Stability of palantir results

[dpeerlab]

Stability of palantir results

[dawe]

Apparently increasing the waypoints (and knn) stabilizes the results, is this expected? I'm now set to n_waypoints=adata.shape[0]//3

[ManuSetty]

The waypoints are indeed randomly sampled and increasing the number of waypoints will indeed stabilize the results. Alternatively, setting the random seed using np.random.seed will ensure that waypoint samples are consistent.

[dawe]

Setting the random seed ensures results are always the same, but it doesn't grant their consistency, I may be in some local minima. I'll go with more waypoints at the cost of runtime (not much, <15m for 15k cells, <1h for ~40k cells). Thanks!

[rahulnutron]

Dear Palantir team,
Thanks for the tool!
I am facing the same problem where branch probabilities are not stable. For me branch probabilities are important and ideally should be fixed in all runs for all cells using same parameters. I set the random seed, which doesn't seem to solve it. Do you recommend to increase the waypoints in this case?
Rahul

[ManuSetty]

Hello - you could increase the number of waypoints as a solution. Alternatively, you can save the set of waypoints from one of the runs and use it as input if you rerun Palantir. The waypoints are stored in pr_res.waypoints and the num_waypoints parameter can be set to a set of cells rather than the number.