To view and reproduce the experiment, please follow the colab notebook. The notebook has already been setup to replicate the require code in the VM with the specified random state from results/rstate.pickle that will reproduce the results from the paper.
With almost every problem that involves similarity detection or categorical classification, the subject of classification in the collected data must be statistically significant for a model to perform relatively well. Such classification or similarity detection becomes much harder when the statistical presence of information of interest in the dataset is much lesser. For such anomaly detection, we propose a different paradigm than typical classification and clustering techniques. Instead of explicitly classifying the anomalies, we focus on modeling the predominant normality of the dataset and hence endeavor to detect deviations in the form of anomalies.
In typical classifications and classification models, the specific classes of interest to be labeled usually predominate the data. In contrast, Anomaly detection techniques focus on labelling and classifying datasets where the occurrence of subject is statistically rare. In certain scenarios where such occurrences in datasets are extremely rare, even traditional anomaly detection techniques fail to detect occurrences accurately. Additionally, for such scenarios data labelling can prove even more difficult.
This project focuses on anomaly detection in time series problems. We make two assumptions about the dataset. First, we assume that most of the given data is normal and only a very small percentage is abnormal. Second, we anticipate abnormal data is statistically various from normal. According to these two assumptions, data groups of similar instances which appear frequently are assumed to be normal, while infrequently instances which considerably various from the majority of the instances are regarded anomalous. For our experimentation we employ a different approach to anomaly detection. Instead of following in the path of traditional techniques and modeling the infrequent anomalies of interest in the dataset, we focus on models that train on the predominant "normal" scenarios of dataset to detect deviations in the datasets in hopes to detect the anomalies.
One such problems in anomaly detection of time series is Daphnet Freezing of Gait (FOG) [1]. The FOG dataset is devised to benchmark automatic methods to recognize gait freeze from wearable acceleration sensors placed on legs and hips. The dataset was recorded in the lab with emphasis on generating many freeze events. Users performed three kinds of tasks: straight line walking, walking with numerous turns, and finally a more realistic activity of daily living (ADL) task, where users went into different rooms while fetching coffee, opening doors, etc.
For this project, we mainly work with proven time series ensembles such as neural networks including LSTM and convolutional techniques as well as support vector machines. We experiment with both supervised and unsupervised methodologies, while focusing on unsupervised techniques as they best fit the problem description. The goal of the project is to train these models on time series datasets in order to detect anomalies and deviations in the time series.
[1] Bächlin, M., Plotnik, M., Roggen, D., Maidan, I., Hausdorff, J. M., Giladi, N., & Tröster, G. (2010). Wearable assistant for Parkinson's disease patients with the freezing of gait symptom. IEEE transactions on information technology in biomedicine : a publication of the IEEE Engineering in Medicine and Biology Society, 14(2), 436–446. https://doi.org/10.1109/TITB.2009.2036165
[2] Omar, S., Ngadi, M., Jebur, H., & Benqdara, S. (2013). Machine Learning Techniques for Anomaly Detection: An Overview_. International Journal of Computer Applications, 79._ http://dx.doi.org/10.5120/13715-1478
[3] Teng, M. (2010). Anomaly detection on time series. 2010 IEEE International Conference on Progress in Informatics and Computing, 1, 603–608. https://doi.org/10.1109/PIC.2010.5687485
[4] Ordóñez, F., & Roggen, D. (2016). Deep Convolutional and LSTM Recurrent Neural Networks for Multimodal Wearable Activity Recognition. Sensors, 16(1), 115. https://doi.org/10.3390/s16010115