Temporal decision making using unsupervised learning
Omar A. Ibrahim · 2021
With the explosion of ubiquitous continuous sensing, on-line streaming clustering continues to attract attention. The requirements are that the streaming clustering algorithm recognize and adapt clusters as the data evolves, that anomalies are detected, and that new clusters are automatically formed as incoming data dictate. In this dissertation, we develop a streaming clustering algorithm, MU Streaming Clustering (MUSC), that is based on coupling a Gaussian mixture model (GMM) with possibilistic clustering to build an adaptive system for analyzing streaming multi-dimensional activity feature vectors. For this reason, the possibilistic C-Means (PCM) and Automatic Merging Possibilistic Clustering Method (AMPCM) are combined together to cluster the initial data points, detect anomalies and initialize the GMM. MUSC achieves our goals when tested on synthetic and real-life datasets. We also compare MUSC's performance with Sequential k-means (sk-means), Basic Sequential Clustering Algorithm (BSAS), and Modified BSAS (MBSAS) here MUSC shows superiority in the performance and accuracy. The performance of a streaming clustering algorithm needs to be monitored over time to understand the behavior of the streaming data in terms of new emerging clusters and number of outlier data points. Incremental internal Validity Indices (iCVIs) are used to monitor the performance of an on-line clustering algorithm. We study the internal incremental Davies-Bouldin (DB), Xie-Beni (XB), and Dunn internal cluster validity indices in the context of streaming data analysis. We extend the original incremental DB (iDB) to a more general version parameterized by the exponent of membership weights. Then we illustrate how the iDB can be used to analyze and understand the performance of MUSC algorithm. We give examples that illustrate the appearance of a new cluster, the effect of different cluster sizes, handling of outlier data samples, and the effect of the input order on the resultant cluster history. In addition, we investigate the internal incremental Davies-Bouldin (iDB) cluster validity index in the context of big streaming data analysis. We analyze the effect of large numbers of samples on the values of the iCVI (iDB). We also develop online versions of two modified generalized Dunn's indices that can be used for dynamic evaluation of evolving (cluster) structure in streaming data. We argue that this method is a good way to monitor the ongoing performance of online clustering algorithms and we illustrate several types of inferences that can be drawn from such indices. We compare the two new indices to the incremental Xie-Beni and Davies-Bouldin indices, which to our knowledge offer the only comparable approach, with numerical examples on a variety of synthetic and real data sets. We also study the performance of MUSC and iCVIs with big streaming data applications. We show the advantage of iCVIs in monitoring large streaming datasets and in providing useful information about the data stream in terms of emergence of a new structure, amount of outlier data, size of the clusters, and order of data samples in each cluster. We also propose a way to project streaming data into a lower space for cases where the distance measure does not perform as expected in the high dimensional space. Another example of streaming is the data acivity data coming from TigerPlace and other elderly residents' apartments in and around Columbia. MO. TigerPlace is an eldercare facility that promotes aging-in-place in Columbia Missouri. Eldercare monitoring using non-wearable sensors is a candidate solution for improving care and reducing costs. Abnormal sensor patterns produced by certain resident behaviors could be linked to early signs of illness. We propose an unsupervised method for detecting abnormal behavior patterns based on a new context preserving representation of daily activities. A preliminary analysis of the method was conducted on data collected in TigerPlace. Sensor firings of each day are converted into sequences of daily activities. Then, building a histogram from the daily sequences of a resident, we generate a single data vector representing that day. Using the proposed method, a day with hundreds of sequences is converted into a single data point representing that day and preserving the context of the daily routine at the same time. We obtained an average Area Under the Curve (AUC) of 0.9 in detecting days where elder adults need to be assessed. Our approach outperforms other approaches on the same datset. Using the context preserving representation, we develoed a multi-dimensional alert system to improve the existing single-dimensional alert system in TigerPlace. Also, this represenation is used to develop a framework that utilizes sensor sequence similarity and medical concepts extracted from the EHR to automatically inform the nursing staff when health problems are detected. Our context preserving representation of daily activities is used to measure the similarity between the sensor sequences of different days. The medical concepts are extracted from the nursing notes using MetamapLite, an NLP tool included in the Unified Medical Language System (UMLS). The proposed idea is validated on two pilot datasets from twelve Tiger Place residents, with a total of 5810 sensor days out of which 1966 had nursing notes.