Individuals have distinctive ways of speaking and writing, and there exists a long history of linguistic and stylistic investigation into authorship attribution. Most authorship identification approaches are exclusively based on lexical measures such as vocabulary richness and lexico-syntactic features, or substantially generate relevant features for different machine learning approaches. These techniques are inefficient without suitable feature selection or large corpus. In this paper, we introduce three graph based models for the task of authorship identification, that consider the interaction of character sequences, phraseological patterns and structure of the sentences in the document to construct graphs separately for an author. For each model, all such graphs for different authors are aggregated to generate a combined weighted training graph. Then a simple graph traversal algorithm is used to compare testing graph with the training graph. The experiment is conducted on the documents of six authors collected from Bengali literature. Experimental results show that our models significantly outperform four state-of-the-art models (9.89% higher than the best baseline model) even for short training dataset.