The rise of e-commerce:
A growing number of customers are switching to e-commerce for their retail needs instead of brick and mortar. This change is driven by the convenience of making decisions from the comfort of one’s home, access to products not available locally, and the time it saves. With growing customers, e-commerce firms are multiplying and the competition in the e-commerce industry has also increased significantly. Machine learning in the form of recommender system algorithms are widely used in these e-commerce firms, which aspire to give as accurate recommendations to their customers, with regards to their preferences, with as little information as possible to gain a competitive edge. These recommender systems not only influence user acquisition and user engagement but also directly affect the revenue and the survival of the company.
Importance of recommendation systems:
The main objective of a recommender is to know a customer and tempt that person with an offer he or she can’t refuse. The importance of recommendation systems is also evident by the number of research papers on the topic of “recommendation system,” which has been increasing exponentially. A few years ago, Netflix awarded prize money of $1M USD for a better performing recommendation algorithm, and in Netflix’s own words, “The $1M Prize delivered a great return on investment for us.” This highlights the importance of recommendation systems in surviving in today’s cut-throat competition. Realizing the value of recommendation systems, they are being widely applied in the domain of e-platforms and are increasingly being adopted by upcoming e-commerce firms.
Why incorporate a recommender system?
A recommender system (RS) can help achieve various objectives that are relevant to any e-commerce firm. A recommender system can typically improve the user experience and increase revenue in the following ways:
- Increased conversion rate: A good recommender system makes an offer that the customer can’t refuse and increases the conversion ratio (refers to the ratio of purchases made and the items browsed) drastically.
- Increased customer loyalty: If a user realizes that the recommendations made on a particular e-commerce platform are of optimum quality and are suited to their likes, the user is more likely to trust it in the future too and become a loyal customer.
- Improved user experience: Making recommendations can also save time for customers who might be looking to buy products that they might not know a lot about, as when, for example, a customer is looking to buy a camera for the first time. In this case, recommendations can list the technical features of cameras in a similar price range from different companies. This makes it easier for the person to make an informed decision.
- Understanding customer base: RS can also identify popular items in certain areas or demographics. This can help e-commerce firms make recommendations for new customers for whom there’s no other known information.
A lot of recommendation systems have been developed for e-commerce over the past two decades. Some of the successful use cases of recommender systems can be found in today’s top companies, including Netflix, Amazon, YouTube. Since the most common types of recommendation systems are content-based, collaborative (user-based or item-based), hybrid, demographic-based, and knowledge-based, these will be briefly described.
Content-based recommender systems:
This recommendation system depends solely on the similarity of items bought or browsed in the past. The content-based recommendation system is made of three component processes: a) a content analyzer that extracts features of the item/product, b) a profile learner that describes the interests of a user, and c) filtering components that match the user profile (created from a profile learner) to an item based on its attributes (resulting from a content analyzer).
A typical example of a content-based RS was developed for the recommendation of coupons on Rakuten (an e-commerce marketplace) to recommend personalized coupons to customers based on their shopping history. Another interesting application of the content-based RS is one of the entries of the Netflix Prize competition that employed an innovative method to estimate the similarity of the movies by using the associated Wikipedia pages. This was used in combination with the user ratings.
Collaborative filtering:
A collaborative recommendation system relies on users who have bought or browsed similar items in the past (user-based) or the similarity of items that the user bought/browsed in the past (item-based). Collaborative filtering RS can be broadly classified as: a) memory-based, also called neighborhood model or “word of mouth,” or “people-to-people correlation,” and b) model-based. These two types differ in their approach. The memory-based approach can only make recommendations based on the memory of the user actions; that is, it will recommend items that have been rated well by users similar to a given user (user-based) or items similar to the items that a given user has already rated well (item-based). The model-based approach doesn’t depend solely on the memory of rated items.
Collaborative filtering RS is the most commonly used technique. There are many examples of successful collaborative filtering RS that work on customer data in order to make recommendations. Some popular examples of collaborative filtering RS are the movie-recommending site MovieLens and email-filtering Tapestry. A community-based collaborative filtering RS was created for sites like Last.fm, Delicious, and Epinions that allow users to form social communities on their platforms. In this case, the recommendations are made based on community attributes and is found to improve on the traditional collaborative filtering RS.
Hybrid and other recommendation systems:
Hybrid RS combines two or many of the above-mentioned approaches to achieve better performing RS, for instance, demographics-based recommenders or knowledge-based recommenders. As the name suggests, a demographic recommender uses information on user demographics to make recommendations. The idea here is to use demographic information to classify similar users and then apply the user ratings to make recommendations. Knowledge-based recommenders take products inferred from the user’s preferences as input.
An interesting example of implementation of the demographic-based algorithm combines user demographic data and posts from the microblogging site Sina Weibo with product information from thee-commerce company Jingdong to detect a user’s intention to purchase. Another example of knowledge-based RS is a recommendation of homes for sale based on customer-specified attributes like marital status, city, family, etc.
The future of recommendation systems:
Having looked at the conventional recommendation systems, it is important to note that these are still not perfect, due to many unforeseen challenges. Among the known challenges, the biggest issue is that of data reliability. A recommendation system can only be as good as the data supplied to it. Fake reviews and bots bias and pollute the input data, thus diminishing the quality of recommendations. Thus, good recommendations not only require an accurate recommendation system but also continuous improvement to tackle the relevant challenges. This continuous need for improvement drives the evolution of recommendation systems, especially in the retail sector.
To get an idea of recommendations in the food industry, check out this article!
