ETD

• deep learning
• graph classification
• graph enconding
• link prediction
• machine learning
• probabilistic model

The Contextual Graph Markov Model (CGMM) is a constructive methodology to build a deep architecture comprising layers of probabilistic models that learn to encode graph-structured information in an incremental fashion.
This thesis extends the original CGMM model to admit the presence of continuous edge attributes while maintaining scalability and efficiency. Besides node and graph encodings, this new architecture also provides edge encodings, which can be used in combination with discriminative models to address edge classification and link prediction tasks.
We show that the proposed model achieves empirical performances that are comparable with the state-of-the-art on graph classification tasks and exhibits significant improvements on link prediction task with respect to CGMM.