Search results for “SHAP”

With a soaring number of scientific publications and rapid emergence of new directions and approaches, the scientific community faces the task of timely identification of trends. By a trend, we mean a semantically homogeneous topic characterized by a steady lexical kernel and a sharp, often exponential increase in the number of publications [1]. Examples of trends in machine learning are “LSTM,” “deep learning,” “word2vec,” “BERT,” and “fake news detection.” For real-time detection of trend topics from a stream of scientific publications, we use incremental methods of probabilistic topic modeling. An ARTM-based approach to early trend detection has been shown to outperform popular classical and neural network approaches to this task. A dataset of 91 trends for performance evaluation has been manually collected and made available for public use.

Accurate prediction of the generalization ability of a learning algorithm is an important problem in computational learning theory. The classical Vapnik-Chervonenkis (VC) generalization bounds are too general and therefore overestimate the expected error. Recently obtained data-dependent bounds are still overestimated. To find out why the bounds are loose, we reject the uniform convergence principle and apply a purely combinatorial approach that is free of any probabilistic assumptions, makes no approximations, and provides an empirical control of looseness. We introduce new data-dependent complexity measures: a local shatter coefficient and a nonscalar local shatter profile , which can give much tighter bounds than the classical VC shatter coefficient . An experiment on real datasets shows that the effective local measures may take very small values; thus, the effective local VC dimension takes values in [0, 1] and therefore is not related to the dimension of the space.