How can Retailers Leverage Product Recommendation to Deliver Personalized Customer Experience?
Matching the Right Customers to the Right Products
In an e-commerce setting with hundreds of thousands of products, it can be difficult for customers to find what they want. Customers often have to guide themselves through a multi-layer menu of categories and scroll through the page to reach the right products. In an effort to drive up sales revenue, marketing executives will carefully hand-pick popular products to show on the landing page, with the hope that these products can appeal to as many customers as possible.
Is there a more efficient way of presenting your products?
Unlike in a physical store where you must display the best products with careful visual merchandising, your e-commerce website can be personalized and display a different set of products to each customer. This is an efficient use of website space as customers do not have to see products they are not interested in and is a great way of achieving precision marketing.
At its core, product recommendation is the act of matching the right customers to the right products. A recommendation system performs complex algorithm to extract pattern from customer data and behaviour and predicts customer interests. Studies show that it can increase conversion rate by as much as 70%, and uplift by as much as 3 times.
Product Recommendation as Personalized Customer Experience
A retail company’s ability to deliver personalized customer experience is a vital part of staying relevant in today’s world. According to the Sailthru Retail Personalization Index, 71% of customers agree that personalization is important, and 80% are willing to share data in exchange for deals and offers that deliver a personalized customer experience.
There is also an interesting gap where Sailthru found 71% of retailers think they have done a good job at personalized marketing, while only 34% of customers agree. This indicates a lot of retailers realize that personalization is important, but they simply do not know how to do it right.
Setting up a recommendation system is a great way to show you know your customers well and to win loyalty. A good recommendation system can cater to the entire customer journey starting from awareness to retention.
For a first-time visitor who just came across your website, there is little personalization to be done. This is known as cold start problem for new users. It is reasonable to recommend the most popular items to maximize the probability of appeal. On the other hand, your recommendation system should be able to learn the needs and preferences gradually of the new visitor like a human salesperson.
A personalized experience also means providing recommendations via different channels according to customer preference.
What are the Product Recommendation Algorithms?
Product recommendation is an active area of academic research. There are numerous types of algorithms, and the three most popular ones are content-based recommendation, collaborative filtering-based recommendation, and hybrid recommendation system.
Content-based filtering utilizes the content information such as product attributes to make recommendation, while collaborative filtering one looks for people that have purchases the same products to make recommendation.
Content-Based Filtering
Given that we know what a customer already likes, we can identify products that are similar to those products and make relevant recommendations. This approach is called content-based filtering.
There are a lot of user preference signals. Some can be “implicit”, for example, a transaction, an add-to-cart or a session click. There are also more “explicit” signals like a product rating from the customer, or when a customer selects a particular product category in a survey.
A very common approach to measuring product similarities is by their categories, brands, and many others. For example, if a user has browsed the following shoes on the website, it is a clear indication that she is interested in flats suitable for a working environment.
It is very straightforward to recommend relevant products to users with this method, and requires no data from other users, making it very easy to scale. However, you might have noticed that we are not cross-selling any product to the user and she will only see more similar shoes in the recommendation. This method also requires manual input of domain knowledge regarding the products, for example, a pair “flats” for “business occasion”.
Collaborative Filtering
On the other hand, collaborative filtering is a more sophisticated method that relies on the similarities between both the users and the products.
Using the previous example of shoes, say now we would like to know whether user D will like a pair of Flats. From the user preference data that we have, we know that user D and user B are quite similar — they both like Heels and Loafers. We also know that user B likes Flats! It is then very reasonable to predict that user D will also like Flats, which we will recommend to her.
As we can see, we did not rely on any intrinsic information regarding the shoes themselves. We did not input any explicit domain knowledge. Our recommendation is solely based on user B’s preference on Flats and her similarity to user D.
In practice, our model can automatically learn the embeddings of each product and each user based on the user-product interaction data. Then, we can very easily compute the probability of a user liking a product by simply calculating the similarity between the user embedding and the product embedding.
This is a very powerful algorithm as we do not need any domain knowledge. This model is capable of giving diverse recommendations and helping users discover new interests.
However, this model cannot handle new products because there is simply no interaction between any user and this product. This is known as the cold start problem for new items. The model can also take a long time to compute as the product catalog and user base get larger.
Comparison
The following table summarizes the key differences between content-based and collaborative filtering recommendations.
These are very flexible algorithms with a lot of variants, and can operate on various data such as:
- Transaction data
- Website click data
- Demographic data
- Location data
- Traffic source data
- Product information
Value your Customer’s Data
Recommendation models are data-hungry and require a lot of customer personal and behavioral data in order to accurately predict customer interests. It is important not to take it for granted and to value your customers’ privacy. A study by Accenture found that apart from receiving relevant recommendations, most customers also expect retailers to be transparent about how their data is being used and to not leak or abuse their data. Retailers will need to realize that this is a two-way exchange of data for service.
Conclusion
Product recommendation is a vital part in creating a personalized experience for your retail customers and utilizing your marketing resources efficiently. As the Age of Big Data gives us more and more data, recommendation system will become more sophisticated. The quest of dealing with the increasing computational complexity and ensuring data privacy will remain an active area of research in the following years.
