P. Bagavathi Sivakumar

Human Gait Recognition and Classification Using Time Series Shapelets

Human gait is the main activity of daily life. Gait can be used for applications like human identification (in medical field etc). Since gait can be perceived from a distance it can be used for human identification. Gait recognition means identifying the person with his/her gait. Human identification using gait can be used in surveillance. A method is proposed for gait recognition using a technique which uses time series shapelets. First, for a gait video a preprocessing is done to extract the silhouette images from the video. From these silhouette images features like joint angle and swing distance are extracted which can be represented as the time series data. From this time series data, time series shapelets are extracted. Shapelets are subsequence of time series data which can discriminate between classes. Shapelets are maximally representative of the class. These time series shapelets can be used to identify human by their gait. Shapelets can also be used for classification. After extracting the shapelets, the prediction is done using the decision tree. In that it can be used for classifying normal and abnormal human gait.

A Novel IoT-Based Energy Management System for Large Scale Data Centers

The high energy consumption in data centers is becoming a major concern because it leads to increased operating costs and also, pollution, as fuel is burnt to produce the required energy. While many techniques and methods have been proposed by various organizations and researchers to minimize the energy consumption, there has been considerably less work done in making a smart-energy management system that is capable of collecting the data available and make decisions based on the energy consumption patterns. In this work, a smart system is proposed that uses Internet of Things to gather data and a machine learning algorithm for decision making.

A Machine Learning Approach to Cluster the Users of Stack Overflow Forum

Online question and answer (Q&A) forums are emerging as excellent learning platforms for learners with varied interests. In this paper, we present our results on the clustering of Stack Overflow users into four clusters, namely naive users, surpassing users, experts, and outshiners. This clustering is based on various metrics available on the forum. We use the X-means and expectation maximization clustering algorithms and compare the results. The results have been validated using internal, external, and relative validation techniques. The objective of this clustering is to be able to trace and predict the activity of a user on this forum. According to our results, majority of users (71 % of 40,000 users under consideration) fall in the ‘experts’ category. This indicates that the users in Stack Overflow are of high quality thereby making the forum an excellent platform for users to learn about computer programming.

A Crowdsourcing-Based Platform for Better Governance

The world’s population has been increasing as every year passes by, and Governments across the world face a stupendous challenge of governing each country. These challenges include providing proper sanitation facilities, efficient disaster management techniques, effective resource allocation and management, etc. Crowdsourcing methodologies, which empower the common man to provide valuable information for better decision making, have gained prominence recently to tackle several challenges faced by several governments. In this paper, we introduce a crowdsourcing-based platform that makes use of information provided by the common man for better governance. We illustrate how this platform can be used in several instances to attend to the problems faced by people.

Modeling and predicting stock returns using the ARFIMA-FIGARCH

Modeling of real world financial time series such as stock returns are very difficult, because of their inherent characteristics. ARIMA and GARCH models are frequently used in such cases. It is proven of late that, the traditional models may not produce the best results. Lot of recent literature says the successes of hybrid models. The modeling and forecasting ability of ARFIMA-FIGARCH model is investigated in this study. It is believed that data such as stock returns exhibit a pattern of long memory and both short term and long term influences are observed. Empirical investigation has been made on closing stock prices of S&P CNX NIFTY. The obtained statistical result shows the modeling power of ARFIMA-FIGARCH. The performance of this model is compared with traditional Box and Jenkins ARIMA models. It is proven that, by combining several components or models, one can account for long range dependence found in financial market volatility. The results obtained illustrate the need for hybrid modeling.

Evaluating the Predictability of Financial Time Series

A discrete- time signal or time series is set of observations taken sequentially in time, space, or some other independent variable. Examples occur in various areas including engineering, natural sciences, economics, social sciences and medicine. Financial time series in particular are very difficult to model and predict, because of their inherent nature. Hence, it becomes essential to study the properties of signal and to develop quantitative techniques. The key characteristics of a time series are that the observations are ordered in time and that adjacent observations are related or dependent. In this paper a case study has been performed on the BSE and NSE index data and methods to classify the signals as Deterministic, Random or Stochastic and White Noise are explored. This pre-analysis of the signal forms the basis for further modeling and prediction of the time series.

Smart Energy Management System Based on Image Analytics and Device Level Analysis

Large educational institutes, organization, and industries face large challenges on energy utilities, consumption and its management strategies. But smart energy management technology solutions help the high energy consumption complexities during while putting the best and greenest foot forward. Smart Energy Management technology solutions, improve and respond quickly to power spikes at times of demand, expedite data gathering, reporting and regulatory compliance, automate services to control operating costs and enable to save energy. Connecting smart meters to data stores requires a reliable, intelligent network. The Smart Energy Management System introduced here deals with a device level analysis that gives information of the device that has caused the peak rise in the total power consumption of the organization. Predictive analysis technique is used on the database to predict the future maximum demand and load balancing technique is applied to reduce the consumption of power from generator source. Therefore the total power consumption from exceeding the maximum demand can be avoided and the maximum demand of the power supply for the organization can be maintained. Further, on application of AI techniques this system control becomes fully automated.

Use of Predictive Analytics Towards Better Management of Parking Lot Using Image Processing

As more and more smart cities are planned in India, there is a growing need for smart parking and smart transportation. Parking has been identified as a major challenge to traffic network and urban life quality. Already most of the cities are facing the problem of pollution. Due to drivers struggling for finding the parking area, 30% of traffic congestion occurs according to industry data. There is also a need for secure, efficient, intelligent and reliable systems that can be used for searching the unoccupied parking facilities, guide towards the parking facilities, and negotiate the parking fee. This would help in the proper management of the parking facility. There is no publically available data on parking in India. This work would be useful in creation of such datasets. Image based model has been proposed to identify the slot occupancy status. A prediction model has also been incorporated in the system to predict the occupancy rate and thereby help the management in better management of parking lots. One of the machine learning method, linear regression is used for predicting the number of car parked every hour. A slot based approach was used and the performances of prediction algorithms were compared.

Forecasting the Stability of the Data Centre Based on Real-Time Data of Batch Workload Using Times Series Models

Forecasting has diverse range of applications in many fields like weather, stock market, etc. The main highlight of this work is to forecast the values of the given metric for near future and predict the stability of the Data Centre based on the usage of that metric. Since the parameters that are being monitored in a Data Centre are large, an accurate forecasting is essential for the Data Centre architects in order to make necessary upgrades in a server system. The major criteria that result in SLA violation and loss to a particular business are peak values in performance parameters and resource utilization; hence it is very important that the peak values in performance, resource and workload be forecasted. Here, we mainly concentrate on the metric batch workload of a real-time Data Centre. In this work, we mainly focused on forecasting the batch workload using the auto regressive integrated moving average (ARIMA) model and exponential smoothing and predicted the stability of the Data Centre for the next 6 months. Further, we have performed a comparison of ARIMA model and exponential smoothing and we arrived at the conclusion that ARIMA model outperformed the other. The best model is selected based on the ACF residual correlogram, Forecast Error histogram and the error measures like root mean square error (RMSE), mean absolute error (MAE), mean absolute scale error (MASE) and p-value of Ljung-Box statistics. From the above results we conclude that ARIMA model is the best model for forecasting this time series data and hence based on the ARIMA models forecast result we predicted the stability of the Data Centre for the next 6 months.

Bayesian nonparametric Multiple Instance Regression

Multiple Instance Regression jointly models a set of instances and its corresponding real-valued output. We present a novel multiple instance regression model that infers a subset of instances in each bag that best describes the bag label and uses them to learn a predictive model in a unified framework. We assume that instances in each bag are drawn from a mixture distribution and thus naturally form groups, and instances from one of this group explain the bag label. The largest cluster is assumed to be correlated with the label. We evaluate this model on the crop yield prediction and aerosol depth prediction problems. The predictive accuracy of our model is better than the state of the art MIR methods.