The Confusion Matrix

To clarify this point, let’s explore the idea of a "Confusion Matrix." This concept is well-known among Data Scientists who have created predictive models for classification tasks. While some may find it new, the methodology and its implications for human and business interactions serve as an enlightening case study on the terminology surrounding accuracy in machine learning. It acts as a valuable visual aid for comprehending the nuances and trade-offs involved.

When discussing total accuracy, we refer to the number of correct predictions (the sum of the green boxes) compared to all predictions (the sum of the four boxes). For instance, a statement like “Our pregnancy test is 99% accurate” pertains to the accuracy of all test predictions, regardless of whether they indicate a positive or negative result.

The complexity arises when we attempt to determine where the inaccurate predictions fall among the remaining categories represented by the red boxes. For infrequent events, one might achieve a high accuracy merely by predicting that the event never occurs (which requires no model). However, the stakes associated with inaccuracy can vary significantly depending on different models and use cases.

In simpler terms, a model with lower accuracy might be deliberately designed this way to minimize the frequency of misclassifications in one direction or another, necessitating a compromise on overall accuracy.

Consider these critical questions:

• Is it riskier to mistakenly predict that someone is pregnant?
• Is it more dangerous to falsely diagnose someone as cancer-free?
• Is it worse to classify a statement as hate speech and remove it or allow it to remain?

While some scenarios have clear answers, others can lead to disagreement, illustrating the complexity involved in decision-making around inaccurate predictions. What one might consider a flaw, another could view as a feature.

Chatbots and Large Language Models (LLMs)

Shifting our focus from straightforward classification models, there is considerable discourse surrounding "hallucinations" in outputs from Large Language Models (LLMs). For some users, these inaccuracies have become significant enough to abandon the tools, fearing they might not recognize these errors. However, some experts argue that such outputs are an inherent aspect of AI design. An article in Scientific American emphasizes that chatbots are programmed to respond, and even if those responses are incorrect, they are fulfilling their intended function. Unfortunately, they often deliver incorrect answers with the same confidence as accurate ones, much like the humans they emulate.

The swift rise of ChatGPT has brought these LLM discussions into the public sphere, offering a platform for broader conversations about the realities of accuracy in predictions. This exposure allows for a dialogue regarding their advantages and disadvantages, which was not as prevalent for other types of models.

Trade-offs in Model Accuracy

The fundamental aspect to grasp when creating, deploying, or utilizing AI outputs is understanding the primary objective. By comprehending these goals, we can enhance our capacity to achieve them responsibly without relying solely on technology. Moreover, understanding the rationale behind a model's use helps users engage with the outputs more responsibly.

Every AI instance is underpinned by a data optimization challenge. Depending on the data characteristics, it’s possible to develop remarkably accurate models that align with specific optimization goals. A prevalent example is the automated advertising technology employed by platforms like Meta and Google. When configuring a campaign, you specify a desired outcome. If you request clicks, that’s what you'll receive, even if those clicks don't lead to valuable results for your business.

Recommender systems are another common model that many interact with daily. From Amazon’s "customers like you" suggestions to TikTok’s content recommendations and Netflix's homepage, these systems cater to what the algorithms believe users want. However, the question remains: is this truly what we desire, or is it merely what aligns with the company's objectives? For instance, Amazon aims for purchases—preferably high-margin ones—while TikTok seeks to maximize viewer engagement for ad revenue, and Netflix wants to quickly connect users with enjoyable content to encourage binge-watching.

Returning to the Confusion Matrix

When Data Scientists or Machine Learning engineers assess the Confusion Matrix of various models, they must keep the model's goal at the forefront. What are we aiming to accomplish, and what does success look like?

As previously noted, many find accuracy to be an intuitive concept, which can lead to misconceptions. For example, when encountering a model with less than 50% accuracy, it’s common to hear, “That’s worse than flipping a coin.” While superficially accurate, this perspective can be misleading. If, due to event rarity, our baseline accuracy is 1% (or less) by random guessing, achieving 10% accuracy represents a significant improvement.

We must evaluate accuracy in relative terms, considering the added value versus having no model (or the previous model). Additionally, we should decide where we prefer our erroneous predictions to occur, factoring in associated risks and costs. This involves determining whether a False Positive is preferable to a False Negative.

The True Positive rate, or model sensitivity, represents the capacity to minimize False Negatives (Type 2 errors) while maximizing the likelihood of detecting relevant predictions. Conversely, the True Negative rate, or specificity, aims to reduce False Positives (Type 1 errors) and enhance the precision of our predictions. A more specific model is less likely to incorrectly identify non-existent predictions but may miss actual relevant instances.

So, What’s the Bottom Line?

Whether this process is executed intentionally by a Data Scientist, overlooked by an inexperienced practitioner, or automated by AI, it underscores the underlying dynamics of accuracy refinement. AI cannot serve all purposes for everyone, so it ultimately returns to its design and the definition of success.

A single statistic regarding an AI's accuracy does not provide the complete picture. Context is crucial, encompassing not only user perception but also the designer's perspective. Higher accuracy does not inherently equate to better performance without an understanding of the decision-making process behind it.

Understanding how our Data and AI products align with our business strategies is essential for unlocking their true value. If you or your leadership team need assistance in this area, consider exploring my offerings at kate-minogue.com. Through a unique focus on People, Strategy, and Data, I provide a range of consulting and advisory services to help you address Business, Data, and Execution challenges effectively. Follow me here or on LinkedIn for more insights.