Outstanding Presentation Paper Award at the Korea Software Congress 2023 (KSC 2023)

- Big Data Processing and DB Laboratory -

The paper 'A Graph Autoencoder-based Data Embedding Technique for Clustering Mixed-Type Data' by Lee Jung-min, graduate student in the Department of Computer Science & Engineering (supervisor: Jung Sung-won), received the Outstanding Presentation Paper Award at the Korea Software Congress 2023 (KSC 2023), held over three days from 20 to 22 December.

▲ From left: Lee Jung-min (master's), Jung Sung-won (supervisor)

▲ Certificate of the Outstanding Presentation Paper Award, Korea Software Congress 2023 (KSC 2023)

The attributes that make up the objects in a dataset fall broadly into numerical attributes, categorical attributes and mixed attributes. Mixed attributes refers to objects that have numerical and categorical attributes at the same time. For mixed-type data with two kinds of attributes that require different distance measures, the usual approaches are either to convert categorical attributes into numerical attributes and compute similarity, or to compute an integrated distance from the similarity of each attribute type and cluster on that basis.

Both methods, however, have drawbacks: loss of the original features during the conversion of the original attribute types, and an increase in the number of hyperparameters needed to build the integrated distance. Taking these problems as its starting point, the paper presents a method for defining similarity between objects in mixed-type datasets so that clustering can be performed by building a graph from the connections between objects in the dataset, passing that graph through a graph autoencoder (GAE), and using the resulting node embeddings as object coordinates for Euclidean similarity and distance measurement.