This article presents a position highlighting the importance of combining artificial intelligence (AI) approaches with natural intelligence, in other words, involvement of humans. To do so, we specifically focus on problems of societal significance, stemming from complex, sensitive domains. We first discuss our prior work across a series of projects surrounding social media and mental health, and identify major themes wherein augmentation of AI systems and techniques with human feedback has been and can be fruitful and meaningful. We then conclude by noting the implications, in terms of opportunities as well as challenges, that can be drawn from our position, both relating to the specific domain of mental health, and those for AI researchers and practitioners.

For individuals with anxiety disorders, maladaptive feelings and negative beliefs can interfere with daily activities and importantly, social relationships. Literature has examined both direct and indirect influences of an individual's anxiety on their social interactions, however, how they co-vary temporally remains less explored. As individuals appropriate social media platforms more pervasively, can anxiety play an equally significant role in impacting one's \textit{online} social interactions? This paper seeks to answer this question. Employing a dataset of 200 Twitter users, their timeline, and social network data, we examine the relationship between manifested anxiety and various attributes of social interaction of a user by employing Granger causality and time series forecasting approaches. We observe that increases in anxiety levels of an individual result in increased future interaction with weak ties, indicating a tendency to seek support from the broader online community. We discuss how our findings provide novel insights and practical lessons around the impact of an individual's mental health state on their online social interactions.

Student deaths on college campuses, whether brought about by a suicide or an uncontrollable incident, have serious repercussions for the mental wellbeing of students. Consequently, many campus administrators implement post-crisis intervention measures to promote student-centric mental health support. Information about these measures, which we refer to as "counseling recommendations", are often shared via electronic channels, including social media. However, the current ability to assess the effects of these recommendations on post-crisis psychological states is limited. We propose a causal analysis framework to examine the effects of these counseling recommendations after student deaths. We leverage a dataset from 174 Reddit campus communities and ~400M posts of ~350K users. Then we employ statistical modeling and natural language analysis to quantify the psychosocial shifts in behavioral, cognitive, and affective expression of grief in individuals who are "exposed" to (comment on) the counseling recommendations, compared to that in a matched control cohort. Drawing on crisis and psychology research, we find that the exposed individuals show greater grief, psycholinguistic, and social expressiveness, providing evidence of a healing response to crisis and thereby positive psychological effects of the counseling recommendations. We discuss the implications of our work in supporting post-crisis rehabilitation and intervention efforts on college campuses.

Sentiment classification has been a well-investigated research area in the computational linguistics community. However, most of the research is primarily focused on detecting simply the polarity in text, often needing extensive manual labeling of ground truth. Additionally, little attention has been directed towards a finer analysis of human moods and affective states. Motivated by research in psychology, we propose and develop a classifier of several human affective states in social media. Starting with about 200 moods, we utilize mechanical turk studies to derive naturalistic signals from posts shared on Twitter about a variety of affects of individuals. This dataset is then deployed in an affect classification task with promising results. Our findings indicate that different types of affect involve different emotional content and usage styles; hence the performance of the classifier on various affects can differ considerably.

Emotional states of individuals, also known as moods, are central to the expression of thoughts, ideas and opinions, and in turn impact attitudes and behavior. As social media tools are increasingly used by individuals to broadcast their day-to-day happenings, or to report on an external event of interest, understanding the rich ‘landscape’ of moods will help us better interpret and make sense of the behavior of millions of individuals. Motivated by literature in psychology, we study a popular representation of human mood landscape, known as the ‘circumplex model’ that characterizes affective experience through two dimensions: valence and activation. We identify more than 200 moods frequent on Twitter, through mechanical turk studies and psychology literature sources, and report on four aspects of mood expression: the relationship between (1) moods and usage levels, including linguistic diversity of shared content (2) moods and the social ties individuals form, (3) moods and amount of network activity of individuals, and (4) moods and participatory patterns of individuals such as link sharing and conversational engagement. Our results provide at-scale naturalistic assessments and extensions of existing conceptualizations of human mood in social media contexts.

Social media and networking websites, such as Twitter and Facebook, generate large quantities of information and have become mechanisms for real-time content dissipation to users. An important question that arises is: how do we sample such social media information spaces in order to deliver relevant content on a topic to end users? Notice that these large-scale information spaces are inherently diverse, featuring a wide array of attributes such as location, recency, degree of diffusion effects in the network and so on. Naturally, for the end user, different levels of diversity in social media content can significantly impact the information consumption experience: low diversity can provide focused content that may be simpler to understand, while high diversity can increase breadth in the exposure to multiple opinions and perspectives. Hence to address our research question, we turn to diversity as a core concept in our proposed sampling methodology. Here we are motivated by ideas in the "compressive sensing" literature and utilize the notion of sparsity in social media information to represent such large spaces via a small number of basis components. Thereafter we use a greedy iterative clustering technique on this transformed space to construct samples matching a desired level of diversity. Based on Twitter Firehose data, we demonstrate quantitatively that our method is robust, and performs better than other baseline techniques over a variety of trending topics. In a user study, we further show that users find samples generated by our method to be more interesting and subjectively engaging compared to techniques inspired by state-of-the-art systems, with improvements in the range of 15--45%.

Platforms such as Twitter have provided researchers with ample opportunities to analytically study social phenomena. There are however, significant computational challenges due to the enormous rate of production of new information: researchers are therefore, often forced to analyze a judiciously selected “sample” of the data. Like other social media phenomena, information diffusion is a social process–it is affected by user context, and topic, in addition to the graph topology. This paper studies the impact of different attribute and topology based sampling strategies on the discovery of an important social media phenomena–information diffusion. We examine several widely-adopted sampling methods that select nodes based on attribute (random, location, and activity) and topology (forest fire) as well as study the impact of attribute based seed selection on topology based sampling. Then we develop a series of metrics for evaluating the quality of the sample, based on user activity (e.g. volume, number of seeds), topological (e.g. reach, spread) and temporal characteristics (e.g. rate). We additionally correlate the diffusion volume metric with two external variables–search and news trends. Our experiments reveal that for small sample sizes (30%), a sample that incorporates both topology and user context (e.g. location, activity) can improve on naive methods by a significant margin of ~15-20%.