What is human in the loop machine learning and how to use that for business?

Kristjan Jansons
March 18th, 2021

Read more
Human in the loop machine learning

This article discusses how to make your business processes more efficient with humans in the loop of machine learning. In addition, it helps you find suitable processes and teaches how to start with these. All in all, it hopes to be a guide for anyone wanting to bring machine learning into their business.

What is Machine Learning and its connection with business?

Machine learning (ML) is often pitched as an amazing solution to a lot of problems. Yes, ML is pretty much the closest thing we have to magic and it can solve many problems or at least improve many situations. On the other hand, as Gartner has predicted, 85% of AI projects fail. Meaning, there are obviously some skills required to tame this magic. One of those skills is understanding how to make your business work with ML.

So what is machine learning? Machine learning takes input (e.g. information from your CRM, databases, spreadsheets) and produces an output (e.g. finding fraudsters, handling claims, classifying what the customer asked). For the machine to be able to produce the required output from input, it needs data to learn from. For most businesses, that’s it and there is no need to delve deeper. It makes sense to just talk to your ML partner for specifics as needed.

ML cannot solve the problem if the problem statement is my business is not profitable, which is sad, but let’s try to stay positive. When we come back to the definition of machine learning, we see that there is an input that the output is based on. Almost like a production line, it sounds like a process. Does the business problem you are trying to solve look like a process? Is there a clear beginning and an end with logical steps in between? Is there somebody who could show how it is done at the moment if it is an existing process? How are the decisions made and what kind of data is being used to make those decisions? If you answer these questions and find that there is a process and it is done more or less in a similar way (even complex things might be solved following clear rules), there is repetition and routine, then there is a good chance that you have found a process that can be (semi-)automated with machine learning.

That you can automate something with ML does not mean that you should. There might be easier ways, there might not be a positive ROI and other issues, but this is not within the scope of this article. Let’s focus on the scenario where everybody gives this project a green light.

What does a suitable business process look like?

As said above, the process should have a clear beginning and an end with logical steps in between. The beginning could be somebody or some system looking at a set of input data (e.g. incoming messages, claims, declarations) at certain times, either regularly or when the data comes in. In Figure 1 below, it’s shown as input data. The logical steps in between could be, for example, somebody looking at an input (e.g. customer message) and making a decision based on that (Figure 1). There might also be rules which help a human to make the decision (Figure 2) or in some cases even automate it. If the process can be automated well with a few rules, do not use ML. On the other hand, if there are a lot of data sources, it is not realistic to come up with all the rules or you would be too dependent on experts writing the rules (who might leave the company for one reason or another), it’s probably a good idea to bring in ML. Last but not least, the output should be something clear – in the examples below, it is whether somebody is a fraudster or not. This is important because the machine should learn from the clearest possible examples. This does not mean that the machine could not pick up multiple things. For example, the machine could pick up that the person has committed multiple types of fraud. It also does not mean that the machine could not learn to predict suspiciousness. However, it does mean that you should review the process (possibly together with your ML partner) and decide whether the process is as standardized and clear as possible. In such a case, you might consider replacing a free-form decision box with a dropdown menu to make it clearer.

Auditor - Fraudster
Figure 1. Auditor making a decision based on input
Rules-Auditor-Fraudster
Figure 2. Auditor making a decision with the help of rules

As stated above, both processes might be suitable for humans in the loop of machine learning, and the devil is in the details. Make sure it makes sense by discussing it both with the business side and your ML partner. And remember, ML can automate a lot of different processes; it can handle database fields, free form text, images, speech data and much more, so you can be bold when going to your ML partner.

How does a machine learn and what to keep in mind?

Making a machine learn is often very difficult. However, a business person can simply forget that and just remember the following process as shown in Figure 3. We need training data to learn from, there is a training process where some magic happens, and finally, we have a machine learning model that is able to produce output from input.

ML model
Figure 3. The simplified process of creating a machine learning model

Yet, there is a lot that a business should know about the challenges of creating, developing and maintaining the machine learning model. From the business side we should keep an eye on at least the following:

  • Collecting training data as easily as possible. As ML learns from examples and the world is often very diverse, the simple rule is “the more, the better”. Of course, this might be easier said than done (for example, it might not be realistic with rare diseases), so one always has to make sure that it is not too costly or does not take too long. But in many cases, there are easy wins in remaking the process or streamlining it, especially with a human-in-the-loop system where we just need to make sure that we are collecting/recording the activities of our human experts to learn from their decisions.
  • Making sure we minimize mistakes. ML model output often has “confidence” attached to it. With high confidence, we can take it as an automated process (if the business rules/risks allow it). When the confidence level is low, we should let a human review the output. This way the process is efficient and we minimize the number of mistakes.
  • Making sure we are getting better over time. Whether with only people or together with machines, getting better should always be a goal for every organization. Together with machines, it is slightly easier. The easiest cases can probably be automated relatively quickly. Usually, the more complicated a case, the more examples are needed. Thankfully, the previous point helps with this. Since humans are handling cases that are more complicated for the machine, they are also generating training data for the machines to learn from.

As we will see, if done correctly, all of these things will be taken care of automatically if a human is in the loop system.

Putting the two together

Putting the two worlds together is not complicated. First, you need to bring an ML model into the process. Secondly, you need to create a loop that creates training data for the ML model to learn from. Yes, there are multiple variations of how to make it happen, but in essence, that’s it. So let’s look at a few examples of how we can automate the process as depicted in Figures 1 and 2. In all of these examples, the ML model is trained on historical data – from which the machine can learn to produce output based on different input examples.

Example 1 in Figure 4. This is the most efficient loop, but it might not always be usable because of regulations or risk averseness. First, the ML model looks at the input data and produces an output based on it. If the confidence level of that output is high, then that decision is fully automated. If the confidence level is not high enough, then the auditor will review those cases and make a human decision. Manually reviewed cases are added to the training data so that the machine could get better over time with more complicated examples. In theory, it might be possible to reach full automation with this setup. However, often life is so complicated that new edge cases keep popping up which need human revision.

Automated Human Decision
Figure 4. Human in the loop of machine learning where high confidence outputs are automated and low confidence outputs are added to the training data

Example 2 in Figure 5. This is for those cases where, because of regulations or risk averseness, everything needs to be at least reviewed by a human. In this system, the ML model can produce an output for every case, whereas the output can contain the decision, the confidence level, and in many cases also what influenced the made decision the most. The human involved can accept or reject the output proposed by the model. Even if a human needs to review every case, the speedup might still be significant because of the considerably reduced investigative work. The decisions made by a human will be added to the training data so that the model will become increasingly better at giving suggestions to the human.

Human rejects or approves
Figure 5. Human in the loop of machine learning where a human reviews all the output produced by the model

Examples 3 in Figure 6. This is also for those cases where, because of regulations or risk averseness, everything needs to be at least reviewed by a human. The difference lies in the fact that there are so many potential cases that only a handful can possibly be reviewed. And let’s be realistic, in the real world we only have a handful of fraudsters (at least in a well-functioning society). So the ML model prioritizes the cases and the auditors focus on the top of the priority list. The human decision will be added to the training data. Again, a huge boost to productivity. What should be noted, is that some attention needs to be given to catching new types of fraud. For this, auditors should also review some random cases and/or anomalous cases, but that is a technicality that can be overcome.

auditor human decision
Figure 6. Human in the loop of machine learning where a human reviews cases from the top of the priority list

This symbiosis is beneficial for both sides. From the business side, the processes get (semi-)automated, mistakes are reduced, and people get to focus on more challenging work. The ML model does not get bored, does not require vacation or even lunch. The processes become more reliable, predictable and efficient. From the machine learning side, the system will gain a reliable source of training data which helps it to make fewer mistakes and become better over time. The result is a cycle that reinforces itself over time as shown in Figure 7.

ml model
Figure 7. Reinforcing the cycle of human in the loop of machine learning systems

How to start?

If based on the section “What does a suitable business process look like?”, you think that there is some potential in the idea, talk first with an ML partner. Somebody who truly knows what machine learning is, has implemented it into real-life processes and is sincerely interested in your business goals. The partner should be able to tell you if it makes sense from the technical perspective and what is needed to make it happen. For example, it might turn out that the process needs to be split up and only part of the process can be automated with machine learning, or that quality data needs to be collected before deciding upon suitability. In the end, your ML partner should be able to give a rough estimation about the amount of work – in some cases, this might require a pilot or exploration of data. After understanding the feasibility and drawing up a rough budget, you can make sure that applying an ML model makes sense for your business.

Words of caution: unless you have a well-functioning ML team in-house, do not even consider hiring a data scientist. Data science is not “regular IT”, you need a lot of skill sets, the work is unevenly distributed, and most probably you don’t have enough work for a whole team or perhaps even a single data scientist once the projects are finished. If you want to get results in the coming years, make a reliable partner deliver them.

Once you have approval from the technical and business side, you need to put together a team suitable for this type of project and get going.

Conclusion

ML can automate processes that have a clear beginning and an end with logical steps in between. ML needs training data to learn from and this can be generated by having a human in the loop. If done correctly, a loop will be created which only reinforces itself to be better over time. A loop system with a human involved can be set up in multiple ways and to understand the specifics, please talk to your ML partner. On that note, why not have MindTitan as that partner.

Want to learn more?

Want to learn more?

Konstantin Sadekov
HEAD OF GROWTH AND MARKETING

Konstantin has graduated from the Estonian Business School major in economics and finance and is currently doing his MBA degree in the USA. Before joining MindTitan he had international business management experience for more than 5 years and overall more than 9 years of international B2B sales and marketing experience

Want to learn more?

Want to learn more?
konstantin sadekov mindtitan