multiple runs with same random seeds?

[culurciello]

HI guys, I am trying to run t-SNE over 2 sets of data in 2 separate runs.
If I do that, does the embedding space remain the same? Or do I get 2 different projections?
How can I keep the projection the same if I want to run a second set of data that I need consistent with the 1st one?

[lvdmaaten]

If you do two separate runs on two different data sets, there is no reason to expect that both embeddings will be somehow aligned. To obtain such alignment, one should presumably learn both embeddings jointly and add a loss term that penalizes non-aligned solutions (perhaps by measuring the divergence between density estimates on both embeddings). This is not currently implemented.

[culurciello]

what do you recommend me to do to get the alignment? I need to be able to project on the same space with 2 runs of data.

[lvdmaaten]

You would have to do the fitting on all data sets simultaneously, while adding some kind of regularizer to the loss that is lower when the maps are aligned. How exactly to implement this regularizer depends on the problem: it could be density-based if no supervision is available, correspondence-based if you know (some) correspondences between data in the data sets, or label-based if you have class labels for both data sets (over the same set of classes). Either way, this would require substantial changes to the code.

[culurciello]

thanks for the input!