We thank the reviewers for their careful reading and thoughtful comments.  Please find responses to the major comments below.$
Reviewer 4 Detailed Comments, First Question: The data sets are already public. The algorithms are not that significant a modification (for example, changing the probability of choosing different dimensions) so we were not thinking about making a big deal about the code. We can revisit this issue if there is a strong feeling about it.

Second Question: We did not try audio or video data. We did experiment with multi-dimensional data (for example, the dataset studied in the batch setting is 16-dimensional), so multiple dimensions can be handled.  As with many other techniques, with higher dimensions, more data will be required.  One can convert audio/video to a stream of sparse features and then feed the resulting sparse stream to the algorithm. However, for these domains, the effort required to design the best feature streams will likely be significant and may require innovation.  We did not want to focus on any specific domains in this article.

Reviewer 5 Detailed Comments, First comment: The reviewer is correct, indeed scaling is an issue.  In a batch setting, we can normalize each dimension by its range in that dimension.  In a streaming setting, if there are no deletions, it should be feasible to re-normalize this scale and recompute parts of the tree (as opposed to changing a small number of node/parent/sibling relation, as currently), but more research is needed. In the presence of deletions in a streaming setting, other ideas will be needed.

Second Comment: We will include a graph of the RRCF on a Gaussian in the appendix.

Third Comment: We are looking into running OCSVM and SVDD on the data sets considered in this paper.  We probably will not have space to have detailed and informative discussions about other datasets.