Machine learning project with BERT GSDMM, and MinCut Pooling
Budget: €250 – €750 EUR
I have a hybrid model that takes embedding from Bert and a hierarchical GSDMM.
GSDMM is a tree in which the items are saved in its nodes using GSDMM as a clustering method at each level, and I then have the embeddings from all of its items.
The hybrid model was then applied to different ratios (0.0 to 1.0) on both embeddings to determine which ratio performed the best.
What I want now:
Extend this excited hierarchical, which uses GSDMM, so that it uses GSDMM or MinCut Pooling as a clustering method.
How to decide which one to use each time:
The hybrid model that compares the performance between GSDMM (not hierarchical GSDMM anymore) and BERT would decide which clustering method to use (best ration is toward Bert then using the BERT embedding with MinCut Pooling to cluster it; best ration is toward GSDMM then using the GSDMM embedding with GSDMM to cluster it).
End Result: use the existing tree with all its functionalities that are right now integrated with GSDMM to make it integrated with the chosen clustering method (GSDMM or MinCut) from the hybrid model. Therefore, the hybrid model would exist in each node in the hierarchical tree and decide which clustering method to use and then cluster the data, passing it to the next nodes where the next hybrid models receive it.
use case I have: clustering from short texts
GSDMM is a tree in which the items are saved in its nodes using GSDMM as a clustering method at each level, and I then have the embeddings from all of its items.
The hybrid model was then applied to different ratios (0.0 to 1.0) on both embeddings to determine which ratio performed the best.
What I want now:
Extend this excited hierarchical, which uses GSDMM, so that it uses GSDMM or MinCut Pooling as a clustering method.
How to decide which one to use each time:
The hybrid model that compares the performance between GSDMM (not hierarchical GSDMM anymore) and BERT would decide which clustering method to use (best ration is toward Bert then using the BERT embedding with MinCut Pooling to cluster it; best ration is toward GSDMM then using the GSDMM embedding with GSDMM to cluster it).
End Result: use the existing tree with all its functionalities that are right now integrated with GSDMM to make it integrated with the chosen clustering method (GSDMM or MinCut) from the hybrid model. Therefore, the hybrid model would exist in each node in the hierarchical tree and decide which clustering method to use and then cluster the data, passing it to the next nodes where the next hybrid models receive it.
use case I have: clustering from short texts