Memorizing Transactional Databases Compressively in Deep Neural Networks for Efficient Itemset Support Queries

摘要

Can a deep neural network memorize a database? Though deep artificial neural networks are remarkable for large memory capacity that makes fitting any dataset possible, memorizing a database is a novel learning task unlike other popular tasks which intrinsically model mappings rather than "memorize" information internally. We give a positive answer to the question by showing that through training with maximal/minimal and frequent/infrequent patterns of a transactional data-base, a dynamically constructed deep net can support random itemset support queries with relatively high precision in regard to data compression ratio. Due to the compressive memorization, the amount of transactions in the database becomes irrelevant to the query time cost in our efficient method. We further discuss the potential interpretation of learnt database representation by analyzing corresponding statistical features of the database and activation patterns of the neural network.