IFODHD: Improved Feature Selection Based Outlier Detection using Hyperdimensional Computing

Outlier detection, also called anomaly detection, is commonly applied to diverse applications, e.g., finance, healthcare, and the Internet of Things (IoT). Current approaches for anomaly detection range from traditional statistical methods and modern parametric methods to machine learning techniques and deep learning approaches. prior approaches to outlier detection using hyperdimensional computing (HDC) incur long latency, require significant memory, and achieve low receiver operating characteristic area under curve (ROC-AUC). This paper proposes feature selection using minimum Redundancy and Maximum Relevance (mRMR) followed by hyperdimensional computing (HDC) for outlier detection. The integration of HDC and mRMR feature selection technique reduces the number of features for anomaly detection, resulting in more robust and accurate performance. We test our model over eight outlier detection datasets that are publicly available. The performance is evaluated using three metrics: accuracy, F1 score, and ROC-AUC. We also evaluate execution time and memory usage of the proposed models. Experimental results indicate that our proposed models outperform the current traditional state-of-the-art algorithms, and state-of-the-art HDC algorithms such as ODHD, for anomaly detection over eight datasets for every metric. Over seven datasets (not including Optdigits), our proposed models obtain improvements up to \(13\%\) $13\%$ in accuracy, \(33\%\) $33\%$ in F1 score, and \(75\%\) $75\%$ in ROC-AUC; they achieve 13X reduction in latency and up to 190X reduction in memory usage. For Optdigits, our proposed models obtain improvements of \(29\%\) $29\%$ in accuracy, \(9,219\%\) $9,219\%$ in F1 score, and \(87\%\) $87\%$ in ROC-AUC; they achieve 3X reduction in latency and 190X reduction in memory usage.