Pinterest is a Social Network Site (SNS) centered around the curation and sharing of visual content. The site encourages users to form ties with (follow) other users based on mutual interests, and use these ties to discover and share content. In this work, we examine the efficacy and relevance of the Pinterest follow mechanism in driving content discovery and curation. We collect a sample of user activity and find that the vast majority (88%) of the unique users who interact with an average user's content are non-followers. Conversely, only 12.3% of a user's followers interact with any of their pins. Users who discover and repost content from outside their follow network also do not subsequently follow the contributors of that content. Our results strongly suggest that following is neither heavily utilized nor strongly effective for driving content discovery and sharing on Pinterest.

During broadcast events such as the Superbowl, the U.S. Presidential and Primary debates, etc., Twitter has become the de facto platform for crowds to share perspectives and commentaries about them. Given an event and an associated large-scale collection of tweets, there are two fundamental research problems that have been receiving increasing attention in recent years. One is to extract the topics covered by the event and the tweets; the other is to segment the event. So far these problems have been viewed separately and studied in isolation. In this work, we argue that these problems are in fact inter-dependent and should be addressed together. We develop a joint Bayesian model that performs topic modeling and event segmentation in one unified framework. We evaluate the proposed model both quantitatively and qualitatively on two large-scale tweet datasets associated with two events from different domains to show that it improves significantly over baseline models.

This paper describes preliminary findings related to a system for live collaborative tagging of enterprise meetings taking place on an audio bridge between distributed participants. Participants can apply to different points of the interaction as it is ongoing and can see, in near real-time, the flow of as they are being contributed. Two novel types of are proposed: deep tags that apply to a portion of the interaction and tags that apply to an instant of the interaction. Our system is being used by enterprise users and we analyze a corpus of 737 live-tags collected from 16 conversations that took place over several months. We found that the live-tags for audio have slightly different characteristics from Web 2.0 tags: they are longer and confer affordances on the audio like description and summarization. Some observations on the cognitive cost of live-tagging are offered.

This article analyzes communication within a set of individuals to extract the representative prototypical groups and provides a novel framework to establish the utility of such groups. Corporations may want to identify representative groups (which are indicative of the overall communication set) because it is easier to track the prototypical groups rather than the entire set. This can be useful for advertising, identifying “hot” spots of resource consumption as well as in mining representative moods or temperature of a community. Our framework has three parts: extraction, characterization, and utility of prototypical groups. First, we extract groups by developing features representing communication dynamics of the individuals. Second, to characterize the overall communication set, we identify a subset of groups within the community as the prototypical groups. Third, we justify the utility of these prototypical groups by using them as predictors of related external phenomena; specifically, stock market movement of technology companies and political polls of Presidential candidates in the 2008 U.S. elections. We have conducted extensive experiments on two popular blogs, Engadget and Huffington Post. We observe that the prototypical groups can predict stock market movement/political polls satisfactorily with mean error rate of 20.32%. Further, our method outperforms baseline methods based on alternative group extraction and prototypical group identification methods. We evaluate the quality of the extracted groups based on their conductance and coverage measures and develop metrics: predictivity and resilience to evaluate their ability to predict a related external time-series variable (stock market movement/political polls). This implies that communication dynamics of individuals are essential in extracting groups in a community, and the prototypical groups extracted by our method are meaningful in characterizing the overall communication sets.

In this paper we develop a recommendation framework to connect image content with communities in online social media. The problem is important because users are looking for useful feedback on their uploaded content, but finding the right community for feedback is challenging for the end user. Social media are characterized by both content and community. Hence, in our approach, we characterize images through three types of features: visual features, user generated text tags, and social interaction (user communication history in the form of comments). A recommendation framework based on learning a latent space representation of the groups is developed to recommend the most likely groups for a given image. The model was tested on a large corpus of Flickr images comprising 15,689 images. Our method outperforms the baseline method, with a mean precision 0.62 and mean recall 0.69. Importantly, we show that fusing image content, text tags with social interaction features outperforms the case of only using image content or tags.

Rich media social networks promote not only creation and consumption of media, but also communication about the posted media item. What causes a conversation to be interesting, that prompts a user to participate in the discussion on a posted video? We conjecture that people participate in conversations when they find the conversation theme interesting, see comments by people whom they are familiar with, or observe an engaging dialogue between two or more people (absorbing back and forth exchange of comments). Importantly, a conversation that is interesting must be consequential - i.e. it must impact the social network itself.

We have developed a computational framework to characterize social network dynamics in the blogosphere at individual, group and community levels. Such characterization could be used by corporations to help drive targeted advertising and to track the moods and sentiments of consumers. We tested our model on a widely read technology blog called Engadget. Our results show that communities transit between states of high and low entropy, depending on sentiments (positive / negative) about external happenings. We also propose an innovative method to establish the utility of the extracted knowledge, by correlating the mined knowledge with an external time series data (the stock market). Our validation results show that the characterized groups exhibit high stock market movement predictability (89%) and removal of 'impactful' groups makes the community less resilient by lowering predictability (26%) and affecting the composition of the groups in the rest of the community.

Blogs, wikis, tagging, podcasts, and social networking websites such as MySpace, Facebook, Flickr and YouTube have radically changed user interactions on the World Wide Web from a static, one-way, consumption model to a dynamic, multi-way, participation model. Broad user power and flexibility have changed how people engage in and experience their interconnections, interests, and collaborations. Online social interactions will evolve in the next decade to address the growing needs of its user community and make entries into many aspects of our lives. This evolution may very well be among the most exciting ones of our times where the individual and collective power of people to contribute and share content, experiences, ideas, expertise etc. may be even more enhanced than it is today. The enhancement may be shaped through a better understanding of user needs and behavior and it may be enabled through the seamless convergence of multi-modal technologies, new applications, new domains, data mining and better navigational and search capabilities. Some of these changes will also permeate into the workplace and change the way we work. This panel will discuss how online social interactions may evolve in the next decade and what impact it may have on diverse dimensions in our world.

The paper develops a novel computational framework for predicting communication flow in social networks based on several contextual features. The problem is important because prediction of communication flow can impact timely sharing of specific information across a wide array of communities. We determine the intent to communicate and communication delay between users based on several contextual features in a social network corresponding to (a) neighborhood context, (b) topic context and (c) recipient context. The intent to communicate and communication delay are modeled as regression problems which are efficiently estimated using Support Vector Regression. We predict the intent and the delay, on an interval of time using past communication data. We have excellent prediction results on a real-world dataset from MySpace.com with an accuracy of 13-16%. We show that the intent to communicate is more significantly influenced by contextual factors compared to the delay.

DDoS attacks can cause extreme performance degradation at network elements when a large number of malicious flows collaborate to cause congestion, resulting in a denial of service to legitimate users. The flows form a logical aggregate which is typically characterized by similar values in several fields in their packets. The fields and their similar values form a signature for the attack. The focus of this paper is on providing protection to legitimate users against such attacks by detecting significant signatures in network traffic and controlling aggregates of flows carrying these signatures. The paper proposes a system PISA* for deployment at a network element. The detection algorithm in PISA* is based on an improved version of our earlier randomized algorithm. A new control mechanism based on a drop probability function over an attribute named RED Drop Aggregate (RDA) is incorporated in PISA*. RDA is the normalized count of the number of RED (Random Early Detection) drops suffered by the flows carrying a signature. This paper discusses results from an implementation of PISA* that demonstrates that it is effective in detecting, isolating, and controlling offending traffic and providing protection to non-offending traffic.

This paper describes an algorithm to control non-adaptive sources in a simple and lightweight fashion that does not maintain per-flow state information. Using experimental results from simulations, we demonstrate that the algorithm outperforms current well-known algorithms for buffer management in providing better protection for TCP flows under varying degrees of attack from non-adaptive flows.

The paper describes an algorithm to control flows from nonadaptive sources in the Internet in a simple and lightweight fashion. The algorithm, called XCHOKe, is an extension to CHOKe and RED. The paper presents experimental results from simulations that demonstrate that the algorithm outperforms current well-known algorithms for buffer management in providing better protection for TCP flows under varying degrees of attack from non-adaptive flows.

An important aspect of network management is fault management-determining, locating, isolating and correcting faults in the network. The paper deals with the algorithms for detecting faults, i.e., behavior of the network different from specifications. It is important for communication networks to detect faults "in-process" i.e., while the network is in its normal operation. Thus, we detect faults by examining the input-output behavior without forcing the system to specialized inputs explicitly for testing. Such testing is commonly called passive testing. We model the network as a finite state machine and develop procedures for passive testing including the required data structure, efficient implementations and the complexity of our procedures. We start with fully observable and deterministic machines and then study more realistic models: partially observable and nondeterministic machines. We also discuss extensions to communicating finite state machines and machines extended with parameters and variables. We apply our techniques to management of a signaling network operating under the Signaling System 7 (SS7) and report experimental results, which show the feasibility of applying passive testing to practical systems.

This paper describes the first results from research on the compilation of constraint systems into task level parallel programs in a procedural language. This is the only research of which we are aware which attempts to generate efficient parallel programs for numerical computation from constraint systems. Computations are expressed as constraint systems. A dependence graph is derived from the constraint system and a set of input variables. The dependence graph, which exploits the parallelism in the constraints, is mapped to the language CODE, which represents parallel computation structures as generalized dependence graphs. Finally, parallel C programs are generated. To extract parallel programs of appropriate granularity, the following features are included. (i) modularity, (ii) operations over structured types as primitives, (iii) sequential C functions. A prototype of the compiler has been implemented. The domain of matrix computations is targeted for applications. Initial results are very encouraging.

The paper presents a simple and effective method for the concurrent manipulation of linearly ordered data structures on hypercube systems. The method Is based on the existence of an augmented binomial search tree, called the pruned binomial tree, rooted at any arbitrary processor node of the hypercube such that; every edge of the tree corresponds to a direct link between a pair of hypercube nodes; and the tree spans any arbitrary sequence of n consecutive nodes containing the root, using a fanout of at most [log/sub 2/ n] and a depth of at most [log/sub 2/ n]+1. Search trees spanning nonoverlapping processor lists are formed using only local information, and can be used concurrently without contention problems. Thus, they can be used for performing operations such as broadcast and merge simultaneously on sets with nonuniform sizes. Extensions of the tree to k-ary n-cubes and faulty hypercubes are presented. Applications of this concurrent data structure to low- and intermediate-level image processing algorithms, and for dictionary operations involving multiple keys, are also outlined.>