Source-linked AI summary
Automated data processing and feature engineering for deep learning and big data applications: a survey
Alhassan Mumuni, Fuseini Mumuni
TL;DR
Large, complex, heterogeneous datasets make manual data preparation and low-level processing difficult, while prior surveys have insufficiently covered automated preprocessing and feature engineering. This survey reviews automated preprocessing, augmentation, feature engineering, and end-to-end AutoML approaches, finding that many tasks can be automated but context-dependent search and competing requirements remain challenging.
Problem
Large, complex, heterogeneous datasets and stringent requirements make data processing difficult, while low-level automated preprocessing and feature engineering remain insufficiently surveyed.
Method
The survey reviews automated preprocessing, data augmentation, feature engineering, holistic deep learning frameworks, and generic AutoML tools for Big Data applications.
Results
Many data processing tasks can already be seamlessly automated in state-of-the-art AutoML pipelines, but full automation remains unresolved for some tasks.
Takeaways & Limitations
Automated processing methods, especially AutoML approaches, have simplified and accelerated deep learning model development as data volume and problem complexity increase.
Takeaways & Limitations
Automated systems must balance competing requirements such as data reduction, predictive accuracy, interpretability, robustness, and fairness, which may require human input.
Abstract
from arXiv · showhide
Modern approach to artificial intelligence (AI) aims to design algorithms that learn directly from data. This approach has achieved impressive results and has contributed significantly to the progress of AI, particularly in the sphere of supervised deep learning. It has also simplified the design of machine learning systems as the learning process is highly automated. However, not all data processing tasks in conventional deep learning pipelines have been automated. In most cases data has to be manually collected, preprocessed and further extended through data augmentation before they can be effective for training. Recently, special techniques for automating these tasks have emerged. The automation of data processing tasks is driven by the need to utilize large volumes of complex, heterogeneous data for machine learning and big data applications. Today, end-to-end automated data processing systems based on automated machine learning (AutoML) techniques are capable of taking raw data and transforming them into useful features for Big Data tasks by automating all intermediate processing stages. In this work, we present a thorough review of approaches for automating data processing tasks in deep learning pipelines, including automated data preprocessing--e.g., data cleaning, labeling, missing data imputation, and categorical data encoding--as well as data augmentation (including synthetic data generation using generative AI methods) and feature engineering--specifically, automated feature extraction, feature construction and feature selection. In addition to automating specific data processing tasks, we discuss the use of AutoML methods and tools to simultaneously optimize all stages of the machine learning pipeline.
1 INTRODUCTION
The survey addresses the growing difficulty of preparing complex, heterogeneous Big Data for machine learning and the limited coverage of automated low-level processing in prior surveys. It focuses on automating preprocessing, data augmentation, and feature engineering within deep learning pipelines.
- Machine learning development remains time-consuming because experts must collect, prepare, engineer, train, evaluate, and iteratively refine models.
- This survey examines automated preprocessing, data augmentation, and feature engineering, including extraction, construction, and selection.
- Big Data applications aggregate multimodal data from diverse sources for processing and downstream deep learning tasks.
- Prior surveys have emphasized end-to-end AutoML tools, while low-level preprocessing and feature engineering received comparatively insufficient coverage.
- The paper also covers end-to-end workflows, generic AutoML tools, industrial implications, and future prospects for automated data processing.
2 AUTOMATED DATA PROCESSING IN MACHINE
Traditional pipelines implement processing stages manually, whereas deep learning integrates feature extraction and other processing into end-to-end training. Automation increasingly uses learned mechanisms and AutoML to handle preprocessing, augmentation, feature engineering, model construction, and optimization.
- Traditional machine learning performs data preparation, augmentation, feature extraction, construction, selection, model choice, and hyperparameter optimization as manual standalone stages.
- Deep learning integrates processing and model training end-to-end, allowing tunable neural parameters to adjust feature extraction during learning.
- Data augmentation increases training-data quantity and variability by manipulating existing samples, addressing bias, imbalance, and domain shift.
- Explicit feature engineering remains useful for complex, data-scarce, noisy, heterogeneous, or interpretability-sensitive applications.
- Automated processing ranges from hand-crafted routines and learned mechanisms to AutoML methods that optimize processing and modeling choices.
- End-to-end AutoML pipelines combine preprocessing, augmentation, feature engineering, model construction, hyperparameter setting, algorithm selection, and evaluation.
The concept of data preprocessing
Data preprocessing transforms raw data into forms usable by machine-learning models and includes cleaning, labeling, encoding, and missing-data imputation. Automation increasingly uses machine learning and AutoML to select context-appropriate operations and optimize their ordering, but categorical encoding and labeling remain difficult to automate fully.
- Data preprocessing transforms raw data into a form useful for machine-learning models through cleaning, labeling, categorical encoding, and missing-data imputation.
- Automated preprocessing selects and applies operations that maximize predictive performance, while also determining transformation strength and operation order.
- Preprocessing is highly sensitive to both data type and machine-learning model, making exhaustive operation combinations impractical for very large datasets.
- Automated data imputation: Automated imputation uses learned data distributions or AutoML search across multiple models and optimization techniques to detect and correct missing values.
- Categorical data encoding: Complete automation of categorical encoding remains challenging, so many systems use traditional encoders and may require human supervision to define categories.
- Automated data labeling: Data labeling remains difficult to automate fully, with most annotations produced manually or semi-automatically and refined or validated by humans.
4 AUTOMATED DATA AUGMENTATION
Automated data augmentation creates task-appropriate training variations through learned, generative, or optimized transformations, often improving performance over manual approaches. Its benefits are constrained by computational cost and risks that transformations may alter semantic meaning.
- Data augmentation generates variations of existing training data to increase quantity and variability, addressing limited representation, bias, imbalance, and domain shift.
- Generative methods can synthesize novel data from scratch when training data do not exist, while GANs commonly generate realistic images or transfer knowledge across domains.
- Automated augmentation has mainly succeeded in computer vision, whereas semantic-preserving transformations remain difficult for time series forecasting and natural language processing.
- Automated augmentation searches transformation operations and their magnitudes, using analytical rules, learned mechanisms, generative models, and optimization techniques.
- Automated augmentation can require enormous computational resources, and unguided transformations may change data semantics or labels, creating erroneous training signals.
5 AUTOMATED FEATURE ENGINEERING
Feature engineering aims to improve data representation by transforming extracted features and retaining compact subsets that preserve predictive information. Traditional approaches are computationally expensive, domain-dependent, and not always optimal, motivating automated alternatives.
- Feature engineering manipulates extracted feature vectors to increase representativeness while seeking a compact subset that preserves necessary predictive information.
- Traditional feature engineering uses assumptions, feature statistics, and analytical transformations, but it is computationally expensive, domain-dependent, and often suboptimal.
5.1 General concept of automated feature engineering
Automated feature engineering constructs a search space of feature-processing operations and optimizes it for a specific dataset and task. Methods can decompose large spaces into reusable components and are supported by specialized AutoML tools.
- Automated feature engineering creates a large space of possible operations and searches for the combination that performs best on a specific dataset and task.
- The formal problem transforms an original feature set F into additional features Φ, then selects a subset over Ω = F + Φ to maximize performance on target task Tsk.
- Large datasets make exhaustive search difficult because many operations and their interactions must be evaluated, so smaller candidate spaces can be tested first.
- Hierarchical decompositions represent search spaces as atomic subspaces that can be combined according to the task, dataset, computational budget, and training constraints.
- Dedicated AutoML utilities automate feature extraction, particularly for time series and tabular data, including TSFEL, tsfresh, tsfeaturex, and Featuretools.
5.2 Common feature engineering tasks and approaches for automation
Automated feature engineering commonly covers extraction, synthesis or construction, and selection, although their boundaries and relationships vary across the literature. Approaches use transformations, neural architecture search, iterative selection, and meta-learning while facing combinatorial, memory, and computational challenges.
- Common feature engineering tasks: The literature distinguishes feature extraction, feature synthesis or construction, and feature selection, but definitions and relationships among them are inconsistent.
- Feature extraction: Feature extraction produces compact, representative features with reduced computational and memory demands, and can improve interpretability while preserving encoded information.
- Feature extraction: Automated extraction methods can combine mathematical operations with selection algorithms or use neural architecture search and transferred architectures to identify useful features.
- Feature synthesis: Feature synthesis generates complementary features by optimizing transformations and hyperparameters, either expanding the feature set before selection or processing operations iteratively.
- Feature synthesis: Aggressive feature expansion requires high memory, whereas iterative expansion reduces memory demand but can impose evaluation overhead and discard useful intermediate features.
- Feature selection: Feature selection is combinatorial because feature usefulness depends on combinations, and different applications may require one optimal subset or all minimal useful subsets.
5.3 End-to-end automated feature engineering
End-to-end automated feature engineering combines feature extraction, construction, and selection to produce useful transformations while controlling feature growth. Approaches include neural transformation search and differentiable optimization.
- Holistic systems automate feature extraction, synthesis, and selection rather than treating these engineering tasks separately.
- Neural transformation search selects effective transformations, then applies further engineering operations to improve the resulting features.This expansion-reduction process limits redundant features that increase computational and memory demands.
- SAFE reduces the initial feature set, expands it through feature combinations, and filters the resulting features according to combination performance.
- NFS uses reinforcement-learning RNN controllers to learn transformation policies, while DIFER maps discrete features into continuous space and optimizes them by gradient ascent.
5.4 Automated feature engineering based on NAS frameworks
Neural Architecture Search is used to automate feature selection by searching for effective feature subsets, including separately optimized spatial and channel-wise features.
- NAS-based feature engineering searches for effective feature subsets in multimodal person re-identification using RGB and infrared images.
- The method constructs a two-level search space that independently selects spatial and channel-wise features.
5.5 Performance of automated feature engineering methods
The survey compares automated and traditional feature engineering, describes end-to-end AutoML pipelines, and reports that automated methods generally outperform basic traditional approaches, though gains vary by dataset.
- 5.5 Performance of automated feature engineering methods: The evaluation compares raw, random, one-operation, brute-force, and automated feature engineering methods using F-1 scores across 23 datasets.The automated methods include LFE, NFS, AutoFeat, and DIFER.
- 5.5 Performance of automated feature engineering methods: Automated feature engineering techniques clearly outperform basic traditional paradigms, but improvements are limited or absent on spambase, autos, and convex.
- End-to-end AutoML frameworks: End-to-end AutoML frameworks can combine data collection, preprocessing, error correction, feature engineering, hyperparameter optimization, and model architecture construction.
- End-to-end AutoML frameworks: AutoSmart integrates preprocessing, table merging, feature synthesis and selection, ensemble learning, and hyperparameter tuning with computational resource control.
- End-to-end AutoML frameworks: AutoPrognosis processes patient data end-to-end to support prognosis and provide explanations for predictions.
7 GENERIC AUTOML TOOLS FOR DATA PROCESS-
Generic AutoML tools automate workflow stages for nonexpert users by evaluating alternative pipelines across tasks and data types. They range from open-source to proprietary systems and increasingly include visualization for interpretation and trust assessment.
- Generic AutoML tools take raw data through processing and model generation, allowing nonexpert users to develop machine-learning solutions without manually building pipelines.
- Frameworks create and validate alternative pipelines with different algorithms and hyperparameter settings, often using specialized components for image, text, and tabular data.
- Common tool functions include task analysis, problem recognition, model evaluation, validation, and user-facing graphical interfaces.
- Visualization supports interpretability, robustness checks, knowledge discovery, and trust by exposing model behavior and enabling feature-removal analysis.
- Integrated platforms combine technologies and data pipelines, speeding big-data solution development by eliminating additional development, testing, and deployment stages.
- AutoML tools include open-source systems built by developer communities and proprietary systems targeting business users with technical knowledge but limited data-science expertise.
8 IMPLICATIONS FOR INDUSTRY AND COMMERCE
Automated data processing and AutoML reduce manual preparation and model-development work across industries, enabling large-scale heterogeneous data analysis. Applications described include business, healthcare, agriculture, finance, retail, and manufacturing.
- Automated techniques eliminate manual data preparation, model selection, tuning, and evaluation, reducing errors and improving productivity and result accuracy.
- Automation lets business users move from raw inputs directly to results, shortening development-to-market time and supporting analysis of very large heterogeneous datasets.
- AutoML and generative AI tools lower businesses’ reliance on scarce AI specialists by presenting complex implementations through intuitive interfaces.
- 8.1 Healthcare: Healthcare systems can process extensive medical data and knowledge bases to support diagnosis, treatment planning, and medical question answering.
- 8.1 Healthcare: Automated medical data processing is described as reducing human factors that can produce costly errors through distraction and fatigue.
- Applications extend to personalized treatment and drug discovery, agricultural prediction, financial analysis, customer services, and manufacturing operations.
9 DISCUSSIONS
The discussion presents automated data processing as a way to accelerate machine-learning development while emphasizing unresolved trade-offs, reliability concerns, scope limitations, and data-quality requirements. It identifies human-in-the-loop, explainability, and progressive synthetic-data refinement as future directions.
- AutoML-based processing simplifies and accelerates deep-learning development as datasets and machine-learning problems become larger and more complex.
- Future systems are expected to automate the full range of tasks from data acquisition through model construction and validation.
- Automated systems must balance performance, complexity, cost, accuracy, interpretability, robustness, and fairness according to user priorities.
- 9.1 Limitations: Generative models may produce insufficiently varied, generic, unreliable, or harmful data, limiting their utility in high-stakes domains.
- 9.1 Limitations: Automatically generated data may perform well on tuned tasks yet generalize poorly, while transferability and adaptability remain insufficiently investigated.
- 9.1 Limitations: Black-box automation makes it difficult to guarantee intended performance and alignment with safety, privacy, ethics, and broader user objectives.
- 9.1 Limitations: Generic AutoML tools primarily support tabular classification and regression, with limited coverage of images, audio, video, point clouds, time series, and natural language processing.
- 9.2 Future directions: Future directions include human feedback, privacy-aware infrastructure, progressive self-refinement of synthetic data, and explainable end-to-end processing.
10 CONCLUSION
Automated data processing is increasingly important because machine learning applications must handle growing data volumes. The survey reviews automation across preprocessing, augmentation, feature engineering, and end-to-end pipelines, while finding that full automation remains challenging.
- Growing data volumes and demand for machine learning solutions have increased the importance of automated data processing.
- The survey covers preprocessing, data augmentation, and feature engineering as the three main categories of automated processing tasks.Feature engineering includes feature extraction, construction, and selection.
- Reviewed preprocessing methods include data cleaning, imputation, labeling, and categorical encoding, alongside automated augmentation and feature engineering.
- The survey also examines end-to-end deep learning frameworks and generic AutoML frameworks that address processing steps holistically.
- Many processing tasks can already be automated in state-of-the-art AutoML pipelines, but full automation remains unresolved for some tasks.The survey identifies context-dependent collection and exploitation of relevant data as a particular challenge for effective search mechanisms.