regression: linear regression, kernel regression O(N2), Gaussian process regression O(N3)

4. Classification: decision tree, nearest-neighbor classifier O(N2), nonparametric Bayes classifier O(N2), support vector machine O(N3)

5. Dimension reduction: principal component analysis, non-negative matrix factorization, kernel PCA O(N3), maximum variance unfolding O(N3)

6. Outlier detection: by density estimation or dimension reduction7. Clustering: by density estimation or dimension reduction, k-means, mean-

shift segmentation O(N2), hierarchical clustering O(N3)8. Time series analysis: Kalman filter, hidden Markov model, trajectory

tracking O(Nn)9. Feature selection and causality: LASSO, L1 SVM, Gaussian graphical

models, discrete graphical models10.Fusion and matching: sequence alignment, bipartite matching O(N3), n-

point correlation 2-sample testing O(Nn)

Now pretty fast (2011)…

1. Querying: spherical range-search O(logN)*, orthogonal range-search O(logN)*, spatial join O(N)*, nearest-neighbor O(logN), all-nearest-neighbors O(N)

2. Density estimation: mixture of Gaussians, kernel density estimation O(N), kernel conditional density estimation O(Nlog3)*

3. Regression: linear regression, kernel regression O(N), Gaussian process regression O(N)*

4. Classification: decision tree, nearest-neighbor classifier O(N), nonparametric Bayes classifier O(N)*, support vector machine

5. Dimension reduction: principal component analysis, non-negative matrix factorization, kernel PCA O(N)*, maximum variance unfolding O(N)*

6. Outlier detection: by density estimation or dimension reduction7. Clustering: by density estimation or dimension reduction, k-means, mean-

shift segmentation O(N), hierarchical clustering O(NlogN)8. Time series analysis: Kalman filter, hidden Markov model, trajectory

tracking O(Nlogn)*9. Feature selection and causality: LASSO, L1 SVM, Gaussian graphical

models, discrete graphical models10.Fusion and matching: sequence alignment, bipartite matching O(N)**, n-

point correlation 2-sample testing O(Nlogn)*

Things we made fastfastest, fastest in some settings

1. Querying: spherical range-search O(logN)*, orthogonal range-search O(logN)*, spatial join O(N)*, nearest-neighbor O(logN), all-nearest-neighbors O(N)

2. Density estimation: mixture of Gaussians, kernel density estimation O(N), kernel conditional density estimation O(Nlog3)*

3. Regression: linear regression, kernel regression O(N), Gaussian process regression O(N)*

4. Classification: decision tree, nearest-neighbor classifier O(N), nonparametric Bayes classifier O(N)*, support vector machine O(N)/O(N2)

5. Dimension reduction: principal component analysis, non-negative matrix factorization, kernel PCA O(N)*, maximum variance unfolding O(N)*

6. Outlier detection: by density estimation or dimension reduction7. Clustering: by density estimation or dimension reduction, k-means, mean-

shift segmentation O(N), hierarchical (FoF) clustering O(NlogN)8. Time series analysis: Kalman filter, hidden Markov model, trajectory

tracking O(Nlogn)*9. Feature selection and causality: LASSO, L1 SVM, Gaussian graphical

models, discrete graphical models10.Fusion and matching: sequence alignment, bipartite matching O(N)**, n-

point correlation 2-sample testing O(Nlogn)*

Core computational problems

What are the basic mathematical operations making things hard?

• Alternative to speeding up each of the 1000s of statistical methods: treat common computational bottlenecks

• Divide up the space of problems (and associated algorithmic strategies), so we can examine the unique challenges and possible ways forward within each

The “7 Giants” of data

1. Basic statistics

2. Generalized N-body problems

3. Graph-theoretic problems

4. Linear-algebraic problems

5. Optimizations

6. Integrations

7. Alignment problems

The “7 Giants” of data

1. Basic statistics•e.g. counts, contingency tables, means, medians, variances, range queries (SQL queries)

2. Generalized N-body problems•e.g. nearest-nbrs (in NLDR, etc), kernel summations (in KDE, GP, SVM, etc), clustering, MST, spatial correlations

The “7 Giants” of data

3. Graph-theoretic problems•e.g. betweenness centrality, commute distance, graphical model inference

4. Linear-algebraic problems•e.g. linear algebra, PCA, Gaussian process regression, manifold learning

5. Optimizations•e.g. LP/QP/SDP/SOCP/MINLPs in regularized methods, compressed sensing

The “7 Giants” of data

6. Integrations•e.g. Bayesian inference

7. Alignment problems•e.g. BLAST in genomics, string matching, phylogenies, SLAM, cross-match

Back to our listbasic, N-body, graphs, linear algebra, optimization, integration, alignment

1. Querying: spherical range-search O(N), orthogonal range-search O(N), spatial join O(N2), nearest-neighbor O(N), all-nearest-neighbors O(N2)

2. Density estimation: mixture of Gaussians, kernel density estimation O(N2), kernel conditional density estimation O(N3)

3. Regression: linear regression, kernel regression O(N2), Gaussian process regression O(N3)

4. Classification: decision tree, nearest-neighbor classifier O(N2), nonparametric Bayes classifier O(N2), support vector machine O(N3)

5. Dimension reduction: principal component analysis, non-negative matrix factorization, kernel PCA O(N3), maximum variance unfolding O(N3)

6. Outlier detection: by density estimation or dimension reduction7. Clustering: by density estimation or dimension reduction, k-means, mean-

shift segmentation O(N2), hierarchical clustering O(N3)8. Time series analysis: Kalman filter, hidden Markov model, trajectory

tracking O(Nn)9. Feature selection and causality: LASSO, L1 SVM, Gaussian graphical

models, discrete graphical models10.Fusion and matching: sequence alignment, bipartite matching O(N3), n-

point correlation 2-sample testing O(Nn)

5 settings

1. “Regular”: batch, in-RAM/core, one CPU

2. Streaming (non-batch)

3. Disk (out-of-core)

4. Distributed: threads/multi-core (shared memory)

5. Distributed: clusters/cloud (distributed memory)

4 common data types

1. Vector data, iid

2. Time series

3. Images

4. Graphs

3 desiderata

1. Fast experimental runtime/performance*

2. Fast theoretic (provable) runtime/performance*

3. Accuracy guarantees

*Performance: runtime, memory, communication, disk accesses; time-constrained, anytime, etc.

7 general solution strategies

1. Divide and conquer (indexing structures)

2. Dynamic programming

3. Function transforms

4. Random sampling (Monte Carlo)

5. Non-random sampling (active learning)

6. Parallelism

7. Problem reduction

1. Summary statistics

• Examples: counts, contingency tables, means, medians, variances, range queries (SQL queries)

• What’s unique/challenges: streaming, new guarantees

• Promising/interesting: – Sketching approaches– AD-trees– MapReduce/Hadoop (Aster,Greenplum,Netezza)

2. Generalized N-body problems

• Examples: nearest-nbrs (in NLDR, etc), kernel summations (in KDE, GP, SVM, etc), clustering, MST, spatial correlations

• What’s unique/challenges: general dimension, non-Euclidean, new guarantees (e.g. in rank)

• Promising/interesting: – Generalized/higher-order FMM O(N2) O(N)

– Random projections

– GPUs

3. Graph-theoretic problems

• Examples: betweenness centrality, commute dist, graphical model inference

• What’s unique/challenges: high interconnectivity (cliques), out-of-core

• Promising/interesting: – Variational methods– Stochastic composite likelihood methods– MapReduce/Hadoop (Facebook,etc)

4. Linear-algebraic problems

• Examples: linear algebra, PCA, Gaussian process regression, manifold learning

• What’s unique/challenges: probabilistic guarantees, kernel matrices

• Promising/interesting: – Sampling-based methods– Online methods– Approximate matrix-vector multiply via N-body

5. Optimizations

• Examples: LP/QP/SDP/SOCP/MINLPs in regularized methods, compressed sensing

• What’s unique/challenges: stochastic programming, streaming

• Promising/interesting: – Reformulations/relaxations of various ML forms– Online, mini-batch methods– Parallel online methods– Submodular functions– Global optimization (non-convex)

6. Integrations

• Examples: Bayesian inference

• What’s unique/challenges: general dimension

• Promising/interesting: – MCMC– ABC– Particle filtering– Adaptive importance sampling, active learning

7. Alignments

• Examples: BLAST in genomics, string matching, phylogenies, SLAM, cross-match

• What’s unique/challenges: greater heterogeneity, measurement errors

• Promising/interesting: – Probabilistic representations– Reductions to generalized N-body problems

Reductions/transformationsbetween problems

• Gaussian graphical models linear alg• Bayesian integration MAP optimization• Euclidean graphs N-body problems• Linear algebra on kernel matrices N-body

inside conjugate gradient• Can featurize a graph or any other structure

matrix-based ML problem• Create new ML methods with different

computational properties

General conclusions

• Algorithms can dramatically change the runtime order, e.g. O(N2) to O(N)

• High dimensionality is a persistent challenge• The non-default (e.g. streaming, disk…)

settings need more research work• Systems issues need more work, e.g.

connection to data storage/management• Hadoop does not solve everything

General conclusions

• No general theory for the tradeoff between statistical quality and computational cost (lower/upper bounds, etc)

• More aspects of hardness (statistical and computational) are needed