I have a large amount of data which has been reduced to two dimensions using t-SNE. Additional data points keep arriving, which I would like two-dimensional embeddings for, but this cannot be achieved without re-running the t-SNE algorithm which is not feasible for every new data point.
Are there any ways to approximate the t-SNE embedding for new data points? I've tried taking $k$-nearest neighbours in the full data space, getting their corresponding t-SNE embeddings, and averaging them to produce an embedding for the new data point. This works okay, but I'm wondering if anyone knows more rigorous methods to approximate this.