This blog post was originally published at Tryolabs’ website. It is reprinted here with the permission of Tryolabs.
Artificial intelligence (AI) and machine learning (ML) significantly impact the retail world, particularly for companies that rely on online sales, where using some kind of AI technology is very common nowadays.
Big players like eBay, Amazon or Alibaba have successfully integrated AI across the entire sales cycle, from storage logistics to post-sale customer service.
However, you do not have to be a big company or sell exclusively online to take advantage from the tremendous power of Machine Learning.
In this guide we will see how online retailers and brick and mortar stores of any size can integrate Machine Learning technology to stay ahead of their competitors, by increasing sales and reducing costs.
From clothes to groceries to household items, the possibilities in the retail space are full of promise. The use cases presented in this guide are a fraction of the feasible Machine Learning projects and serve as examples of what can be done today in the Retail space. That being said, many companies have very unique needs that could be served with data and custom Machine Learning development. We hope you find this guide useful.
Retail intelligence
Pricing strategy
How you price an item and scale pricing operations intelligently, consistently, and efficiently is critical for retail and e-commerce businesses. AI and data intelligence dramatically improve pricing decisions.
At Tryolabs, we have created AI solutions for retail that come in different flavors depending on the pricing strategy objectives.
Auto-pricing
Sometimes, the price of an item is defined by a person based on how she values a given item. This is usually the case for shopping and resale platforms. This manual process sometimes leads to inconsistent results, as different people value the same item differently. The human bottleneck is also a factor: every piece published on a site needs to be priced by a human, impacting costs and publication speed. It is also difficult for a person to consider the various factors influencing demand and price.
At Tryolabs, we have extensively worked in Machine Learning models that automatically and consistently price every new item by looking at data. The model leverages image recognition technology to understand the product’s look and consider other aspects such as brand, condition, fabric, category, description, historical sales data, and past pricing decisions of similar items. With these data points, it computes an optimal price for each SKU. It is also possible to expand the model to include other features, including competitor price if it’s available.
Some of our largest customers have used such models to automate the vast majority of their pricing operations and scale to more items effortlessly.
Pricing optimization
Pricing is a critical predictor of profitability. Based on econometric science, a Machine Learning algorithm can take key pricing variables into account to define an automatic pricing strategy with real-time, dynamic prices.
The willingness to pay can be estimated from the customer’s behavior, for example, considering the items they look at and purchase, or the time they spend on each web page.
The goal is to understand how clients react to items at different price points and estimate the right price for each product if they want to sell it in a certain period, ultimately maximizing profit.
The model forecasts demand and considers many factors to advise on dynamic price adjustments. By including supply, seasonality, external events related to your business (e.g. a concert, a match, a festival) and market demand and offer, an automatic pricing system can test and learn from past experiences which is the most profitable or estimated “optimal” price.
Such a system needs historical sales data to be trained, but unlike auto-pricing, the system benefits from variability in the past prices used for each product. This is because it needs “exploration” in order to estimate the price-demand relation (usually referred to as demand elasticity).
Suppose there’s no variability in past prices. In that case, the first iterations will be mainly in “exploration mode,” which consists of introducing some randomness to the prices and evaluating how the sales data for each product reacts.
Significant deviations from previous prices are not needed, the range can be expanded gradually as the system gains information about what’s the right direction to move in each case.
In our experience, there’s an enormous potential for improvement in the pricing strategy compared to other approaches, and the benefits are usually for both sellers and customers, since low prices in some products can end up driving a surprising increase in profitability.
Product matching
Product matching is the process of identifying and linking identical items between two or more catalogs.
Some of the reasons why you might want to do that linking are: price comparison with competitors or different providers of the same products, merging multiple offers from different sellers into a single product page (e.g: Amazon’s buy box), among other less obvious uses like performing assortment gap analysis against your competitors.
To understand the business value of solving this problem, let’s consider the first example mentioned above. In retail, 96% of customers compare prices before making a purchase. Applying product matching allows any business to track the price their competitors are offering for the same products.
As businesses grow, product catalogs tend to become large and messy. If you’re lucky, products sometimes have codes that uniquely identify them and make the matching trivial. But more often than not, these codes are not available or sellers do not expose them. Sometimes, the codes are even incorrect! Another thing that may happen is that the products may indeed be the same, but their presentation may not. For example, single bottle vs. pack of 6.
Because of all this, you may be left with a complex problem: using whatever information of the products is available to do the matching. This can be a product title, a free-text description, some other metadata, and images.
This matching task can be done manually up to a point, but it escalates out of control rather quickly when catalogs grow in size. Even worse if you are matching multiple catalogs (multiple competitors or providers).
In our experience with retailers, we have faced this problem quite often. From this, we have built a solution that leverages AI through natural language processing (NLP) and computer vision techniques to find the most similar pairs by comparing each product’s image or unstructured text fields, like title, attributes and description.
In most cases, it’s not possible to guarantee that the matching pairs are perfect, so we use a human-in-the-loop process in which the task becomes a quick verification of the suggested matches that gets progressively easier with time. This is how it works:
- The first batch of manually verified matches is leveraged to re-train the system (using them as a list of “verified matching examples”).
-
- The AI learns what features it should use to identify actual matches for each type of product.
- Now we can perform a second iteration for the remaining items and the proposed matches will be more trustworthy.
- The system is re-trained again.
This process can be repeated as needed depending on the confidence achieved and requirements of the particular problem.
Marketing Campaigns Optimization
Like pricing strategy, Marketing Campaigns, besides being essential for retail business, are complex and require a deep understanding of the market. Machine Learning models augment the decision-making process by leveraging historical data to forecast the ROI and provide optimal parameters for its execution. The model predictions aid decision makers in the process of balancing the costs and profits when designing the campaign.
At Tryolabs, we have created Machine Learning models customized to target specific business metrics when designing Marketing Campaigns such as the net revenue or margin contribution. Primarily we have worked on optimizing product promotion and replenishment campaigns.
As these solutions are employed, they produce more data so that we can evaluate them by measuring business KPIs. This in turn produced more data for improving the models, eventually creating a virtuous learning cycle.
Product Promotion optimization (discounts to offer, similar to pricing optimization but not the same)
Product promotion campaigns can have different flavors, for example, offering discounts to increase the net revenue. However, how can we optimize the discount offers provided so that margin contributions are maximized and therefore, maximize the overall net revenue?
For this task, we can define a function modeling the discount offering problem and optimize it to maximize profit using Machine Learning. We then combine demand forecasting together with the unit costs of each SKU as well as their price data to train a model that recommends the decision makers what discount to offer according to the business goals.
By employing Machine Learning to optimize Product Promotion Campaigns, Marketing Teams are able to use their time more efficiently and use it where their creativity and experience can be leveraged the most.
Product Replenishment optimization (recommend what to buy for clients in b2b2c)
In retail, especially in the case of B2B2C businesses, the goal of Product Replenishment Campaigns is to increase repeat buys and decrease purchasing cycles by using specialized recommendations. Achieving these goals requires a deep understanding of the market and a thorough analysis to target the campaign successfully.
By utilizing market and business data, we employed a customized Machine Learning Clustering model to perform customer segmentation and enable the business to optimize their recommendations to their buyers with better targeting. This did not just increase their own net revenue, it also increased their buyers’ for the case of B2B2C businesses. Moreover, this kind of customer segmentation can be used for better targeting other Marketing Efforts downstream.
Predicting customer behavior
The goal of a predicting customer behavior system is to estimate how buyers will behave in the future based on data of previous behaviors. These systems allow retailers to segment customers and perform personalized marketing actions that are more effective than general approaches. Moreover, taking actions based on predicted customer needs increases loyalty and retention.
A typical application is to predict purchases. For example, to know which customers are likely to make a purchase in the next seven days. More complex predictions may have to do with important events in people’s lives. For example, to predict marriage or pregnancy, and then send custom offers.
The predictive models basically need consumer behavior data. That is, for example, purchasing history or buying trends, but it could also include social media activity and domain specific knowledge.
Predicting the needs of consumers is a challenging task where Machine Learning algorithms are of great help. How often do your customers make transactions? Do they buy during sales time or before their birthday? How many items do they usually buy? What are they currently buying? What topics are they talking about on social media? All this kind of information is used by the models to predict future behavior.
And last but not least, retailer experience is very important to choose business specific criteria and fine tune the models.
Retail Stocking and Inventory
Optimizing inventory planning and predictive maintenance is a key issue and a very important logistic concern for retailers. Moreover, this could affect the UX of end users of e-commerce platforms if a product they buy is actually out-of-stock and their order cannot be completed.
Predicting inventory
Machine Learning algorithms can exploit purchase data to predict inventory needs in real-time. Based on the day of the week, the season, nearby events, social media data and customer past behavior, these algorithms can provide a daily dashboard of suggested orders to a purchasing manager.
From a different perspective, we have built similar solutions to predict if an item a user is purchasing might be out of stock in scenarios where the e-commerce platform is not managing inventories directly. A predictive model leverages historical data such as past stockout events, demand, price, stock level for each product. It also uses real-time data such as price updates and date-time features to make accurate predictions. In this way, e-commerce platforms could advise users their items might be out of stock in a timely manner.
The inventory planning models need consumer behavior data. That is, for example, purchasing history or buying trends, but it could also include social media activity and domain specific knowledge.
Also, the algorithms used for pricing optimization, which need a sales forecasting model (as a function of the price, in order to calculate what’s the optimal one) can also be leveraged for stock management purposes, and they usually work integrated in order to avoid prices that will create an early out-of-stock event (since that’s always a suboptimal solution).
e-Commerce
Tagging and copywriting automation
Searching for items on a crowded site can be a daunting experience. That is why retail teams spend days writing descriptive copy for items and tagging them according to their primary attributes and preferred taxonomy. Traditionally, this is a manual process that relies on each person’s time and criteria, often resulting in a lack of consistency in data.
With the traditional manual process, it is common to face some of these issues:
- Inconsistent descriptions among similar items (some too long and detailed, some too short)
- Missing attributes
- Lack of human alignment in criteria for specific attributes or values
Tryolabs has developed an AI solution that uses computer vision technology to automate item tagging and descriptions, obtaining consistent and standardized data. The information comes directly from the image. The machine automatically fills product attributes and tags them accordingly to sit in the correct category and hierarchy on the website by simply uploading the photo. This can be tuned for each particular taxonomy unique for each retailer.
Image retouching
Images ensure online shoppers get as close as possible to fully understanding product features. Professional-looking photos can give an edge to an e-commerce site, making customers more likely to buy, and brands happier with the portrayal of their products.
Usually, pictures are taken in a photo studio and then sent out for retouching. The retouching process includes several different steps, depending on each particular site. The most common ones are calibrating colors for accurate depiction (you don’t want customers returning items because the colors don’t match their expectations!), removing backgrounds, aligning/rotating items, smoothing creases, and erasing mannequins. Unfortunately, this process can take several days, delaying the whole item uploading process.
But AI is increasingly used to cut corners and automate most of this time-consuming process.
Tryolabs has worked on automated image retouching using AI that cuts both time and costs to a fraction. The system replicates the traditional “Photoshop pipeline” done by artists but on a much larger scale. The opportunity is huge: one of our largest clients reduced the turnaround time for retouching hundreds of thousands of images from three days to only 30 minutes.
Browsing experience
Browsing online catalogs can be exhausting like a needle in a haystack situation, or an enjoyable process in which you may end up learning a lot more about the products you’re looking for. It depends mostly on the relevancy of the results that the system provides while you interact with it in various ways. The problem is known in computer science as Information Retrieval, and it’s not limited to introducing a search term and hitting Enter. In general, we can think of it as: given a query, provide the most relevant results, usually defined -in this context- as those which the user is most likely to click, so it’s evaluated by measuring the click-through rate.
Let’s see some examples that follow this pattern, in which there’s a query (e.g: a search term or product that the user is currently looking at) and a list of results that might be relevant to the user. The interaction with such elements ends up defining to a great extent the browsing experience for the users.
Visual search
Customers tend to search for visual content prior to making a purchase. However, in some cases they cannot easily find good keywords to describe what they want. The goal of visual search is to make it much easier for consumers to find exactly what it is they’re looking for.
Instead of typing a query such as ‘cordless combo kit with soft case’, which will probably return a lot of general results, customers can upload an image to help narrow the search down to more specific items. With the huge and increasing amount of snapping and sharing images, Machine Learning algorithms can currently achieve amazing results.
This is one of the most trending use cases in online content. Leading companies like Microsoft, Google or eBay have presented in 2017 Bing Visual Search, Google Lens and Image Search.
Regarding the e-commerce market, Pinterest introduced Lens Your Look, a visual search engine that allows you to find outfit ideas inspired by your wardrobe. So if you are looking for new ways to wear your favorite jean or blazer, you can add a photo of it to your search to find outfit ideas that you eventually can buy.
The solution to this problem is actually very related to finding similar products (see next section), because the visual search operation consists of finding the most visually similar products to your query image, hence it’s a sub-problem of the next one. If you have the best deep learning model to perform visual search, you also want to use it to find similar products in general, maybe adding other attributes and text descriptions.
Similar products
As mentioned above, finding similar products makes use of the visual information (if it’s available), but also the text information and relevant attributes.
The background models are also similar to those used in product matching (described above), but the purpose here is not to find exact matches, but those that are visually similar and also have certain attributes that make them similar.
The exact features that need to be prioritized in order to suggest similar products, depend a lot on the use case and the business strategy. For example, some retailers may want to prioritize products from the same brand or manufacturer, while others may prefer the exact opposite, and show diverse brands when there are many similar items for the same product category. Another example is the color in the case of fashion, is it important that the suggestions have similar colors? Those things can be configured as needed, using heuristics defined with some business strategy in mind, or leveraging domain knowledge from the clients.
Having a similar products system deployed is also a robust step before a good recommendation engine. The reason is that in order to retrieve similar products, there’s no need for prior user interaction data. But it’s precisely a way to generate that data, because these suggestions will result in a lot of interaction between users and products, and that information can be then used to train a recommendation system that takes into consideration usage metrics like click-through rate.
Recommendation engines
Many e-commerce and retail companies are leveraging the power of data and boosting sales by implementing recommender systems on their websites.
Companies using recommender systems focus on increasing sales as a result of very personalized offers and an enhanced customer experience. Recommendations typically speed up searches and make it easier for users to access content they’re interested in, and surprise them with offers they would have never searched for.
What is more, companies are able to gain and retain customers by sending out emails with links to new offers that meet the recipients’ interests, or suggestions of films and TV shows that suit their profiles.
The user starts to feel known and understood and is more likely to buy additional products or consume more content. By knowing what a user wants and showing it right away, it is less likely that it leaves the platform. This translates into a higher chance of purchase and a decrease in the threat of losing a customer to a competitor.
Providing that added value to users by including recommendations in systems and products is appealing. Furthermore, it allows companies to position ahead of their competitors and eventually increase their earnings.
Catalog tree optimization
Ecommerce platform catalog trees serve various purposes including facilitating search by buyers, categorizing products for the sellers, and allowing the platform to personalize the experience and provide better recommendations for the buyers among others. Additionally, they enable better SEO.
The main issue platforms face comes from dealing with uncategorized products which end up in overdimensioned Other categories. These end up being the go to option for users categorizing products that don’t fit the existing tree, be it because the category was overlooked or it is a new kind of product. This in turn makes it more difficult for users to find the products they are looking for, it affects the website’s SEO, the buying user experience, and eventually the sales volume.
With a combination of recommender systems and NLP techniques, we built a solution that allows ecommerce platforms to reduce the size of their Other categories by predicting the category of new products for users and recommending operators new categories to move uncategorized products over to. This brought better SEO, better discoverability of items, better buying experience, and therefore, an increase in revenue.
Brick and mortar
Behavioral Tracking via Video Analytics
A good thing about physical stores is that the behavior and interaction of humans with products can generate valuable insights in ways that online retail cannot. Computer Vision algorithms can recognize faces and people’s characteristics such as gender or range of age, generating precious exploitable data.
Human flow counting
Being able to measure human flow in stores allows retailers to know not only the effectiveness of their stores when it comes to converting customers, but also the distribution of people entering the stores during the day. This is useful to optimize the use of operational resources and provide better customer service.
Using computer vision algorithms, we built a video analytics solution that counts human flow in stores – the number of people entering and exiting physical stores at each given time, based on videos from pre-installed security cameras. This allows retailers to leverage their existing CCTV infrastructure to gain further insight into their business with the aid of Machine Learning.
Analyzing navigational routes
Where to put different items is a crucial matter for physical retailers, who always look for additional ways to understand the customer’s path to purchase.
Computer vision algorithms can track customers’ journey in stores to understand how they are interacting with it. These algorithms can detect the walking patterns and the direction of the gaze of the customers. Retailers can use this information to restructure store layouts or to measure the interest in their products. They can also discover locations that get a lot of traffic and visual attention.
Do elderly people shop more on weekdays? Do teenagers tend to cover only part of the store, for example the front part? Is the store more visited in winter? Variables such as age, day of the week or season could be used to generate insights that help to dynamically modify product placements and create efficient promotions.
Theft prevention
Theft Prevention is a common problem in retail with a strong ROI, where Machine Learning technologies can go beyond the typical use of video cameras to detect shoplifters.
Facial recognition algorithms can be trained to spot known shoplifters when they enter the store. Walmart has tested this technology in 2015 as an anti-theft mechanism. In the same way computer vision can detect if someone picks an item, it could detect if someone hides an item in their backpack or jacket. Moreover, the same approach can be used to detect when checkout clerks skip scanning items, either inadvertently or on purpose.
A system based on Machine Learning can alert in real-time security personnel or managers and send them video excerpts so they can judge by themselves before confronting the individual in the store.
Product tracking and gesture recognition
Brick and mortar retailers usually have no information about the items that customers pick up, glance at, and put back on the shelf. They do not have any information either about what customers look at next.
A computer vision algorithm can monitor shoppers’ facial and hand gestures to estimate how interesting an item is. These kinds of applications generate precious data about how many times an item is picked up from the shelf, and then purchased or put back on the shelf.
Conclusions
The retail industry keeps growing and finding new challenges in an ever-evolving consumer paradigm. Serving these necessities with state-of-the-art technology can determine whether you succeed or become outdated.
At Tryolabs we help retailers leverage the power of AI by providing expertise in all the phases of AI initiatives, from ideas to execution: from definition of opportunities in an AI Roadmap, up until implementation in production, iteration and improvements.
We build data-driven solutions to improve company’s KPIs. This means that we partner up with retailers to conceptualize & build machine learning systems that either increase revenue or reduce costs.
If you’re thinking of ways to integrate AI into your strategy, get in touch with us and we’d love to discuss your case.
Javier Couto
Research Scientist, Tryolabs
