Abstract

The goal of data mining is to discover knowledge and review new, interesting and previously unknown information to the user. One of the important data mining techniques is association rules. Let I= {i1, i2, i3, … in} be a set of n binary attributes called items. Let D= {t1, t2, t3, … tm} be a set of transactions that form a database where each transaction in D has a unique transaction ID and contains a subset of the items in I. A set of items that occur together is an itemsets and a frequent itemset is an itemset whose support is greater than some user-specified minimum support. For event X and Y, an association rule is a rule of type X =>Y, where X, Y I are set of values such that XY=with certain support and confidence. Support determines how often a rule is applicable to a given dataset, while confidence determines how frequently items in Y appear in transactions that contain X. Classical use of association rules is with market-basket data resulting in rules such as “People who buy beer also buy diapers with confidence 0.67 and support 0.4”. Association rules discover patterns and correlations that may be buried deep inside a database. Most of the association rules are based on categorical data- where notions of items and transactions are very clear. For quantitative data with many such attributes, it is not clear to define items and transactions. Therefore, an empirical study is conducted to comprehend how quantitative data can be mapped to items and transactions to get association rules. From this study, we can have an approach towards a range of possibilities for discovering quantitative association rules. The key concept in this work is to give labels to ranges of quantitative data and the use of a combination of labels to decide on items and transactions. We conducted a large number of experiments on three different datasets to present varying results and understanding of quantitative association rules, and issues therein.

Abstract

Poetry is one of the oldest and most widely practised art forms, by far the most popular medium to represent aesthetics in text. However, over the last few decades, a rumour has grown in strength that poetry, perhaps, is dying. Some studies do suggest that readership of poems has declined in the second half of the twentieth century as different forms of prose have emerged to be the preferred forms of expression of our age. However, with the proliferation of social media and Web 2.0, poems are witnessing a revival. Prose emerged on the literary scene parallel to the rise of modernity and the market. Since then, it has grown in both extent and variety, and today is the dominant artform for textual aesthetics. Parallel to this rise of prose, poetry has witnessed a radical transformation both in its form and content, but also in the role it plays in society. The democratization of literature that the rise of prose represents reflects in poetry losing its traditional scaffolding of meter, rhyme and other poetic devices. The emergent forms of poetry allow it to be practised more widely and in a variety of forms. However, this also meant that the poem lost its ontological precision. This thesis is an exploration of the travels of the poem as a form and a genre in the age of mechanical reproduction; an age which democratizes as well as breaks restrictive structures in its drive towards its prime tenets of uniformity and accessibility. The first part of this thesis studies the various definitions of poetry and identifies their inadequacy in setting ontologically precise boundaries. Today, it is nearly impossible for us to identify a poem by any currently accepted definition. Based on this, the thesis proposes a hypothesis that poetry is merging into prose at the level of its basic ontology. The second part tries to prove this hypothesis by computationally showing the reducing distance between prose and poetry. It does this by using novel features and machine learning techniques. Consequently, we try to work towards a new definition of a poem which tries to preserve some elements of form, such that a poem retains ontological coherence. The fourth part concludes that for the revival of the craft of poetry-writing, a better alternative to computational poetry generation, is a personalized poetry assistant created using machine learning techniques.

Abstract

In the course of evolution, human genome has been disrupted, modified and rearranged as a consequence of various genetic events, such as single nucleotide mutations, insertions,deletions, copy number variations, translocations and inversions, making each individual’s genome unique. The accumulated genetic variations are known to play a significant role in shaping population-specific phenotypes and pre-disposition to various lethal diseases including cancer. Detection and annotation of these variants is the first step in understanding their role in disease initiation. Advancements in next generation sequencing (NGS) technologies have made it possible to carry out variant profiling of individuals, slowly replacing traditional and hybridization-based techniques. However, NGS techniques demonstrate inherent challenges due to short read length, incomplete coverage, GC composition, mappability bias, etc. To address some of these issues, we have developed two open-source, integrated platforms, SeqVItA and iCopyDAV, for the detection of small sequence variations (SSVs) and copy number variations (CNVs), respectively, in NGS data. Our tools are tailored to quickly and efficiently identify and functionally annotate genetic variations from large NGS data. In this thesis work we have carried out the detection and functional analysis of variants in Diffuse Large B-cell Lymphoma whole genome sequence data using these tools. Diffuse Large B-cell Lymphoma (DLBCL) is the most aggressive form of haematological malignancies and mostly develops as a result of accumulation of several genetic variations affecting important cellular processes and immuno-oncogenic pathways. The cell-of-origin classification system categorized DLBCL into Germinal Center B-cell (GCB) and Activated B-cell (ABC) subtypes, identified with distinct outcomes when treated with standard immunochemotherapy. Despite several studies carried out to characterize subtype-specific genetic variations, the complete spectrum of these alterations and their relationship with clinicopathological characteristics remains to be elucidated. Detection of subtype-specific genetic variations would assist in understanding the pathogenesis in each case and help in identifying molecular biomarkers for clinical testing, and novel targets for efficient therapy. In this study, we focus on genome-wide analysis of small sequence variants (SNVs and INDELs) and copy number variations (CNVs) in the two subtypes of DLBCL with the objective of identifying molecular markers implicated in the initiation and progression of tumour in these subtypes. The reason for considering these two category of genetic variants is that small sequence variants are widely studied in association studies for disease susceptibility and treatment outcome, while, CNVs affect a large fraction of the genome through focal or armlength amplifications and deletions contributing to the aggressiveness of the cancer. Detection, analysis and annotation of genetic variations is carried out in 12 DLBCL cell lines (7 GCB and 5 ABC subtype). We observe large differences in the variation profiles of the cell lines indicating differences in the genes and associated pathways affected across the cell lines. Detailed analysis revealed a novel set of recurrent genetic variations that may play a key role in lymphomagenesis in a subtype-specific manner. A set of novel genetic variations with potential prognosis are also uncovered. Integration of SSVs and CNVs revealed subgroups of shared variations among the cell lines analysed which help in understanding their shared effect on lymphomagenesis and prognosis. To summarize, in this thesis we present the workflow for detection and functional characterization of genetic variations in cancer genome and provide novel insights into the molecular nature of DLBCL tumorigenesis

Abstract

The rapid increase of textual content on the internet has made efficient text processing very important for various applications ranging from product analysis and implementing business strategies to inculcating public opinion in governance processes or creating biased news to gain public support. In this context, Sentiment Analysis has become one of the important fields of research. Sentiment analysis can be used to know about public opinions expressed over various platforms using NLP and machine learning techniques. A lot of research has been done in the English language in these fields due to the availability of a large number of tools and resources. Though a significant amount of progress is being made in Indian languages in recent years, there is a lot to explore in Telugu language with the machine and deep learning techniques. This thesis aims to enhance the sentiment prediction and the detection of political bias aiming to alter the sentiment across people using various approaches. For this, the need to develop a corpus first was felt essential. For the development of sentiment classifiers in Telugu, we have created corpora “Sentiraama” for different domains like movie reviews, song lyrics, product reviews and book reviews in Telugu language with the text written in Telugu script as there were no such datasets available previously. We describe the process of creating the corpora and assigning polarities to them. Typically, sentiment classifier is trained using data from the same domain it is intended to be tested on. But there may not be sufficient data available in the same domain and additionally using data from multiple sources and domains may help in creating a more generalized sentiment classifier that enhances sentiment prediction. We showed how this generalized sentiment classifier can help in handling low domain-specific data scenarios. To improve sentiment analysis in a low resource language, sentiment labeled corpora are translated from English into the focus language and use them as additional resources for sentiment analysis research in the focus language. But when the text is translated from one language into another, the sentiment is preserved to varying degrees. We use product and book reviews in English as a stand-in for source language text and to determine the loss in sentiment and sentiment predictability, we used manually and automatically determined sentiment labels of the English text as a benchmark. We show that sentiment analysis of Telugu manual translations of English text produces competitive results w.r.t English sentiment analysis. We show that machine translation significantly reduces the human ability to recover sentiment. In the process, we created a Telugu-English parallel corpus that is independently annotated for sentiment using a 5-value scale by Telugu and English speakers. We also created a Telugu lexicon annotated at both sentiment and emphasis level. Another kind of text which is different from the ones we analyzed is ‘song lyrics’. Songs are important to sentiment analysis since the songs and mood are mutually dependent on each other. Song lyrics are a rich source of dataset that is helpful in the analysis and classification of sentiments generated from it. Nowadays we observe a lot of inter-sentential and intra-sentential code-mixing in songs which has a varying impact on the audience. To study this impact we created a Telugu songs dataset which contained both Telugu-English code-mixed and pure Telugu songs. We classified sentiment of the songs based on its arousal rather than routine valence level analysis and enhanced sentiment prediction by introducing code-mixing features. On the other hand, in the domain of news, media houses and journalists try to affect the sentiment or polarity of people towards some political parties by shrewd means such as misinterpreting reality and distorting viewpoints. This can be mainly observed during elections where media houses try to alter people’s sentiment by providing bias news. Detection of such biased news is very useful as unrestricted access to unbiased information is crucial for forming a well-balanced understanding. So along with sentiment analysis, we also aimed at detecting biased news that aims to cause sentiment shifts. As there was no existing dataset available in Telugu language, we created a dataset comprising of 1329 news articles collected from various Telugu newspapers and marked them for bias towards a political party. We also proposed a headline attention model which enhances the bias detection over various methods by a substantial margin.

Abstract

With the current fast pace growing trend towards AI, machine learning, gaming, social media scale of processing, there is great interest in the related algorithms and the constant need to improve on the efficiency of these fundamental algorithms such as matrix multiplications and graph algorithms. The method of sampling the input instance to gain insights and leverage them to build more efficient and quicker solutions is a well explored area and is gaining popularity. It involves creating an intermediate representation of the input which is processed for obtaining insights on the entire input. Applications of this technique can be seen from Fischler et al. [37], who proposed the RANSAC (Random sample consensus) in 1981 for estimating the parameters of a mathematical model from some observed data containing outliers. Sparse matrix computations like multiplying a sparse matrix with another sparse matrix (spmm) and multiplying a sparse matrix with a dense matrix (csrmm) are fundamental to the linear algebra computations and hence is a core part of major scientific problems at scale. Similarly, graph algorithms, such as finding the connectivity of graphs, have been studied for centuries and have been an area of extensive research and have widespread applications ranging from linguistics and computational biology to social sciences and network topologies. Finding biconnected components of a graph is used in planarity testing [47], isomorphism in planar graphs [50], network analytics [26, 41, 80] and the like. The trend towards heterogeneity leads us to customize algorithms that use properties of the input graph for splitting the workload among the computing units in a heterogenous setting. The majority of the algorithms propose a mechanism of splitting the workload into the right pieces and assigning the right pieces to the right computational units for processing based on various thresholds. Computing these thresholds for each of the inputs might not be straightforward always and generally requires an empirical search across the possible range of threshold values, which leads to the time taken to find the threshold overshadowing the benefits gained from the Heterogenous efforts. There has been increasing interest in finding biconnectivity properties of a given graph given its varied applicability and is reflected in the research progress made in terms of speeding up the existing approaches using various techniques with the newly emerging computing platforms. If we have a quick and straightforward method to find the certificate of the graph, we can test it for properties like connectivity and claim that it also holds for the actual graph itself. Hence, finding a certificate for a graph is an interesting problem with immense impact.In this thesis, we propose a simple and effective sampling-based technique for the work partitioning of heterogenous algorithms. We find a nearly optimal threshold value to distribute workload among the computation units for multiplying two unstructured sparse matrices(spmm). We validate our approach on a variety of real-world graphs and show that we can find the best possible threshold value ±10.6% away from the best possible threshold value. We also propose a sequential sampling-based technique for computing the biconnected components of a graph. We validate our approach on real-world graphs and report a speedup of 3.4 times when sampling technique is used

Abstract

Two view stereo problem is a subset of multiview stereo problem where only two views or orientations are available for estimation of depth or disparity. Given the constrained nature of the setup, traditional algorithms assume either the intrinsic or extrinsic parameters to be available in advance in order to build the homographies between the views to rectify the images. Stereo rectification allows epipolar constraint to be enforced such that the corresponding projections of a 3D point could be searched in one dimension along the epipolar lines. When both calibration matrices are not available, the two view stereo problem reduces to estimating the fundamental matrix. A condition number which measures the instability of a function when input conditions change, is high for a fundamental matrix when estimated using an 8 point algorithm. Deep learning methods have been the sought after solutions to numerous computer vision problems as the state of the art research have exposed the power in terms of learning capability of neural networks in general. We explore stereo correspondences in an uncalibrated setting in general by estimating a depthmap given a pair of unrectified stereo images. An end-to-end solution is sought after in a setting where the relative depths of pixels with respect to a single view point could be extracted with the aid of another view. Extending the capabilities of the correlation layer as devised by the flownet architecture, a modified flownet architecure is designed to regress depthmaps with an extension of multiscale correlations for handling textureless surfaces and repetitive textured surfaces. Due to unavailability of dataset for deep learning of unrectified images, a constrained setup of turn table sequences is constructed for this purpose using Google 3D warehouse models.Following the concepts of Attention modelling, we implement an architecture for combining correlations computed at multiple resolutions using a simple element-wise multiplication of the correlations to aid the architecture to resolve correspondences for textureless and repeated textured surfaces. Our experiments show both qualitative and quantitative improvements of depth maps over the original Flownet architecture and provide a solution to the unrectified stereo depth estimation which in literature, most algorithms work on stereo rectified image pairs to compute depthmaps/disparities.

Abstract

Contours are critical inhumanperception of a scene. Theyprovide information about object boundaries, surface planes and surface intersections. This information helps to isolate objects from a scene. In computer vision, contours have similar importance. It has been shown that labelled edges can contribute to segmentation, reconstruction and recognition problems. This thesis has addressed edge labeling of images in indoor and outdoor scenes using depth and RGB data. We classify the contours as occluding, planar (depth discontinuity), and convex, concave (surface normal discontinuity). This task is not straightforward and it is one of the fundamental problems in computer vision.We propose a novel algorithm using random forest for classifying edge pixels into occluding, planar, convex and concave entities.We approach the problem by first focusing on indoor images where we use depth information fromKinect. We release an indoor data set withmore than 500 RGBD images with pixel-wise ground labels. Our method produces promising results and achieves an F-score of 0.84. We also test the approach onmore complex images from from NYU kinect data set and we obtain F-Score of 0.74. While addressing this problem in outdoor images where we use depth from stereo, we realise the need for additional features. Stereo depth of outdoor scenes has artifacts and errors which cannot confidently represent an edge type locally.We show that a simple feature based on semantic classes helps improving the labeling. On Kitti outdoor driving stereo data set, we obtain occluding and planar average F-Score of 0.77 while the approach works poorly to classify curvature edges i.e convex and concave edges. We find this to be because of stereo depth errors and low resolution depth at far distance, which gives poor feature extraction. However,we acknowledge the potential of using semantic classes to improve edge labeling and with large amount of ground truth edgelabels andbetter semantic segmentation, there is ahope of improving the classification.

Abstract

How can we assess the quality of human driving using AI? Driver inattention is one of the leading causes of vehicle crashes and incidents worldwide. Driver inattention includes driver fatigue leading to drowsiness and driver distraction, say due to the use of cellphone or rubbernecking, all of which leads to a lack of situational awareness. Hitherto, techniques presented to monitor driver attention evaluated factors such as fatigue and distraction independently. However, to develop a robust driver attention monitoring system, all the factors affecting a driver’s attention needs to be analyzed holistically. In this thesis, we present two novel approaches for driver attention analysis on the road using driver video and fusion of driver and road video. In the first approach, we propose the driver attention rating system that leverages the front camera of a windshield-mounted smartphone to monitor the driver attention by combining several features. We derive a driver attention rating by fusing spatio-temporal features based on the driver state and behavior such as head pose, eye gaze, eye closure, yawns, use of cellphones, etc. We present a few architectures for feature aggregation like AutoRate and Attention-based AutoRate. We perform an extensive evaluation of feature aggregation networks on real-world driving data and also data from controlled, static vehicle settings with 30 drivers in a large city. We compare the proposed method’s automaticallygenerated rating with the scores given by 5 human annotators. We introduce the kappa coefficient, an evaluation metric to compute the inter-rater agreement between the generated rating and the rating provided by human annotators. We observe that Attention-based AutoRate outperforms other proposed designs for feature aggregation by 10%. Further, we use the learned temporal and spatial attention to visualize the key frame and the key action, which justifies the model’s predicted rating. Finally, to provide driver-specific results, we fine-tune the Attention-based AutoRate model using the specific driver data to give personalized driver experience. In the second approach, we propose driver gaze mapping on the road using the fusion of driver and road videos as input. The proposed approach is used to estimate driver attention and determine which objects the driver is focusing on while driving. To solve such a task, we introduce a new dataset called DGAZE, which is an image dataset that contains the driver view and road view annotated with the driver gaze point on the road. The data is collected in a lab setting, mimicking road conditions using low-cost mobile phone cameras. It has a total of 100,000 images collected with 20 drivers and 103 unique objects on the road belonging to 7 classes, including cars, pedestrians, traffic signals, auto-rickshaw, etc. We also present I-DGAZE, a fused convolutional neural network trained on the DGAZE dataset for predicting driver gaze on the road. Our architecture combines facial features such as facial key-point location and head pose along with the image of the left eye to get optimum results. Our model achieves an error of 94.5 pixels without calibration and 45 pixels with calibration. We compare our model with state-of-the-art eye gaze works and present extensive ablation results. Overall, in this thesis, we propose two methods for driver attention analysis on the road. These approaches provide feedback about the quality of driver attention using driver video and fusion of driver and road video. We introduced dataset for driver attention rating and driver gaze mapping on the road. We also introduced two novel architectures, Attention-based AutoRate and I-DGAZE, corresponding to each proposed task. The evaluation metric and experimental results prove the efficacy of the same. Our two significant contributions include the proposal of a rating system for measuring driver inattention on the road and the dataset consisting of both driver and road view along with driver gaze location on the road.

Abstract

IoT (Internet of Things) is the foundation and infrastructure of any ‘Smart Solution’. It is the network of physical objects or ‘things’ embedded with electronics, software, sensors and connectivity. Things, in the IoT, can refer to a wide variety of devices such as your smartphone, tracking beacon, heart monitoring implants, biochip transponders on farm animals, electric clams in coastal waters, automobiles with built-in sensors, or field operation devices that assist fire-fighters in search and rescue. These devices are termed as smart devices. Internet of Things (IoT) solutions uses smart devices that have the properties of being programmable, sensor-based and connected. Today’s smart applications leverage IoT technologies that enable collaboration between different entities involved in the application. Further, smart applications provide local intelligence attached to each device/physical object. A smart solution constantly gets data and events at different levels of abstraction from underlying IoT devices, sensors, and environmental parameters to make decisions and perform actions. An IoT smart application interconnects various smart objects and interacts with the environment for a specific purpose. Smart solutions require fusion of heterogeneous sensors’ data from a variety of IoT applications and orchestration of data, events, control, and process flows arising from individual IoT applications. Data from IoT devices and sensors are an integral part of smart solutions. The data requirements at the application level are usually different from the data that is captured by the Sensors. IoT driven smart solutions typically use the raw sensor data that is aggregated, filtered and processed to drive the smart solution. Such data has different semantics in contrast to sensor data. There is a need for proper data logistics for the movement of data from the sensors producing data to smart solutions consuming it. Further, new smart devices appear and disappear on-the-fly during the smart solutions life-cycle. Smart solutions must be resilient and adapt to the changes in the environment seamlessly. Due to multiple data flows from Sensors to Smart Solutions, the modelling, and the enactment of IoT processes is a challenging task. In this work, we present IoT data logistics flows as workflows that are orchestrated by the events raised during the execution of a smart solution to create, update and deploy the workflow instances. The data and event handling through workflow events support resilient, smart solutions. As the execution of the smart solutions is mostly dependent on the context (such as location, time and events), there is a dependency between the context for the data generated and the data requirements for the smart solutions. Thus, there is a need for updating the workflows as well as adding new workflows as the execution of the smart solution is progressing for the effective vi vii processing of IoT data. Our approach uses Execution Workflows for the creation, modification, and the execution of IoT workflows on-the-fly. In this work, we also present an implementation framework of our approach and demonstrate the proposed approach for different scenarios in a smart office solution. Workflow management systems drive most of the complex business applications or processes. The workflow management system has a specification component, and an execution component along with other functional components for data flow and control flow, along with orchestration and management of tasks and worklists. Thus, the workflow management system becomes monolithic or composed of tightly coupled modules that enable the specification and execution of workflow instances. Therefore, any modification of execution components of a workflow system requires considerable effort and cost. We address this issue by using a meta execution workflow system that gives us control over specifying various workflow definitions to execute workflow instances as per our needs. The meta execution workflow system allows us to specify user workflows according to the smart application requirements and the workflow engine consists of modules required to execute the tasks of the workflows. The data, control and event flow in the smart solution is handled by the execution workflow system. We present a data logistics system that includes 1) A workflow model for a data logistics system 2) Architecture to support the movement of data from data producers to consumers. Our solution uses the meta execution workflows and user-specified workflows, to support the data fulfilment requirements for the enactment of smart solutions.

Abstract

Semantic Role Labelling (SRL) is a sub-task of Natural Language Understanding. It helps us understand the function of a particular token, with respect to its associated verb in a sentence. SRL identifies the event (verb) of the sentence and the associated participants in the sentence. SRL is used as an intermediate step in various tasks such as information extraction, question answering systems, machine translation, to name a few. In the past decade, social media has become increasingly popular. As a result, there is a huge amount of data available on forums like Facebook and Twitter which can’t be processed or analysed by traditional Natural Language Processing (NLP) tools. This data usually doesn’t follow grammatical rules, is spelt incorrectly and often, for multilingual users, contains a mixture of languages. In this thesis, we discuss our effort towards building a semantic role labeller for such ‘mixed language - hinglish’ data, i.e. Hindi-English code-mixed data. This is the first attempt at creating a SRL tool for Hindi-English code-mixed data, or for that matter, any code-mixed data, to the best of our knowledge. First, we created a dataset which is annotated with semantic roles. We released this dataset of 1460 Hindi-English code-mixed tweets consisting of 20,949 tokens, which is parsed and annotated with semantic roles (Proposition Bank labels). We also created required frame files to take into account mixed language complex predicate formations and certain special constructions which are unique to code-mixed data. We then built a rule-based baseline model for automated labelling of this data. Research shows that there is a strong relation between dependency labels and semantic roles and we leverage this information to build rules for our model. This model can further reduce annotation effort in future. The task is carried out in two steps. The first step is to identify the participants (arguments) of the event (verb) in the sentence. The second step is to classify these identified arguments according to their function, i.e. mark their semantic roles. We then did a thorough error analysis. Further, we worked on improving the accuracy scores obtained by our baseline model. For this we explored statistical methods used in the past for SRL in English and Indian Languages. Before that, we described our effort towards pre-processing and normalising the data to ensure consistency and uniformity of tokens across the corpus. Then, we formulated and extracted various features which work towards capturing semantic information from our data. These are a combination of features traditionally used for SRL, features used in Hindi SRL and features specific to the code-mixed nature of our data. We experimented with and compared the performance of our model by using Paninian dependency labels and Universal Dependency (UD) labels as well. We also trained our model on Hindi monolingual corpus and tested on code-mixed data. We compared and analysed the performance of our models for the various experiments.