Tanvir Ahmad

Feature and Opinion Mining for Customer Review Summarization

In this paper, we present an opinion mining system to identify product features and opinions from review documents. The features and opinions are extracted using semantic and linguistic analysis of text documents. The polarity of opinion sentences is established using polarity scores of the opinion words through Senti-WordNet to generate a feature-based summary of review documents. The system is also integrated with a visualization module to present feature-based summary of review documents in a comprehendible way.

Securing multimedia colour imagery using multiple high dimensional chaos-based hybrid keys

The rapid exchange of multimedia data over worldwide available internet and shared networks has encouraged its unauthorised access, illegal usage, disruption, alteration. Therefore, it demands efficient methods to provide security to multimedia contents for ensuring authorised access, preventing illegal disruption, alteration, etc. To meet the aforesaid needs, the authors propose an efficient encryption method to secure the multimedia colour imagery. Complex dynamic responses of multiple high-order chaotic systems are utilised to carry out image pixels shuffling and diffusion processes under the control of secret key. The pixels diffusion is done by randomly picking the actual encryption keys out of nine hybridised keys that are extracted from complex sequences of Chen, Rossler and Chua chaotic systems. Eventually, a high encryption effect is turned up in the encrypted multimedia colour images. Moreover, the shuffling and diffusion processes are made plain-image information dependent to resist the potential chosen-plaintext, chosen-ciphertext and known-plaintext attacks. The simulation results validate that the proposed method has great encryption performance and practicableness.

Opinion Mining Using Frequent Pattern Growth Method from Unstructured Text

In the last one decade, the area of opinion mining has experienced a major growth because of the increase in online unstructured data which are contributed by reviewers over different topics and subjects. These data sometimes become important for users who want to take their decision based on opinions of actual users of the product. In this paper, we present the FP-growth method for frequent pattern mining from review documents which act as a backbone for mining the opinion words along with their relevant features by experimental data over two different domains which are very different in their nature.

Neighborhood Topology to Discover Influential Nodes in a Complex Network

This paper addresses the issue of distinguishing influential nodes in the complex network. The k-shell index features embeddedness of a node in the network based upon its number of links with other nodes. This index filters out the most influential nodes with higher values for this index, however, fails to discriminate their scores with good resolution, hence results in assigning same scores to the nodes belonging to same k-shell set. Extending this index with neighborhood coreness of a node and also featuring topological connections between its neighbors, our proposed method can express the nodes influence score precisely and can offer distributed and monotonic rank orders than other node ordering methods.

Review ranking method for spam recognition

E-commerce websites are becoming popular among customers who are buying products online. Online reviews play a major role in selling of online products. Reviews give the customer a complete overview of the product thus making it popular or unpopular among buyers thus increasing its sales. In order to increase sales of a product, reviewers are writing fake reviews. In this paper, a review ranking method is proposed. This method assigns a score to each review based on different parameters. The reviews having high score are considered to be more helpful or genuine and thus are ranked higher than the reviews having lower score. The lower ranked reviews are fake reviews and thus they are non-useful to the users. The proposed approach is an effective approach which avoids heavy computation of learning. Evaluation on real-life flipkart review dataset shows a precision of 83.3% thus showing the effectiveness of proposed model.

Random Forest for the Real Forests

A forest is a vast area of land covered predominantly with trees and undergrowth. In this paper, adhering to cartographic variables, we try to predict the predominant kind of tree cover of a forest using the Random Forests (RF) classification method. The study classifies the data into seven classes of forests found in the Roosevelt National Forest of Northern Colorado. With sufficient data to create a classification model, the RF classifier gives reasonably accurate results. Fine-tuning of the algorithm parameters was done to get promising results. Besides that a dimensionality check on the dataset was conducted to observe the possibilities of dimensionality reduction.

Ranking of Fuzzy Similar Faces Using Relevance Matrix and Aggregation Operators

Abstract In perception based imaging, Sketching With Words (SWW) is a well-established methodology in which the objects of computation are fuzzy geometric objects (f-objects).The problem of facial imaging of criminal on the basis of onlooker statement is not lack of method and measures but the modeling of onlooker(s) mind set. Because the onlooker has to give statements about different human face parts like forehead, eyes, nose, and chin etc.The concept of fuzzy similarity (f-similarity) and proper aggregation of components of face may provide more flexibility to onlooker(s). In proposed work onlooker(s) statement is recorded. Thereafter it is compared with existing statements. The f-similarity with different faces in database is estimated by using ‘as many as possible’ linguistic quantifier. Three types of constraints over size of parts of face ‘small’, ‘medium’, and ‘large’ are considered. Possibilistic constraints with linguistic hedges and negation operator like ‘very long’, ‘not long’, ‘not very long’ etc. are used. Moreover we have generated ranking of alike faces in decreasing order by using the concepts of f-similarity and relevance matrix.

Improving Classification by Outlier Detection and Removal

Most of the existing state-of-art techniques for outlier detection and removal are based upon density based clustering of given dataset. In this paper we have suggested a novel approach for iteratively pruning of outliers based upon the non-alignment with model created in a n-dimensional hyperspace. The technique could be used for any classification problem as a pre-processing step, regardless of the classifier used. We have tested our hypothesis with Support Vector and RandomForest classifiers and have obtained significant improvement in results for both these classifiers. The effectiveness of this novel method has also been verified by improvements in results while performing classification using these standard classifiers. When pruned with our method standard classifiers like SVC and RandomForest classifier showed an improvement up to 4 percent.

Satire Detection from Web Documents Using Machine Learning Methods

Satire exposes humanitys vices and foibles through the use of irony, wit, and sometimes sarcasm too. It is also frequently used in online communities. Recognition of satire can help in many NLP applications like dialogue system and review summarization. In this paper we filter online news articles as satirical or true news documents using SVM (Support Vector Machine) classification method combined with machine learning techniques. With ample training documents SVM tends to give good classification results. For obtaining promising results with SVM an understanding of its working and ways to influence its accuracy is required. We also use various feature extraction strategies and conclude that TF-IDF-BNS feature extraction gives maximum accuracy for detection of satire in web content.

Mining Local Association Rules from Temporal Data Set

In this paper, we present a novel approach for finding association rules from locally frequent itemsets using rough set and boolean reasoning. The rules mined so are termed as local association rules. The efficacy of the proposed approach is established through experiment over retail dataset that contains retail market basket data from an anonymous Belgian retail store.