Abstract
We introduce a means of automating machine learning (ML) for big data tasks, by performing scalable stochastic Bayesian optimisation of ML algorithm parameters and hyper-parameters. More often than not, the critical tuning of ML algorithm parameters has relied on domain expertise from experts, along with laborious hand-tuning, brute search or lengthy sampling runs. Against this background, Bayesian optimisation is finding increasing use in automating parameter tuning, making ML algorithms accessible even to non-experts. However, the state of the art in Bayesian optimisation is incapable of scaling to the large number of evaluations of algorithm performance required to fit realistic models to complex, big data. We here describe a stochastic, sparse, Bayesian optimisation strategy to solve this problem, using many thousands of noisy evaluations of algorithm performance on subsets of data in order to effectively train algorithms for big data. We provide a comprehensive benchmarking of possible sparsification strategies for Bayesian optimisation, concluding that a Nystrom approximation offers the best scaling and performance for real tasks. Our proposed algorithm demonstrates substantial improvement over the state of the art in tuning the parameters of a Gaussian Process time series prediction task on real, big data.
Introduction
Hyperparameter tuning is often considered the "dark art" of machine learning. Selecting optimal learning rates, kernel parameters, or architecture depths typically requires significant domain expertise or exhaustive, computationally expensive searches. While Bayesian Optimization (BO) has emerged as a powerful framework for automating this process by using Gaussian Process (GP) surrogate models, it faces two significant bottlenecks in big data contexts:
- Computational Complexity: Standard GPs scale cubically ($O(N^3)$) with the number of evaluations, making thousands of iterations infeasible.
- Evaluation Overhead: Evaluating a single hyperparameter configuration on a massive dataset is prohibitively slow.
To address these challenges, we present STOAT (STOchastic Algorithm Tuning), a framework designed to bring scalable, automated machine learning to big data environments.
The STOAT Framework
STOAT overcomes the limitations of traditional BO through two primary innovations: scalable sparse GPs and stochastic evaluations.
Scalable Sparse Gaussian Processes
Instead of a full GP, STOAT utilizes sparse approximations to maintain computational efficiency. We evaluated several techniques, including Fully Independent Training Conditional (FITC), Laplacian-GPs, and Nyström approximations. While Laplacian-GPs are effective in low dimensions, they suffer from an exponential growth in the number of required basis functions as dimensionality increases, particularly for functions with short lengthscales.
Our analysis identifies the Nyström approximation as the superior choice. By intelligently selecting kernel eigenpairs, it provides a robust balance between predictive accuracy and computational scaling, especially in high-dimensional parameter spaces.
Stochastic Bayesian Optimization
Rather than evaluating the objective function on the entire dataset, STOAT performs "noisy" evaluations using stochastically selected mini-batches. Although these samples are noisier, the underlying GP surrogate is specifically designed to handle this uncertainty. This allows the optimizer to perform thousands of rapid, inexpensive evaluations, exploring the parameter space far more effectively than a few costly "exact" evaluations.
Experimental Evaluation
We benchmarked STOAT against standard methods using both synthetic functions and real-world big data.
Surrogate Model Performance and the "Uncertainty Trap"
A critical insight from our benchmarking involves the behavior of different regressors in unexplored regions. As shown in Figure 2, while Random Forest Regressors (RFRs) are computationally fast, they often produce misleadingly low uncertainty estimates in flat or unexplored areas. This "uncertainty trap" causes the optimizer to prematurely stop exploring. In contrast, GP-based methods (Full, Nyström, Laplacian, and FITC) maintain principled uncertainty budgets, ensuring continued exploration until the global minimum is found.
Real-World Application: AgentSwitch Energy Data
We applied STOAT to the AgentSwitch dataset, comprising over 34,000 observations of household electricity usage. The task involved tuning a complex "double periodic" time-series model. STOAT successfully tuned the model on a standard laptop in approximately one hour, achieving a test log-likelihood of -7.25. This significantly outperformed the previous state-of-the-art (-7.40) and simpler baseline models (-9.22), demonstrating STOAT's efficacy on real-world big data.
Conclusion
STOAT provides a scalable solution for automated machine learning on large datasets. By combining Nyström-based sparse GPs with stochastic evaluations, it avoids the $O(N^3)$ scaling wall while remaining robust to noisy performance metrics. Future work will explore removing the need for initial pre-sampling by marginalizing hyperparameters through MCMC or Bayesian Quadrature, further streamlining the automation process.
Summary updated