I have chosen to do project2 in the context of a lab. In our case, we are trying to predict parent-child relations between yeast cells.
The dataset are raw microscopy movies, which we can segment in order to gather information about position, shape, eccentricity, etc.
We have manually labeled about 300 division events (ie : we have a graph with 300 vertices, but the graphs features change in time (for instance, if a cell moves, the feature encoding cell-to-cell distance changes, even if no new vertices are created)
The main idea right now is to use the graph representation of the cells (each cell is a node (features : eccentricity, age, etc.), and each vertex a parent-child relation (features : distance between membranes, relative position, etc.)) in order to make an edge prediction when a new unconnected node (cell) appears.
I've started to learn about GNNs, but am still confused as to how
time-dependent information can be taken into account
one can make predictions on a new edge appearing (for now I've only seen examples of graph classifications, and node feature prediction)
Asking around, I've had people suggesting that we use GATs. Is this a good idea ?
Project 2 - Resources for learning about GNNs
Hello,
I have chosen to do project2 in the context of a lab. In our case, we are trying to predict parent-child relations between yeast cells.
The dataset are raw microscopy movies, which we can segment in order to gather information about position, shape, eccentricity, etc.
We have manually labeled about 300 division events (ie : we have a graph with 300 vertices, but the graphs features change in time (for instance, if a cell moves, the feature encoding cell-to-cell distance changes, even if no new vertices are created)
The main idea right now is to use the graph representation of the cells (each cell is a node (features : eccentricity, age, etc.), and each vertex a parent-child relation (features : distance between membranes, relative position, etc.)) in order to make an edge prediction when a new unconnected node (cell) appears.
I've started to learn about GNNs, but am still confused as to how
Asking around, I've had people suggesting that we use GATs. Is this a good idea ?
Do you have resources for this?
Cheers
1
Add comment