Sophie Pichler
Supervisor(s): Dr. Astrid Berg
TU Wien
Abstract: Many deep learning applications are based on graph data in order to explore relationships or to analyze structures. Labeling this data is expensive and often requires expert knowledge. For the application of graph clustering to neuron data, the SOTA method GraphDINO generates self-supervised graph embeddings combined with the downstream task of clustering these embeddings. We observe on a particularly challenging neuron dataset that this method does not lead to satisfying clustering results. Therefore we use the graph embeddings generated by GraphDINO as an initial starting point to improve the network and to guide the network training. To achieve this, we developed the visual analytics framework NetDive. The user can analyze the graph embeddings and label single neurons that are falsely clustered. This annotation information is then used to train a semi-supervised model. To this end, we developed a network architecture, titled GraphPAWS, that assembles components of GraphDINO and of the semi-supervised network architecture PAWS. The model training can be started from within the visual analytics application NetDive and the resulting graph embeddings are available in NetDive as soon as the retraining is completed. We demonstrate how we iteratively improve the model performance using NetDive and GraphPAWS and evaluate our model against the self-supervised SOTA for our dataset.
Keywords: Human-Computer Interaction, Scientific Visualization
Full text: Year: 2024
