Not all sources are created equal: Explicit vs implicit feedback in Machine Learning
The pros and cons of explicit vs. implicit feedback in machine learning has become an important topic in recent times; yet, with many factors at play, it can still seem difficult to understand which approach is preferable for a given context.
To help you gain that understanding, we’ll explore the key differences between the two techniques, as well as the cost, transparency, and integrity related to both.
Increased accuracy, at a lower cost?
The cost of explicit feedback is a core difference between the two techniques. When humans provide explicit feedback (e.g. image labeling), they need to set aside time to provide the feedback and make sure it's accurate. This can be time-consuming and costly, especially for large-scale applications.
In addition, humans can get "survey fatigue" when they're asked to constantly provide feedback, and this can lead to lazy, inaccurate or incomplete data. "Garbage in, garbage out." In the machine learning world, if the data is low-quality, the model will be low-quality too.
For example, imagine a company asking customers to rate their satisfaction with a service on a 1-10 NPS scale after every single interaction. Over time, customers may get tired of this and start selecting random numbers or reflexively click the first option to get end the process as quickly as they can.
With explicit feedback, a clear advantage is that the model can easily use the data to update itself. On the flipside, humans can get tired of providing so much feedback, and will not reliably provide accurate or truthful feedback.
Implicit feedback, on the other hand, can be collected automatically as humans go about their normal activities. This makes it cheaper and easier to collect, with the clear advantage that the process is "invisible" to the humans, so they don't have to even think about giving feedback. That said, an application based on implicit feedback can occasionally be challenged to capture the full picture of an interaction (e.g. was the recommendation followed?) and interpret the relevance of its observations. For example, how should an AI adapt to a user skipping a song on a music service?
Data Privacy and Transparency Concerns
Transparency is a desirable characteristic in mass-market consumer applications. With explicit techniques, users providing feedback know exactly how their inputs are used, and how it will impact the system.
For example, if you're asked to rate a service on a scale of 1 to 5, you understand that your rating will go into an average score that the company will use to measure customer satisfaction.
This kind of transparency can be beneficial because it gives humans control over the system and helps them feel like their feedback is valued.
When it comes to implicit feedback, achieving transparency becomes more complex, especially in consumer (non-expert) applications. Feedback providers may not know that their actions are being used to train a machine learning system.
For example: imagine a company is tracking the time that customers spend on their website, without telling them this information is being used for training. The customers are still "providing feedback" through their actions, but they don't have any transparency into the process. This can be problematic because it can lead to privacy concerns, (Microsoft has a helpful hax toolkit that applies here, especially guidelines 13 and 16).
That said, with implicit feedback from human experts, the system functions in a way that’s more akin to an apprenticeship or a mentoring relationship: the experts teach the system through their actions, and the system learns and improves based on the experts’ knowledge and behaviors. Beyond the more complete and complex feedback that can be gathered, it entirely avoids the cases of "survey fatigue" to which explicit feedback is prone.
In both types of feedback, trust is therefore built between humans and the system; humans feel like they some have control over the system and mostly understand how their feedback is used. However, the nature and depth of human influence and understanding deeply varies according to the chosen approach.
With great power comes great… blind spots?
When human feedback is involved, one always runs the risk of reflecting groupthink or otherwise incorporating biases in the system, as the humans providing feedback often have similar backgrounds and perspectives.
Even beyond unconscious biases, implicit feedback can lead to much more objective data; for example, it's much easier to collect data on the amount of time someone spends on a specific task than to ask them to manually file the time spent.
Yet, an even bigger advantage of gathering implicit feedback when it comes to human expertise (most notably in non-consumer contexts) is that it’s able to capture much higher quality training data, as human experts have often internalized their knowledge beyond their own ability to explicitly and fully transfer it to neophytes.
For example, if you were to train an ML system to identify diseases based on medical images, the first step would be to seek feedback from experts, such as experienced radiologists. As the radiologists interact with the system, it would rapidly get better and better from leveraging their expertise. With outcomes such as more accurate diagnoses and better patient care, these same radiologists would immediately grasp the value of the system and deepen their engagement with it, which would naturally lead to a virtuous cycle of continual improvement.
Gaming the system
The possibility, however, that humans can provide unhelpful feedback in a way that "games" the system in a way remains a challenge for both methods.
In an explicit feedback system, users might optimize for a score in a way that doesn't actually improve the AI's performance in the real world - it could in fact be detrimental to its abilities. Indeed, humans may often be tempted to optimize for something that requires less effort/thought, but doesn't actually result in the AI being more useful.
With the implicit feedback system, users might focus on acting in ways that guide the system to mimic their own approach, rather than focusing on actions that actually help the AI learn best practices
Getting the best of both worlds
Both methodologies are appropriate and beneficial for their target applications. Make sure you know which is right for yours. Our platform Cogment can help by enabling you to find the right quantity and type of human feedback for your ML training needs. Cogment supports the continuous injection of human feedback - explicit and/or implicit - into any and all phases of the ML lifecycle, from design to development, deployment, operations, and evaluation.