Georgios Drakopoulos

Higher order graph centrality measures for Neo4j

Graphs are currently the epicenter of intense research as they lay the theoretical groundwork in diverse fields ranging from combinatorial optimization to computational neuroscience. Vertex centrality plays a crucial role in graph mining as it ranks them according to their contribution to overall graph communication. Specifically, within the social network analysis context centrality identifies influential indivduals, whereas in the bioinformatics field centrality locates dominant proteins in protein-to-protein interaction. In recent years graph databases, part of the rising NoSQL movement, have been added to the graph analysis toolset. An implementation of eigenvector centrality, a prominent member of the broad class of spectral centrality, in Java and NetBeans designed for use with Neo4j, a major schemaless graph database, is outlined and the findings resulting from its application to a real world social graph are discussed.

Evaluating Twitter Influence Ranking with System Theory

A considerable part of social network analysis literature is dedicated to determining which individuals are to be considered as influential in particular social settings. Most established algorithms, such as Freeman and KatzBonacich centrality metrics, place emphasis on various structural properties of the social graph. Although this makes centrality metrics generic enough to be applied in virtually any setting, they are oblivious to the functionality of the underlying social network. This paper examines five social influence metrics designed especially for Twitter and their implementation in a Java client retrieving network information from a Neo4j server. Additionally, a sceheme is proposed for evaluating the performance of an influence ranking based on estimating the exponent of a Zipf model fitted to the ranking score.

Tensor-based document retrieval over Neo4j with an application to PubMed mining

PubMed mining is currently at the epicenter of intense interdisciplinary research. Text mining methodologies provide a way to retrieve and analyze emotionally charged words, punctuation, and syntax. Moreover, they can analyze scientific literature and process document collections. Moving beyond traditional document-term matrix representation, an architecture for content based retrieval from PubMed is proposed whose core is a document-term-author third order tensor. This methodology has been implemented in Python over Neo4j and has been applied to a PubMed document article collection.

Tensor fusion of social structural and functional analytics over Neo4j

For large directed graphs ranking vertices is an algorithmic as well as a computational challenge. Also it is a cornerstone problem in fields so diverse such as online social media, combinatorial optimization, deep learning, econometrics, and computational neuroscience. Gell-point centrality metric is a common case of structural vertex ranking which can be efficiently computed through power method. This paper proposes harmonic centrality, also a structural ranking which can be computed equally efficiently. When the graph represents a network, both methods are oblivious to its functionality. To address that issue, a tensor based methodology is proposed for combining functional characteristics with these structural metrics. As a concrete demonstration, the tensor fusion methodology was implemented in Java over Neo4j and Tensor toolbox and was applied to a Twitter subgraph. The functional features were directly related to affective computing, while the hashtags in this subgraph were relared to a current controversial political topic, namely #brexit.

On the weight sparsity of multilayer perceptrons

Approximating and representing a process, a function, or a system with an adaptive parametric model constitutes a major part of current machine learning research. An important characteristic of these models is parameter sparsity, an indicator of how succintly a model can codify fundamental properties of the approximated function. This paper investigates the sparsity patterns of a multilayer perceptron netwrok trained to mount a man-on-the-middle attack on the DES symmetric cryptosystem. The notions of absolute and effective synaptic weight sparsity are introduced and their importance to network learning procedure is explained. Finally, the results from the training of the actual multilayer perceptron are outlined and discussed. In order to promote reproducible research, the MATLAB network implementation has been posted in GitHub.

A Space Efficient Scheme for Persistent Graph Representation

Graph mining is currently the focus of intense research. Major driving factors include social media, opinion mining, and the schemaless noSQL databases. Time evolving or dynamic graphs are the primary data structures in these fields. Often dynamic graphs must support persistency, meaning that from any given graph state past states can be accessed. Within the graph database context, persistency enables rollback capability, whereas in social media several phenomena such as friend deletion can be modeled. A novel, efficient, and persistent data structure based on tries is proposed. Its potential is displayed by added persistency to the deterministic Kronecker graph model.

Augmenting fMRI-generated brain connectivity with temporal information

fMRI is an established neuroimaging technique providing a snapshot sequence of brain activity with very high spatial resolution. An efficient data mining algorithm is proposed for agumenting voxel activation detection, best performed spatially, with task-specific information, which is primarily temporal. Thus, the observation of evolution of brain region co-activation through fMRI is enhanced. Also, based on the AAL2 standard brain atlas, a neurophysiological interpretation of the findings during a motor task is offered, strengthening the validity of the proposed algorithm.

Regularizing large biosignals with finite differences

In the biomedical analytics pipeline data preprocessing is the first and crucial step as subsequent results and visualization depend heavily on original data quality. However, the latter often contain a large number of outliers or missing values. Moreover, they may be corrupted by noise of unknown characteristics. This is in many cases aggravated by lack of sufficient information to construct a data cleaning mechanism. Regularization techniques remove erroneous values and complete missing ones while requiring little or no information regarding either data or noise dynamics. This paper examines the theory and practice of a regularization class based on finite differences and implemented through the conjugate gradient method. Moreover, it explores the connection of finite differences to the discrete Laplace operator. The results obtained from applying the proposed regularization techniques to heart rate time series from the MIT-BIH dataset are discussed.

Eventually consistent cardinality estimation with applications in biodata mining

Large set cardinality estimators and other streaming oriented operations are the cornerstone of big data processing. Cardinality estimators combined with in-memory based storage systems provide a fast framework for keeping valuable application data easily queryable and maintanable. This has a plethora of applications. For instance, a common use case is to maintain a number of counters for monitoring application statistics for real time dashboard purposes. Another such case is large set size estimation for big data systems in internal operations like counting. In this paper is addressed the issue of scaling the computation of a cardinality estimator in the presence of node failures in a distributed setting. Moreover, for the proposed estimation technique eventual consistency is proved, which is adequate for most cases in distributed applications. To the best of the authors knowledge, this functionality is not currently provided by commonly used commercial and open source systems. Additionally, the proposed approach is generic enough to be applied to other algorithms, which can help build a basic framework for more complex operations in the big data field. We demonstrate this with graph metric calculation applications in the large scale biodata mining field.

Graph Community Discovery Algorithms in Neo4j with a Regularization-based Evaluation Metric.

Community discovery is central to social network analysis as it provides a natural way for decomposing a social graph to smaller ones based on the interactions among individuals. Communities do not need to be disjoint and often exhibit recursive structure. The latter has been established as a distinctive characteristic of large social graphs, indicating a modularity in the way humans build societies. This paper presents the implementation of four established community discovery algorithms in the form of Neo4j higher order analytics with the Twitter4j Java API and their application to two real Twitter graphs with diverse structural properties. In order to evaluate the results obtained from each algorithm a regularization-like metric, balancing the global and local graph self-similarity akin to the way it is done in signal processing, is proposed.