First Presentation

[wagpa]

• Motivation
• Theme/Goal
• Next Steps

[wagpa]

• different ways to generate embeddings
• high/low-dimensional
• ML mostly high dimensional
• Algo mostly low dimensional
• Differences/Advantages
• Create Embedding Evaluator
• Evaluates if an embedding is good
• How to evaluate? (without Downstream-Task)
• For Embedding and Graph, makes "perfect" Decoder, if the Decoder is "good", then the embedding (and encoder) was "good"
• What makes a good embedding (embedder)?
• What does the Decoder need, to be "good"?

[wagpa]

Structure

Motivation

1. Why are we looking into embeddings, what are they used for? TODO
2. What is an embedding? TODO
3. How do you generate an embedding, what are key features to look for? (Deterministic/ML) TODO

Theme/Goal

1. What are differences/advantages between the two methods?
2. Create an embedding evaluator for evaluating generated embeddings?
3. Create an evaluator framework that evaluates, if a generated embedding is "good".
4. The evaluator takes the original graph (graph with nodes and edges) and an embedding (node positions).
5. It tries to generate a decoder with ML that is capable of predicting graph edges.
6. Idea: if the decoder-setup is generally chosen "good", then it should be able to generate a decoder, that is able to predict edges. If the generated decoder cannot predict edges, then the embedding was chosen poorly. (is it easy to "read" the embedding?)
7. This workflow will be presented in a visual application.

Next Steps

1. Question: How to make a good decoder-setup?
2. Decide on an application structure.