I am currently working with a sparse matrix dataset and employing the Isolation Forest algorithm for outlier detection. Given the nature of my data (high dimensionality and sparsity), I am contemplating whether increasing the subsampling size for the Isolation Forest algorithm could yield better results.
- The default subsampling size for Isolation Forest is typically set to 256, but with sparse data, I wonder if this might not capture the nuances in the dataset effectively.
- Does it make sense to increase the subsampling size for Isolation Forest when dealing with sparse matrix data?
- Would a larger subsample provide a more accuraterepresentation of the data's structure, thereby improving outlierdetection accuracy? Is there a risk of overfitting or significant increases in computational time?
- Are there any established best practices or guidelines in the literature or industry regarding the subsampling size for Isolation Forest, particularly in the context of sparse data?
Thank you in advance for your time.
