Representation and Evolution of Knowledge Structures to Detect Anomalies in Financial Statements

Abstract

Deep learning, has delivered a variety of practical uses in the past decade. It has revolutionized customer experience and machine translation. It has made language recognition, autonomous vehicles and computer vision a reality. A multitude of other AI applications are common now. With Deep Learning we gain insights about hidden correlations. We extract features and distinguish categories. But we lack transparency of reasoning behind these conclusions. Most importantly, there is the absence of common sense. Deep learning models might be the best at perceiving patterns. Yet they cannot comprehend what the patterns mean. And they lack the ability to model their behaviors and reason about them.We present a new approach to augment Deep Learning using model based Deep Reasoning and its application to address fraud detection using financial statements. Recent theoretical models of computing structures with cognizing agents go beyond neural networks to provide models of observations, abstractions and generalizations from experience and create time dependent evolution and history to provide reasoning and predictive. We use Knowledge Structures defined therein to represent relevant domain knowledge. In this case, in a company’s financial statements. We analyze their history to detect potential fraud based on specific rules and observations. We use information from governance and compliance rules and experience of past violations. We analyze SEC 10-K statements using Deep Learning and model based Deep Reasoning. We use the Knowledge Structures to identify red flags and anomalies.