Jeon-Hyung Kang

Asking questions of targeted strangers on social networks

When people have questions, they often turn to their social network for answers. If the answer is obscure or time sensitive however, no members of their social networks may know the answer. For example, it may be difficult to find a friend who has experience with a particular feature or model of digital camera or who knows the current wait time for security at the local airport. In this paper, we explore the feasibility of answering questions by asking strangers. In this approach, strangers with potentially useful information are identified by mining the public status updates posted on Twitter, questions are sent to these strangers, and responses are collected. We explore feasibility in two ways: will users respond to questions sent by strangers and, if they do respond, how long must we wait for a response? Our results from asking 1159 questions across two domains suggest that 42% of users will respond to questions from strangers. 44% of these responses arrived within 30 minutes.

Centrality metric for dynamic networks

Centrality is an important notion in network analysis and is used to measure the degree to which network structure contributes to the importance of a node in a network. While many different centrality measures exist, most of them apply to static networks. Most networks, on the other hand, are dynamic in nature, evolving over time through the addition or deletion of nodes and edges. A popular approach to analyzing such networks represents them by a static network that aggregates all edges observed over some time period. This approach, however, under or overestimates centrality of some nodes. We address this problem by introducing a novel centrality metric for dynamic network analysis. This metric exploits an intuition that in order for one node in a dynamic network to influence another over some period of time, there must exist a path that connects the source and destination nodes through intermediaries at different times. We demonstrate on an example network that the proposed metric leads to a very different ranking than analysis of an equivalent static network. We use dynamic centrality to study a dynamic citations network and contrast results to those reached by static network analysis.

Recommending items to group of users using Matrix Factorization based Collaborative Filtering

Group recommender systems are becoming very popular in the social web owing to their ability to provide a set of recommendations to a group of users. Several group recommender systems have been proposed by extending traditional KNN based Collaborative Filtering. In this paper we explain how to perform group recommendations using Matrix Factorization (MF) based Collaborative Filtering (CF). We propose three original approaches to map the group of users to the latent factor space and compare the proposed methods in three different scenarios: when the group size is small, medium and large. We also compare the precision of the proposed methods with state-of-the-art group recommendation systems using KNN based Collaborative Filtering. We analyze group movie ratings on MovieLens and Netflix datasets. Our study demonstrates that the performance of group recommender systems varies depending on the size of the group, and MF based CF is the best option for group recommender systems.

Analyzing the quality of information solicited from targeted strangers on social media

The emergence of social media creates a unique opportunity for developing a new class of crowd-powered information collection systems. Such systems actively identify potential users based on their public social media posts and solicit them directly for information. While studies have shown that users will respond to solicitations in a few domains, there is little analysis of the quality of information received. Here we explore the quality of information solicited from Twitter users in the domain of product reviews, specifically reviews for a popular tablet computer and L.A.-based food trucks. Our results show that the majority of responses to our questions (>70%) contained relevant information and often provided additional details (>37%) beyond the topic of the question. We compare the solicited Twitter reviews to other user-generated reviews from Amazon and Yelp, and found that the Twitter answers provided similar information when controlling for the questions asked. Our results also reveal limitations of this new information collection method, including its suitability in certain domains and potential technical barriers to its implementation. Our work provides strong evidence for the potential of this new class of information collection systems and design implications for their future use.

Analyzing microblogs with affinity propagation

Recently, there has been a great deal of interest in analyzing inherent structures in posts on microblogs such as Twitter. While many works utilize a well-known topic modeling technique, we instead propose to apply Affinity Propagation [4] (AP) to analyze such a corpus, and we hypothesize that AP may provide different perspective to the traditional approach. Our preliminary analysis raises some interesting facts and issues, which suggest future research directions.

Characterising emergent semantics in twitter lists

Twitter lists organise Twitter users into multiple, often overlapping, sets. We believe that these lists capture some form of emergent semantics, which may be useful to characterise. In this paper we describe an approach for such characterisation, which consists of deriving semantic relations between lists and users by analyzing the co-occurrence of keywords in list names. We use the vector space model and Latent Dirichlet Allocation to obtain similar keywords according to co-occurrence patterns. These results are then compared to similarity measures relying on WordNet and to existing Linked Data sets. Results show that co-occurrence of keywords based on members of the lists produce more synonyms and more correlated results to that of WordNet similarity measures.

VIP: Incorporating Human Cognitive Biases in a Probabilistic Model of Retweeting

Information spread in social media depends on a number of factors, including how the site displays information, how users navigate it to find items of interest, users’ tastes, and the ‘virality’ of information, i.e., its propensity to be adopted, or retweeted, upon exposure. Probabilistic models can learn users’ tastes from the history of their item adoptions and recommend new items to users. However, current models ignore cognitive biases that are known to affect behavior. Specifically, people pay more attention to items at the top of a list than those in lower positions. As a consequence, items near the top of a user’s social media stream have higher visibility, and are more likely to be seen and adopted, than those appearing below. Another bias is due to the item’s fitness: some items have a high propensity to spread upon exposure regardless of the interests of adopting users. We propose a probabilistic model that incorporates human cognitive biases and personal relevance in the generative model of information spread. We use the model to predict how messages containing URLs spread on Twitter. Our work shows that models of user behavior that account for cognitive factors can better describe and predict user behavior in social media.

Analyzing Answers in Threaded Discussions Using a Role-Based Information Network

Online discussion boards are an important medium for collaboration. The goal of our work is to understand how messages and individual discussants contribute to Q&A discussions. We present a novel network model for capturing in-formation roles of messages and discussants, and show how we identify useful answers to the initial question. We first classify information seeking or information providing roles of messages, such as question, answer or acknowledgement. We also identify user intent in the discussion as an information seeker or a provider. We capture such role information within a reply-to discussion network, and identify messages that answer seeker questions and how answeres are acknowledged. Message influences are analyzed using B-centrality measures. User influences across different threads are combined with message influences. We use the combined score in identifying the most useful answer in the thread. The resulting ranks correlate with human provided ranks with an MRR score of 0.67.

Structural and cognitive bottlenecks to information access in social networks

Information in networks is non-uniformly distributed, enabling individuals in certain network positions to get preferential access to information. Social scientists have developed influential theories about the role of network structure in information access. These theories were validated through numerous studies, which examined how individuals leverage their social networks for competitive advantage, such as a new job or higher compensation. It is not clear how these theories generalize to online networks, which differ from real-world social networks in important respects, including asymmetry of social links. We address this problem by analyzing how users of the social news aggregator Digg adopt stories recommended by friends, i.e., users they follow. We measure the impact different factors, such as network position and activity rate; have on access to novel information, which in Diggs case means set of distinct news stories. We show that a user can improve his information access by linking to active users, though this becomes less effective as the number of friends, or their activity, grows due to structural network constraints. These constraints arise because users in structurally diverse position within the follower graph have topically diverse interests from their friends. Moreover, though in most cases users friends are exposed to almost all the information available in the network, after they make their recommendations, the user sees only a small fraction of the available information. Our study suggests that cognitive and structural bottlenecks limit access to novel information in online social networks.

Using graphical models to classify dialogue transition in online Q&A discussions

In this paper, we examine whether it is possible to automatically classify patterns of interactions using a state transition model and identify successful versus unsuccessful student Q&A discussions. For state classification, we apply Conditional Random Field and Hidden Markov Models to capture transitions among the states. The initial results indicate that such models are useful for modeling some of the student dialogue states. We also show the results of classifying threads as successful/unsuccessful using the state information.