Deduplication is the process of removing replicated data content from storage facilities like online databases, cloud datastore, local file systems, etc. It is commonly performed as part of data preprocessing to eliminate redundant data that requires extra storage spaces and computing power and is crucial for data storage management in cloud computing. Deduplication is essential for file backup systems since duplicated files will presumably consume more storage space, especially with a short backup period such as daily. A common technique in this field involves splitting files into chunks whose hashes can be compared using data structures or techniques like clustering. This paper explores the possibility of performing such file chunk deduplication leveraging an innovative reinforcement learning approach to achieve a high deduplication ratio. The proposed system is named SegDup, which achieves 13% higher deduplication ratio than Extreme Binning, a state-of-the art deduplication algorithm.

As society grows increasingly more online with each passing year, the problem of cyberbullying becomes more and more prominent, with such incidents having the capacity to negatively impact mental health in a major way, especially among children and teenagers. The proposed approach builds on our previous work that established multi-modal detection of cyberbullying on Twitter, and restructures the multi-modal approach by incorporating social media features such as time-related features and social network information. As a result, the new models reach a classification accuracy between 94.4% and 94.6%, from the previous accuracy of 93%. The proposed new approach affirms the use of context-based data in addition to more directly-related features when analyzing cyberbullying and other interactions with promising improvements. We believe that this work contributes significantly to the study of cyberbullying detection, which is an imminent problem with growing importance in the post-COVID society.

The growth in both frequency and severity of gun violence in the United States has necessitated increased research into prevention, despite the lack of funding. Comprising more than 60k gun violence media articles with a total data size of 520 MB, the gun violence database (GVDB) was developed to assist natural language processing researchers in developing and testing prevention methods. Original research based on the GVDB utilized a span-selection model to extract shooter and victim information, but their works might potentially trim out important span candidates. We proposed a new approach to improve identification accuracy and recognize every token in a sentence using a sequence tagging technique. We implemented a BIO sequence tagging model at the token-level using BERT, then further classified each token using LSTM, BiLSTM, and CRF. We found that utilizing BERT as an embedding layer, and decoding word representation as a sequence tagging task, improved shooter/victim identification compared to a span-selection model. We believe that if this improved model is combined with gun violence related keywords, automated techniques could be implemented to identify precursors/risks to gun violence on social media, allowing for intervention by law enforcement or community agencies before escalation to deaths.

Online job search and talent procurement have given rise to challenging match and search problems in the e-recruitment domain. Existing systems perform direct keyword matching of technical skills which misses out a closely matching candidate on account of it not having the exact skills. This results in substandard results which ignores the relationships between technical skills. In an attempt to improve relevancy, this paper proposes a semantic similarity measure between IT skills using a knowledge based approach. The approach builds an ontology using DBpedia and uses it to derive a similarity score using feature based similarity measures. The proposed approach performs better than the Resumatcher system in finding the similarity between skills.

Object detection plays a vital role in many real-world computer vision applications such as self-driving cars, human-less stores and general purpose robotic systems. Convolutional Neural Network(CNN) based Deep Learning has evolved to become the backbone of most computer vision algorithms, including object detection. Most of the research has focused on detecting objects that differ significantly e.g. a car, a person, and a bird. Achieving fine-grained object detection to detect different types within one class of objects can be crucial in tasks like automated retail checkout. This research has developed deep learning models to detect 200 types of similar birds. The models were trained and tested on CUB-200-2011 dataset. To the best of our knowledge, by attaining a mean Average Precision (mAP) of 71.5% we achieved an improvement of 5 percentage points over the previous best mAP of 66.2%.

The online apparel retail market size in the United States is worth about seventy-two billion US dollars. Recommender systems on retail websites generate a lot of this revenue. Thus, improving recommender systems can increase their revenue. Traditional recommendations for clothes consisted of lexical methods. However, visual-based recommendations have gained popularity over the past few years. This involves processing a multitude of images using different image processing techniques. In order to handle such a vast quantity of images, deep neural networks have been used extensively. With the help of fast Graphics Processing Units, these networks provide results which are extremely accurate, within a small amount of time. However, there are still ways in which recommendations for clothes can be improved. We propose an event-based clothing recommender system which uses object detection. We train a model to identify nine events/scenarios that a user might attend: White Wedding, Indian Wedding, Conference, Funeral, Red Carpet, Pool Party, Birthday, Graduation and Workout. We train another model to detect clothes out of fifty-three categories of clothes worn at the event. Object detection gives a mAP of 84.01. Nearest neighbors of the clothes detected are recommended to the user.

Despite clinical trials by pharmaceutical companies as well as current FDA reporting systems, there are still drug side effects that have not been caught. To find a larger sample of reports, a possible way is to mine online social media. With its current widespread use, social media such as Twitter has given rise to massive amounts of data, which can be used as reports for drug side effects. To process these large datasets, Apache Spark has become popular for fast, distributed batch processing. In this work, we have improved on previous pipelines in sentimental analysis-based mining, processing, and extracting tweets with drug-caused side effects. We have also added a new ensemble classifier using a combination of sentiment analysis features to increase the accuracy of identifying drug-caused side effects. In addition, the frequency count for the side effects is also provided. Furthermore, we have also implemented the same pipeline in Apache Spark to improve the speed of processing of tweets by 2.5 times, as well as to support the process of large tweet datasets. As the frequency count of drug side effects opens a wide door for further analysis, we present a preliminary study on this issue, including the side effects of simultaneously using two drugs, and the potential danger of using less-common combination of drugs. We believe the pipeline design and the results present in this work would have great implication on studying drug side effects and on big data analysis in general.

One challenge in text classification is that it is difficult to make feature reductions based on the definition of the features. An ineffective feature reduction may even worsen the classification accuracy. Word2Vec, a word embedding method, has recently been gaining popularity due to its high precision rate of analyzing the semantic similarity between words at relatively low computational cost. However, there is limited research about feature reduction using Word2Vec. In this project, we developed a method using Word2Vec to reduce the feature size while increasing the classification accuracy. We achieved feature reduction by loosely clustering similar features using graph search techniques. Similarity thresholds above 0.5 were used in our method to pair and cluster the features. Finally, we utilized Multinomial Naïve Bayes classifier, Support Vector Machine, K Nearest Neighbor and Random Forest classifier to evaluate the effect of our method. Four datasets with dimensions up to 100,000 feature size and 400,000 document size were used to evaluate the result of our method. The result showed that around 4-10% feature reduction was achieved with up to 1-4% improvement of classification accuracy in terms of different datasets and classifiers. Meanwhile, we also succeeded in improving feature reduction and classification accuracy by combining our method with other classic feature reduction techniques such as chi-square and mutual information.

Extensive clinical trials are required before a drug is placed on the market; yet it is difficult to discover all the side effects for any approved drugs. The United States Food and Drug Administration actively monitors approved medications to identify adverse events. The FDA Adverse Event Reporting System contains a database of adverse drug events (ADE) reported by the healthcare providers and consumers. The pervasive online social networks, such as Twitter, can provide additional information ADE. Concurrently, advancements in social media technology have resulted in the booming of massive public data; the availability of these huge datasets offers numerous research opportunities for extracting ADEs. Towards this purpose, in this paper a simple, effective computation pipeline is proposed, which uses simple drug-related classification and sentiment analysis to extract ADEs on Twitter. The pipeline is described in detail, and is implemented into an automatic process. Experiments are carried out based on 4-months of Twitter data collected. Comparing with an existing pipeline, the new design is able to successfully capture 5 times more valid ADEs, among them 20% are new ADEs. The proposed method may be applied to other areas such as food, beverages, and other daily consumer products for identifying side effects and user opinions.

Sentiment analysis is considered to be a category of machine learning and natural language processing. It is used to extricate, recognize, or portray opinions from different content structures, including news, audits and articles and categorizes them as positive, neutral and negative. It is difficult to predict election results from tweets in different Indian languages. We used Twitter Archiver tool to get tweets in Hindi language. We performed data (text) mining on 42,235 tweets collected over a period of a month that referenced five national political parties in India, during the campaigning period for general state elections in 2016. We made use of both supervised and unsupervised approaches. We utilized Dictionary Based, Naive Bayes and SVM algorithm to build our classifier and classified the test data as positive, negative and neutral. We identified the sentiment of Twitter users towards each of the considered Indian political parties. The results of the analysis for Naive Bayes was the BJP (Bhartiya Janta Party), for SVM it was the BJP (Bhartiya Janta Party) and for the Dictionary Approach it was the Indian Nathional Congress. SVM predicted a 78.4% chance that the BJP would win more elections in the general election due to the positive sentiment they received in tweets. As it turned out, BJP won 60 out of 126 constituencies in the 2016 general election, far more than any other political party as the next party (the Indian National Congress) only won 26 out of 126 constituencies.

As social media has become increasingly popular in the modern world, people are using these platforms to express their opinions about products, businesses, and services. The need for categorizing these consumer reviews has been prominent. One effective solution is sentiment analysis (SA), which has been an active research topic. The goal of SA is to automatically extracting and classifying user opinions. Pervious research works however have not shown satisfied results. In this paper, a multilayer architecture is proposed to increase the performance of multiclass classification. The framework includes data-preprocessing, feature extraction and selection, and classifier building. The framework is a two-layer classification, choosing from Naïve Bayes, Support Vector Machine, Random Forest, and Logistic Regression as base models, and using a voting scheme to obtain the final predicted class. The proposed model is applied to more than 1.3 million restaurant reviews from the Yelp Challenge dataset. We have achieved a high accuracy of 86% for cross validation, and using real-world online review data as test data, we have achieved an accuracy of 80%. The results show that the proposed framework has greatly improved classification accuracy while comparing with those using single-layer architectures. We believe that the proposed method may be applied to, and would have significant contributions to other areas of opinion mining.

Efforts have been made to introduce an extra layer of security on mobile devices, including a good amount of research initiated in the behavioral biometrics domain. However, all prior research approaches for mobile gesture-based authentication has been carried out uni-directionally. Despite of the fact that there are many devices with their own configurations, the study of mobile authentication based on behavioral biometrics has been done only with the Android operating system and devices. In this paper, a novel approach to identifying the owner of a mobile device based on Behavioral Biometrics Mobile Gestures Recognition is presented. This research takes the first step towards implementing behavioral biometrics identification for iOS based iPhone devices. In this research work, it is shown that a user can be identified as the true owner or an imposter of such a device based on the interactive behavior and gestures of the user. In this way continuous identification or authentication of an owner can be done based on the interaction of the user and the device. It is shown that a continuous authentication mechanism can be established using a self-learning model based on machine learning classification approaches such as Random Forests, Gradient Boosting Machine, Deep Learning, and Naive Bayes. The results in this paper show that, with behavioral biometrics, automated user authentication mechanism, EER (Equal Error Rate) can be improved to around 27%, clearly demonstrating that the chances of authenticating the user are good.

Although rigorous clinical studies are required before a drug is placed on the market, it is impossible to predict all side effects for the approved medication. The United States Food and Drug Administration actively monitors approved drugs to identify adverse events. The FDA Adverse Event Reporting System (FAERS) contains a database of adverse drug events reported by the healthcare providers and consumers. The ubiquitous online social networks, such as Twitter, can provide complementary information about adverse drug events. Short Twitter postings, or tweets, are often used to express an opinion about drugs, as well as solicit and receive feedback from consumers of a drug. Thus, adverse drug events can be discovered by extracting from tweets users' opinions about drugs. Here, we developed a computational pipeline for collecting, processing, and analyzing tweets to find signals about adverse drug reactions, defined as drug side effects caused by a drug at a normal dose during normal use. Manual examination of processed tweets identified several known side effects of four drugs.

Blogs have become a major communication media, and have recently been extremely effective in making political and social changes. It is therefore vital to recognize powerful, persuasive bloggers in a web community. In this work we examine existing models for identifying influential bloggers, and construct an improved model based on two new factors: uniqueness and FacebookCount. The former measures the originality of a post, combining with outlink count they represent the novelty of the post. The latter reflects the influence of emerging social network platforms, and can be extended to include twitter share, G +1's, etc. The proposed model also adopts other effectual factors including the number of inlinks, outlinks and comments, the timing of posts and of comments, and the influence of commenters. In addition, to capture the true influence of a post we mine through each comment on a post to identify the sentiment, or the tone of the comment. The experiments show that, comparing with an existing model [3], the proposed approach is able to capture the true influence, and to give bloggers distinctive rankings. We believe that the two new factors are timely and are significant for identifying influential bloggers in the blogosphere.

Search engines give so many results that a user cannot handle them all in a short period of time. Many approaches have been proposed to alleviate this problem. For example, some approaches try to add personalized features and some try to group the results into different categories. The latter one is called a clustering engine, which is the emphasis of this paper. It first reviews several existing approaches such as STC, SHOC, LINGO and SnakeT. It then gives a new approach called HICSEC (Hierarchical Clustering Search Engine with Correlation) to improve the accuracy of clustering by the Correlation calculated with the Singular Value Decomposition.