AutoML

AutoML (Automated Machine Learning) is the application of automation to the process of designing, building, and optimising machine learning pipelines. It encompasses the automatic selection of the most suitable ML algorithm for a given dataset, the engineering and selection of input features, the tuning of model hyperparameters, and the configuration of the full end-to-end pipeline from raw data to deployed model. The goal of AutoML is to reduce the manual expertise, time, and computational effort required to develop effective ML models, making machine learning more accessible to practitioners without deep specialisation and accelerating the work of experienced researchers. AutoML addresses what has been called the full model selection problem: given a dataset and a task objective, find the best combination of preprocessing steps, model architecture, and hyperparameter settings to maximise predictive performance. This problem is computationally intractable to solve exhaustively, so AutoML systems use principled search strategies to find high-quality solutions efficiently.

Core Components

Automated Feature Engineering

Feature engineering — the process of transforming raw data into informative representations suitable for ML algorithms — has traditionally been one of the most labour-intensive and knowledge-dependent steps in building effective models. Automated feature engineering tools search over a space of transformations (aggregations, encodings, interaction terms, polynomial features) and select those that improve model performance. Tools such as Featuretools and the feature engineering components of platforms like H2O Driverless AI and DataRobot automate large portions of this process, drawing on a library of transformation primitives applied to structured data. For unstructured data such as images, text, and audio, transfer learning from pre-trained deep learning models largely replaces hand-crafted feature engineering.

Hyperparameter Optimisation

Most ML algorithms have hyperparameters — settings that are fixed before training and control model complexity, regularisation, and learning dynamics. Examples include the learning rate and depth in gradient boosted trees, the number of layers and neurons in a neural network, and the regularisation strength in a support vector machine. Finding good hyperparameter values manually requires expert intuition and extensive trial and error. AutoML systems automate this search using several strategies. Grid search exhaustively evaluates a predefined grid of hyperparameter values but scales poorly with the number of hyperparameters. Random search samples randomly from the hyperparameter space and is more efficient than grid search for high-dimensional spaces. Bayesian optimisation builds a probabilistic surrogate model of the objective function and uses it to select the next hyperparameter configuration most likely to improve on the best result so far. Successive halving and Hyperband strategies allocate compute budgets adaptively, terminating poorly performing configurations early to focus resources on promising ones.

Neural Architecture Search

Neural Architecture Search (NAS) applies AutoML principles to the design of deep neural network architectures. Instead of relying on human-designed architectures, NAS algorithms search over a predefined space of architectural choices — number of layers, layer types, skip connections, activation functions — to find architectures that achieve high performance on the target task. Early NAS methods were computationally expensive, requiring thousands of GPU hours, but efficient NAS approaches including DARTS (Differentiable Architecture Search) and weight-sharing methods have substantially reduced this cost.

Pipeline Combination (Combined Algorithm Selection and Hyperparameter Optimisation)

The full AutoML problem — selecting both the algorithm and its hyperparameters jointly — is known as the Combined Algorithm Selection and Hyperparameter Optimisation (CASH) problem. Systems such as Auto-sklearn and TPOT (Tree-based Pipeline Optimisation Tool) address CASH by treating the full pipeline configuration as a structured search problem, using meta-learning (learning from results on prior datasets) to warm-start the search with promising configurations.

Major Platforms

Cloud-based AutoML platforms include Google Cloud AutoML (now integrated into Vertex AI AutoML), Azure Machine Learning's automated ML, and Amazon SageMaker Autopilot. These services provide point-and-click or API-based access to AutoML for tabular data, image classification, and text classification, handling data preprocessing, training, and deployment as managed services. Open-source AutoML frameworks include Auto-sklearn (built on Scikit-learn), AutoKeras (built on Keras/TensorFlow), Auto-PyTorch, FLAML from Microsoft Research, and TPOT. H2O Driverless AI and DataRobot are commercial platforms with especially strong support for enterprise tabular data use cases.

Limitations and Considerations

AutoML does not eliminate the need for domain expertise. Problem framing — deciding what to predict, how to define success, and how to handle domain-specific data quality issues — remains a fundamentally human task. AutoML also tends to produce models that are optimised for a single metric on a held-out validation set, without necessarily considering interpretability, fairness, or operational constraints. Computational cost is another practical limitation. Full AutoML searches can require substantial compute, and cloud AutoML services incur costs that scale with search time. Efficient AutoML strategies and early stopping are important for practical deployment. AutoML is particularly relevant to Malaysia's AI democratisation agenda. The country has a significant shortage of experienced ML engineers relative to the demand created by its ambitions in the MyDigital Blueprint and Malaysia's National AI Roadmap. AutoML platforms reduce the barrier to entry for companies without large data science teams, enabling business analysts and domain experts in manufacturing, finance, and agriculture to build useful ML models without deep technical specialisation. MDEC (Malaysia Digital Economy Corporation) and the Malaysia Productivity Corporation (MPC) have both promoted the adoption of AI tools including AutoML as part of their SME digitalisation initiatives. Small and medium enterprises in sectors such as retail, logistics, and palm oil processing can potentially apply AutoML to predictive maintenance, demand forecasting, and supply chain optimisation without needing to hire specialist ML engineers. In Malaysian financial services, AutoML has been used to accelerate credit scoring model development. Banks and licensed digital lenders operating under BNM's regulatory framework must validate and document their credit models, and AutoML platforms that produce explainable, auditable pipelines are better aligned with BNM's model risk management requirements than opaque custom models. Universiti Teknologi Malaysia (UTM) and Universiti Putra Malaysia (UPM) have incorporated AutoML into their data science and AI curricula, recognising that graduates must be familiar with automation tools that are standard in industry practice. Research groups at these institutions have applied AutoML to local agricultural prediction problems, including crop yield forecasting for paddy and palm oil using sensor and satellite data. HRD Corp has co-funded industry-linked training programmes in Malaysia that include modules on AutoML tools, ensuring that Malaysian data practitioners are familiar with both the capabilities and limitations of automated approaches. This is part of a broader push to build a data-literate workforce capable of applying AI practically in business contexts.

See Also

1. Hutter, F., Kotthoff, L., and Vanschoren, J. (Eds.) (2019). Automated Machine Learning: Methods, Systems, Challenges. Springer.
2. Feurer, M. et al. (2015). Efficient and Robust Automated Machine Learning. Advances in Neural Information Processing Systems 28 (NeurIPS 2015).
3. Zoph, B., and Le, Q.V. (2017). Neural Architecture Search with Reinforcement Learning. Proceedings of ICLR 2017.
4. Google Cloud. (2025). Vertex AI AutoML Documentation. cloud.google.com/vertex-ai/docs/automl.
5. MDEC. (2023). SME AI Adoption Playbook. Malaysia Digital Economy Corporation, Cyberjaya.