Research Projects

A list of current and previous projects.


Finding prize-collecting subgraphs in activity networks [under submission]

S. M. Nikolakaki, C. Mavroforakis, A. Ene and E. Terzi

Code

Multi-FPGA Architectures for Streaming Computations of Information Transfer [under submission]

S. M. Nikolakaki, K. Iordanou, P. Malakonakis and A. Dollas

Report

Towards Accelerating the LIBSVM Software on the Reconfigurable Maxeler System

S. M. Nikolakaki, P. Malakonakis, E. Sotiriades and A. Dollas

HiStencils 2016, 3rd International Workshop on High-Performance Stencil Computations

Academic Projects


Twitter Sentiment Analysis: U.S. Election 2016 (We identified the Donald Trump trend on 04/16!)

Priya Ayyappan, S. M. Nikolakaki, G. Kollios, K. Zhao

Poster

This project used sentiment analysis and clustering techniques on Twitter data to estimate the citizens’ voting intention for the 2016 U.S. election. More specifically, our goal is to cluster Twitter users based on their political sentiment towards a party candidate and their U.S. residence location. This presumes performing sentiment analysis on an input Twitter data instance to extract an emotion and identifying the state from where the data were posted. Furthermore, classification was performed to create a class for each sentiment-political candidate party pair, that is 16 classes; we are considering the top four political candidates and four feelings (Positive/Negative/Neutral/Undecided). In addition to classification, we apply the well-known K-Means clustering method on the emotion and the location of the data instance to extract the voting tendency per state. Our results were compared to real prediction voting polls in order to identify and interpret possible divergence between the two sources.


Spatial and Temporal Analysis of Crime for the Discovery of Hot Spots in Road Networks: the case of Boston

S. M. Nikolakaki, E. D. Kolaczyk

Report

The formation of criminal hot spots is a result of high concentration in criminal activity in a specific area. Restricting their discovery to human reasoning reduces the possibility of confining this phenomenon. A major challenge in hot spot analysis research is the lack of withstanding definitions and techniques thus making the specific task application-oriented.

This work studied the dynamic behavior of criminality in the example of Boston with the goal to discover crime hot spots within road network neighborhoods. In order to do so we follow a two-step methodology. The first step involves temporal criminal analysis and dynamic crime mapping using Geographic Information System (GIS) techniques. We extract and integrate spatial, temporal and criminal occurrence information to construct the visual product of the spatio-temporal criminality dynamics on the road network of Boston. We utilize graduated symbols and color gradient dots to illustrate data points. The second step quantifies the "dangerousness" of an area surrounding a location to allow hot spot mining. For this purpose we follow a statistical approach using crime counts and the Location Quotient (LQC) crime measure as well as a clustering analysis method with the use of the well-known agglomerative hierarchical clustering algorithm. Both procedures successfully discover high-crime density places, with the former confirming the results with numerical values and the latter with community visualizations.


A Minimax Implementation of a Blokus Duo Agent

S. M. Nikolakaki, I. Demertzis, P. Malakonakis, G. Chrysos, A. Dollas

Report

This work presented an FPGA-based Blokus Duo player architecture and its implementation. The Technical University of Crete (TUC) arxhitecture is based on a depth-first search Minimax algorithm with alpha beta pruning and it was fully implemented on a Xilinx Spartan 6 FPGA. The operational system complies with the rules of the 2013 ICFPT Design Contest, and when operating under the strict computation time requirements it performs up to level 3 against Pentobi, operating under no restrictions on computational resources or time.