I am currently looking into using gdsl for a node classification task in which context i have three questions.
1. Is it possible to use relationships instead of properties to train a model ?
2. Is it only possible to use numbers to train a model or can strings be used ?
(regarding those two questions: I would have to use keywords which are each represented as a separate node which are connected to another node based on a relationship. If it is not possible to use relationships and strings a would probably have to create properties for them and represent them as a number )
3. Is multi-class classification possible ? Because a node could be classified as more than one class so more than one relationship would have to be created. (Example: Person could be overweight and sick in any other way at the same time)
If you use strings you need to categorize them e.g. via one-hot-encoding into numbers.
How would you want to use relationships? Tomaz had some examples for graph feature extraction that turn relationship data into attributes on the projected graph
Afaik you can use relationship-weight properties for your training.
Not sure about multi-classification
Thanks for your answer, I looked into One Hot Encoding and it could be helpful but an alpha state is probably not the best solution for production later on. But as mentioned further down in the comments there are also node embeddings.
Towards the relationships, basically I have keywords and one idea was to represent them each as an node so there would be relationships between them and the node the keywords belong to. Primarily based on those keywords i have to do a node classification.
But I am slowly coming to the conclusion that it would be better to use python for that matter (or Java if possible)