Differences between paper and code

[nguyenhungquang]

Hi,
While reading your code, I realized some slight differences between your algorithm in the paper and what was implemented here, and found it really difficult to understand, could you help me to clarify?

• In the paper, barcodes at different values of a factor are combined by Wasserstein barycenter. However, function barymean here computes between different runs of RLT algorithm at a specific value (which are totally different because we are using witness complex). And in this repo barcodes at different values of a factor are aggregated by simple mean operator (as shown in here).
• The formulation of (unsupervised) evaluation metric is $\mu = \rho_c - \rho_{\c}$, in which $\rho_c$ are summation of W distance of mean barcodes of different factor in a same cluster, and $\rho_{\c}$ are those at different clusters minus $\rho_c$. When I look at bicluster_score function, it seems to be $\rho_{\c}$ only, which makes me really confused.