Why are recommendation engines important?
As the competition in all industries is increasing, keeping their customers engaged is an important goal for organizations. Recommendation engines enable organizations to increase their sales by upselling (selling a higher volume of products that they buy) or cross-selling (selling new products) to existing customers. Here are some recommendation engine examples from tech leaders:
- 35% of Amazon.com’s revenue is generated by its recommendation engine.
- 75% of users in Netflix choose movies/tv series according to recommendation engine suggestions. Netflix executives Carlos A. Gomez-Uribe and Neil Hunt state that recommendations reduce the churn rate by several percentage points. This increases the lifetime value of existing customers that’s why they believe recommendations save them more than $1B per year.
- Spotify first released Discover Weekly playlist recommendations in 2015 and they experienced an 80% revenue increase with 40 million Discover Weekly users(40% of total users by that time) in 2016.
How does it work?
Recommendation engines have three basic steps to make recommendations:
Core of a recommendation engine is consumer data. These engines collect implicit and explicit data.
- Implicit data is the information that is gathered unintentionally from customers by checking their website history. Examples are web search history, clicks and order history.
- Explicit data is created by customers’ inputs such as ratings and likes/dislikes.
As the amount of data you store increases, you provide better recommendations for your customers. Organizations need to keep as much data as possible on the cloud or enterprise to analyze customers and divide them into segments.
Data Analysis and Recommendation
Recommendation engines analyze data by filtering it to extract relevant insights to make the final recommendations.
What are common approaches used in recommendation engines?
Recommendations depend on a combination of similar users' actions (collaborative filtering), products similar to those consumed by the user (content based filtering) or the context of the user (context aware filtering):
Content based filtering, as its name refers, is recommending a product that is similar to products the customer liked before. Below is an example of a movie recommendation content based filtering. Rabin is a user who mostly watches commercial dram movies and the system provides Movie A and Movie B as a recommendation. The downside of content-based filtering is product mappings are manual and depend on labelers’ bias.
Another example is if the user rated a song from an artist, system recommends him another song from the same album.
Collaborative filtering methods are divided into two categories:
- User-based collaborative filtering: Engine recommends a product if the product has been liked by users similar to the user.
- Item-based collaborative filtering: Based on users’ previous ratings, system identifies similar items. For example, if users A,B and C rated books X and Y, then when a new user purchases book Y, systems recommend purchasing book X as well due to the pattern created by A,B and C users.
Context aware filtering adjusts recommendations based on the time, place, the users' consumption right before the recommendation. For example, ice creams should be recommended more often in summer.
What are its use cases?
Recommendation systems can be useful and applicable to various industries in the B2C environment. We’ve explained those use cases before, feel free to check our article.
What are its benefits?
Recommendation engines are generally used to boost sales processes along with the relationship between the organization and customers. To learn all benefits, check out our article.
What are its challenges?
- Cold start problem: What should you recommend to new users? Should you recommend the most commonly recommended items or should you try to understand more about the user? Answers to such questions depend on the specific application.
- Obvious recommendations: With no long tail data, recommendation systems make quite obvious recommendations which could easily be programmed by a few rules. Data is crucial for a recommendation system.
- Static recommendations that become outdated with changing tastes: If the system is not continously learning, such a scenario is inevitable. Companies are advised to invest in continuously learning systems
- Recommendations that violate personal privacy: Consumption data is personal data and using such data for recommendations requires care even when the recommendation is only shared by the user. A NY Times article from 2012 includes an anecdote about how Target predicted a teen's pregnancy before her father. Though we don't know if such a thing really happened, it is indeed an example of how innocent looking recommendations can violate personal privacy.