Graph network models for Data Science

Machine learning approaches extract information from data with generalized optimization methods. However, besides the knowledge brought by the data, extra a-priori knowledge of the modeled phenomena is often available. Hence an inductive bias can be introduced from domain knowledge and physical constraints, as proposed by the emerging field of Theory-Guided Data Science.
Within this broad field, the candidate will explore solutions exploiting the relational structure among data,  represented by means of Graph Network approaches.
Relational structure is present in many real-life settings, both in physical conditions, such as among actors in supply chains or users in social networks, and logical processes performed by humans, such as industrial procedures.
The structure of the data can be exploited to directly build the network graph itself, incorporating hierarchies and relationships among the different elements.
Analogous approaches can be exploited for logical processes where domain experts separate the overall procedure in connected subtasks for their decision making.
Hence, a graph-like structure can be crafted to design an ensemble architecture consisting of different building-blocks, each connecting a network node representing the sub-problem and the corresponding domain-driven knowledge and constraints.
The candidate should explore such a set of approaches to design and evaluate innovative learning strategies able to blend domain-expert behaviors, a-priori knowledge, and physical or theoretical constraints with the traditional data-driven training.

The aim of the research is to define new methodologies for semantics embedding, propose novel algorithms  and data structures, explore applications, investigate limitations, and advance the solutions based on  different emerging Theory-guided Data Science approaches.
The final goal is to contribute to improving the Machine Learning model performance by reducing the learning space thanks to the exploitation of existing domain knowledge in addition to the (often limited) available training data, pushing towards more unsupervised and semantically richer models.
To this aim, the main research objective is to exploit the Graph Network frameworks in deep-learning architectures by addressing the following issues:

• Improving state-of-the-art strategies of organizing and extracting information from structured data.
• Overcoming the Graph-Network model limitation in training very deep architectures, with a consequent loss in expressive power of the solutions.
• Advancing the state-of-the-art solutions to dynamic graphs, which can change nodes and mutual connections over time. Dynamic Networks can successfully learn the behavior of evolving systems.
• Experimentally evaluate the novel techniques in large-scale systems, such as supply chains, social networks, collaborative smart-working platforms, etc. Currently, for most graph-embedding algorithms, the scalability of the structure is difficult to handle since each node has a peculiar neighborhood organization.
• Applying the proposed algorithms to natively graph-unstructured data, such as texts, images, audio, etc.
• Developing techniques to design ensemble graph architectures to capture domain-knowledge relationships and physical constraints.

1st year. The candidate will explore the state-of-the art techniques of dealing with both structured and unstructured data, to integrate domain-knowledge strategies in network model architectures.
Applications to physics phenomena, images and text, taken from real-world networks such as social platforms and supply chains will be considered.

2nd year. The candidate will define innovative solutions to overcome the limitations described in the research objectives, by experimenting the proposed techniques on the identified real-world problems. The development and the experimental phase will be conducted on public, synthetic, and possibly real-world
datasets. New challenges and limitations are expected to be identified in this phase.

During the 3rd year the candidate will extend the research by widening the experimental evaluation to more complex phenomena able to better leverage the domain-knowledge provided by the Graph Networks. The candidate will perform optimizations on the designed algorithms, establishing limitations of the developed solutions and possible improvements in new application fields.

Expected target publications

IEEE TKDE (Trans. on Knowledge and Data Engineering)
ACM TKDD (Trans. on Knowledge Discovery in Data)
ACM TOIS (Trans. on Information Systems)
ACM TOIT (Trans. on Internet Technology)
ACM TIST (Trans. on Intelligent Systems and Technology)
IEEE TPAMI (Trans. on Pattern Analysis and Machine Intelligence)
Information sciences (Elsevier)
Expert systems with Applications (Elsevier)
Engineering Applications of Artificial Intelligence (Elsevier)
Journal of Big Data (Springer)
ACM Transactions on Spatial Algorithms and Systems (TSAS)

IEEE Transactions on Big Data (TBD)
Big Data Research
IEEE Transactions on Emerging Topics in Computing (TETC)
Information sciences (Elsevier)

Funded projects of the proposer related to the proposal

Research contract “Data Science and Machine Learning techniques for clinical supply chains”.

Industries/companies that are involved in the proposal

XelionTech, FBK, Istituto Mario Negri.