Georgios Pitsilis

Clustering Recommenders in Collaborative Filtering Using Explicit Trust Information

In this work, we explore the benefits of combining clustering and social trust information for Recommender Systems. We demonstrate the performance advantages of traditional clustering algorithms like k-Means and we explore the use of new ones like Affinity Propagation (AP). Contrary to what has been used before, we investigate possible ways that social-oriented information like explicit trust could be exploited with AP for forming clusters of high quality. We conducted a series of evaluation tests using data from a real Recommender system Epinions.com from which we derived conclusions about the usefulness of trust information in forming clusters of Recommenders. Moreover, from our results we conclude that the potential advantages in using clustering can be enlarged by making use of the information that Social Networks can provide.

Abstracting massive data for lightweight intrusion detection in computer networks

Abstract Anomaly intrusion detection in big data environments calls for lightweight models that are able to achieve real-time performance during detection. Abstracting audit data provides a solution to improve the efficiency of data processing in intrusion detection. Data abstraction refers to abstract or extract the most relevant information from the massive dataset. In this work, we propose three strategies of data abstraction, namely, exemplar extraction, attribute selection and attribute abstraction. We first propose an effective method called exemplar extraction to extract representative subsets from the original massive data prior to building the detection models. Two clustering algorithms, Affinity Propagation (AP) and traditional k-means, are employed to find the exemplars from the audit data. k-Nearest Neighbor (k-NN), Principal Component Analysis (PCA) and one-class Support Vector Machine (SVM) are used for the detection. We then employ another two strategies, attribute selection and attribute extraction, to abstract audit data for anomaly intrusion detection. Two http streams collected from a real computing environment as well as the KDD’99 benchmark data set are used to validate these three strategies of data abstraction. The comprehensive experimental results show that while all the three strategies improve the detection efficiency, the AP-based exemplar extraction achieves the best performance of data abstraction.

Does Trust Matter for User Preferences? A Study on Epinions Ratings

Recommender systems have evolved during the last few years into useful online tools for assisting the daily e-commerce activities. The majority of recommender systems predict user preferences relating users with similar taste. Prior research has shown that trust networks improve the performance of recommender systems, predominantly using algorithms devised by individual researchers. In this work, omitting any specific trust inference algorithm, we investigate how useful it might be if explicit trust relationships (expressed by users for others) are used to select the best neighbours (or predictors), for the provision of accurate recommendations. We conducted our experiments using data from Epinions.com, a popular recommender system. Our analysis indicates that trust information can be helpful to provide a slight performance gain in a few cases especially when it comes to the less active users.

Abstracting audit data for lightweight intrusion detection

High speed of processing massive audit data is crucial for an anomaly Intrusion Detection System (IDS) to achieve real-time performance during the detection. Abstracting audit data is a potential solution to improve the efficiency of data processing. In this work, we propose two strategies of data abstraction in order to build a lightweight detection model. The first strategy is exemplar extraction and the second is attribute abstraction. Two clustering algorithms, Affinity Propagation (AP) as well as traditional k-means, are employed to extract the exemplars, and Principal Component Analysis (PCA) is employed to abstract important attributes (a.k.a. features) from the audit data. Real HTTP traffic data collected in our institute as well as KDD 1999 data are used to validate the two strategies of data abstraction. The extensive test results show that the process of exemplar extraction significantly improves the detection efficiency and has a better detection performance than PCA in data abstraction.

An Efficient Authorship Protection Scheme for Shared Multimedia Content

Many electronic content providers today like Flickr and Google, offer space to users to publish their electronic media(e.g. photos and videos) in their cloud infrastructures so that they can be publicly accessed. Features like including other information, such as keywords or owner information into the digital material is already offered by existing providers. Despite the useful features made available to users by such infrastructures, the authorship of the published content is not protected against various attacks such as compression. In this paper we propose a robust scheme that uses digital invisible watermarking and hashing to protect the authorship of the digital content and provide resistance against malicious manipulation of multimedia content. The scheme is enhanced by an algorithm called MMBEC, that is an extension of an established scheme MBEC towards higher resistance.

Effective hate-speech detection in Twitter data using recurrent neural networks

This paper addresses the important problem of discerning hateful content in social media. We propose a detection scheme that is an ensemble of Recurrent Neural Network (RNN) classifiers, and it incorporates various features associated with user-related information, such as the users’ tendency towards racism or sexism. This data is fed as input to the above classifiers along with the word frequency vectors derived from the textual content. We evaluate our approach on a publicly available corpus of 16k tweets, and the results demonstrate its effectiveness in comparison to existing state-of-the-art solutions. More specifically, our scheme can successfully distinguish racism and sexism messages from normal text, and achieve higher classification quality than current state-of-the-art algorithms.This paper addresses the important problem of discerning hateful content in social media. We propose a detection scheme that is an ensemble of Recurrent Neural Network (RNN) classifiers, and it incorporates various features associated with userrelated information, such as the users’ tendency towards racism or sexism. These data are fed as input to the above classifiers along with the word frequency vectors derived from the textual content. Our approach has been evaluated on a publicly available corpus of 16k tweets, and the results demonstrate its effectiveness in comparison to existing state of the art solutions. More specifically, our scheme can successfully distinguish racism and sexism messages from normal text, and achieve higher classification quality than current state-of-the-art algorithms.

Trust-Enhanced Recommender Systems for Efficient On-Line Collaboration

Trust has been explored by many researchers in the past as a solution for assisting the process of recommendation production. In this work we are examining the feasibility of building networks of trusted users using the existing evidence that would be provided by a standard recommender system. As there is lack of models today that could help in finding the relationship between trust and similarity we build our own that uses a set of empirical equations to map similarity metrics into Subjective Logic trust. In this paper we perform evaluation of the proposed model as being a part of a complete recommender system. Finally, we present the interesting results from this evaluation that shows the performance and benefits of our trust modeling technique as well as its impact on the user community as it evolves over time.