| dc.description.abstract | Time-series outlier detection reveals uncommon points or patterns with abnormal behaviors within time-series datasets and settings. It is a crucial research area to explore because it can be helpful for many real-world scenarios. Some popular fields, such as fraud detection, healthcare, cancer detection, cybersecurity attack detection, and fault detection, could benefit from time-series outlier detection. For example, in real-world fraud detection, millions of transactions are fed into the database daily. The outlier detection system needs to recognize a suspicious transaction or pattern as soon as possible. If we manually download the data daily to make predictions on it, this would take too much time and effort, and most importantly, it could potentially be too late to detect the fraud case. Therefore, in these real-world databases, time-series data becomes a real challenge for researchers to explore.
Oftentimes, engineers take a dataset and manually build a fixed-designed neural network for a specific task to predict and recognize outliers. However, this isn’t the optimal strategy for treat-ing time-series data. A fixed-designed neural network will not have enough power to capture all the details inside a time-series data. Each time-series outlier detection task will require different network architectures to detect outlier points and patterns accurately.
In general machine learning, researchers have found a way to search for the best model accord-ing to the unique behaviors of each dataset, which is Automated Machine Learning (AutoML). AutoML automates machine learning tasks and workflows with different techniques, which bene-fits non-experts to use machine learning models more quickly. Some standard methods for AutoML include hyperparameter optimization, meta-learning, and neural architecture search (NAS).
In this thesis proposal, by adapting and modifying the traditional NAS strategies, we propose a new method to construct a suitable search space with the proper size and combine the power of different deep learning time-series outlier detection algorithms with AutoML searching methods to search an effective neural network for time-series outlier detection automatically. | |
