means being highly related to my personal research interest.
NeurIPS 2019-Breaking the Glass Ceiling for Embedding-Based Classifiers for Large Output Spaces
NOTE: Not available yet.
AISTATS 2019-Stochastic Negative Mining for Learning with Large Output Spaces
NOTE: In this paper we specifically consider retrieval tasks where the objective is to output the k most relevant classes for an input out of a very large number of possible classes. Training and test examples consist of pairs (x, y) where x represents the input and y is one class that is relevant for it. This setting is common in retrieval tasks: for example, x might represent a search query, and y a document that a user clicked on in response to the search query. The goal is to learn a set-valued classifier that for any input x outputs a set of k classes that it believes are most relevant for x, and the model is evaluated based on whether the class y is captured in these k classes.
To this end, we first define a family of surrogate losses and show that they are calibrated and convex under certain conditions on the loss parameters and data distribution, thereby establishing a statistical and analytical basis for using these losses. Furthermore, we identify a particularly intuitive class of loss functions in the aforementioned family and show that they are amenable to practical implementation in the large output space setting (i.e. computation is possible without evaluating scores of all labels) by developing a technique called Stochastic Negative Mining. We also provide generalization error bounds for the losses in the family. Finally, we conduct experiments which demonstrate that Stochastic Negative Mining yields benefits over commonly used negative sampling approaches.