The E-Commerce space is constantly abuzz with news about the importance and impact of AI and ML. Lately, we’ve been hearing a lot of confusion regarding how exactly an individual business owner can take advantage of these tools and what problems they can be used to solve. In this post, we’re going to present a brief summary of several potential applications of AI/ML in the e-commerce space and discuss why these applications are worth looking into for the average store owner.
What is AI/ML?
ML stands for Machine Learning, and it describes software tools that do just that: allow a machine to learn (from data). Using an ML tool typically has two phases: training, where you (the user of the tool) provide the tool with inputs and outputs and it tries to construct an internal model of how to consistently get from an input to its corresponding output, and inference, where you provide an input and the model tries to predict the corresponding output. These two phases can occur one after the other, or they can both occur simultaneously. There are also certain types of ML that don’t require you to provide outputs, but for the most part, ML workflows follow the steps described above. For more information on what exactly ML entails, read the relevant section of our last blog post here.
AI (Artificial Intelligence) is more of an umbrella term, used to describe any kind of research or tool that tries to give computers the power to think, reason, and/or learn. It’s a vast field, and ML is just one component of it. However, in today’s typical parlance, when a tool is described as AI, it’s almost always an example of ML. Within ML, there are tons of subfields and types of tools; neural networks and deep learning have been especially popular as of late, but many valuable examples of e-commerce ML rely on simpler methods such as collaborative filtering. If you’re interested in a bit more detail on the differences between these different types of ML, check back later — a blog post on that topic is in the works! For the time being, though, let’s dive in to the potential applications of these tools.
The most obvious application of AI/ML in e-commerce, and the one that you undoubtedly interact with every single day, is the targeting and retargeting of advertising. Companies that sell advertising, such as Google and Facebook, rely heavily on ML models to make sure the ads on their platforms reach the right target audience. Historically, advertising has actually driven a significant amount of research in the ML community. In this use case, the inputs for training the model are the details about a customer, which can include their browsing history, their search history, their Likes, their demographic info, and countless other details. The output, meanwhile, is whether or not that customer is a good target for a certain ad campaign. Being good at this predictive task is essentially the core of Google’s business model, and it has proven to be incredibly lucrative for numerous internet companies. Without ML, targeted advertising would be much less effective.
The typical e-commerce catalog requires some kind of search functionality, to make it easy for users to find the particular product they’re looking for. This problem is much more difficult to solve than it might seem. Customers have a tendency to phrase what they’re looking for in ways that you didn’t expect, and there are countless different ways to describe any given product (many of which will not be the ones you chose to use in the product’s description). In addition, for e-commerce brands that have extremely large catalogs (Amazon is the archetypal example), ML can help make searches speedy and efficient. Several companies work to provide more effective ML-based search for brands; we’re a fan of Algolia. These ML-based tools let your search feature understand the meaning behind your user’s query, so that the most relevant results can be reliably surfaced. At Toucan, we also work w/ similar tech; though we aren’t quite a search provider, we’re also focused on using ML to understand the meaning behind your customers’ messages.
This use case is much more recent, but ML is also showing great promise at optimizing marketing campaigns, especially those that span several different channels. Essentially, ML tools in this space work by accepting a bunch of different assets and marketing material from a marketer and then optimizing the delivery of this material by carefully monitoring the metrics that the store owner cares about. For example, an ML-based optimization engine might be able to decide that if certain content is delivered over email, while other content is used in banner ads, the store’s revenue numbers go up substantially. In a more controlled use case, the ML tool might just provide you, as a store owner, with more accurate attribution metrics regarding the impacts of your different marketing campaigns. Practically, applying this sort of ML tool is often somewhat finnicky; however, companies such as ConversionLogic and Amplero are making great strides in the space, so we’re definitely going to be watching it carefully!
The flight and hotel booking industries are especially notorious for using ML for this use case, but it’s accessible to e-commerce stores of all sizes and scales. ML price optimization tools can essentially take in as much information as possible, including user data, time of day, weather, time of year, competitor pricing, and length of session, and use it all to spit out the best possible price for a given product, i.e. the price that will maximize your revenue in the long term. A complex optimization process like that usually requires development of a custom ML model, by an ML consultancy of some kind, but simpler models that take fewer factors into account are often available off-the-shelf. If you’re looking to get started in Price Optimization, the best starting point is probably a non-ML approach to dynamic pricing, such as the one provided by Prisync.
Product Recommendation is another classic use case of ML, one that you’re likely familiar with already. Essentially, ML tools for product recommendation try to predict what products you might be interested in buying based on your purchasing/browsing history (by comparing it, essentially, to the history of other users and data on what they ended up buying). There are countless software providers in this space, ranging from small startups to software giants like Adobe. An overview of the space would require its own blogpost — if you’d be interested in reading that, shoot us an email at firstname.lastname@example.org, and we’ll make it happen! We do Product Recommendation research at Toucan as well, though our models need to make recommendations in the context of conversations, so they’re not quite the same as the ones described above.
Financial Fraud Detection
Fraudulent transactions can have a crippling effect on e-commerce stores, and are estimated to cost e-commerce stores tens of billions of dollars in revenue. Scamming and fraud is a growing problem, and those who seek to perpetrate fraud are using increasingly sophisticated approaches. Luckily, methods for detecting and stopping fraud are also growing more intelligent. The instant nature of many e-commerce transactions means that relying on humans to flag fraud is often infeasible. However, ML models can analyze a number of signals in real-time and catch fraud as soon as it occurs. The fact that ML models are capable of learning patterns from incomplete data means that they can recognize fraud even when it doesn’t exactly fit a mold that they’ve seen before. By searching for anomalies, these models can flag potentially fraudulent transactions and make sure they are addressed before they cause damage to a store’s bottom line. Our favorite ML tool for fraud detection is actually one that you might already be using without knowing it: Stripe Radar, a part of Stripe’s fantastic payment-processing platform.
Fake Review Detection
In a similar vein, fake reviews are a problem that plagues e-commerce stores across the board, in ways that can have strongly negative impacts on sales. ML tools in the fake-detection space can identify which reviews exhibit the hallmarks of being auto-generated or fake and flag or remove them from review listings. We haven’t seen any compelling companies that provide this sort of review-flagging as a service to store owners, but Fakespot does a pretty good job of providing it to consumers, and there are plenty of freely available ML projects on Github that a store owner could incorporate into their storefront if they so chose.
This list is a survey of several common ML applications, but it’s far from exhaustive. Wondering if you can apply ML to a different business problem that you don’t see listed here? Basically, see if you can re-frame the problem as one of either predicting or classifying something. If you can, and if there’s sufficient data available, there’s a pretty high chance that an ML-based solution is possible. Even if data seems scarce, there might be ways to work around it that allow you to still utilize ML. There’s no doubt that in the coming years, ML-based tools will play an increasingly large role in e-commerce and serve as a key investment priority for many store owners. Might as well get a head start!
- The Toucan Team