Aruna Malapati

Effect of Feature Selection on Kinase Classification Models

Classification of kinases will provide comparison of related human kinases and insights into kinases functions and evolution. Several algorithms exist for classification and most of them failed to classify when the dimension of feature set large. Selecting the relevant features for classification is significant for variety of reasons like simplification of performance, computational efficiency, and feature interpretability. Generally, feature selection techniques are employed in such cases. However, there has been a limited study on feature selection techniques for classification of biological data. This work tries to determine the impact of feature selection algorithms on classification of kinases. We have used forward greedy feature selection algorithm along with random forest classification algorithm. The performance was evaluated by selecting the feature subset which maximizes Area Under the ROC Curve (AUC). The method identifies the feature subset from the datasets which contains the physiochemical properties of kinases like amino acid, dipeptide, and pseudo amino acid composition. An improvised performance of classification is noted for feature subset than with all the features. Thus, our method indicates that groups of kinases are classifiable with maximum AUC, if good subsets of features are used.

TelNEClus: Telugu Named Entity Clustering Using Semantic Similarity

Semantic similarity plays a significant role in many of natural language processing (NLP) and information retrieval (IR) applications. Most IR methodologies represent the documents using the vector space model (VSM) traditionally known as bag-of-words (BoW) hypothesis. The main disadvantage of BoW is that the grammatical and the structural information of words is not captured. In this paper, we have attempted to cluster named entities (NEs) extracted from Telugu corpus based on semantic similarity. We contend that for this sort of work, more suited VSM is distributional hypothesis which is usually applied for measuring word similarity using word-context matrix. In the word-context matrix, the row vector is words given in the corpus; here, it is a proper noun as most of NEs are proper noun, and column vector is context such as windows of words, grammatical information. The row vector in word-context matrix is constructed in two ways with two different feature sets: The first way is to represent each NE with unique row vector (Row Vector1) without considering different occurrences in a corpus, and the second way is to represent each NE with a set of row vectors (Row Vector2) considering different occurrences in a corpus. For Row Vector1 representation, classical similarity functions like cosine, scalar product, Jaccard can be utilized, but for Row Vector2 representation, we have generalized similarity functions to Sum-of-Sum and Sum-of-Max. Experimentally, Row Vector2 representation enhances the clustering results.

CapAct: A Wordnet-Based Summarizer for Real-World Events from Microblogs

Short messages from microblog streams often contain information about real-world events. Streams of related messages can be clustered and classified as events or non-events. Summarizing events from clusters of event related messages is a challenging task as the summary needs to be concise yet informational. We present a novel method of summarization of events from short messages. We also propose a method of creating a set of extensive reference summaries from manually created summaries for effective evaluation. We used standard ROUGE based metrics to compare the proposed summarizer with many existing baselines including a strong Hybrid-tfidf method. Our summarizer consistently outperformed others in F1-score with a margin of 11% in ROUGE-1 and 5% in ROUGE-2 over Hybrid-tfidf.

A Comparative Study of Named Entity Recognition for Telugu

In this paper, we apply three classification learning algorithms to Telugu Named Entity Recognition (NER) task and we present a comparative study between these three learning algorithms on Telugu dataset (NER for South and South-East Asian Languages (NERSSEAL) Competition). The empirical results show that Support Vector Machine achieves the best F-measure of 54.78% on the dataset.

A novel similarity measure: Voronoi audio similarity for genre classification

One of the major challenges in genre classification, recommender systems is to find similarity between the query song and songs in a database. In this paper, we propose a novel similarity measure called Voronoi audio similarity (VAS). We extracted the Content-based features from the audio signal of the song split in frames over a particular time period and we represented each song as a point in 2D space. The proposed system is a two-level classification process, where songs are first clustered by K-means clustering and then a Voronoi diagram is created using centroids from the resulting K-means, which is called the template Voronoi diagram (TVD). This approach learns the decision boundary used for genre classification. The genre of the song could thus be predicted as the genre with the maximum normalised area overlap. Empirical results performed with 10 cross-fold validations on million song subsets of 500 songs showed 78% accuracy.

Software Defect Prediction Using Augmented Bayesian Networks.

Prediction models are built with various machine learning algorithms to identify defects prior to release to facilitate software testing, and save testing costs. Naive Bayes classifier is one of the best performing classification techniques in defect prediction. It assumes conditional independence of features and for defect prediction problem some of the features are not actually conditionally independent. The interesting problem is to relax these conditional independence assumptions and to check whether there is any improvement in performance of classifiers. We have built Bayesian Network structures using different classes of algorithms namely score-based, constraint-based and hybrid algorithms. We propose an approach to augment these Bayesian Network structures with class node. Bayesian Network classifiers along with Random Forests, Logistic Regression and Naive Bayes classifiers are then evaluated using measures like AUC and H-measure. We observe that RSMAX2 and Grow-Shrink classifiers (after augmentation) perform consistently better in defect prediction.

Teaching Computer Programming Using MOOCs in Multiple Campuses: Challenges and Solutions

In recent times MOOCs has become a technology enabled platform which allows delivery of content in locations geographically separated out. In Birla Institute of Technology and Science (BITS), Pilani the EdX platform has been successfully used to deliver content over multiple campuses namely Pilani Campus, Goa Campus, and Hyderabad Campus. We present in this paper challenges faced in implementing this pedagogy, solutions attempted, and other tools required to make learning of Computer Programming course, a better experience. In particular, we have described how it was handled in BITS Pilani, Hyderabad Campus

Towards next-generation alert management of data centers

The performance of todays enterprise IT systems depends upon the accurate generation of alerts to identify any anomalous behavior. Todays solutions however suffer from several shortcomings. The configurations are manual and do not adapt to any system changes. A large volume of redundant alerts are generated leading to inefficient resolution. Moreover, the generated alerts are reactive in nature and provide less time to take corrective actions. In this paper, we address the issues of setting the correct configurations, aggregating redundancies and generating proactive alerts. We demonstrate the effectiveness of our approach on a real-world case-study.

A Study on Feature Selection and Classification Techniques of Indian Music

In this paper we present the effect of four feature selection algorithms namely genetic algorithm, Forward feature selection, information gain and correlation based on four different classifiers (Decision tree C4.5, K-Nearest neighbors, neural network and support vector machine). The feature sets used in this paper are extracted features from the preprocessed songs using MIR Toolbox in MATLAB, which encompass rhythm based, timbre based, pitch based, tonality based and dynamic features. Feature vectors are extracted from music segments from first 30 seconds and last thirty seconds of the music signal (time-decomposition). Experiments were carried out on the three dominant genres of Indian music: Carnatic, Hindustani and Bollywood. Our dataset is small with 290 songs, trimmed to extract the first and the last 30 second percepts. As pure Carnatic and Hindustani music being more prevalent in traditional settings, have limited work done to make their digital copies available but the collection of music we have used consists of songs of some of the most profound singers contributing to each of these genres. For high-dimensional feature sets, the feature selection provides a compact but discriminative feature subset which has an interesting trade-off between classification accuracy and computational effort. The experimental results have shown that the common features selected by each of the feature selection algorithms with respect to classifiers and percentage of classification accuracies for all the classification algorithms. Furthermore, it can be observed from our experiment that information gain based feature selection gives better and consistent accuracies than other feature selection algorithms and Neural network and SVM classifiers are the best suited classifiers for Indian Song dataset.

Playlist generation based on user perception of songs

Large online music collections often frustrate users and have increased the importance of recommender systems. This has led to interesting problem of automated playlist generation. Most of the existing playlists compare a pair songs based on low-level/mid-level features and calculate the similarity. These systems lack user perception of music. This work supplements such existing systems by providing user perception of songs conveyed in Twitter messages. The proposed system combines audio based features and sentiment associated with the song. This unique fusion not only yields better results but also better user satisfaction. Further a validation on 200 users who used our playlist showed that atleast 67% of the songs in the playlist were liked by the user.