Transparency in Decision-Making: Advantages of a Rule-Based Approach — Part 2/2
by Tobias Schaefer, Zachary Hervieux-Moore, Norvard Khanaghyan, and Dmitry Lesnik
The Clarity of Rules
In the previous post we discussed the importance of transparency for automated decision-making. We also presented limitations of other approaches like LIME that add local explainability to black-box algorithms. In this post, we present a different way to think about explainability globally using rule-based decision algorithms, like Stratyfy’s Probabilistic Rule Engine (PRE). Let’s revisit the data from the first post, presenting the outcomes of a two-dimensional sigmoid classifier such that the positive outcomes cluster in quadrant I and III.
What happens if we use the Stratyfy’s Probabilistic Rule Engine (PRE) for the above data? The PRE will analyze the data and mine 4 simple rules that describe the data globally. The 4 rules are:
Rule R_1: “IF (x > 0 AND y > 0 ) THEN outcome=1 (conf 0.9)”
Rule R_2: “IF (x < 0 AND y < 0 ) THEN outcome=1 (conf 0.9)”
Rule R_3: “IF (x < 0 AND y > 0 ) THEN outcome=0 (conf 0.9)”
Rule R_4: “IF (x > 0 AND y < 0 ) THEN outcome=0 (conf 0.9)”
Clearly, this collection of rules is a high-level description of the data, similar to how we as humans would intuitively describe the data: One in quadrant I and III and zero in quadrant II and IV. The number (0.9) measures the confidence level of each rule in the decision process which also indicates how much the rule contributes to the overall decision. In this case, the rule engine shows that we can be 90% confident about the statement expressed in this rule.
Transparency in Credit Risk
If we look, as an example, at credit risk evaluations, transparency is one of the most important aspects of the decision process. Banks have to report to regulators and show that all customers are treated equally using an interpretable decision system. And customers might ask the bank for their reasoning, in particular if they are denied a line of credit. In addition, they might also wonder what to change in order to become creditworthy.
As a direct consequence from these requirements imposed by the credit risk market, many institutions are still using deterministic rules in order to decide whether to grant credit to a particular customer. Commonly, these rules are formulated in terms of a scorecard, or simply a set of rigid knock-out rules. The customer’s data serves as an input, and, after applying the rules, the bank determines whether the client is creditworthy or not. If the bank decides to change the rules, this change will affect all customers equally. The rules in the scorecard are human-readable and transparent. If a regulator asks the bank to provide details about their decision-making process, the bank can simply respond by showing these rules. If a customer is denied credit because of a particular rule, the bank can provide this information to the customer and explain which rule or rules triggered the credit denial. Clearly, in terms of transparency, a decision-making algorithm comprised of deterministic rules is a great decision engine.
What is the drawback of using a scorecard for making risk decisions? Let us look at two main disadvantages. First, and this is well-known in the credit industry, rule-based systems are not very accurate. To limit the financial risks, banks put the threshold bar very high. As a result, many people are denied credit who are, in fact, good, low-risk candidates. In financial terms, this constitutes an enormous lost opportunity which is estimated to be of the order of billions of dollars in the United States alone. A second drawback of rule-based classifiers is more subtle, but equally painful, namely scaling. When using a system of deterministic logical rules, if you need to add another rule, you must first check if the rule that is being added does not contradict any of the existing rules. For a large number of rules, this procedure becomes complicated, tedious, and costly.
With Stratyfy’s solutions, a simple first step for any decision process that currently uses rules is to run these expert rules on the probabilistic engine. This simple step, already, increases accuracy significantly. While this might seem surprising at first, there is no magic at all. This process simply imitates the human decision process consisting of giving different factors different weights when making a decision. On the other hand, what is surprising, is that using rule mining (step IV. in the workflow below) to supplement the engine with additional rules, can create models that are even superior in terms of accuracy and can even achieve levels of accuracy that are on par with ANNs.
If we had a wish list regarding transparency of decision-making, what would be at the very top?
- Each individual factor that contributed to the decision needs to be human-interpretable.
- We need to be able to assess in which way each factor contributed to the decision. In particular in the case of a yes/no decision, we need to know whether a factor was in favor or against, and we need to formulate a measure for the importance of the factor in the decision process.
- We would probably give preference to a simpler model of decision making than to a very complex model.
This is the type of decision making Stratyfy’s probabilistic rules deliver.
From Data to Rules
Transparency is one major advantage of a probabilistic rules. The basic idea behind probabilistic rules is to mimic the high-level human decision process. A particular decision is made after contemplating factors which impact the decision with various degrees of importance. Stratyfy is a pioneer in building solutions around this probabilistic rules approach.
The rules used to build any model can be specified directly by experts representing their knowledge of the field or found from data through a rule mining algorithm. Combining both approaches — expert rules from humans and mined rules from data — is possible as well. To give a more concrete example of a rule in the context of credit risk, consider the following statement:
Rule R_1: “IF (DTI < 20 AND IntRate < 2.5) THEN (Loan is safe)” (conf. 0.74)
Here, the variables “DTI” and “IntRate” signify “debt-to-income ratio” and “interest rate”. The rule expresses that the two variables need to be below a certain threshold simultaneously to indicate that the loan is safe and that the confidence level of the rule is 74%.
What is the algorithm behind Stratyfy’s Probabilistic Rule Engine (PRE) that is used to build models? The main idea is to first create a set of rules and then to use reference data to determine the optimal weight for each rule. The typical workflow to work with Stratyfy’s Rule Engine is
I. Specify a set of rules (“expert knowledge”) — optional
II. Provide reference data as training set — optional
III. Calibrate the Rule Engine and evaluate performance
IV. Use Rule Mining to find additional rules
V. Calibrate Rule Engine on the new set of rules and evaluate performance
VI. Use calibrated model for default prediction on new data
As a result, we obtain a decision model consisting of rules and their weights. This resulting model is entirely transparent and interpretable.
Rules are Human-Interpretable
Let us look again at rule-based models like those developed by Stratyfy. A common way to think of factors in a decision process is to look at a rule. Most of our daily life is organized by the presence of rules (and laws, which can be thought of as more formalized rules). These rules are supplemented by rules formulated by ourselves that constantly help us to make decisions. A great many of these rules are expressed using the structure of logical formulas with a set of variables as input. Often, to trigger a decision, rules assess if the input parameters lie below or above certain thresholds or if they are contained in particular intervals. A simple example would be the rule “If my body temperature is above a certain threshold, this signifies fever and I need to rest”.
If we ask someone why a certain decision was made, a response in terms of rules, together with an indication of their importance (or weight) in the decision process, is a satisfactory answer. Note that this form of explanation can serve as the starting point of a discussion, for example, about the applicability of a rule or its impact. Even if we do not agree with the final decision, it is comforting to know that at least we understand how the decision was made.
CONTACT INFO: If you’d like to learn more about how Stratyfy can help you develop transparent rule-based models, please reach out to me or info@stratyfy.com to see how we can help.
