National Technical University of Athens (NTUA)
Τεχνικές μηχανικής μάθησης για την ανίχνευση ομοιότητας/ανομοιότητας: εφαρμογές εντοπισμού κοινοτήτων σε κοινωνικά δίκτυα και ελέγχου λογοκλοπής σε κείμενα
Abstract
dc:descriptionIn recent years, the amount of information transmitted online has greatly increased, which dictates the creation and usage of new systems, capable of handling large volumes of information. Machine Learning and Data Mining are two fields of study, which facilitate the analysis and classification of information. Machine learning algorithms "learn" directly from data, by discovering meaningful patterns, without the use of explicit instructions. In this thesis, three separate, but related, approaches for Community Detection and Intrinsic Plagiarism Detection, which utilize machine learning techniques, were studied and implemented. Community Detection, or graph clustering, is one of the most popular topics in modern science of networks, which aims to solve the problem of identifying the community structure in networks. Most networks display community structure, i.e. the vertices are organized in groups, called communities, groups or clusters. Community detection is not a well defined problem, as there is no strict and universally accepted definition of community. The definition often depends on the application, the research question at hand or on the specific system under study. In this thesis, the problem of community detection in social networks was studied and a methodology for identifying similar users on Twitter was proposed. Communities are defined as groups of users that are more densely connected to each other than to the rest of the network, interact more between them and share common interests. Therefore, this methodology does not solely depend on the network topology in order to group the users into communities, but also takes into account the text that users share, as well as their interactions. Initially, six different similarity metrics are defined, which are based on Twitter attributes that provide information regarding the interactions between users. These metrics are combined, and their combination is used for the clustering of users into communities. Additionally, a new method for extracting the topics discussed in each community is presented, which helps identify the users' interests. Also, a method for the elimination of the trivial topics is proposed, and a process for automatically generating labels for the topics is described. Secondly, graph embedding and the extraction of node embeddings are being studied. Graph embedding methods have been proposed as an alternative to traditional graph mining techniques. The objective is to convert a graph into a low dimensional representation, where each node of the graph would be mapped to a low dimensional vector. These vectors, also called node embeddings or feature vectors, can then be presented as input to any supervised learning algorithm, thus simplifying the original problem. Therefore, these methods can be useful in a variety of real-world applications, such as node classification, community detection, link prediction and network visualization. Based on the above, the second approach is proposed, which, contrary to previous approaches, which only take into account the edges of the graph when exploring the graph through random walks, also considers the similarities between the nodes. Plagiarism is the act of taking or closely imitating someone else's work and presenting it as original, without proper citation or acknowledgment. Plagiarism detection in text documents is divided into two major categories, extrinsic and intrinsic methods. Extrinsic methods detect the suspicious similarities between a collection of potential source documents and a set of suspicious documents, while in intrinsic methods the objective is to identify which of the passages of an investigated document are plagiarized by observing the variation of the writing style within the document. Intrinsic plagiarism detection is based on the idea that every author has its own personal and unique writing style, which can be detected and quantified using stylistic and/or semantic means. Based on the above, an intrinsic plagiarism detection approach for text documents is presented. Initially, a set of novel stylistic features, which help quantify the author's writing style for the whole document and each suspicious passage, is introduced. These features are then combined with a number of supervised learning methods, in order to classify the passages into plagiarized or non-plagiarized. Finally, the unbalanced nature of the datasets is examined, which is considered a crucial parameter for this task. As a result, over-sampling and under-sampling techniques are used in order to examine whether the performance of the proposed system is improved. The proposed approaches were evaluated on publicly available datasets. Due to the nature of the problem of community detection, the results of the first methodology are not evaluated in comparison to an already existing approach. On the contrary, the second and third approaches were compared to the state-of-the-art algorithms for each respective field of study. The experimental results demonstrate the satisfactory behavior of the proposed methodologies, which in many cases outperform the algorithms they are compared to.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2019
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Vathi, Eleni
- Βάθη, Ελένη
Subjects
dc:subject × 20- Μηχανική μάθηση
- Ανίχνευση κοινοτήτων
- Ανάλυση κοινωνικών δικτύων
- Μοντελοποίηση θεμάτων
- Διανυσματικές παραστάσεις κόμβων
- Μάθηση χαρακτηριστικών
- Εγγενής ανίχνευση λογοκλοπής
- Μη ισορροπημένα δεδομένα
- Machine learning
- Community detection
- Social networks analysis
- Topic modeling
- Node embeddings
- Feature learning
- Intrinsic plagiarism detection
- Imbalanced data
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Natural Sciences
- Computer and Information Sciences
Rights
- Language dc:language
- gre
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/47095