Sho Tsugawa

Recognizing Depression from Twitter Activity

In this paper, we extensively evaluate the effectiveness of using a users social media activities for estimating degree of depression. As ground truth data, we use the results of a web-based questionnaire for measuring degree of depression of Twitter users. We extract several features from the activity histories of Twitter users. By leveraging these features, we construct models for estimating the presence of active depression. Through experiments, we show that (1) features obtained from user activities can be used to predict depression of users with an accuracy of 69%, (2) topics of tweets estimated with a topic model are useful features, (3) approximately two months of observation data are necessary for recognizing depression, and longer observation periods do not contribute to improving the accuracy of estimation for current depression; sometimes, longer periods worsen the accuracy.

Community Structure and Interaction Locality in Social Networks

Research on social network analysis (SNA) has been actively pursued. Most SNAs focus on either social relationship networks (e.g., friendship and trust networks) or social interaction networks (e.g., email and phone call networks). It is expected that the social relationship network and social interaction network of a group should be closely related to each other. For instance, people in the same community in a social relationship network are expected to communicate with each other more frequently than with people in different communities. To the best of our knowledge, however, there is not much understanding on such interaction locality in large-scale online social networks. This paper aims to bridge the gap between intuition about interaction locality and empirical evidences observed in large-scale social networks. We investigate the strength of interaction locality in large-scale social networks by analyzing different types of data: logs of mobile phone calls, email messages, and message exchanges in a social networking service. Our results show that strong interaction locality is observed equally in the three datasets, and suggest that strength of the interaction locality is invariant with regard to the scale of the community. Moreover, we discuss practical implications as well as possible applications.

Influence Maximization Problem for Unknown Social Networks

We propose a novel problem called influence maximization for unknown graphs, and propose a heuristic algorithm for the problem. Influence maximization is the problem of detecting a set of influential nodes in a social network, which represents social relationships among individuals. Influence maximization has been actively studied, and several algorithms have been proposed in the literature. The existing algorithms use the entire topological structure of a social network. In practice, however, complete knowledge of the topological structure of a social network is typically difficult to obtain. We therefore tackle an influence maximization problem for unknown graphs. As a solution for this problem, we propose a heuristic algorithm, which we call IMUG (Influence Maximization for Unknown Graphs). Through extensive simulations, we show that the proposed algorithm achieves 60-90% of the influence spread of the algorithms using the entire social network topology, even when only 1-10% of the social network topology is known. These results indicate that we can achieve a reasonable influence spread even when knowledge of the social network topology is severely limited.

Ambient Suite: enhancing communication among multiple participants

We propose a room-shaped information environment called Ambient Suite that enhances communication among multiple participants. In Ambient Suite, the room itself works as both sensors to estimate the conversation states of participants and displays to present information to stimulate conversation. Such nonverbal cues as utterances, positions, and gestures are measured to sense participant states. The participants are surrounded by displays so that various types of information can be given based on their states. Although this system is adaptable to a wide range of situations where groups talk with each other, our implementation assumed standing-party situations as a typical case. Using this implementation, we experimentally evaluated the performance of input, output, and whether our system can actually stimulate conversation. The results showed that our system measured sensor data to recognize the conversational states, presented information, and adequately encouraged participant conversations.

Negative Messages Spread Rapidly and Widely on Social Media

We investigate the relation between the sentiment of a message on social media and its virality, defined as the volume and the speed of message diffusion. We analyze 4.1 million messages (tweets) obtained from Twitter. Although factors affecting message diffusion on social media have been studied previously, we focus on message sentiment, and reveal how the polarity of message sentiment affects its virality. The virality of a message is measured by the number of message repostings (retweets) and the time elapsed from the original posting of a message to its Nth reposting (N-retweet time). Through extensive analyses, we find that negative messages are likely to be reposted more rapidly and frequently than positive and neutral messages. Specifically, the reposting volume of negative messages is 1.2--1.6-fold that of positive and neutral messages, and negative messages spread at 1.25 times the speed of positive and neutral messages when the diffusion volume is large.

Effectiveness of Link Prediction for Face-to-Face Behavioral Networks

Research on link prediction for social networks has been actively pursued. In link prediction for a given social network obtained from time-windowed observation, new link formation in the network is predicted from the topology of the obtained network. In contrast, recent advances in sensing technology have made it possible to obtain face-to-face behavioral networks, which are social networks representing face-to-face interactions among people. However, the effectiveness of link prediction techniques for face-to-face behavioral networks has not yet been explored in depth. To clarify this point, here we investigate the accuracy of conventional link prediction techniques for networks obtained from the history of face-to-face interactions among participants at an academic conference. Our findings were (1) that conventional link prediction techniques predict new link formation with a precision of 0.30–0.45 and a recall of 0.10–0.20, (2) that prolonged observation of social networks often degrades the prediction accuracy, (3) that the proposed decaying weight method leads to higher prediction accuracy than can be achieved by observing all records of communication and simply using them unmodified, and (4) that the prediction accuracy for face-to-face behavioral networks is relatively high compared to that for non-social networks, but not as high as for other types of social networks.

On the robustness of centrality measures against link weight quantization in social networks

In social network analysis, individuals are represented as nodes in a graph, social ties among them are represented as links, and the strength of the social ties can be expressed as link weights. However, in social network analyses where the strength of a social tie is expressed as a link weight, the link weight may be quantized to take only a few discrete values. In this paper, expressing a continuous value of social tie strength as a few discrete value is referred to as link weight quantization, and we study the effects of link weight quantization on centrality measures through simulations and experiments utilizing network generation models that generate synthetic social networks and real social network datasets. Our results show that (1) the effects of link weight quantization on the centrality measures are not significant when determining the most important node in a graph, (2) conversely, a 5–8 quantization level is needed to determine other important nodes, and (3) graphs with a highly skewed degree distribution or with a high correlation between node degree and link weights are robust against link weight quantization.

Robustness of centrality measures against link weight quantization in social network analysis

Research on social network analysis has been actively pursued. In social network analysis, individuals are represented as nodes in a graph and social ties among them are represented as links, and the graph is therefore analyzed to provide an understanding of complex social phenomena that involve interactions among a large number of people. However, graphs used for social network analyses generally contain several errors since it is not easy to accurately and completely identify individuals in a society or social ties among them. For instance, unweighted graphs or graphs with quantized link weights are used for conventional social network analyses since the existence and strengths of social ties are generally known from the results of questionnaires. In this paper, we study, through simulations of graphs used for social network analyses, the effects of link weight quantization on the conventional centrality measures (degree, betweenness, closeness, and eigenvector centralities). Consequently, we show that (1) the effect of link weight quantization on the centrality measures are not significant to infer the most important node in the graph, (2) conversely, 5--8 quantization levels are necessary for determining both the most central node and broad-range node rankings, and (3) graphs with high skewness of their degree distribution and/or with high correlation between node degree and link weights are robust against link weight quantization.

On the Effectiveness of Link Addition for Improving Robustness of Multiplex Networks against Layer Node-Based Attack

Recent research trends in network science are shifting from the analysis of single-layer networks to the analysis of multilayer networks. In particular, the robustness of multilayer networks has been actively studied. There exist two popular multilayer network models: one is interdependent network, and the other is multiplex network. We aim to construct a methodology for effectively improving the robustness of multiplex networks against layer node-based attack. As the first step to achieve this goal, in this paper, we examine the effectiveness of existing link addition strategies, which are proposed for interdependent networks, for improving the robustness of multiplex networks. Through the network attack simulations, we show that the strategic link addition can effectively improve the robustness of multiplex networks. Moreover, link addition strategies are suggested to be effective particularly when a large number of nodes are attacked.

Retweets as a Predictor of Relationships among Users on Social Media

Link prediction is the problem of detecting missing links or predicting future link formation in a network. Application of link prediction to social media, such as Twitter and Facebook, is useful both for developing novel services and for sociological analyses. While most existing research on link prediction uses only the social network topology for the prediction, in social media, records of user activities such as posting, replying, and reposting are available. These records are expected to reflect user interest, and so incorporating them should improve link prediction. However, research into link prediction using the records of user activities is still in its infancy, and the effectiveness of such records for link prediction has not been fully explored. In this study, we focus in particular on records of reposting as a promising source that could be useful for link prediction, and investigate their effectiveness for link prediction on the popular social media platform Twitter. Our results show that (1) the prediction accuracy of techniques using reposting records is higher than that of popular topology-based techniques such as common neighbors and resource allocation for actively retweeting users, (2) the accuracy of link prediction techniques that use network topology alone can be improved by incorporating reposting records.