Predict function: the paper says that rules are ordered by F1-measure and the consequent of the first firing rule in the sorted list is used to label the test instance. Do we understand it correctly?
Tuning lambda parameters: The technical report accompanying your KDD paper says that during tuning of lambda functions user-set bounds on interpretability metrics were observed. Could you hint how this "tabu search" was implemented within coordinate ascent?
Scalability: Even after considerable optimization and tuning, we have trouble running the algorithm on small/medium size datasets when the number of input rules exceeds about 50, as seen in Figure 7A of our paper (http://ceur-ws.org/Vol-2438/paper8.pdf). We suspect one of the reasons are the expensive partial objectives f_3 and f_4, which compute overlap in coverage between all pairs of rules. Can you maybe give us some optimization hints?
The main things we are unsure of are:
The reimplementation can be found at https://github.com/jirifilip/pyIDS.