Tweet sentiment classification seeks to identify the emotional polarity of a tweet. One potential way to enhance classification performance is to include emoticons as features. Emoticons are representations of faces expressing various emotions in text. They are created through combinations of letters, punctuation marks and symbols, and are frequently found within tweets. While emoticons have been used as features for sentiment classification, the importance of their inclusion has not been directly measured. In this work, we seek to determine if the addition of emoticon features improves classifier performance. We also investigate how high dimensionality impacts the addition of emoticon features. We conducted experiments testing the impact of using emoticon features, both with and without feature selection. Classifiers are trained using four different learners and either emoticons, unigrams, or both as features. Feature selection was conducted using five filter based feature rankers with four feature subset sizes. Our results showed that the choice of feature set (emoticon, unigram or both) had no significant impact in our initial tests when using no feature selection; however, with any of the tested feature selection techniques, augmenting unigram features with emoticon features resulted in significantly better performance than unigrams alone. Additionally, we investigate how the addition of emoticons changes the top features selected by the rankers.

Attacks on Recommender Systems (RS) tend to bias predictions and corrupt datasets, which may cause user distrust in the recommendations and dissatisfaction with the RS. Attacks on RSs are mounted by malicious users to "push" or promote an item, "nuke" or disparage an item, or simply to disrupt the recommendations; typically, attacks are motivated by financial gains, by a desire to "game" the system, or both. Although attack research indicates that item-based recommenders are resistant to a wide variety of push and nuke attacks, in previous work we have shown that push attacks on item-based recommenders can be effective using a multiple-target approach. In this paper, we explore nuke attacks on item-based recommenders using a multiple-target approach and variations on the Pearson Correlation calculation. We show that nuke attacks using a multiple-target approach can be configured to be effective against item-based recommenders. To evaluate the effectiveness of these attacks, we use new and existing robustness metrics and an experimental design that includes a variety of attack models, attack sizes, target item types, number of target items, and datasets.

Argument mining systems for student essays need to be able to reliably identify argument components independently of particular essay topics. Thus in addition to features that model argumentation through topic-independent linguistic indicators such as discourse markers, features that can abstract over lexical signals of particular essay topics might also be helpful to improve performance. Prior argument mining studies have focused on persuasive essays and proposed a variety of largely lexicalized features. Our current study examines the utility of such features, proposes new features to abstract over the domain topics of essays, and conducts evaluations using both 10-fold cross validation as well as cross-topic validation. Experimental results show that our proposed features significantly improve argument mining performance in both types of cross-fold evaluation settings. Feature ablation studies further shed light on relative feature utility.

Elderly patients, aged 65 or older, make up 13.5% of the U.S. population, but represent 45.2% of the top 10% of healthcare utilizers, in terms of expenditures. Middle-aged Americans, aged 45 to 64 make up another 37.0% of that category. Given the high demand for healthcare services by the aforementioned population, it is important to identify high-cost users of healthcare systems and, more importantly, ineffective utilization patterns to highlight where targeted interventions could be placed to improve care delivery. In this work, we present a novel multi-level framework applying machine learning (ML) methods (i.e., random forest regression and hierarchical clustering) to group patients with similar utilization profiles into clusters. We use a vector space model to characterize a patient’s utilization profile as the number of visits to different care providers and prescribed medications. We applied the proposed methods using the 2013 Medical Expenditures Panel Survey (MEPS) dataset. We identified clusters of healthcare utilization patterns of elderly and middle-aged adults in the United States, and assessed the general and clinical characteristics associated with these utilization patterns. Our results demonstrate the effectiveness of the proposed framework to model healthcare utilization patterns. Understanding of these patterns can be used to guide healthcare policy-making and practice.

Providing an adequate help to a learner remains a challenge. In this paper we aim to find how to provide learners with real time help in an educational game. Detecting that a player is engaged or motivated is a good sign that he is progressing. For these reasons we need to assess learner’s states while learning. In this study we gather a variety of data using three types of sensors (electroencephalography, eye tracking and automatic facial expression recognition) to build a reliable user adaptation system. The data result from an interaction of 40 players with LewiSpace game, that we built for experimental purpose to learn construction of Lewis diagrams. We used machine learning algorithms in order to identify the most important features gathered from each sensor. Two models were trained with these data: a generalized model, trained on all data available, and a personalized model, trained only on the current user during an early phase of the game experience. The predictive results showed that personalized model could outperform the generalized model.

Active learning, a technique in which a learner self-selects the most important unlabeled examples to be labeled by a human expert, is a useful approach when labeled training data is either scarce or expensive to obtain. While active learning has been well-documented in the offline pool-based setting, less attention has been paid to applying active learning in an online streaming setting. In this paper, we introduce a novel generic framework called ART (Availability-aware active leaRning in data sTreams). We examine the multiple-oracle active learning environment and present a novel method for querying multiple imperfect oracles based on dynamic availability schedules. We introduce a flexible availability-based definition of labeling budget for data streams, and present a mechanism to automatically adapt to implicit changes in oracle availability based on past oracle behavior. Compared to the baseline approaches, our results indicate improvements in accuracy and query utility using our availability-based multiple oracle framework.

Accurate multiagent social simulation requires a computational model of how people incorporate their observations of real-world events into their beliefs about the state of their world. Current methods for creating such agent-based models typically rely on manual input that can be both burdensome and subjective. In this investigation, we instead pursue automated methods that can translate available data into the desired computational models. For this purpose, we use a corpus of real-world events in combination with longitudinal public opinion polls on a variety of opinion issues. We perform two experiments using automated methods taken from the literature. In our first experiment, we train maximum entropy classifiers to model changes in opinion scores as a function of real-world events. We measure and analyze the accuracy of our learned classifiers by comparing the opinion scores they generate against the opinion scores occurring in a held-out subset of our corpus. In our second experiment, we learn Bayesian networks to capture the same function. We then compare the dependency structures induced by the two methods to identify the event features that have the most significant effect on changes in public opinion.

User-generated content in social media is increasingly acknowledged as a rich resource for research into health problems. One particular area of interest is in the semantics individuals’ evoke because they can influence when health-related information is disclosed. While there have been multiple investigations into why self-disclose occurs, much less is known about when individuals choose to disclose information about other people (e.g., a relative), which is a significant privacy concern. In this paper, we introduce a novel framework to investigate how semantics influence disclosure routines for 34 health issues. This framework begins with a supervised classification model to distinguish tweets that communicate personal health issues from confounding concepts (e.g., metaphorical statements that include a health-related keyword). Next, we annotate tweets for each health issue with linguistic and psychological categories (e.g. social processes, affective processes and personal concerns). Then, we apply a non-negative matrix factorization over a health issue-by-language category space. Finally, the factorized basis space is leveraged to group health issues into natural aggregations based around how they are discussed. We evaluate this framework with four months of tweets (over 200 million) and show that certain semantics correspond with whom a health mention pertains to. Our findings show that health issues related with family members, high medical cost and social support (e.g., Alzheimer's Disease, cancer, and Down syndrome) lead to tweets that are more likely to disclose another individual's health status, while tweets with more benign health issues (e.g., allergy, arthritis, and bronchitis) with biological processes (e.g., health and ingestion) and negative emotions are more likely to contain self-disclosures.

One goal of online social recommendation systems is to harness the wisdom of crowds in order to identify high quality content. Yet the sequential voting mechanisms that are commonly used by these systems are at odds with existing theoretical and empirical literature on optimal aggregation. This literature suggests that sequential voting will promote herding---the tendency for individuals to copy the decisions of others around them---and hence lead to suboptimal content recommendation. Is there a problem with our practice, or a problem with our theory? Previous attempts at answering this question have been limited by a lack of objective measurements of content quality. Quality is typically defined endogenously as the popularity of content in absence of social influence. The flaw of this metric is its presupposition that the preferences of the crowd are aligned with underlying quality. Domains in which content quality can be defined exogenously and measured objectively are thus needed in order to better assess the design choices of social recommendation systems. In this work, we look to the domain of education, where content quality can be measured via how well students are able to learn from the material presented to them. Through a behavioral experiment involving a simulated massive open online course (MOOC) run on Amazon Mechanical Turk, we show that sequential voting systems can surface better content than systems that elicit independent votes.

We define On-Average KL-Privacy and present its properties and connections to differential privacy, generalization and information-theoretic quantities including max-information and mutual information. The new definition significantly weakens differential privacy, while preserving its minimalistic design features such as composition over small group and multiple queries as well as closeness to post-processing. Moreover, we show that On-Average KL-Privacy is **equivalent** to generalization for a large class of commonly-used tools in statistics and machine learning that samples from Gibbs distributions---a class of distributions that arises naturally from the maximum entropy principle. In addition, a byproduct of our analysis yields a lower bound for generalization error in terms of mutual information which reveals an interesting interplay with known upper bounds that use the same quantity.