6 Ways Retailers Use Machine Learning
Machine Learning, a sub-field of artificial intelligence (AI), is a field that is at the forefront of computing today. In fact, it is omnipresent in the computing world! Even if you’ve never heard of Machine Learning before, you use it many times a day without probably even realizing it!
Simply put, Machine Learning, or ML for short, is a set of statistical modelling techniques, which applied to data, use algorithms to detect behaviors in the past and apply that knowledge to situations in the future. The concept of ML, in essence, is very much like we humans learn. Think of how we learn to walk.
As infants, we found it difficult to walk, but as we grew, we learnt from our experiences of falling and taking bad steps. But eventually, after some time, our brains learn to walk comfortably, and thus we become the faultless walkers that we are now. Similarly in ML, such experiences serve as data, for e.g., data from a car’s sensors as it learns to drive. Thus, ML uses past data to make the computer learn. Based on this “learning”, it makes predictions about future events.
Machine learning is being used today to make driver-less cars, predict emergency room waiting times, identify whales in oceans based on audio recordings so that ships avoid hitting them, make intelligent recommendations on which movie one should watch next on streaming services such as Amazon Prime and Netflix, and provide personalized shopping experiences in online retail.
Machine Learning In The Retail Industry
Online retail, like traditional retail involves taking in a lot of data and trying to make the best decisions based on the data that you have. In a retail store data is relatively limited, you know how many customers come in the store each day (or are able to estimate it), but it’s hard to keep track of how many customers try things on, how long the average customer stays in a store, or how many items they look at during each visit.
Unlike brick-and-mortar retail where the complexity comes in when you’re trying to collect data, online the complexity is making sense of all the data this is coming in, it’s like a fire hose and if your business is growing, so is the data. Advanced machine learning algorithms are key, as they help the retailer to use this data to deliver against specific KPIs, learn from patterns to predict demand, automate decisions and deliver profitability.
Machine learning is enabling retailers pinpoint critical action areas within an avalanche of possibilities, analyse and consume data at scale, and enabling retailers to look into the future: to plan for scenarios and consider options. To get a better idea, let’s take a look at how retailers are using machine learning.
Matching Products With People
People engage with brands in infinitely different ways, and they certainly don’t fit neatly into standard audience personas. It has historically been a challenge for marketers to predict what campaign strategies will be most effective for each individual customer. Machine learning is being used by retailers to identify patterns in purchase behavior to develop relevant offers that are predicted to result in high response rates.
Retailers, by and large, already own large sets of customer and transaction data. By applying predictive models to these datasets, which can include campaign, transaction, response, on-boarding, social media and even sales data, retailers are able to gain foresight into how customers may respond to marketing, product or other stimuli they experience. As new data comes about through market engagement, retailers are able to make near real-time adjustments to campaigns and strategies and achieve better results through more personalized and targeted marketing.
More granular and dynamic audience segments can be discovered and defined through machine learning capabilities. Target’s infamous machine learning incident where the company learned of a teenage girl’s pregnancy before her own family even knew, is a case in point.
Product recommendations — “if you like product x, you will probably also like product y” — have been shown to work remarkably well, and serve as a valuable tool to guide users through the ever-increasing options available to them. Traditionally, recommendations have been added by hand based on hard-coded product categories, but this is extremely time-consuming, error-prone, and quickly out-dated.
Modern recommendation algorithms are classified into two categories:
- collaborative filtering In collaborative filtering, recommendations for a given customer are based on what similar customers have chosen in the past. Simplifying somewhat, if customer A has bought product X and Y, and customer B has only bought product X, then product Y could be recommended to customer B.
- content-based filtering. In content-based filtering, recommendations try to match the content of customer profiles (e.g. gender, brand preference, or age) to the content of products (e.g. category, price, or color).
Studies from companies like Forrester and Gartner have found that a product recommendation strategy, if deployed effectively, can increase revenue by up to 300%, boost the conversion rate by 150% and the average order value by 50%. Little wonder that 70% of Amazon’s homepage is dedicated to product recommendations!
Different product recommendations can be aligned to different stages of the customer journey. For example, new customer acquisition can focus on special offer product recommendations, while existing customers can be targeted with recommendations based on past purchases.
Cross-selling and Up-selling
Machine learning, along with analytics, helps simplify retail operations at the back end and cross-sell and up-sell. It does a lot of the heavy lifting and through analytics, retailers can find useful information to
- Offer best prices to customers based on past purchases
- Suggest complementary products for current purchase
- Recognize when a consumer might need a past product purchased again and provide more context-aware choices to them
This provides many cross-selling and up-selling opportunities to retailers with a higher chance of conversion into actual sales.
Airlines were among the first companies to embrace the concept of automatically adapting prices. On a basic level, this can simply mean to increase prices when the demand is high and decrease them when the demand is low. But there are plenty of other variables that can also be used to estimate optimal prices, such as prices of competitors, time of day, warehouse stock, or season.
Pricing algorithms cannot be painted with a broad brush and need to be adapted for specific products to accommodate factors like marketing strategies (e.g. whether to give particular competitor prices more weight than others, or whether or not to keep prices low after introducing a new product to boost sales.
From the perspective of the customer, these techniques can have positive as well as negative effects. Some customers will regret their purchase and stop using a shop when they see the price dropping only minutes later. For other customers, dynamic prices can turn out to be an exciting game in which they can try to hunt the best prices. It remains to be seen what the large-scale effects will be, but if the rate at which retailers adopt these new techniques is any indication, then dynamic pricing is here to stay.
The more data you have, the harder it is to check for inconsistencies. One way to handle this problem is automatic anomaly detection. The idea is to have an algorithm that can identify patterns in the data to learn what is ‘normal’ and then send alerts as soon as data points exceed that range. From the machine learning perspective, the main challenge of this problem is to train a robust model despite having a heavily imbalanced dataset, since there are far fewer cases labelled as ‘anomaly’ than ‘normal’.
A popular application of this approach is fraud detection. Retailers frequently have to deal with abusive customers that use stolen credit cards to make excessive orders, or customers that retract payments via their credit card company once products have already been delivered. Besides cases of fraud, anomaly detection can also be used to ensure a high level of data quality for product information. Large databases in the e-commerce sector often contain errors like incomplete product titles, missing images, or products sorted in the wrong categories. Detecting these cases quickly and efficiently can therefore save companies a lot of time, money, and effort.
Trying to get help when you have trouble with a service can often be a frustrating experience. Customers frequently complain about exceedingly long waiting times, having to explain and re-explain their problem multiple times, unqualified advice, or stressed out employees.
Machine learning can help automatize this process through robots that can answer phone calls. Whereas previous systems were only able to deal with a narrow range of problems and had frequent misunderstandings, recent advances in speech recognition and natural language processing via deep learning have made it possible to have a more flexible and natural interaction with robots.
These methods have shown improvements in taking contextual information into account. Instead of analyzing a speech sound or a single word in isolation, modern approaches take information from the whole input into account and compare it against frequently occurring patterns, which has boosted the accuracy of machine learning models.
Machine learning can also add to other support channels, such as automatically answering emails, categorizing emails (e.g. complaint vs. question vs. request), or providing support via chat bots. Chat bots in particular have inspired a variety of AI startups that want to revolutionize communication channels for marketing, consulting, or recruiting.
The applications mentioned above are only a small selection of what machine learning can do for retail, but there are plenty of other options, such as:
- Customer Segmentation: Identify systematic groups of customers to make marketing more precise.
- Product Categorization: Automatically sort products into categories to speed up inventory management and improve customer navigation.
- Churn Prediction: Predict when customers will stop using a service to analyze potential reasons and allow for countermeasures.
- Sentiment Analysis: Evaluate the public perception of a product based on sources like social media.
- Inventory Forecasting: Make production and distribution more efficient by predicting market demands.
Based on the vast number of possible applications, it is important for companies to define strategic priorities to get the most value out of machine learning.
While it might sound dramatic, it’s no secret that retail is in trouble and those that want to survive in this brave new world of retail need to embrace AI and ML technologies, NOW. Retailers have struggled to get comfortable with AI and ML technologies which take time to learn and tune. We’re entering an era of machines that think for themselves and learn over time. The key to success lies in understanding that it’s a long play, not a quick win.