In the American Banker article When AI suspects money laundering, banks struggle to explain why by Carter Pape, the question arises: can we automate "explainability" in machine learning models? This question inquire into the ongoing debate of black box versus white box ML models. While rules-based models are inherently explainable due to their transparent nature, ML models often operate in a probabilistic realm, introducing classification errors. The "explainable" ML models fall into the category where you can interpret the feature space used for classification, such as color, shape, and other parameters. Conversely, black box models are considered "not explainable" because they obscure the process of feature extraction.
In light of this, when we don't rely on ML at all, decisions become automatically explainable, as we can clearly articulate the underlying rules. When opting for ML models, one has the option to use either classical explainable models or ML models rooted in deep learning. These models, to varying degrees, can provide explanations for why a particular classification was made based on the dataset used.
Furthermore, a valuable approach involves storing the outcomes and the models employed. This enables the revisitation of decisions at a later time, affording the opportunity for corrections if necessary. However, it's important to note that making deep learning black box models entirely explainable remains a challenging task.
Synergizing Traditional Rules and Machine Learning for Enhanced Decision Quality and Outcome Optimization
In the quest for precision and reliability, the fusion of established rules with the capabilities of Machine Learning (ML) has emerged as a formidable strategy. This harmonious integration aims to bolster the quality of decisions and ultimately improve the outcomes, thereby reducing both false positives and false negatives.
A pivotal facet of this approach involves the fine-tuning of decision thresholds within ML algorithms. These thresholds, which play a crucial role in classification and predictive modeling, are adjusted based on insights gained from the performance of predefined rules. This dynamic adaptation leverages the strengths of both rule-based and ML approaches, enhancing the accuracy of decision-making.
To ensure the integrity and accountability of this process, the results of these decisions are documented and stored during the processing. This repository encompasses not only the decision outcomes but also the rule policies and ML models employed in each scenario. This archive not only facilitates performance audits but also serves as a vital resource for future model refinement.
The importance of data quality cannot be overstated in this context. As the need for a robust foundation for ML models grows, there is an ongoing emphasis on curating and expanding high-quality datasets. These datasets, marked by their continual improvement, serve as the lifeblood for training and evolving ML models:
In summary, the strategic fusion of rules and ML, coupled with adaptive threshold adjustments and a comprehensive decision repository, ushers in an era of heightened decision quality and outcomes optimization. By maintaining a strong commitment to data quality and continuous improvement, organizations are poised to navigate the evolving landscape of decision-making with confidence and efficacy.