Small-business owners are always on the hunt for opportunities that can give their businesses a leg up. In the past, small business owners have eagerly adopted software-as-a-service products, transitioned to cloud infrastructure and embraced self-service digital advertising. Today, one of the most exciting opportunities is the potential to leverage machine learning (ML) to give your business a competitive advantage. ML solutions automate workflows, enhance data-driven decisions and facilitate interactions with customers.
Two factors that are making ML accessible to small businesses are the commoditization of machine learning algorithms and the democratization of pre-trained ML models. These two shifts have allowed my team to help several small businesses build custom ML solutions that would have been pipe dreams even five years ago. We've built a range of solutions from automatically flagging fraudulent online survey submissions to a HIPAA compliant content recommendation engine. Throughout this process, I've realized that there are broad misunderstandings around what machine learning is and how businesses can effectively build ML solutions. So as a small business owner, how can you leverage these ML opportunities to benefit your business?
What Is Machine Learning?
Before delving into how SMBs can build machine learning solutions, it’ll be helpful to understand what ML really is. Contrary to corporate marketing, machine learning isn’t black magic. Instead, it's a class of algorithms that allow computers to perform pattern matching extremely efficiently. Many of these algorithms have been around since the 1970s. Newer ones have emerged from academia more recently. Interestingly, many of these algorithms aren’t particularly complicated to implement but they are computationally expensive. These computational costs are one of the factors that have made ML algorithms difficult to use in the past. So how do companies use ML today?
Let’s jump back to before machine learning and consider how your credit card company might have evaluated that a given transaction is fraudulent. They may have pulled your payment history, the details of your previous transactions and then evaluated the details of the potentially fraudulent one against this history. An analyst may have seen that you never missed a payment, had an average transaction size of $100 and only used your card at a handful of merchants. Then they could evaluate the potentially fraudulent transaction within these parameters, using their gut instinct decide that it was fraudulent.
Now, with machine learning, the credit card company would be able to automate this decision-making process using a ML model. Using the ML approach, the company would still pull the same information. But they would feed it into a ML model which would decide if the transaction was fraudulent. At its core, the ML model would recognize a common pattern between fraudulent transactions and be able to flag new transactions correctly.
Algorithms For Everybody
At a high level, building any machine learning solution involves a similar set of steps. A developer needs to feed data through an algorithm to train a model and then use the model to analyze new, unknown samples. One driver for making ML readily accessible has been the increased availability of open source ML algorithms. Today, there is a wide range of open source tools like scikit-learn and Apache Mahout that feature industry-vetted implementations of dozens of algorithms to tackle a wide range of problems.
As noted above, in order to train a ML model, you need to have a substantial amount of labeled training data. In the context of training a model, “labeled training data” would consist of data that is labeled as the goal of the model. Taking the credit card example, labeled data would be sets of transactions along with their metadata that have been labeled “fraud” or “not fraud.”
Today, small business have a wealth of first-party data they could leverage to train ML models. Because of their adoption of SaaS products, small businesses now have hundreds of rich data points about their clients and customers. Data points like the frequency that a customer opens MailChimp emails, how many times they’ve swiped a credit card in store and if they’ve left a review on Yelp could all easily be used to train an ML model. Using this data, you can build ML solutions to do things like identify brand champions, predict churn or preemptively identify anomalies.
Open Access Pre-Trained Models
For almost all machine learning solutions, the hard part is training the models. This has to do with how much compute time and data they require. Access to cloud computing resources has helped lower the barrier on the compute side but for certain problems, access to a large volume of data is still a challenge. For problems like speech recognition or decoding the intent of natural language, it would be infeasible for small companies to effectively train their own ML models. However, larger companies like Amazon or Google have access to both the data and compute resources to effectively train complex models for these use cases.
Large cloud vendors have made these pre-trained models available as services on their cloud platforms. For example, Amazon has made the model that Alexa uses to recognize what you’re asking it to do available as a service on the AWS platform. With access to these tools, small businesses can build solutions that used to be only available to large enterprises. They can now do things like automatically identify the content of customer uploaded photos, automatically flag negative customer support interactions and build custom voice-enabled products.
It’s Time To Build
Now is a great time for SMBs to use machine learning to give their businesses a leg up. The availability of quality algorithms and the wealth of data generated by SaaS products have made it possible for small business to effectively train their own ML models. In addition, cloud vendors have begun making several of their own proprietary ML models available as a service.
