Automated Feature Engineering for Deep Neural Networks with Genetic Programming

Thesis (PDF Available) · April 2017with 681 Reads
DOI: 10.13140/RG.2.2.22508.16002
Thesis for: PhD - Computer Science, Advisor: James Cannady
Cite this publication
Abstract
Feature engineering is a process that augments the feature vector of a machine learning model with calculated values that are designed to enhance the accuracy of a model’s predictions. Research has shown that the accuracy of models such as deep neural networks, support vector machines, and tree/forest-based algorithms sometimes benefit from feature engineering. Expressions that combine one or more of the original features usually create these engineered features. The choice of the exact structure of an engineered feature is dependent on the type of machine learning model in use. Previous research demonstrated that various model families benefit from different types of engineered feature. Random forests, gradient-boosting machines, or other tree-based models might not see the same accuracy gain that an engineered feature allowed neural networks, generalized linear models, or other dot-product based models to achieve on the same data set. This dissertation presents a genetic programming-based algorithm that automatically engineers features that increase the accuracy of deep neural networks for some data sets. For a genetic programming algorithm to be effective, it must prioritize the search space and efficiently evaluate what it finds. This dissertation algorithm faced a potential search space composed of all possible mathematical combinations of the original feature vector. Five experiments were designed to guide the search process to efficiently evolve good engineered features. The result of this dissertation is an automated feature engineering (AFE) algorithm that is computationally efficient, even though a neural network is used to evaluate each candidate feature. This approach gave the algorithm a greater opportunity to specifically target deep neural networks in its search for engineered features that improve accuracy. Finally, a sixth experiment empirically demonstrated the degree to which this algorithm improved the accuracy of neural networks on data sets augmented by the algorithm’s engineered features.
Figures - uploaded by Jeffrey Theodore Heaton
Author content
All content in this area was uploaded by Jeffrey Theodore Heaton
Content may be subject to copyright.
A preview of the PDF is not available
  • Thesis
    Full-text available
    In order to ensure the validity of sensor data, it must be thoroughly analyzed for various types of anomalies. Traditional machine learning methods of anomaly detections in sensor data are based on domain-specific feature engineering. A typical approach is to use domain knowledge to analyze sensor data and manually create statistics-based features, which are then used to train the machine learning models to detect and classify the anomalies. Although this methodology is used in practice, it has a significant drawback due to the fact that feature extraction is usually labor intensive and requires considerable effort from domain experts. An alternative approach is to use deep learning algorithms. Research has shown that modern deep neural networks are very effective in automated extraction of abstract features from raw data in classification tasks. Long short-term memory networks, or LSTMs in short, are a special kind of recurrent neural networks that are capable of learning long-term dependencies. These networks have proved to be especially effective in the classification of raw time-series data in various domains. This dissertation systematically investigates the effectiveness of the LSTM model for anomaly detection and classification in raw time-series sensor data. As a proof of concept, this work used time-series data of sensors that measure blood glucose levels. A large number of time-series sequences was created based on a genuine medical diabetes dataset. Anomalous series were constructed by six methods that interspersed patterns of common anomaly types in the data. An LSTM network model was trained with k-fold cross-validation on both anomalous and valid series to classify raw time-series sequences into one of seven classes: non-anomalous, and classes corresponding to each of the six anomaly types. As a control, the accuracy of detection and classification of the LSTM was compared to that of four traditional machine learning classifiers: support vector machines, Random Forests, naive Bayes, and shallow neural networks. The performance of all the classifiers was evaluated based on nine metrics: precision, recall, and the F1-score, each measured in micro, macro and weighted perspective. While the traditional models were trained on vectors of features, derived from the raw data, that were based on knowledge of common sources of anomaly, the LSTM was trained on raw time-series data. Experimental results indicate that the performance of the LSTM was comparable to the best traditional classifiers by achieving 99% accuracy in all 9 metrics. The model requires no labor-intensive feature engineering, and the finetuning of its architecture and hyper-parameters can be made in a fully automated way. This study, therefore, finds LSTM networks an effective solution to anomaly detection and classification in sensor data.
  • Conference Paper
    A great deal of research has focused on algorithms for learning features from un- labeled data. Indeed, much progress has been made on benchmark datasets like NORB and CIFAR by employing increasingly complex unsupervised learning al- gorithms and deep models. In this paper, however, we show that several very sim- ple factors, such as the number of hidden nodes in the model, may be as important to achieving high performance as the choice of learning algorithm or the depth of the model. Specifically, we will apply several off-the-shelf feature learning al- gorithms (sparse auto-encoders, sparse RBMs and K-means clustering, Gaussian mixtures) to NORB and CIFAR datasets using only single-layer networks. We then present a detailed analysis of the effect of changes in the model setup: the receptive field size, number of hidden nodes (features), the step-size (stride) be- tween extracted features, and the effect of whitening. Our results show that large numbers of hidden nodes and dense feature extraction are as critical to achieving high performance as the choice of algorithm itselfso critical, in fact, that when these parameters are pushed to their limits, we are able to achieve state-of-the- art performance on both CIFAR and NORB using only a single layer of features. More surprisingly, our best performance is based on K-means clustering, which is extremely fast, has no hyper-parameters to tune beyond the model structure it- self, and is very easy implement. Despite the simplicity of our system, we achieve performance beyond all previously published results on the CIFAR-10 and NORB datasets (79.6% and 97.0% accuracy respectively).
  • Article
    Machine learning models, such as neural networks, decision trees, random forests and gradient boosting machines accept a feature vector and provide a prediction. These models learn in a supervised fashion where a set of feature vectors with expected output is provided. It is very common practice to engineer new features from the provided feature set. Such engineered features will either augment, or replace portions of the existing feature vector. These engineered features are essentially calculated fields, based on the values of the other features. Engineering such features is primarily a manual, time-consuming task. Additionally, each type of model will respond differently to different types of engineered features. This paper reports on empirical research to demonstrate what types of engineered features are best suited to which machine learning model type. This is accomplished by generating several datasets that are designed to benefit from a particular type of engineered feature. The experiment demonstrates to what degree the machine learning model is capable of synthesizing the needed feature on its own. If a model is capable of synthesizing an engineered feature, it is not necessary to provide that feature. The research demonstrated that the studied models do indeed perform differently with various types of engineered features.
  • Chapter
    This article presents SimpleDS, a simple and publicly available dialogue system trained with deep reinforcement learning. In contrast to previous reinforcement learning dialogue systems, this system avoids manual feature engineering by performing action selection directly from raw text of the last system and (noisy) user responses. Our initial results, in the restaurant domain, report that it is indeed possible to induce reasonable behaviours with such an approach that aims for higher levels of automation in dialogue control for intelligent interactive systems and robots.
  • Chapter
    We introduce fully automated process for sentiment analysis in short texts in Polish language. Process consists of (a) generation of emotion lexicon using Twitter annotated messages (b) building sentiment data set using annotated messages and the generated lexicon, (c) training NEAT genetic algorithm using previously prepared data set and (d) the final evaluation using 10 fold cross validation. We show that this method provides good results and can be used to simplify sentiment analysis processes for Polish language content.