Machine Learning for your Business!

It’s easy to see the massive rise in popularity for venture investment, conferences, and business-related queries for “machine learning” since 2012 – but most technology executives often have trouble identifying where their business might actually apply machine learning (ML) to business problems.

With new AI buzzwords being created weekly, it can seem difficult to get a hold of what applications are viable, and which are hype, hyperbole or hoax.

In this article, we’ll break down categories of business problems that are commonly handled by ML, and we’ll also provide actionable advice to begin a ML initiative with the right approach and perspective (even it’s the first such project you’ve undertaken at your company).

Best of all, we’ll reference real business use cases, along with quotes and perspectives about “how to solve business problems with ML” from our network of AI researchers and executives. By the end of this article, you’ll have a good idea as to whether any of your present business challenges might be handled well with ML.

Note: At the bottom of this article, I’ve listed a basic glossary of ML terms in simple language. If you find a phrase or term in this article that you don’t understand, see the glossary below, or contact us if you’d like us to be more clear about a concept in this piece.

1 . Is the prediction you’re trying to make (or decision you’re trying to make) complex enough to warrant ML in the first place?

If it’s possible to structure a set of rules or “if-then scenarios” to handle your problem entirely, then there may be no need for ML at all. Also, if there is no precedent for any successful outcome applying machine learning to the specific problem to which you’re developing, it may not be the best foray into the ML world.

For illustrative purposes, it will be helpful to list a number of well-established business use-cases for machine learning so that you (the reader) can churn up your own application ideas:

• Face detection: It’s incredibly difficult to write a set of “rules” to allow machines to detect faces (consider all the different skin colors, angles of view, hair / facial hair, etc), but an algorithm can be trained to detect faces, like those used at Facebook. Many tools for facial detection and recognition are open source. Below is a video of facial recognition using MATLAB : (EXAMPLE 01)
• Email spam filters – Some spam filtering can be done by rules (IE: by overtly blocking IP addresses known explicitly for spam), but much of the filtering is contextual based on the inbox content relevant for each specific user. Lots of email volume and lots of user’s marking “spam” (labeling the data) makes for a good supervised learning problem.
• Product / music / movie recommendation – Each person’s preferences are different, and preferences change over time. Companies like Amazon, Netflix and Spotify use ratings and engagement from a huge volume of items (products, songs, etc) to predict what any given user might want to buy, watch, or listen to next.
• Speech recognition – There is no single combination of sounds to specifically signal human speech, and individual pronunciations differ widely – machine learning can identify patterns of speech and help to convert speech to text. Nuance Communications (maker of Dragon Dictation) is among the better known speech recognition companies today.
• Real-time bidding (online advertising) – Facebook and Google could never write specific “rules” to determine which ads a given type of user is most likely to click on. Machine learnings helps to identify patterns in user behavior and determine which individual advertisements are most likely to be relevant to which individual user.
• Credit card purchase fraud detection – Like email spam filters, only a small portion of fraud detection can be done using concrete rules. New fraud methods are constantly being used, and systems must adapt to detect these patterns in real time, coaxing out the common signals associated with fraud.

EXAMPLE 01

2 . Do you have new data and clean data?

“Clean data is better than big data” is a common phrase among experienced data science professionals. If you have reams of business data from years ago, it may have no relevance today, particularly in fields where the basic business processes change drastically year-over-year, such as mobile eCommerce). If you have reams of unstructured and disjointed data, you may have too much “cleaning” to do before you can ever get around to learning from the information collected.

UBER’s Head of Machine Learning Danny Lange once recommended that companies just starting out in machine learning should begin by applying supervised machine learning to historical data. Find data that’s already clean and relatively recent, and use labelled training data to start finding insights.

Note that in a rapidly-changing field, newer data is positively required. For example, if you run a door delivery service for pet supplies, and your app, prices, product offerings, and service areas have changed significantly over the last six months, you will need much more recent data to learn from than, say, a company selling homeowners’ insurance in Montana. If data is not related to the relevant trends and nuances of your current business, it is unlikely to glean predictive value.

3 . Does your data have existing labels to help a machine make sense of it?

While unsupervised learning (see glossary below) allows for a wide degree of applications in making sense of data without labels, it’s usually not advised for companies to “jump into” ML with a first application in unsupervised learning. The low-hanging fruit for an ML use case is likely to spawn from its historical, labelled data. Below are some examples that might help a reader garner new ideas:

• Facebook had millions and millions of tagged human faces on its platform, faces that were already associated with an individual person. This gave Facebook the ability to train algorithms on a tremendous volume of labeled data, with millions of faces in all kinds of lighting conditions and from various angles, allowing the algorithms to be highly refined and attuned to identifying specific human faces.
• Google serves billions and billions on search results, and can gauge the usefulness and relevance of its search results based on click-through rate of its top lists, page -load time, time-on-page from a specific visitor, and many other factors. It would be impossible to find a set of hard and fast rules for showing the right search results, so Google’s algorithms learn what the best options will be based on real-time engagement from billions of daily searches.
• Credit card companies like CapitalOne are faced with a huge volume of chargebacks and reportedly fraudulent purchases each day. By finding connections and patterns across types of purchases, locations of purchases, and types of customers, CapitalOne can use the “labelled” instances of fraud to predict other transactions that are most likely to be fraud. Anomaly detection plays an important role in various security and fraud applications; below is a short “anomaly detection primer” video from Numenta’s YouTube channel: (EXAMPLE 02)
• An eCommerce company with a massive volume of customer support emails will have a lengthy record of support tickets that were labelled “refund requests”, “technical issues”, “delivery issues”, among others. The company may choose to develop an ML system to instantly label incoming emails, transcribed phone calls, and chat requests with the proper support issue “type.”

EXAMPLE 02

4 . Can your solution to this problem afford for some allowance of error?

ML might be thought of as a kind of “skill”, in the same sense that one might apply the word to human beings. A skill that’s alive, adapting, growing and informed by experience. For this reason, an ML solution will often be incorrect a certain percentage of the time, especially when it’s informed by new or varied stimuli. If your task absolutely cannot allow for any error, ML is likely to be the wrong tool for the job.

An example of an application that cannot allow for error might be an application that aims to read the amount of an invoice or bill and then pay that invoice or bill. One letter difference or one number difference could mean overpaying your bill by 10x the original amount (if the decimal was interpreted to be in the wrong place), or sending money to the wrong company (if an invoicing company name isn’t registered exactly).

In a case like above, some degree of ML might help with “bucketing” different types of bills or invoices, but the final decision to enter the payment amount and send a payment would likely require an accountable human.

As an interesting caveat, there is a San Francisco-based startup called Roger.ai which is aiming to use natural language processing and machine vision to real and pay bills, albeit it pulls humans into the loop before sending funds.

1 . Begin with a priority problem, not a toy problem

In an off-mic conversation with Dr. Charles Martin (AI consultant in the Bay Area), he mentioned that many companies read about ML with enthusiasm and decide to “find some way to use it.” This leads to teams without the real motivation or gusto (or committed resources) to drive an actual result. Pick a business problem that matters immensely, and seems to have a high likelihood of being solved

UBER’s Danny Lange has mentioned from stage that there is one thought process that’s highly likely to yield fruitful machine learning use case ideas: “If we only knew ____.”

Ask yourself, what mission-critical business information are you dying to know, but can’t currently access? Maybe it’s understanding the lead sources most likely to yield the highest customer lifetime value, or the user behavior most indicative of expected churn.

2. You can give it data, but all of the context must come from you

Thinking through what information to “feed” your algorithm is not as easy as one might presume. While ML algorithms are adept in identifying correlations, they won’t understand the facts surrounding the data that might make it relevant or irrelevant. Here are some examples of how “context” could get in the way of developing an effective ML solution:

• Predicting eCommerce customer lifetime value: An algorithm could be given data about historical customer lifetime value, without taking into account that many of the customers with the highest lifetime value were contacted via a phone outreach program that ran for over two years but failed to break even, despite generating new sales. If such a telephone follow-up program will not be part of future eCommerce sales growth, then those sales shouldn’t have been fed to the machine.
• Determining medical recovery time: Data might be provided to a machine in order to determine treatment for people with first- or second-degree burns. The machine may predict that many second-degree burn victims will need only as much time as first-degree burn victims, because it doesn’t take into account the faster and more intensive care that second-degree burn victims received historically. The context was not in the data itself, so the machine simply assumes that second degree burns heal just as fast as first degree.
• Recommending related products: A recommendation engine for an eCommerce retailer over-recommends a specific product. Researchers only discover later that this product was promoted heavily over a year ago, so historical data showed a large uptick in sales from existing buyers; however, these promotional purchases were sold more based on the “deal” and the low price, and less so by the actual related intent of the customer.

3. Expect to tinker, tweak, and adjust to find ROI

Building a ML solution requires careful thinking and testing in selecting algorithms, selecting data, cleaning data, and testing in a live environment. There are no “out-of-the-box” machine learning solutions for unique and complex business use cases. Even for extremely common use cases (recommendation engines, predicting customer churn), each application will vary widely and require iteration and adjustment. If a company goes into an ML project without resources committed to an extended period of tinkering, it may never achieve a useful result.