Fast Algorithms for Learning with Long N-grams via Suffix Tree Based Matrix Multiplication

This matrix format is inefficient when storing frequency data since we know all entries in x are non-negative integers. Moreover, the number of bits needed to store each index in the jc array is dlog2 nze which can be significantly larger than dlog2 Ue where U is the largest number of non-zero elements in any column. Our modified CSC format simply replaces the jc array with an integer array of size N that stores the number of non-zero elements in each column and it replaces x by an integer array of frequency counts. This modifications can lead to substantial savings when appropriate.