Section: New Results

Stochastic Decision Trees for Similarity Computation

We have designed a method to compute similarities on unlabeled data using stochastic decision trees [20]. The main idea of Unsupervised Extremely Randomized Trees (UET) is to randomly and iteratively split the data until a stopping criterion is met. Pairwise similarity values are computed based on the co-occurrence of samples in the leaves of each generated tree. We evaluate our method on synthetic and real-world datasets by comparing the mean similarities between samples with the same label and the mean similarities between samples with distinct labels. Empirical studies show that the method effectively gives distinct similarity values between samples belonging to distinct clusters, and gives indiscernible values when there is no cluster structure. We also assessed some interesting properties such as invariance under monotone transformations of variables and robustness to correlated variables and noise. Our experiments show that the algorithm outperforms existing methods in some cases, and can reduce the amount of preprocessing needed with many real-world datasets. We plan to study the application of this “global” pairwise similarity computation to quantify protein structural similarities. Two interesting problems will concern the representation of the protein structure and how to tackle extra constraints such as invariance under rotational and translational transformations.