How to decide what level of Machine Learning software is right for your business

What are retailers asking for, and how can they evaluate machine learning vendors?

This article was orignally published in the Technology section of FEAST Magazine. Written by our CEO and co-founder, Robin Mellstrand.

The retail industry is well aware of all the hype surrounding Machine Learning (ML) and those magic words, “Artificial Intelligence” (AI). Supposedly, it can read the minds of customers and predict what they are about to do. But, ML is so riddled with misconceptions and marketing-speak, that retailers can become blinded by it all, which can result in poor decision making in the process. Many retailers simply don’t understand the purpose of the technology and if it will make a difference.

While the success of Google’s Deepmind has set high expectations for ML (having defeated world champions in the game, GO), it’s important e-commerce leaders understand the difference between the “controlled” environment of a game like GO - where the rules are known and limited, and outcomes that are finite - and the world of online commerce. In e-commerce, the number of variables are nearly infinite and the available data the machine can learn from is often sparse.

Despite the complexity of the challenge, some ML vendors are capable, to some degree, of delivering on those high expectations. The most commonly requested is one-to-one personalisation, as opposed to broad customer segments set manually via If-This-Then-That rules.

Unlike in a physical store, where a sales representative can infer context from the situation and quickly gather the information they need to tailor the customer’s experience, the online shopper expects relevant feedback instantaneously, particularly if they’re a returning customer, and even more so if they use the website’s site-search functionality to find what they are looking for. Patients quickly waine when, for example, page load speeds surpass two seconds, site-search results are irrelevant and product recommendations don’t reflect a visitor’s unique taste.

Ultimately, ML software should save the business time through automation, and it should drastically outperform the current solution (it does the job better, measured in increased conversion rates, revenue and/or profitability). If the software meets these two requirements, the next question is - how much better performance is the business willing to pay for? In other words, what is the optimal price-performance ratio for the business (including time-saved as a measure of performance)? Answering this question will help businesses assess what level of Machine Learning is right for them.

Robin Mellstrand, CEO and Founder of Loop54 comments “ML software that is transparent in its goal, is what an e-commerce company should look for. Minimal input with maximum results is something that our newest machine learning algorithm achieves without the client having to manually configure product feeds or have masses of visitor behaviour data. Vendors really need to take a ‘kid-glove’ approach in this sector, as many e-tailers are still scared to take on something they may not understand yet. If the company selling software cannot explain its purpose in its simplest terms, my advice is to shop around and find a better fit as it may cause more problems in the long-run, or even cost the business a lot of money and time.”

What are the levels of Machine Learning software in e-commerce?

Level 0 - Rule-Based Decision Making (i.e. no machine learning): performance gains are limited and the solution actually creates work. Although it may appear to save time because the “action” part of any If-This-Then-That rules are triggered automatically, the alternative is actually doing nothing at all. Therefore, ongoing manual work is required to see big gains in performance. Humans are needed to analyze the data and create new rules for the machine to act upon. Software of this type are typically not very expensive (or shouldn’t be) and the gains made are often worth the initial software and labour costs. Unfortunately, this level doesn’t scale and the retailer will experience diminishing marginal rates of return, forcing them to eventually consider higher, more efficient, levels of ML software.

Level 1 - Basic Machine Learning: using dynamic inputs, like visitor behaviour data, to improve performance automatically. For example, product or category listings re-sort themselves automatically, showcasing the most frequently purchased items higher in the list. The performance improvement at this level, in comparison to level 0, is typically quite substantial since there is a massive time savings (e.g. merchandisers no longer need to pour over purchase data at timed intervals to decide what products should appear on top of listings). The largest drawback at this level is the data sparsity issue. There needs to be enough dynamic data to make sense of and improve the whole system. Therefore, vendors at level 1 typically offer some sort of level 0 functionality (If-This-Then-That rule setting interface) in conjunction with their basic ML features in order to maneuver around the lack of data.

Level 2 - Advanced Machine Learning: at this level, basic one-to-one personalisation is possible. The machine can make sense of the whole system with much less data and can recognise and tailor experiences down to the individual level based a customer's previous interaction with specific products and/or categories. Only machines can deliver one-to-one personalisation online, therefore, gains at this level primarily come in the form revenue. However, since much less data is required to improve performance, there is less reliance on manual rule-setting and more time is saved compared to level 1.

Level 3 - Artificial intelligence: predictive personalisation. The machine will “learn” about each individual customer over time and use that information to build a robust profile of the customer. The machine can predict what a customer might be interested next based on what it knows about them more generally, even in areas where the consumer has never interacted before. It can use gender, brand/style affinities and price-range preferences to continuously customise the shopping experience. (psst! .... this is Loop54's next personalisation release.)

Curious how this applies to on-site product search?