Search

The purpose of this thesis is to investigate and analyze the present status of solution methods for workforce scheduling problems. Furthermore, the thesis addresses two workforce scheduling problems that appear in organizations operating in Finland and proposes approaches to solving those problems. In the first part, the most interesting results of most up-to-date articles regarding workforce scheduling are presented. These results show that workforce scheduling is extensively studied, but at the moment, no general purpose solution method have been developed. The thesis then introduces and describes mathematical formulations for the two problems in consideration. The thesis provides computational experiments which show that these initial formulations are not sufficiently efficient and not solvable in a reasonable amount of time with a state-of-the-art integer programming solver. In the second part, the thesis concentrates on producing tight lower -and upper-bounds on the optimal cost. Lower bounds are obtained with a column generation approach and the upper bounds with a hybrid heuristic. The computational experiments reveal that the methods developed in this thesis outperform the initial formulation in most cases. This thesis was done for Zenopt Ltd as a part of research and development project., Tässä diplomityössä tarkastellaan ja analysoidaan eri työvuorosuunnitteluongelmien ratkaisumenetelmien nykytilannetta. Tämän ohella, työssä tuodaan esille kaksi Suomessa ilmenevää työvuorosuunnitteluongelmaa sekä tapoja, joilla kyseiset ongelmat voidaan ratkaista. Työn ensimmäisessä osassa käydään läpi ajan tasalla olevien sekä työvuoronsuunnittelua tarkastelevien artikkelien kiinnostavimmat tulokset. Nämä tulokset osoittavat, että työvuorosuunnittelu on kattavasti tutkittu aihealue, mutta tällä hetkellä täysin yleiskäyttöistä ratkaisumenetelmää ei ole kehitetty. Seuraavaksi työssä esitellään ja kuvataan ongelmille matemaattinen formulaatio. Laskennalliset testit osoittavat, että kyseiset formulaatiot eivät ole riittävän tehokkaita ja huippuluokan optimointiohjelmistolla ongelmien ratkaisuaika ylittää kohtuulliset rajat. Seuraavaksi työssä keskitytään löytämään ongelmille kapeat ylä- ja alarajat. Alarajat löydetään sarakkeita generoivan algoritmin avulla ja ylärajat hybridi heuristiikalla. Työn tämän osan laskennalliset tulokset vahvistavat, että uudet menetelmät tuottavat parempia tuloksia suurimmalle osalle instansseista alkuperäiseen formulaatioon verrattuna. Tämä työ on tehty Zenopt Oy:lle osana tutkimus- ja tuotekehitysprojektia.

Use of interception proxies is becoming more popular. They are used to audit access and enforce policies and constraints to important servers or whole network segments. The sheer amount of data captured with the devices makes fully manual pruning of the data impractical. Methods to analyze the gathered data to highlight possible attacks or problems would be valuable in freeing up administrator time and resources. This thesis investigates the use of clustering methods to identify anomalous connections, either by identifying them as outliers or bundling them with other connections which have raised alarm in the past. The work shows that a practical approach can be implemented with a DBSCAN-based clustering method, but concluded that an unsupervised approach is not enough. As a semisupervised method the system can have value in production environments., Sieppausvälityspalvelimien käyttö on yleistymässä. Niitä käytetään käytäntöjen ja rajoitusten täytäntöönpanossa sekä kriittisten palvelimien ja verkon osien käytön valvomisessa. Laitteiden kaappaaman tiedon määrä on niin valtava, että tiedon purkaminen manuaalisesti on epäkäytännöllistä. Menetelmät jotka analysoivat dataa mahdollisten hyökkäysten tai ongelmien esiin nostamiseksi olisivat hyvin arvokkaita vapauttamaan järjestelmänvalvojien aikaa ja resursseja. Tässä työssä tutkitaan ryhmittelyalgoritmien käyttökelpoisuutta epätavallisten yhteyksien havainnoimisessa joko tunnistamalla ne poikkeaviksi, koska ne eivät kuulu mihinkään ryhmään tai asettamalla ne samaan ryhmään sellaisen yhteyden kanssa joka on todettu hälyttäväksi aiemmin. Työssä todetaan, että käytännöllinen sovellus järjestelmästä voidaan toteuttaa käyttäen DBSCAN-pohjaista ryhmittelyalgoritmia, mutta täysin valvomattomalla lähestymistavalla ei saada riittävän hyvää tulosta. Osittain valvottuna menetelmästä voi olla hyötyä tuotantojärjestelmien valvonnassa.

Social media has brought a revolution on how people are consuming news. Beyond the undoubtedly large number of advantages brought by social-media platforms, a point of criticism has been the creation of echo chambers and filter bubbles, caused by social homophily and algorithmic personalization. In this paper we address the problem of balancing the information exposure in a social network. We assume that two opposing campaigns (or viewpoints) are present in the network, and that network nodes have different preferences towards these campaigns. Our goal is to find two sets of nodes to employ in the respective campaigns, so that the overall information exposure for the two campaigns is balanced. We formally define the problem, characterize its hardness, develop approximation algorithms, and present experimental evaluation results. Our model is inspired by the literature on influence maximization, but we offer significant novelties. First, balance of information exposure is modeled by a symmetric difference function, which is neither monotone nor submodular, and thus, not amenable to existing approaches. Second, while previous papers consider a setting with selfish agents and provide bounds on best response strategies (i.e., move of the last player), we consider a setting with a centralized agent and provide bounds for a global objective function., Comment: Published at the Thirty-First Annual Conference on Neural Information Processing Systems (NIPS 2017). References updated

We study the evolution of long-lived controversial debates as manifested on Twitter from 2011 to 2016. Specifically, we explore how the structure of interactions and content of discussion varies with the level of collective attention, as evidenced by the number of users discussing a topic. Spikes in the volume of users typically correspond to external events that increase the public attention on the topic -- as, for instance, discussions about `gun control' often erupt after a mass shooting. This work is the first to study the dynamic evolution of polarized online debates at such scale. By employing a wide array of network and content analysis measures, we find consistent evidence that increased collective attention is associated with increased network polarization and network concentration within each side of the debate; and overall more uniform lexicon usage across all users., Comment: accepted at ACM WebScience 2017

Polarization is a troubling phenomenon that can lead to societal divisions and hurt the democratic process. It is therefore important to develop methods to reduce it. We propose an algorithmic solution to the problem of reducing polarization. The core idea is to expose users to content that challenges their point of view, with the hope broadening their perspective, and thus reduce their polarity. Our method takes into account several aspects of the problem, such as the estimated polarity of the user, the probability of accepting the recommendation, the polarity of the content, and popularity of the content being recommended. We evaluate our recommendations via a large-scale user study on Twitter users that were actively involved in the discussion of the US elections results. Results shows that, in most cases, the factors taken into account in the recommendation affect the users as expected, and thus capture the essential features of the problem., Comment: accepted as a short paper at ACM WebScience 2017. arXiv admin note: substantial text overlap with arXiv:1703.10934

We explore how the polarization around controversial topics evolves on Twitter - over a long period of time (2011 to 2016), and also as a response to major external events that lead to increased related activity. We find that increased activity is typically associated with increased polarization; however, we find no consistent long-term trend in polarization over time among the topics we study., Comment: Accepted as a short paper at ICWSM 2017. Please cite the ICWSM version and not the ArXiv version

Polarized topics often spark discussion and debate on social media. Recent studies have shown that polarized debates have a specific clustered structure in the endorsement net- work, which indicates that users direct their endorsements mostly to ideas they already agree with. Understanding these polarized discussions and exposing social media users to content that broadens their views is of paramount importance. The contribution of this demonstration is two-fold. (i) A tool to visualize retweet networks about controversial issues on Twitter. By using our visualization, users can understand how polarized discussions are shaped on Twitter, and explore the positions of the various actors. (ii) A solution to reduce polarization of such discussions. We do so by exposing users to information which presents a contrarian point of view. Users can visually inspect our recommendations and understand why and how these would play out in terms of the retweet network. Our demo (https://users.ics.aalto.fi/kiran/reducingControversy/ homepage) provides one of the first steps in developing automated tools that help users explore, and possibly escape, their echo chambers. The ideas in the demo can also help content providers design tools to broaden their reach to people with different political and ideological backgrounds., Comment: Accepted as a demo at WWW 2017

Motivated by applications that arise in online social media and collaboration networks, there has been a lot of work on community-search and team-formation problems. In the former class of problems, the goal is to find a subgraph that satisfies a certain connectivity requirement and contains a given collection of seed nodes. In the latter class of problems, on the other hand, the goal is to find individuals who collectively have the skills required for a task and form a connected subgraph with certain properties. In this paper, we extend both the community-search and the team-formation problems by associating each individual with a profile. The profile is a numeric score that quantifies the position of an individual with respect to a topic. We adopt a model where each individual starts with a latent profile and arrives to a conformed profile through a dynamic conformation process, which takes into account the individual's social interaction and the tendency to conform with one's social environment. In this framework, social tension arises from the differences between the conformed profiles of neighboring individuals as well as from differences between individuals' conformed and latent profiles. Given a network of individuals, their latent profiles and this conformation process, we extend the community-search and the team-formation problems by requiring the output subgraphs to have low social tension. From the technical point of view, we study the complexity of these problems and propose algorithms for solving them effectively. Our experimental evaluation in a number of social networks reveals the efficacy and efficiency of our methods., Comment: A short version of this paper appeared in the 2017 SIAM International Conference on Data Mining, SDM'17. In this extended version, we discuss the team-formation problem variant, beside the original community-search problem, and include additional experimental results

People are shifting from traditional news sources to online news at an incredibly fast rate. However, the technology behind online news consumption promotes content that confirms the users' existing point of view. This phenomenon has led to polarization of opinions and intolerance towards opposing views. Thus, a key problem is to model information filter bubbles on social media and design methods to eliminate them. In this paper, we use a machine-learning approach to learn a liberal-conservative ideology space on Twitter, and show how we can use the learned latent space to tackle the filter bubble problem. We model the problem of learning the liberal-conservative ideology space of social media users and media sources as a constrained non-negative matrix-factorization problem. Our model incorporates the social-network structure and content-consumption information in a joint factorization problem with shared latent factors. We validate our model and solution on a real-world Twitter dataset consisting of controversial topics, and show that we are able to separate users by ideology with over 90% purity. When applied to media sources, our approach estimates ideology scores that are highly correlated (Pearson correlation 0.9) with ground-truth ideology scores. Finally, we demonstrate the utility of our model in real-world scenarios, by illustrating how the learned ideology latent space can be used to develop exploratory and interactive interfaces that can help users in diffusing their information filter bubble., Comment: Accepted at WSDM 2018. Please cite the WSDM version, not the arxiv version

Many large projects in organizations need to be broken down into manageable tasks for their completion, yet such tasks still demand a diverse set of skills to be completed. Team formation is a way of acquiring workers with different set of skills to complete the tasks and minimizing the cost of assigned workers for the benefit of the organization. It is also found that workers work effectively when they are happy with the task they are working on. In this project, we are given a set of tasks that need to be completed and a set of workers who can be assigned to the tasks. Each task will require a set of skills to be completed. Each worker will possess a set of skills. Each worker has some cost of working on a task. This cost can also be seen as a dissatisfaction factor. And this cost might be different for different tasks. We want to find an assignment of workers to tasks, while making sure that all the skills required to complete a task are covered. In addition to that, we also have to minimize the cost (dissatisfaction) of the workers assigned to the tasks. In real world applications, not all tasks can be completely covered by the available workers. This is because, not all the required skills to complete the tasks are possessed by the workers. In this project, we propose two approaches to overcome these downfalls. One approach is to outsource the entire task for a cheaper cost. And the other approach is to train the workers. We provide algorithmic solutions for both the approaches. We first prove that this team formation problem is NP-Complete. And then we propose and analyze different algorithms for both the approaches. These algorithms are inspired from solutions to matching and set cover problems. We used the data from stackexchange Q \& A discussion forum and bibsonomy social bookmarking and publication-sharing website to model workers and tasks for our experiments. From the results we found that the difference in the performance of the algorithms was very lit

Understanding customers is important. In the case of content providers, this means knowing which groups of people are viewing different types of content, and even more importantly, which groups of people are not yet your active users. However, information about website's visitors is not typically available, as people might visit websites without making an account. Therefore, there is a need for finding out information about the users in other ways. This thesis studied how accurately demographics (age and gender) and motives can be predicted for website visitors. The predictions were made using the analytics data from a single website. This data included what content the users had viewed, at which times of the day they were active, and what browser and operating system they were using. In addition, this thesis studied how much data is needed to make the predictions, and which features in the analytics data are useful for these predictions. It turned out to be possible to predict the demographics with good accuracy. Gender was the easiest to predict with 77.8 % accuracy. Age prediction with four age groups had the weighted F1 score of 0.513. Predicting motives was more difficult, and only few of the motives were predicted more accurately than when choosing the most common motives for everyone. The prediction accuracy depended strongly on how much content the users had viewed. Also, the other features in the analytics data were very weak predictors for the demographics, and therefore they did not increase the prediction accuracy significantly after the user had viewed some content., Asiakkaiden ymmärtäminen on tärkeää. Sisällöntuottajien kohdalla tämä tarkoittaa ymmärrystä siitä, mitkä ihmisryhmät katsovat erilaisia sisältöjä, ja vieläkin tärkeämmin tietoa siitä, mitä ryhmiä ei ole vielä tavoitettu. Verkkosivun kävijöistä ei kuitenkaan ole usein saatavilla tarkempaa tietoa, koska ihmiset vierailevat sivustoilla ilman rekisteröitymistä. Tämän vuoksi tieto käyttäjistä täytyy selvittää muilla tavoin. Tämä diplomityö tutki kuinka tarkasti verkkosivun vierailijoiden demografiat (ikä ja sukupuoli) sekä motiivit voidaan ennustaa. Ennusteet tehtiin yksittäisen verkkosivuston analytiikkadataan pohjautuen. Kyseinen analytiikkadata sisälsi tiedon käyttäjien katsomasta sisällöstä, mihin aikoihin käyttäjät olivat aktiivisia ja mitä selainta sekä käyttöjärjestelmää he käyttivät. Lisäksi tämä diplomityö tutki kuinka paljon käyttödataa ennusteiden tekeminen vaatii sekä mitkä piirteet ovat hyödyllisiä ennustamisessa. Demografioiden ennustaminen osoittautui mahdolliseksi hyvällä tarkkuudella. Sukupuolen ennustaminen oli helpointa, ja sen tarkkuus oli 77,8 %. Iän ennustaminen neljää ikäryhmää käyttäen sai painotetun F1-arvon 0,513. Motiivien ennustaminen oli hankalampaa, ja vain osa motiiveista pystyttiin ennustamaan tarkemmin kuin mitä suosituimman vastauksen valitseminen kaikille olisi antanut. Ennustuksen tarkkuus riippui vahvasti käyttäjän katsoman sisällön määrästä. Lisäksi analytiikkadatassa vain käyttäjien katsoma sisältö oli merkittävä tekijä, eikä muiden piirteiden lisääminen parantanut tarkkuutta merkittävästi, kun käyttäjä oli katsonut jotakin sisältöä.

This thesis applies the explorative double diamond design process borrowed to iteratively frame a research problem applicable in the context of a recruitment web service and then find the best approach to solve it. Thereby the problem focus is laid on multi-class classification, in particular the task of labelling sentences in job advertisements with one of six topics which were found to be covered in every typical job description. A dataset is obtained for evaluation and conventional N-Gram Vector Space models are compared with Representation Learning approaches, notably continuous distributed representations, and Sequential Modeling techniques using Recurrent Neural Networks. Results of the experiments show that the Representation Learning and Sequential Modeling approaches perform on par or better than traditional feature engineering methods and show a promising direction in and beyond research in Computational Linguistics and Natural Language Processing.

Researchers of semi-supervised learning methods have been developing the family of Learning Vector Quantization models which originated from the well-known Self-Organizing Map algorithm. The models of this type can be characterized as prototype-based, self-explanatory and flexible. The thesis contributes to the development of one of the LVQ models – Semi-Supervised Relational Prototype Classifier for dissimilarity data. The model implementation is developed based on the related research work and thesis author findings, and applied to the task of anomaly detection from a real-time air condition data. We propose a regularization algorithm for gradient descent in order to achieve better convergence and a new strategy for initializing prototypes. We develop an innovative framework involving a human expert as a source of labeled data. The framework detects anomalies of environment parameters in both real-time and long-run observations and updates the model according to findings. The data set used for experiments is collected in real-time from sensors installed inside the Aalto Mechanical Engineering building located at Otakaari, 4, Espoo. Installation was done as a part of the project of VTT and Korean National Research Institute. The data consists of 3 main parameters – air temperature, humidity and CO2 concentration. Total number of deployed sensors is around 150. One month recorded data observations contains approximately 1.5M of data points. The results of the project demonstrate the efficiency of the developed regularized LVQ method for classification in given settings. Its regularized version generally overperforms its parent and various baseline methods on air conditioning, synthetic and UCI data. Together with the proposed classification framework, the system has shown its robustness and efficiency and is ready for deployment to a production environment.

Currently, people use different online services (i.e., websites and mobile applications) for different language learning purposes. On the hand, online businesses use usually use different approaches and techniques to process a massive amount of text in different languages, in order to analyze users’ feedback and opinions regarding different entities such as products, movies, or pictures. Although the available online language learning platforms and opinion mining techniques are useful in different cases, they still suffer from some limitations related to slangs in different languages. Slangs (plural of slang) are any informal words or idioms used by people in different languages and countries. These slangs are continuously developed, especially in dense social networks and communities. In addition, people also use slangs to communicate on online services (e.g., Facebook a ) as well as to write reviews on different websites. Accordingly, learning about specific language’s slangs would be useful for many people. The current problems related to slangs are twofold: with people, and with online businesses. For people, there is no any comprehensive online dictionary to learn about worldwide slangs. On the other hand, online businesses have some challenges to understand the semantics (e.g., sentiments and contexts) behind the texts being exchanged on their platforms. Both people and online businesses face the problem of inaccurate translation of slang using current available online translation services. Therefore, in this thesis we propose, design, and develop a novel solution with main objective and vision to build the largest online knowledge base of slangs. In order to achieve this objective, we used different approaches and methodologies both theoretical and empirical. We conducted a literature review in the sentiment analysis domain. In addition, we made different personal interviews as well as an online survey to validate our assumptions and facts about slang develop

Batch processing technologies (Such as MapReduce, Hive, Pig) have matured and been widely used in the industry. These systems solved the issue processing big volumes of data successfully. However, first big amount of data need to be collected and stored in a database or file system. That is very time-consuming. Then it takes time to finish batch processing analysis jobs before get any results. While there are many cases that need analysed results from unbounded sequence of data in seconds or sub-seconds. To satisfy the increasing demand of processing such streaming data, several streaming processing systems are implemented and widely adopted, such as Apache Storm, Apache Spark, IBM InfoSphere Streams, and Apache Flink. They all support online stream processing, high scalability, and tasks monitoring. While how to evaluate stream processing systems before choosing one in production development is an open question. In this thesis, we introduce StreamBench, a benchmark framework to facilitate performance comparisons of stream processing systems. A common API component and a core set of workloads are defined. We implement the common API and run benchmarks for three widely used open source stream processing systems: Apache Storm, Flink, and Spark Streaming. A key feature of the StreamBench framework is that it is extensible -- it supports easy definition of new workloads, in addition to making it easy to benchmark new stream processing systems.

Society is often polarized by controversial issues, that split the population into groups of opposing views. When such issues emerge on social media, we often observe the creation of 'echo chambers', i.e., situations where like-minded people reinforce each other's opinion, but do not get exposed to the views of the opposing side. In this paper we study algorithmic techniques for bridging these chambers, and thus, reducing controversy. Specifically, we represent the discussion on a controversial issue with an endorsement graph, and cast our problem as an edge-recommendation problem on this graph. The goal of the recommendation is to reduce the controversy score of the graph, which is measured by a recently-developed metric based on random walks. At the same time, we take into account the acceptance probability of the recommended edge, which represents how likely the edge is to materialize in the endorsement graph. We propose a simple model based on a recently-developed user-level controversy score, that is competitive with state-of- the-art link-prediction algorithms. We thus aim at finding the edges that produce the largest reduction in the controversy score, in expectation. To solve this problem, we propose an efficient algorithm, which considers only a fraction of all the combinations of possible edges. Experimental results show that our algorithm is more efficient than a simple greedy heuristic, while producing comparable score reduction. Finally, a comparison with other state-of-the-art edge-addition algorithms shows that this problem is fundamentally different from what has been studied in the literature., Comment: Proceedings of the Tenth ACM International Conference on Web Search and Data Mining (WSDM 2017)

Given a large graph, the densest-subgraph problem asks to find a subgraph with maximum average degree. When considering the top-$k$ version of this problem, a na\"ive solution is to iteratively find the densest subgraph and remove it in each iteration. However, such a solution is impractical due to high processing cost. The problem is further complicated when dealing with dynamic graphs, since adding or removing an edge requires re-running the algorithm. In this paper, we study the top-$k$ densest-subgraph problem in the sliding-window model and propose an efficient fully-dynamic algorithm. The input of our algorithm consists of an edge stream, and the goal is to find the node-disjoint subgraphs that maximize the sum of their densities. In contrast to existing state-of-the-art solutions that require iterating over the entire graph upon any update, our algorithm profits from the observation that updates only affect a limited region of the graph. Therefore, the top-$k$ densest subgraphs are maintained by only applying local updates. We provide a theoretical analysis of the proposed algorithm and show empirically that the algorithm often generates denser subgraphs than state-of-the-art competitors. Experiments show an improvement in efficiency of up to five orders of magnitude compared to state-of-the-art solutions., Comment: 10 pages, 8 figures, accepted at CIKM 2017

Network alignment is the problem of matching the nodes of two graphs, maximizing the similarity of the matched nodes and the edges between them. This problem is encountered in a wide array of applications-from biological networks to social networks to ontologies-where multiple networked data sources need to be integrated. Due to the difficulty of the task, an accurate alignment can rarely be found without human assistance. Thus, it is of great practical importance to develop network alignment algorithms that can optimally leverage experts who are able to provide the correct alignment for a small number of nodes. Yet, only a handful of existing works address this active network alignment setting. The majority of the existing active methods focus on absolute queries ("are nodes $a$ and $b$ the same or not?"), whereas we argue that it is generally easier for a human expert to answer relative queries ("which node in the set $\{b_1, \ldots, b_n\}$ is the most similar to node $a$?"). This paper introduces two novel relative-query strategies, TopMatchings and GibbsMatchings, which can be applied on top of any network alignment method that constructs and solves a bipartite matching problem. Our methods identify the most informative nodes to query by sampling the matchings of the bipartite graph associated to the network-alignment instance. We compare the proposed approaches to several commonly-used query strategies and perform experiments on both synthetic and real-world datasets. Our sampling-based strategies yield the highest overall performance, outperforming all the baseline methods by more than 15 percentage points in some cases. In terms of accuracy, TopMatchings and GibbsMatchings perform comparably. However, GibbsMatchings is significantly more scalable, but it also requires hyperparameter tuning for a temperature parameter., Comment: This is a pre-print of an article appearing at CIKM 2017

We introduce and study the problem of computing the similarity self-join in a streaming context (SSSJ), where the input is an unbounded stream of items arriving continuously. The goal is to find all pairs of items in the stream whose similarity is greater than a given threshold. The simplest formulation of the problem requires unbounded memory, and thus, it is intractable. To make the problem feasible, we introduce the notion of time-dependent similarity: the similarity of two items decreases with the difference in their arrival time. By leveraging the properties of this time-dependent similarity function, we design two algorithmic frameworks to solve the sssj problem. The first one, MiniBatch (MB), uses existing index-based filtering techniques for the static version of the problem, and combines them in a pipeline. The second framework, Streaming (STR), adds time filtering to the existing indexes, and integrates new time-based bounds deeply in the working of the algorithms. We also introduce a new indexing technique (L2), which is based on an existing state-of-the-art indexing technique (L2AP), but is optimized for the streaming case. Extensive experiments show that the STR algorithm, when instantiated with the L2 index, is the most scalable option across a wide array of datasets and parameters.

We introduce a framework for analyzing roles of users during a riot. The dataset contains messages from Twitter during the 2014 US Ferguson protests. First, after some preprocessing one the data, we extract topics from a riot by comparing two techniques : k-means and Latent Dirichlet Allocation. Secondly, we focus on the content of the tweets. We train and test a Naive Bayes classifier to predict if a tweet is either supportive or informative about the riot. We also study the medias shared by users in order to improve our classifier and in the end we define and compute user polarity score. Then, we perform graph analysis on the hashtags and users and we visualize these networks through examples. We also define and compute user influence score with degree centrality and Page Rank. Finally, we define the role of a user during a riot (or several) as being the feature vector of its polarity and influence scores for each topic. We perform dimension reduction with PCA and t-SNE for visualization purpose and neighbors research. Through experiments, we show how users with the same job (for example journalists or activists) are grouped together.

Among the topics discussed on social media, some spark more heated debate than others. For example, experience suggests that major political events, such as a vote for healthcare law in the US, would spark more debate between opposing sides than other events, such as a concert of a popular music band. Exploring the topics of discussion on Twitter and understanding which ones are controversial is extremely useful for a variety of purposes, such as for journalists to understand what issues divide the public, or for social scientists to understand how controversy is manifested in social interactions. The system we present processes the daily trending topics discussed on the platform, and assigns to each topic a controversy score, which is computed based on the interactions among Twitter users, and a visualization of these interactions, which provides an intuitive visual cue regarding the controversy of the topic. The system also allows users to explore the messages (tweets) associated with each topic, and sort and explore the topics by different criteria (e.g., by controversy score, time, or related keywords)., Comment: Accepted as demo at CSCW 2016