Due to vast resources brought by social media services, social data mining has

received increasing attention in recent years. The availability of sheer amounts of

user-generated data presents data scientists both opportunities and challenges. Opportunities are presented with additional data sources. The abundant link information

in social networks could provide another rich source in deriving implicit information

for social data mining. However, the vast majority of existing studies overwhelmingly

focus on positive links between users while negative links are also prevailing in real-

world social networks such as distrust relations in Epinions and foe links in Slashdot.

Though recent studies show that negative links have some added value over positive

links, it is dicult to directly employ them because of its distinct characteristics from

positive interactions. Another challenge is that label information is rather limited

in social media as the labeling process requires human attention and may be very

expensive. Hence, alternative criteria are needed to guide the learning process for

many tasks such as feature selection and sentiment analysis.

To address above-mentioned issues, I study two novel problems for signed social

networks mining, (1) unsupervised feature selection in signed social networks; and

(2) unsupervised sentiment analysis with signed social networks. To tackle the first problem, I propose a novel unsupervised feature selection framework SignedFS. In

particular, I model positive and negative links simultaneously for user preference

learning, and then embed the user preference learning into feature selection. To study the second problem, I incorporate explicit sentiment signals in textual terms and

implicit sentiment signals from signed social networks into a coherent model Signed-

Senti. Empirical experiments on real-world datasets corroborate the effectiveness of

these two frameworks on the tasks of feature selection and sentiment analysis.

Background: Accurately prioritizing candidate disease genes is an important and challenging problem. Various network-based methods have been developed to predict potential disease genes by utilizing the disease similarity network and molecular networks such as protein interaction or gene co-expression networks. Although successful, a common limitation of the existing methods is that they assume all diseases share the same molecular network and a single generic molecular network is used to predict candidate genes for all diseases. However, different diseases tend to manifest in different tissues, and the molecular networks in different tissues are usually different. An ideal method should be able to incorporate tissue-specific molecular networks for different diseases.

Results: In this paper, we develop a robust and flexible method to integrate tissue-specific molecular networks for disease gene prioritization. Our method allows each disease to have its own tissue-specific network(s). We formulate the problem of candidate gene prioritization as an optimization problem based on network propagation. When there are multiple tissue-specific networks available for a disease, our method can automatically infer the relative importance of each tissue-specific network. Thus it is robust to the noisy and incomplete network data. To solve the optimization problem, we develop fast algorithms which have linear time complexities in the number of nodes in the molecular networks. We also provide rigorous theoretical foundations for our algorithms in terms of their optimality and convergence properties. Extensive experimental results show that our method can significantly improve the accuracy of candidate gene prioritization compared with the state-of-the-art methods.

Conclusions: In our experiments, we compare our methods with 7 popular network-based disease gene prioritization algorithms on diseases from Online Mendelian Inheritance in Man (OMIM) database. The experimental results demonstrate that our methods recover true associations more accurately than other methods in terms of AUC values, and the performance differences are significant (with paired t-test p-values less than 0.05). This validates the importance to integrate tissue-specific molecular networks for studying disease gene prioritization and show the superiority of our network models and ranking algorithms toward this purpose. The source code and datasets are available at http:/ijingchao.github.io/CRstar/.

Cyberbullying is a phenomenon which negatively affects individuals. Victims of the cyberbullying suffer from a range of mental issues, ranging from depression to low self-esteem. Due to the advent of the social media platforms, cyberbullying is becoming more and more prevalent. Traditional mechanisms to fight against cyberbullying include use of standards and guidelines, human moderators, use of blacklists based on profane words, and regular expressions to manually detect cyberbullying. However, these mechanisms fall short in social media and do not scale well. Users in social media use intentional evasive expressions like, obfuscation of abusive words, which necessitates the development of a sophisticated learning framework to automatically detect new cyberbullying behaviors. Cyberbullying detection in social media is a challenging task due to short, noisy and unstructured content and intentional obfuscation of the abusive words or phrases by social media users. Motivated by sociological and psychological findings on bullying behavior and its correlation with emotions, we propose to leverage the sentiment information to accurately detect cyberbullying behavior in social media by proposing an effective optimization framework. Experimental results on two real-world social media datasets show the superiority of the proposed framework. Further studies validate the effectiveness of leveraging sentiment information for cyberbullying detection.

High-level inference tasks in video applications such as recognition, video retrieval, and zero-shot classification have become an active research area in recent years. One fundamental requirement for such applications is to extract high-quality features that maintain high-level information in the videos.

Many video feature extraction algorithms have been purposed, such as STIP, HOG3D, and Dense Trajectories. These algorithms are often referred to as “handcrafted” features as they were deliberately designed based on some reasonable considerations. However, these algorithms may fail when dealing with high-level tasks or complex scene videos. Due to the success of using deep convolution neural networks (CNNs) to extract global representations for static images, researchers have been using similar techniques to tackle video contents. Typical techniques first extract spatial features by processing raw images using deep convolution architectures designed for static image classifications. Then simple average, concatenation or classifier-based fusion/pooling methods are applied to the extracted features. I argue that features extracted in such ways do not acquire enough representative information since videos, unlike images, should be characterized as a temporal sequence of semantically coherent visual contents and thus need to be represented in a manner considering both semantic and spatio-temporal information.

In this thesis, I propose a novel architecture to learn semantic spatio-temporal embedding for videos to support high-level video analysis. The proposed method encodes video spatial and temporal information separately by employing a deep architecture consisting of two channels of convolutional neural networks (capturing appearance and local motion) followed by their corresponding Fully Connected Gated Recurrent Unit (FC-GRU) encoders for capturing longer-term temporal structure of the CNN features. The resultant spatio-temporal representation (a vector) is used to learn a mapping via a Fully Connected Multilayer Perceptron (FC-MLP) to the word2vec semantic embedding space, leading to a semantic interpretation of the video vector that supports high-level analysis. I evaluate the usefulness and effectiveness of this new video representation by conducting experiments on action recognition, zero-shot video classification, and semantic video retrieval (word-to-video) retrieval, using the UCF101 action recognition dataset.

Measuring node centrality is a critical common denominator behind many important graph mining tasks. While the existing literature offers a wealth of different node centrality measures, it remains a daunting task on how to intervene the node centrality in a desired way. In this thesis, we study the problem of minimizing the centrality of one or more target nodes by edge operation. The heart of the proposed method is an accurate and efficient algorithm to estimate the impact of edge deletion on the spectrum of the underlying network, based on the observation that the edge deletion is essentially a local, sparse perturbation to the original network. Extensive experiments are conducted on a diverse set of real networks to demonstrate the effectiveness, efficiency and scalability of our approach. In particular, it is average of 260.95%, in terms of minimizing eigen-centrality, better than the standard matrix-perturbation based algorithm, with lower time complexity.

Stock market news and investing tips are popular topics in Twitter. In this dissertation, first I utilize a 5-year financial news corpus comprising over 50,000 articles collected from the NASDAQ website matching the 30 stock symbols in Dow Jones Index (DJI) to train a directional stock price prediction system based on news content. Next, I proceed to show that information in articles indicated by breaking Tweet volumes leads to a statistically significant boost in the hourly directional prediction accuracies for the DJI stock prices mentioned in these articles. Secondly, I show that using document-level sentiment extraction does not yield a statistically significant boost in the directional predictive accuracies in the presence of other 1-gram keyword features. Thirdly I test the performance of the system on several time-frames and identify the 4 hour time-frame for both the price charts and for Tweet breakout detection as the best time-frame combination. Finally, I develop a set of price momentum based trade exit rules to cut losing trades early and to allow the winning trades run longer. I show that the Tweet volume breakout based trading system with the price momentum based exit rules not only improves the winning accuracy and the return on investment, but it also lowers the maximum drawdown and achieves the highest overall return over maximum drawdown.

Many neurological disorders, especially those that result in dementia, impact speech and language production. A number of studies have shown that there exist subtle changes in linguistic complexity in these individuals that precede disease onset. However, these studies are conducted on controlled speech samples from a specific task. This thesis explores the possibility of using natural language processing in order to detect declining linguistic complexity from more natural discourse. We use existing data from public figures suspected (or at risk) of suffering from cognitive-linguistic decline, downloaded from the Internet, to detect changes in linguistic complexity. In particular, we focus on two case studies. The first case study analyzes President Ronald Reagan’s transcribed spontaneous speech samples during his presidency. President Reagan was diagnosed with Alzheimer’s disease in 1994, however my results showed declining linguistic complexity during the span of the 8 years he was in office. President George Herbert Walker Bush, who has no known diagnosis of Alzheimer’s disease, shows no decline in the same measures. In the second case study, we analyze transcribed spontaneous speech samples from the news conferences of 10 current NFL players and 18 non-player personnel since 2007. The non-player personnel have never played professional football. Longitudinal analysis of linguistic complexity showed contrasting patterns in the two groups. The majority (6 of 10) of current players showed decline in at least one measure of linguistic complexity over time. In contrast, the majority (11 out of 18) of non-player personnel showed an increase in at least one linguistic complexity measure.

The amount of time series data generated is increasing due to the integration of sensor technologies with everyday applications, such as gesture recognition, energy optimization, health care, video surveillance. The use of multiple sensors simultaneously

for capturing different aspects of the real world attributes has also led to an increase in dimensionality from uni-variate to multi-variate time series. This has facilitated richer data representation but also has necessitated algorithms determining similarity between two multi-variate time series for search and analysis.

Various algorithms have been extended from uni-variate to multi-variate case, such as multi-variate versions of Euclidean distance, edit distance, dynamic time warping. However, it has not been studied how these algorithms account for asynchronous in time series. Human gestures, for example, exhibit asynchrony in their patterns as different subjects perform the same gesture with varying movements in their patterns at different speeds. In this thesis, we propose several algorithms (some of which also leverage metadata describing the relationships among the variates). In particular, we present several techniques that leverage the contextual relationships among the variates when measuring multi-variate time series similarities. Based on the way correlation is leveraged, various weighing mechanisms have been proposed that determine the importance of a dimension for discriminating between the time series as giving the same weight to each dimension can led to misclassification. We next study the robustness of the considered techniques against different temporal asynchronies, including shifts and stretching.

Exhaustive experiments were carried on datasets with multiple types and amounts of temporal asynchronies. It has been observed that accuracy of algorithms that rely on data to discover variate relationships can be low under the presence of temporal asynchrony, whereas in case of algorithms that rely on external metadata, robustness against asynchronous distortions tends to be stronger. Specifically, algorithms using external metadata have better classification accuracy and cluster separation than existing state-of-the-art work, such as EROS, PCA, and naive dynamic time warping.

This thesis studies recommendation systems and considers joint sampling and learning. Sampling in recommendation systems is to obtain users' ratings on specific items chosen by the recommendation platform, and learning is to infer the unknown ratings of users to items given the existing data. In this thesis, the problem is formulated as an adaptive matrix completion problem in which sampling is to reveal the unknown entries of a $U\times M$ matrix where $U$ is the number of users, $M$ is the number of items, and each entry of the $U\times M$ matrix represents the rating of a user to an item. In the literature, this matrix completion problem has been studied under a static setting, i.e., recovering the matrix based on a set of partial ratings. This thesis considers both sampling and learning, and proposes an adaptive algorithm. The algorithm adapts its sampling and learning based on the existing data. The idea is to sample items that reveal more information based on the previous sampling results and then learn based on clustering. Performance of the proposed algorithm has been evaluated using simulations.

Micro-blogging platforms like Twitter have become some of the most popular sites for people to share and express their views and opinions about public events like debates, sports events or other news articles. These social updates by people complement the written news articles or transcripts of events in giving the popular public opinion about these events. So it would be useful to annotate the transcript with tweets. The technical challenge is to align the tweets with the correct segment of the transcript. ET-LDA by Hu et al [9] addresses this issue by modeling the whole process with an LDA-based graphical model. The system segments the transcript into coherent and meaningful parts and also determines if a tweet is a general tweet about the event or it refers to a particular segment of the transcript. One characteristic of the Hu et al’s model is that it expects all the data to be available upfront and uses batch inference procedure. But in many cases we find that data is not available beforehand, and it is often streaming. In such cases it is infeasible to repeatedly run the batch inference algorithm. My thesis presents an online inference algorithm for the ET-LDA model, with a continuous stream of tweet data and compare their runtime and performance to existing algorithms.