Still afraid to use Machine Learning in your business?
Now’s the time to start: high-leverage, low-cost, and a vital key to future data-driven successes
Now’s the time to start: high-leverage, low-cost, and a vital key to future data-driven successes
Remember e-mail in the 90’s and early 00’s? Need an example of brilliant, quality-of-life machine learning? Well, one of the more famous (but invisible) everyday uses of machine learning is spam protection. Imagine NOT having spam protection… 🤮 With this said, you see that machine learning is neither magic nor unspeakably obscure.
Do you think machine learning seems awfully distant and hard to implement in your line of business? It’s actually (probably) not! Here’s at least a few things you won’t need to use machine learning:
- Bank full of money 💰 💴 💶 💷 💸
- A huge IT department computer 🖥 💻 🖥
- A throng of data scientists 🥼🥼🥼
- The world’s hottest product or service 📢💘 📣
Of all the big trendy tech buzzwords flying around, machine learning (or ML, for short) is probably the one being bandied about most intensely for the last few years. This article aims to make ML understandable as a part of your business.
What is Machine Learning?
Probably the biggest issue with explaining machine learning is that it comprises an entire field, and therefore—as a whole—doesn’t translate well into pithy nuggets customized for cheery LinkedIn posts. Further, from a practical standpoint, ML is really just one single part to be integrated into a larger context. It’s not a free-standing, all-knowing function like it sometimes may seem from casual conversation and popular culture.
So how should we describe it, then? Let’s go to Wikipedia and look at the first sentence on the article on machine learning:
Machine learning (ML) is a field of artificial intelligence that uses statistical techniques to give computer systems the ability to “learn” (e.g., progressively improve performance on a specific task) from data, without being explicitly programmed.
— Wikipedia, https://en.wikipedia.org/wiki/Machine_learning
This is a pretty accurate, descriptive way of understanding what ML actually does. No magic whatsoever. Any referentiality to murderous robots can be safely discarded at this point. And Google explains it like:
In basic terms, ML is the process of training a piece of software, called a model, to make useful predictions using a data set. This predictive model can then serve up predictions about previously unseen data. We use these predictions to take action in a product; for example, the system predicts that a user will like a certain video, so the system recommends that video to the user.
— Google, https://developers.google.com/machine-learning/problem-framing/cases
Machine learning’s predictive capacity indeed stems from learning from what the algorithm encounters. This training stage is very often an entirely separate step that happens before deploying and using the model, as the static code is called when it’s taught and built. The most common models for learning are called supervised training—where you teach it what to look for and decide on, with the help of numerous labeled examples—and unsupervised learning, in which it tries to find and sort items based on relations unknown at the time of training. Some problems are best resolved by explicitly teaching it what something is (for example: a sorting machine removing bad apples from a factory line) and sometimes what you need is understanding complex groupings and patterns (for example: finding relations in large, unstructured datasets).
The input to an ML algorithm can for example be text (natural language processing), image (computer vision) and voice. Input data needs to be cleaned and formatted to be accessible by the ML: This is the biggest part of creating an ML application, and it’s also the reason lots of companies employ dedicated data scientists to spend the majority of their time doing this. Note: Depending on the complexity of the task, pure data scientists will or will not be needed. It does takes a fair bit of know-how to understand, read and interpret complex sets of data, but a seasoned non-data scientist can absolutely do this work under many circumstances. Model building is often built on a schedule rather than using so-called online learning which continuously learns, but is hard to implement correctly and is potentially dangerous if done improperly.
The secret sauce is mostly about getting the algorithm correct to some reasonable degree—unless you’re doing something really big or complex, or putting your entire business on the line for the sake of an ML solution. That would not be very smart, however 😅
Machine learning is all about making decisions
What you probably want most of all this is to have a system that helps you make decisions. You know what? ML is really good at this.
Ultimately, ML is a tool for making decisions: Based on a specified type of input, a computer decides what that input is and returns you its decision. These decisions are called predictions.
A prediction can obviously only ever happen before a predicted event. Since machine learning can only function meaningfully with a rather large amount of historical data, we have a good view of what we’ve been tracking, for example transactions, time-to-completion, errors and so on. Using statistical models in the ML field, we can therefore construct patterns from that data and predict events in our tracked domains. Concretely, the prediction delivered from the algorithm is essentially a number which will (usually) stand for a probability. Assuming that there is some value in what you are attempting, business value is created when you know how to use that probability.
Typical business cases
Machine learning is used by a wide variety of companies ranging from ecommerce brands using ML to detect fraudulent payments, to family businesses growing cucumbers 🥒 getting machine-aid to correctly identify and sort according to size and quality.
It may be helpful to look at your own organization and what pain points you can find in it today, such as work moments that involve higher-than-usual risk or heightened failure degree. Bottlenecks are also something that could be optimized.
Automating processes is a very common first step for machine learning. This is where you would think of automatic spam detection, automated customer surveys and robots on an assembly line. Consider a camera connected to ML discover flaws in your production, maybe even a lot earlier than with manual inspection. The manual solution would be to pause production while an inspector controls the quality. Overall, any process where there is a lot of manual handling/passing of materials is good to use ML for. Offload people from dangerous, tedious or attention-craving tasks and let ML do these bits, instead leveraging human skills for higher-level work.
Better customer experience can fall into this category as well. You’ve probably used online customer service funnels that ask you to make multiple categorical choices before you can ask your actual question. What if you wouldn’t need to do so? Chatbots (many of which just happen to be machine learning-driven) are one way of automating the tedious manual category-choices.
Here you want to have a high degree of accuracy, like 95% or more. Automating processes which run continously during a normal day will give you great leverage and are often the first thing a company wants to ML’ify.
In processes where humans still need to be involved, you would want to optimize how those tasks are handled in order to reduce time usage—for example sorting out customer support tickets into correct lanes or doing inventory addition in a faster way by knowing about commonly added stock and their requirements. You can see this type of nice, invisible optimization at work when you write an email and you get contextual suggestions for your text drafting. Any health care context can benefit immensely by making fast, concrete decisions on medical documents or paperwork that is potentially life-saving if handled swiftly.
Predictive policing, while ethically problematic, aims to (among other things) add law enforcement resources to areas that are deemed more likely to require policing. This is one of several areas where ethics and an empathic understanding goes a long way to keep a company’s name clean.
For optimization tasks, the important thing is to get down the average task time. You could perhaps allow for a slightly lesser degree of accuracy, while still retaining the option of confirming unknown/unclear decisions with the help of a qualified person.
Insight: Develop new ideas, business strategies, understand better what you are doing
The broadest field, and the one that’s both the loosest as well as the potentially most lucrative, regards insights gained from ML assistance. Especially at the point when your company has a large amount of data—hopefully collected at various levels and checkpoints—you can use ML to check for patterns, relations and correlations between points that may be very hard for a person to detect.
Also consider crossing multiple datasets to uncover new data: Companies can use multiple sets of data, such as weather and traffic incident data, to better predict accidents in certain weather conditions. Film companies use video analysis of trailers to understand what types of elements that audiences seem to like, in order to make better selections and recommendations for their production.
Typical use cases would for example be:
- Why do users buy this item and that item together?
- Why don’t users re-subscribe?
- How can you provide relevant content to customers?
- How can you better know when equipment is failing?
- How can you better predict periods with high (and low) volumes of orders?
Insights can be fluctuating and since this area is much less clear-cut, manual analysis and trials are likely needed. Don’t expect super-clear answers here without some manual laying-of-hands. However, you have rich potential in this area. A company would initially begin this step by simply collecting and pre-empting the future need for historical data: Don’t collect, and you will miss out on future data-crunching.
How to begin thinking in ML terms
Some pointers, then.
Collect data and be an “ML organization”
Data is the primary driver — without it, there is no way of making use of machine learning. While data is definitely a hard currency, you should consult with experts on what to collect and how to store it. This should be a first step into setting a base level in your organization. Moreover, prepare your organization for the impact of ML and don’t forget to ask your staff to locate their own pain points or problem spaces. Machine learning is meant to enhance our capabilities, so it makes sense that any part of a business can and should get attention. Which brings us to…
What do you want to solve? Is it quantifiable?
Think of tasks with clear business value, maybe areas you are facing issues with right now. Break these tasks down into smaller parts. Decide on how to capture necessary data to inform such decisions. Is there enough quantifiable data to drive a decision?
Don’t overthink ML and don’t force it into a problem that won’t benefit from machine learning—some things just won’t be “ML problems”. Sometimes all you need is a set of reasonable heuristics to guide you (“if this, then that…”).
How are actions expected to be taken on decisions made with ML?
This is where you organization comes in again. Just like employing a new role that you haven’t had in your org previously, the ML will most likely have some kind of mandate that you want to specify.
Who does it answer to? Who is affected most immediately by the ML decision? What’s the consequence of ML decisions in the bigger scheme of things? What other services and technology are involved, before and after, the ML step?
Getting to the starting line
I suggest that you take an hour to go through Google’s Introduction to Machine Learning Problem Framing if you are more curious about how to take this approach a bit further.
At some point you will want to select a cloud-based host for your machine learning operations. The “Big 4” cloud ML providers are Amazon Web Services, Microsoft Azure, Google Cloud AI, and IBM Watson. Some of these offer free credit or trial periods so you can get started without thinking too hard about the initial costs.
A number of services are pre-built and are ready to use right away, for workloads like labeling images, doing sentiment analysis and reading text from images. Some things, like custom classification, will need to be built from scratch. That’s where you need experts.
Don’t hesitate to pick low-hanging fruit with ready-to-use APIs—that’s where I think most of the immediate potential resides when starting out.
Machine learning is just one part of the path to superb user experiences
Finally, don’t forget to focus on the ball: the overall business you’re running. For a digital experience to be top-notch in 2018 you will want to consider the various interacting levels as a holistic entity:
- User Experience: Put the user/customer first and be responsive to their needs, requirements and wants. Proper UX can never be a post-fix band-aid (or “toilet spray” as I call it).
- Agile/lean product development: Promote and streamline production for better results in less time, and in the ML age acting fast in the digital domain is not getting any less important.
- Machine learning: Collect and harness data to inform business decisions and automate and customize the user experience.
Some key observations from knowing anything about ML in 2018 is that many businesses will have a whole lot of low hanging fruit for the next foreseeable years, and it’s getting easier for every week to make a first move.
You should have no reason to believe you have to be a huge corporation to benefit from AI/ML. For all I know, you could be a corner store and still reach great results from machine learning. Why wouldn’t your business get anything out of it?
Mikael Vesavuori is a Technical Designer at Humblebee, a digital product and service studio based in Gothenburg, Sweden. Humblebee has worked with clients such as Volvo (Cars, Trucks, Construction Equipment), Hultafors Group, SKF, Mölnlycke Health Care, and Stena. Our design sprint-based approach and cutting-edge technical platform lets us build what’s needed.