Final Report Summary - MIGRANT (Mining Graphs and Networks: a Theory-based approach)
The MiGraNT project developed a sound theoretical understanding of graph mining, i.e. data mining and machine learning in graph-structured data, both transactional data such as databases of molecules, images, case descriptions,etc. and data networks such as social networks, economic networks, administrative networks, traffic networks, etc. The project has also built algorithms based on these insights and applied them to significant applications.
Two challenges the project considered were the following. First, many operations on graphs (e.g. pattern matching) are computationally very expensive (NP-hard). The project showed despite the NP-hardness of many problems and the ineffectivity of existing approaches, it is in many cases possible to design computationally efficient graph mining algorithms, e.g. by exploiting ideas from fixed parameter tractability.
The second challenge the project considered is a statistical one. In large data networks, objects are connected with each other and are hence not independent (e.g. persons in a social network influence each other). This violates the assumption made by the vast majority of statistics and machine learning approaches, that observations are identically and independently distributed. Here, the project developed methods to model dependencies between objects and take them into account when learning predictive models.
The project results have several applications, e.g. in the fields of proteomics (leading to the ERC PoC SNIPER project), fraud detection and mobility.
Two challenges the project considered were the following. First, many operations on graphs (e.g. pattern matching) are computationally very expensive (NP-hard). The project showed despite the NP-hardness of many problems and the ineffectivity of existing approaches, it is in many cases possible to design computationally efficient graph mining algorithms, e.g. by exploiting ideas from fixed parameter tractability.
The second challenge the project considered is a statistical one. In large data networks, objects are connected with each other and are hence not independent (e.g. persons in a social network influence each other). This violates the assumption made by the vast majority of statistics and machine learning approaches, that observations are identically and independently distributed. Here, the project developed methods to model dependencies between objects and take them into account when learning predictive models.
The project results have several applications, e.g. in the fields of proteomics (leading to the ERC PoC SNIPER project), fraud detection and mobility.