Representation Independent Data Analytics

Over the last years, users’ information needs over structured data expanded from seeking exact answers to precise queries, to performing database analytics tasks, such as finding entities or patterns similar to a given entity or pattern, discovering interesting patterns, or predicting novel relations and concepts. As part of its response, the research community proposed a multitude of algorithms to solve exploration and analytics problems over structured data. Since the properties of interesting and desirable answers are no longer precisely defined in the query, these algorithms use intuitively appealing heuristics to choose, from among all possible answers, those that are most likely to satisfy the user’s information need. Unfortunately, such heuristics typically depend on the precise choice of representation of the underlying database. Generally, there is no canonical representation for a particular set of content and people often represent the same information in different representations. Thus, in order to effectively use database analytics algorithms, users generally have to restructure their databases to some proper representation or change hyper-parameter settings. As a result, today’s database exploration and analytics algorithms and tools are usable only by highly trained data scientists who can predict which algorithms are likely to be effective for particular representations of the underlying database, and under which settings. To cope with the structural heterogeneity in large-scale data, we propose a novel approach to database analytics that considers representation as a first-class citizen. We introduce the concept of representation independence as the ability to deliver the same answers regardless of the choices of structure for organizing the data. Because representation independence may not always be achievable, we also consider representation scalability, the ability to return similar answers over different structures of data. We discuss our work on providing ordinary users with an arsenal of effective database analytics methods that are robust across multiple representations of the same information. We present our ongoing work on creating representation independent analytics