Search

In the present work, the study on class imbalance problems in a distributed setting exploiting sparsity structure in the data has been carried out. We formulate the class-imbalance learning problem as a cost-sensitive learning problem with $L_1$ L 1 regularization. The cost-sensitive loss function is a cost-weighted smooth hinge loss. The resultant optimization problem is minimized within the Distributed Alternating Direction Method of Multiplier (DADMM) framework. We partition the data matrix across samples. This operation splits the original problem into a distributed $L_2$ L 2 regularized smooth loss minimization and a $L_1$ L 1 regularized squared loss minimization. $L_2$ L 2 regularized subproblem is solved via Limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) and random coordinate descent method in parallel at multiple processing nodes using MPI whereas $L_1$ L 1 regularized problem is just a simple soft-thresholding operation. We show, empirically, that the distributed solution approximates the centralized solution on many benchmark data sets. The centralized solution is obtained via Cost-Sensitive Stochastic Coordinate Descent (CSSCD). Empirical results on small and large-scale benchmark datasets show some promising avenues to further investigate the real-world applications of the proposed algorithms such as anomaly detection, class-imbalance learning, etc. To the best of our knowledge, ours is the first work to study class-imbalance in a distributed environment on large-scale sparse data.

Class imbalance learning is the study of problems in which some classes appear more frequently than the others. Most existing works that study this problem assume data set to be dense and do not exploit the rich structure of the data. One such structure is the sparsity. In the present work, we focus on solving the class imbalance problem under the sparsity assumption. More specifically, a well-known Gmean metric for class imbalance learning problem in binary classification setting has been maximized, which results in a non-convex loss function. Convex relaxation techniques are used to convert the non-convex problem to the convex problem. The problem formulation in the present work uses L1 regularized proximal learning framework and is solved via accelerated-stochastic-proximal gradient descent algorithm. Our aim in the paper is to show: (i) the application of proximal algorithms to solve real world problems (class imbalance); (ii) how it scales to Big data; and (iii) how it outperforms some recently proposed algorithms in terms of Gmean, F-measure and Mistake rate on several benchmark data sets.

Account payables are amount owed to vendors for goods and services delivered to a company. Vendors raise invoices which go through several processing steps before they are paid by a company. Companies have contractual obligations with vendors for paying the invoices within a stipulated time. Invoices that exceed this time attract penalty and affect vendor satisfaction to work with the company. It is very critical for large firms dealing with thousands of vendors for their day to day operations to meet the service level agreements with vendors to avoid penalties. Any assistance for practitioners, warning them of potential invoices that can breach the service level agreements, can help them in minimizing the penalties. In this research, we model the problem of identifying delayed invoices as a supervised classification task. There are three characteristics of this problem which are challenging from a classification perspective: (i) the status of an invoice is affected by other invoices that are simultaneously being processed, as there are limited resources to process the huge volume of invoices, (ii) feature engineering to capture the temporal aspect of the invoice and having the optimal representation of the multiple data entries created per invoice, and (iii) the number of paid late invoices are much smaller in percentage compared to paid on time invoices in the training data set, hence the classes are imbalanced. The results obtained by training an ensemble of classifiers show that penalties can be avoided on more than 82% of the invoices which are currently being penalized.

The ubiquitous nature of the Internet and its applicability as an inexpensive advertising media has resulted in Web and mobile platforms to be on target for product advertising. Existing advertising approaches on different social media platforms include sponsored search, e-mail advertising, banner ads, etc., to attract customers. The current systems for advertising on these platforms require the advertisers to manually come up with a catchy tagline, which is time-consuming and a challenging task. To overcome this problem, we propose a novel framework to automatically generate novel and catchy taglines tailored for advertisers' products. The framework consists of a domain-specific knowledge graph, a novel modified long short-term memory (LSTM) (we call it CopyLSTM), and a neural paraphrase model. We also curate and build a database of fashion products with 100K input product features and 300K sentences (three sentences per product data). This dataset is used to populate the knowledge graph and learn the modified LSTM to generate taglines. The proposed framework is evaluated on different metrics for sentence quality and compared with the state-of-art methods for tagline generation. Subjective evaluation was performed to understand the novelty and creativity of the generated tagline.

Anomaly detection is an important task in many real world applications such as fraud detection, suspicious activity detection, health care monitoring etc. In this paper, we tackle this problem from supervised learning perspective in online learning setting. We maximize well known \emph{Gmean} metric for class-imbalance learning in online learning framework. Specifically, we show that maximizing \emph{Gmean} is equivalent to minimizing a convex surrogate loss function and based on that we propose novel online learning algorithm for anomaly detection. We then show, by extensive experiments, that the performance of the proposed algorithm with respect to $sum$ metric is as good as a recently proposed Cost-Sensitive Online Classification(CSOC) algorithm for class-imbalance learning over various benchmarked data sets while keeping running time close to the perception algorithm. Our another conclusion is that other competitive online algorithms do not perform consistently over data sets of varying size. This shows the potential applicability of our proposed approach., This paper is accepted for publication in IC3 2015, Jaypee Noida

Anomaly detection has drawn a slew of attention in recent years, although term has been known as outlier detection in statistics several decades ago. Everyday large volume of data is being generated. For example, flight navigation data, health care monitoring data, social media data, video surveillance data etc. This data contains rare events or anomalous points that needs to be found out-for example less than 2 % of all visitors who visits Amazon website make a purchase. Thus anomaly detection problem can be interesting due to business perspective, security, maintenance etc. The problem becomes challenging because of noise, heterogeneity, high dimensionality of the data. This paper studies a robust algorithm, based on support vector data description, for anomaly detection. We perform extensive experiments on real data coming from nuclear power plant to empirically demonstrate the effectiveness of the algorithm as well as finding anomalies in the data set. We also discuss extensions of the algorithm to find anomalies in high dimension and non linearly separable data.

Conventional statistical analysis of Internet traffic data is often employed to determine traffic distribution, to summarize users behavior patterns, or to predict future network traffic for network management and planning. However, the statistical techniques like autoregressive integrated moving average (ARIMA) models fail to capture some peculiar network traffic characteristics like Self-similarity, Long-range dependency (LRD), etc. With rapid growth of Internet, accurate and reliable forecasting is essential for the network resource management. Therefore, the present work explores to utilize computational intelligence techniques like fuzzy logic to predict Internet traffic data pattern. This paper presents the development of Fuzzy inference system (FIS) for one benchmark dataset and one real dataset. The experimental results of the FIS of the benchmark dataset Mackey-Glass series and real dataset JNU LAN traffic data collected for one hour are presented. The performance of the FIS is estimated using cross-validation method by measuring root mean square error (RMSE).