Amendra Shrestha

Detecting multiple aliases in social media

Monitoring and analysis of web forums is becoming important for intelligence analysts around the globe since terrorists and extremists are using forums for spreading propaganda and communicating with each other. Various tools for analyzing the content of forum postings and identifying aliases that need further inspection by analysts have been proposed throughout literature, but a problem related to this is that individuals can make use of several aliases. In this paper we propose a number of matching techniques for detecting forum users who make use of multiple aliases. By combining different techniques such as time profiling and stylometric analysis of messages the accuracy of recognizing users with multiple aliases increases, as shown in experiments conducted on the ICWSM dataset boards.ie.

Time Profiles for Identifying Users in Online Environments

Many people who discuss sensitive or private issues on web forums and other social media services are using pseudonyms or aliases in order to not reveal their true identity, while using their usual accounts when posting messages on nonsensitive issues. Previous research has shown that if those individuals post large amounts of messages, stylometric techniques can be used to identify the author based on the characteristics of the textual content. In this paper we show how an authors identity can be unmasked in a similar way using various time features, such as the period of the day and the day of the week when a users posts have been published. This is demonstrated in supervised machine learning (i.e., author identification) experiments, as well as unsupervised alias matching (similarity detection) experiments.

Detecting Multipliers of Jihadism on Twitter

Detecting terrorist related content on social media is a problem for law enforcement agency due to the large amount of information that is available. This work is aiming at detecting tweeps that are involved in media mujahideen - the supporters of jihadist groups who disseminate propaganda content online. To do this we use a machine learning approach where we make use of two sets of features: data dependent features and data independent features. The data dependent features are features that are heavily influenced by the specific dataset while the data independent features are independent of the dataset and can be used on other datasets with similar result. By using this approach we hope that our method can be used as a baseline to classify violent extremist content from different kind of sources since data dependent features from various domains can be added. In our experiments we have used the AdaBoost classifier. The results shows that our approach works very well for classifying English tweeps and English tweets but the approach does not perform as well on Arabic data.

Activity profiles in online social media

Analysis and mining of social media has become an important research area. A challenging problem in this area consists in the identification of a group of users with similar patterns. In this paper, we propose the classification of users based on their activity profiles (e.g., periods of the day when the user is most and least active in online communications). Activity profiles can be useful for many purposes, such as marketing and user behavior analysis. They can also serve as a basis for other techniques such as stylometric and time analysis in order to increase the precision and scalability of multiple aliases identification techniques. We have implemented a prototype tool and applied it on a dataset from the ICWSM data set Boards.ie, showing the usefulness of our classification.

Timeprints for identifying social media users with multiple aliases

Many people who discuss sensitive or private issues on social media services are using pseudonyms or aliases in order to not reveal their true identity, while using their usual, non-private accounts when posting messages on less sensitive issues. Previous research has shown that if those individuals post large amounts of user-generated content, stylometric techniques can be used to identify the author based on the characteristics of the textual content. In this article we show how an author’s identity can be unmasked in a similar way using various time features (e.g., period of the day and the day of the week when a user’s posts have been published). We combine several different time features into a timeprint, which can be seen as a type of fingerprint when identifying users on social media. We use supervised machine learning (i.e., author identification) and unsupervised alias matching (similarity detection) in a number of different experiments with forum data to get an understanding of to what extent timeprints can be used for identifying users in social media, both in isolation and when combined with stylometric features. The obtained results show that timeprints indeed can be a very powerful tool for both author identification and alias matching in social media.

Identifying Warning Behaviors of Violent Lone Offenders in Written Communication

Violent lone offenders such as school shooters and lone actor terrorists pose a threat to the modern society but since they act alone or with minimal help form others they are very difficult to detect. Previous research has shown that violent lone offenders show signs of certain psychological warning behaviors that can be viewed as indicators of an increasing or accelerating risk of committing targeted violence. In this work, we use a machine learning approach to identify potential violent lone offenders based on their written communication. The aim of this work is to capture psychological warning behaviors in written text and identify texts written by violent lone offenders. We use a set of features that are psychologically meaningful based on the different categories in the text analysis tool Linguistic Inquiry and Word Count (LIWC). Our study only contains a small number of known perpetrators and their written communication but the results are promising and there are many interesting directions for future work in this area.

A Machine Learning Approach towards Detecting Extreme Adopters in Digital Communities

In this study we try to identify extreme adopters on a discussion forum using machine learning. An extreme adopter is a user that has adopted a high level of a community-specific jargon and therefore can be seen as a user that has a high degree of identification with the community. The dataset that we consider consists of a Swedish xenophobic discussion forum where we use a machine learning approach to identify extreme adopters using a number of linguistic features that are independent on the dataset and the community. The results indicates that it is possible to separate these extreme adopters from the rest of the discussants on the discussion forum with more than 80% accuracy. Since the linguistic features that we use are highly domain independent, the results indicates that there is a possibility to use this kind of techniques to identify extreme adopters within other communities as well.

Automatic detection of xenophobic narratives: A case study on Swedish alternative media

In this work we use text analysis to analyze communication on a set of Swedish immigration critic alternative media sites. Our analysis is focused on detecting narratives containing xenophobic and conspiratorial stereotypes. We are also interested in identifying differences in emotional tone and pronoun use in a comparison with traditional media. For our analysis we have used the text analysis tool LIWC (Linguistic Inquiry and Word Count) and a set of dictionaries made to capture a xenophobic narrative. The results show that there are significant differences between regular media and immigration critic alternative media when it comes to the use of narratives and also in the emotional tone and pronoun use.

Multi-domain Alias Matching Using Machine Learning

We describe a methodology for linking aliases belonging to the same individual based on a users writing style (stylometric features extracted from the user generated content) and her time patterns (time-based features extracted from the publishing times of the user generated content). While most previous research on social media identity linkage relies on matching usernames, our methodology can also be used for users who actively try to choose dissimilar usernames when creating their aliases. In our experiments on a discussion forum dataset and a Twitter dataset, we evaluate the performance of three different classifiers. We use the best classifier (AdaBoost) to evaluate how well it works on different datasets using different features. Experiments show that combining stylometric and time-based features yield good results on our synthetic datasets and a small-scale evaluation on real-world blog data confirm these results, yielding a precision over 95%. The use of emotion-related and Twitter-related features yield no significant impact on the results.

Personalized mobile-assisted smart transportation

Digital media covers larger parts of our daily lives nowadays. Mobile services enable a better connected society where citizens can easily access public services, discover events, and obtain important information in the city. We observe the popularity of mobile car sharing applications, such as Uber and Didi Dache. Mobile social applications provide new ways of developing and optimizing public transportation. In this paper, we present a mobile platform for timetable-free traveling. It can capture the traffic demand of citizens in real-time, and support efficient planning and scheduling for vehicles on-demand. At the moment, the platform is targeted for public bus services, but it has great potential to be extended for self-driving vehicles in the future.