Machine learning in ecommerce changes online shopping

A customized shopping experience

Between 2011 and 2012, Amazon enhanced its purchasing processes with a new recommendation system based on machine learning algorithms. The first effect of its integration was an astonishing 29% increase in sales (equivalent to about $12.83 billion) recorded in the second fiscal quarter of 2012.

But considering such a sales spike as the main benefit ensured by Amazon’s machine learning implementation would be simplistic. After all, this is just a consequence (more than welcome, actually) of something far wider and significant, namely machine learning's ability to connect buyers with the products they want. Including those they didn't realize they wanted.

In traditional shopping, this routing process to guide purchasers towards the right merchandise is based on the interaction between the customers and the sellers. Shopkeepers can ask what customers are looking for, understand their preferences, address their concerns, and offer the right products based on all the information collected and a mix of gut and sales experience.

In technical terms, sellers carry out a simplified, homemade version of the marketing process known as customer or market segmentation, i.e. the classification of potential buyers into sub-groups based on shared characteristics. Then, according to this classification, they offer customized recommendations by targeting customers with the products that best suit their needs.

Unlike physical stores, however, ecommerce portals don't have the luxury of leveraging sellers' experience to operate these two essential tasks, namely segmentation and targeting. And that's where machine learning comes in as part of retail digital transformation.

Machine learning and targeted recommendations

Machine learning is a discipline within the vast realm of AI, specializing in creating computer algorithms that can autonomously improve their performance through experience. Specifically, machine learning algorithms can be trained by processing huge datasets, spotting patterns among all this information, and building mathematical models representing such patterns.

These models, which are refined as the algorithms process more and more data, offer us valuable insights into certain phenomena and the interconnections among all the variables beneath them. Something that has proven extremely useful to analyze current events, forecast future trends and make data-driven decisions. But how can we apply the extraordinary capabilities of machine learning to ecommerce?

Well, for example, we can leverage algorithms to segment customers and target them with the right offer, somehow replicating the customized experience offered by human sales assistants. That’s what happens with machine learning-based recommendation systems, a rather popular marketing solution among major ecommerce platforms.

Recommendation tools can process past sales data, recognize recurring purchase patterns among typical buyer archetypes, and predict the items that might grab the attention of specific users to provide them with personalized suggestions—a mechanism typically used on a larger scale for predictive analytics in marketing.

According to McKinsey's 2019 The Future of Personalization paper, product recommendation systems can easily increase revenues by 5-15% and marketing-spend efficiency by 10-30%. However, not all recommendation engines are the same. Indeed, they may follow two distinct approaches known as content-based and collaborative filtering, as well as some sort of a hybrid system combining them.

Content-based filtering vs collaborative filtering

Content-based filtering provides recommendations by focusing on two main factors:

• The items’ features, such as their price or category. In this regard, each product is described by assigning it some keywords.
• The characteristics of customers, which basically means their purchase preferences and reviews on previously purchased items.

Following this approach, the machine learning algorithm will check a customer's purchase history and recommend other products with similar characteristics to those already bought.

Unfortunately, such a mechanism suffers from some significant drawbacks. First, it requires scrupulous cataloging of all the merchandise present on the ecommerce platform. Secondly, it is a rather "conservative" approach as it tends to suggest the same types of products already purchased previously. And finally, it may have a hard time providing effective suggestions to new users without historical data relating to them.

The other possible solution to segment customers and target them with customized offers is via collaborative filtering. This approach, adopted among others by Amazon, is based on the idea that people with similar preferences in the past will still agree in the future. Hence, an algorithm that follows this strategy will suggest to a customer some new products which have already been ordered and positively rated by other users with similar buying patterns.

In this case, what matters are not the characteristics of a product per se, but the perception that the community has of this product. There are many ways to measure this perception, such as asking users to rate items, probing the merchandise that shoppers check on the ecommerce platform, and so on. The main advantage of this system over the previous one is that collaborative filtering does not require a tedious description of each item in the online store.

However, machine learning-powered recommendation systems using this approach have their flaws too.

• Cold start: once again, it may be difficult to provide recommendations for users with limited or no purchase history.
• Scalability: relating such massive amounts of variables (basically millions of customers and millions of products) requires considerable computational power.
• Sparsity: the most important ecommerce corporations offer a huge variety of merchandise. This means that there may be a relatively small amount of reviews for each product and therefore a not-so-accurate recommendation choice.

More ways to tailor online shopping

When it comes to personalizing the shopping experience and connecting customers to their favorite merchandise, the role of machine learning in ecommerce is not limited to recommendation engines. As you may expect, the same logic examined above can be applied to other aspects of marketing.

• Targeted advertising: Before offering products to users of an ecommerce platform, it may be necessary to attract these users to the aforementioned portal. To accomplish this, potential customers can be targeted with personalized ads based on their social media interactions, previous online purchases (including those made in virtual stores), Google search history, and other types of big data used in ecommerce. As previously described, all this data can be processed by machine learning algorithms to analyze users' behavioral patterns and forecast the products or services they may like.

• Smart search engines: How could customers reach the desired products without an efficient search engine? This may sound like a minor problem until we are dealing with a huge stock of merchandise with literally millions of products to choose from. Traditional systems rely on matching the keywords used for the search and those chosen to describe the products in stock.Machine learning algorithms, combined with other AI-related technologies such as natural language processing, can empower search engines with a deeper degree of context understanding. For example, a machine learning-driven engine can take into account a broader range of synonyms. It may also adjust the search process in real time by prioritizing some results according to each user's purchase habits and taste.

A couple more noteworthy use cases

Let's conclude this overview with two additional ways to take advantage of machine learning in ecommerce, related to logistics and anti-fraud measures respectively.

Restocking and delivery optimization

When it comes to the logistical aspect of ecommerce predictive analytics, machine learning may play a double pivotal role. First, its forecasting capabilities can be leveraged to match supply and demand, particularly by predicting upcoming sales trends and setting up proper restock planning. Therefore, if your ML-solution is integrated with different types of POS, it can analyze inventory data, and, in case some product is going to run short, notify a team that they need to replenish stocks.

In this regard, you might think about the interconnection between purchases and holiday seasons, but it's not always that simple. Luckily, machine learning-driven systems can certainly have a broader picture of what is going on compared to "mere mortals", as they’re able to take into account an immense number of variables (such as economic trends, customer behavioral patterns, social media sentiment, and more).

Second, machine learning in retail proved to be a useful tool for delivery optimization, processing data from prior cases to select and suggest the best expedition methods and conditions according to customer expectations. For example, a machine learning-based eсommerce platform may decide to offer free or one-day delivery as an incentive to purchase if this option is valued as a deciding factor.

Fraud detection

As the world abruptly shifts towards a full digitalization, online fraud and other cybercrimes become an increasingly widespread reality. In such a context, ecommerce platforms can be easy prey for hackers and scammers.

That's why many companies have already turned their gaze to machine learning as a potential weapon to protect their ecommerce portals from a wide range of criminal actions, including identity theft and fraudulent electronic payments.

As we've said earlier, machine learning algorithms can easily spot recurring patterns among the datasets they process. But this implicitly means that they can also notice if something "breaks the rules". Indeed, machine learning in fraud detection is commonly deployed to detect anomalous behaviors among the credit card accounts under scrutiny (such as an increasing frequency of transactions) that may be signs of fraudulent attempts.