Unsupervised feature selection for sensor time-series in pervasive computing applications

Abstract

The paper introduces an efficient feature selection approach for multivariate time-series of heterogeneous sensor data within a pervasive computing scenario. An iterative filtering procedure is devised to reduce information redundancy measured in terms of time-series cross-correlation. The algorithm is capable of identifying nonredundant sensor sources in an unsupervised fashion even in presence of a large proportion of noisy features. In particular, the proposed feature selection process does not require expert intervention to determine the number of selected features, which is a key advancement with respect to time-series filters in the literature. The characteristic of the prosed algorithm allows enriching learning systems, in pervasive computing applications, with a fully automatized feature selection mechanism which can be triggered and performed at run time during system operation. A comparative experimental analysis on real-world data from three pervasive computing applications is provided, showing that the algorithm addresses major limitations of unsupervised filters in the literature when dealing with sensor time-series. Specifically, it is presented an assessment both in terms of reduction of time-series redundancy and in terms of preservation of informative features with respect to associated supervised learning tasks

ADMITS: Architecting Distributed Monitoring and Analytics in IoT-based Disaster Scenarios

“…Based on a threshold value and the trust level rendered by the FD, it can take decisions about the confidence degree of IoT and collected data. AD-MITS investigates alternatives to amend such effects, maintaining the representativeness of collected data under disaster conditions in order to keep the computed models usable [Shcherbakov et al 2017, Bacciu 2016.…”

An ambient intelligence approach for learning in smart robotic environments

“…The former exploits adaptation of the trainable weights of the model in closed form by ridge regression (as described above by Equation ), whereas the latter implements a recursive least squares algorithm for online weights adjustment . Although both mechanisms are available in RUBICON, in this paper, we focus on showing the adaptation abilities of the system using the retraining functionalities. Unsupervised feature selection , based on the iterative cross‐correlation filter algorithm by Bacciu, that allows automatic identification of those input sources that are either redundant or provide irrelevant information. A model selection mechanism with integrated supervised feature selection for RC, allowing automatic selection of the best learning model configuration and parameters and the determination of the most predictive input sources for a computational learning task. In particular, model selection among different LI‐ESN hyperparametrizations is implemented by using a hold‐out cross‐validation strategy (see Section 4.2 for details on data set splitting), and exploring the hyperparameters space by grid search that takes into account the major aspects of LI‐ESN design (see Section 4.2 for details on the hyperparameters and the corresponding values, explored for the experiments described in this paper).…”

DropIn: Making reservoir computing neural networks robust to missing inputs by dropout

“…Each device is equipped with light, temperature, humidity and passive infrared (PIR) presence sensors. For our Kitchen Cleaning experiment, we consider only those inputs which have been deemed relevant by the feature selection analysis in [19], that are the PIR readings of the first five motes plus the information on the x position of the robot in the environment as measured by the rangefinder localization system. The dataset contains information on the robot initiating 104 Kitchen Cleaning tasks: these consist in the robot moving from its base station in the living room to the kitchen, where a user might be present or not or might enter the kitchen during robot navigation.…”

“…In this task, the learning model is required to learn to predict a user preference that was not modelled in the robotic planner domain knowledge, that is the fact of not having the robot cleaning the kitchen when the user is in. This task is part of a larger effort to assess the self-adaptation abilities of a robotic ecology planner developed as part of the RUBICON project, including also automated feature selection mechanisms on time series [19]. The experimental scenario consists of a real-world flat sensorized by an RFID floor, a mobile robot with range-finder localization and a WSN with six mote-class devices.…”