Skip to Main content Skip to Navigation

Mining useful patterns in attributed graphs

Anes Bendimerad 1, 2 
Abstract : We address the problem of pattern discovery in vertex-attributed graphs. This kind of structure consists of a graph augmented with attributes associated to vertices. Vertex-attributed graphs provide a powerful abstraction that can be used to represent many datasets in an intuitive manner. Mining these graphs can be very useful for many applications. When mining vertex-attributed graphs, the principled integration of both graph and attribute data poses two important challenges. First, we need to define a pattern syntax that is intuitive and lends itself to efficient search. Considering that a pattern is generally defined over a subgraph, a pattern can be often huge in terms of vertices, which makes it difficult to grasp. Thus, the assimilation cost of a pattern is an important question that needs to be addressed. The second challenge is the formalization of the pattern interestingness. A pattern is generally relevant if it depicts some local properties that are somehow exceptional, otherwise, it will be already expected from the overall properties of the graph. Furthermore, the interestingness of patterns is subjective in practice, i.e., it significantly depends on the final user, her background knowledge and her preferences. Another common problem related to the interestingness of patterns is the redundancy issue in the result set. In other terms, a data mining approach may return a set of patterns that give redundant information, because these patterns cover very overlapping parts of vertices and attributes. We address these challenges for the problem of mining attributed graphs. We first introduce the task of discovering exceptional attributed subgraphs, which is rooted in Subgroup Discovery. The goal is to identify connected subgraphs whose vertices share characteristics that distinguish them from the rest of the graph. Then, we propose methods that aim to take into account the user and the domain knowledge when assessing the interestingness of patterns. We design a method that makes it possible to incorporate user’s background knowledge and pattern’s assimilation cost. This method is able to identify patterns that are both informative and easy to interpret. Furthermore, we propose another graph mining approach that integrates user’s preferences. This method exploits an interactive process with the user to bias the pattern interestingness. It has been defined for the task of geo-located event detection in social media. Then, we design an approach that is able to incorporate hierarchical attribute dependencies into the pattern interestingness, which allows to avoid redundancy related to this kind of semantic relations between attributes. In other terms, when the attributes are organized as a hierarchy, this method is able to account for the inference that the user would make about some attribute values when she is informed about values of other attributes. Finally, we conclude this thesis by discussing some research perspectives.
Document type :
Complete list of metadata

Cited literature [241 references]  Display  Hide  Download
Contributor : ABES STAR :  Contact
Submitted on : Tuesday, February 25, 2020 - 3:40:28 PM
Last modification on : Tuesday, June 1, 2021 - 2:08:09 PM
Long-term archiving on: : Tuesday, May 26, 2020 - 5:43:03 PM


Version validated by the jury (STAR)


  • HAL Id : tel-02490868, version 1


Anes Bendimerad. Mining useful patterns in attributed graphs. Other [cs.OH]. Université de Lyon, 2019. English. ⟨NNT : 2019LYSEI058⟩. ⟨tel-02490868⟩



Record views


Files downloads