Surfing data mining techniques for representing data sources have specifically attracted much attention among researchers. Given the curse of dimensionality in representing text using the traditional Bag-of-words models, lower-dimensional representation of text has been an important line of research due to its impact on many prediction, and recommendation tasks.
This thesis studies two main different viewpoints in text representation using content and citation information and then, different existing approaches along with their advantages, limitations and drawbacks are reviewed. A novel hybrid distributed technique for text representation is proposed where the textual content of documents is projected into a vector representation using an artificial neural network .
To test the performance of the new proposed technique, the well known link-prediction problem is selected to serve as a benchmark. A comparison is performed with other common techniques by predicting the existence of citation links between tuple of papers in a large citation graph.