The power of data: How recommendation engines are driving business growth

Have you ever wondered how Spotify shows you a list of similar songs that you have liked and listened to on a loop, or how Netflix displays the shows and movies of the same genre that you have been binge-watching? Have you ever wondered what logic forms the “Picked for you” or “Frequently bought together” lists? 

These lists or options are recommendations that are given by recommendation systems, which these organizations tailor-make to understand their users, their preferences, and past interactions with the systems using the power of AI/ML. 

With personalization of products gaining all the hype, this article brings to you the business data drives and nitty-gritty you need to know about recommendation systems to build these Next Generation products as a Product Manager and also as AI product leader Marily Nika, working at Meta’s reality lab, rightly says that “all product managers will be AI product managers eventually.

Numbers don’t lie- Companies maximizing ROI with Recommendation systems: Recommendation systems have changed how people interact with a lot of websites and services. A recommendation engine can help you and your team drive an increase in your customer engagement, retention, and growth by giving customers the products they desire based on their purchasing history. Did you know that Amazon drives about 35% of its eCommerce revenue from product recommendations, while Netflix drives 75% of its revenue from product recommendations? This generates over $1 billion per year of Netflix’s revenue.Stitch Fix has grown from a modest women’s-only subscription-box service to a retail powerhouse. Thanks to Spotify’s “Discover Weekly” recommendation algorithm, its users have listened to over 2.3 billion hours of music! 

Evolving customer preferences and expectations seek improved and enhanced experiences whenever they engage with a digital platform. With a plethora of options available, whether it’s selecting movies on an online cinema or consumer goods in an e-commerce store, users often face the challenge of decision fatigue. The abundance of information can overwhelm them, making it difficult to make informed choices. As product managers, addressing this issue becomes crucial. Leveraging targeted recommendations is a powerful solution to help users navigate through the vast array of options available. By implementing effective recommendation systems, product managers can deliver personalized and relevant suggestions, enabling users to make confident decisions, ultimately enhancing their overall experience.

Building a business model with data-driven insights from the core recommendation can be divided into three strategies: 

1. Global Strategy: This strategy focuses on showcasing popular items and trends to both new and returning customers. For example, Amazon utilizes this strategy by displaying “Customers Who Bought This Also Bought” recommendations, highlighting products that are frequently purchased together or popular within a specific category. Product managers can leverage this approach to provide customers with relevant and trending options. 

2. Contextual Strategy: Moving to the next level, the contextual strategy analyzes product features, customer behaviors, and content type to deliver accurate recommendations. Spotify uses this strategy by considering a user’s listening habits, such as genre preferences and time of day, to create personalized playlists and suggest similar songs. Product managers can implement similar approaches by analyzing user data to provide tailored recommendations based on product attributes, purchase patterns, and content preferences. 
3.   Personalization Strategy: The pinnacle of recommendation strategies, personalization relies on extensive customer data to create tailored experiences. Netflix excels in this area by leveraging user data, including viewing history, ratings, and genre preferences, to curate personalized movie and TV show recommendations. Product managers can strive for a similar level of personalization by leveraging comprehensive behavioral data collected over time to deliver highly customized recommendations based on individual preferences and context.

Strategizing an effective customer-based Recs Engine as an AI Product Manager

The entire recommendation process of this system consists of three basic steps: Collection of user information, algorithm processing of data, and recommendation suggestion. 

The following are some of the best practices you can follow as an AI PM, to draw a seamless customer journey and customer-preferred personalizations/recommendations:

Define the Objectives

Not every problem is appropriate for ML. It’s a Product Manager’s job to remain disciplined and focused on the product, and not to get carried away by the possibilities of exciting new technology. 

ML products don’t have an interface, thus as a PM you need to describe its features differently. So write a story to close this gap – explain how users interact with your product, what are their expectations, and the results to be obtained in different cases. This story will also help synchronize between different teams working with you.

So start by asking what user problem you’re solving, as you would with any product management problem. Ask yourself what specific business outcomes you want to achieve, such as increasing sales, improving customer retention, or enhancing the overall customer experience. Defining the initial OKRs will be helpful. Sometimes, you already have a new objective you want to work for in mind for your customer by intuition while other times you need to know your customer to understand their pain points in defining the objective. 

Understanding your customer

This step works hand in hand with the first step. No one knows your customer segment better than you and your Customer support team. Collaborate with them and understand the key pain points they face. Recommendation systems work majorly on understanding patterns. Therefore, analysing customer data, including demographics, behavior, and historical interactions, to identify patterns and preferences that can drive personalized recommendations.

Data Quality and Availability

 Ensure that you have access to relevant and high-quality data for building your recommendation system. This includes customer data, product information, transactional data, and any other relevant data sources. Clean and preprocess the data to eliminate noise and inconsistencies that could affect the accuracy of your recommendations. You must use real data sets instead of historical or downloaded ones. Building a solution on an actual data dataset will help you understand more in terms of the budget, workforce, and infrastructure you would require instead of working on test data.

Consider using contextual data such as user location, time of day, device, or browsing history to improve the relevance of your recommendations. Contextual data can assist deliver more targeted and timely recommendations, thereby improving the customer journey. Allow users to provide feedback, modify their preferences, or opt out of personalized recommendations if desired. 

Your data team will be your friend here! There are a few questions that you can work on while gathering the data:

1. What data sources are available to my recommendation system? If your data is, say, full of free text boxes where anyone can type what they like, then it’s probably not suitable.
2. Is the data relevant, accurate, and up-to-date? Will it be real time or historical data? For example, are there any inconsistencies or missing values in the dataset that could affect the accuracy of the recommendations?
3. What customer attributes and behaviors should be considered in the recommendation algorithm? For example, should we prioritize recent interactions or take into account the customer’s demographic information?
4. Are there any privacy or legal considerations related to the use of customer data? For example, do we need to obtain explicit consent from users or comply with specific data protection regulations?
5. How can we segment and group customers based on their preferences and behaviors? For example, can we identify distinct customer segments to tailor recommendations accordingly?
6. What are the key metrics or success criteria for evaluating the performance of the recommendation system? For example, are we looking to optimize click-through rates, conversion rates, or customer engagement metrics?
7. How can we handle missing or incomplete data to ensure accurate recommendations? For example, can we use techniques like imputation or develop algorithms that are robust to missing data?
8. Are there any biases in the data that may impact the fairness or inclusivity of the recommendations? For example, do certain customer groups receive disproportionately more or fewer recommendations?
9. Can we leverage contextual information, such as location or time, to enhance the relevance of the recommendations? For example, should we recommend nearby stores or time-sensitive offers based on the user’s current location or time of day?
10. How frequently should the recommendation system be updated to incorporate new data and evolving customer preferences? For example, should recommendations be refreshed daily, weekly, or in real-time?
11. Are there any external factors, such as seasonal trends or industry changes, that should be considered in the recommendation algorithm? For example, should the system adapt recommendations during holiday seasons or account for changing product availability?
12. What testing methodologies can be employed to validate the effectiveness of the recommendations? For example, can we conduct A/B testing or user surveys to compare different recommendation algorithms?
13. How can customer feedback and user interactions be incorporated to improve the recommendation system? For example, can we gather explicit feedback or implicit signals to refine the recommendations over time?
14. What scalability and performance considerations should be taken into account when processing large volumes of data? For example, do we have the infrastructure and computational resources to handle the increasing data size once the MVP launch is complete and we have more customer flow, in case of a new product?
15. Are there any ethical considerations, such as promoting diverse content or avoiding harmful recommendations, that should be addressed? For example, can we implement mechanisms to ensure fairness, transparency, and prevent the amplification of biased content?

Phew! The above list is never ending, the more you understand the Product, the more questions you can ask and the more loose ends you are able to solve. 

• Choosing the right recommendation Algorithm: This is the most important part of your strategy building. Brainstorm with the data and Based on the data and your business model, assess the pros and cons of each to select the RecSys that will suit.There are three types of recs engines: 

• Content Based Filtering: Based on description of the product and user profile preferences. Stitch Fix’s fashion box is another example of content-based recommendation. A user’s attributes are collected (height, weight, etc.) and matched fashion products are put in a box delivered to the user.

1. 1. Collaborative filtering: Gathering or analyzing data based on user’s behaviour. Spotify’s recommendation system is based on collaborative filtering. This is because it uses the user’s past song preferences to make recommendations to the new ones. 
   2. Hybrid Filtering: It’s a combination of both Content and Collaborative filtering. Netflix is an excellent illustration of how hybrid recommender systems may be used. The website creates recommendations by comparing comparable users’ viewing and searching behaviours (i.e., collaborative filtering) and by suggesting films that share features with films that a user has rated highly (content-based filtering).
   3. Iterative Testing and Optimization: Building a recommendation system is an iterative process. Test and optimize the system continuously to improve its performance. Monitor key metrics such as click-through rates, conversion rates, and customer feedback to evaluate the effectiveness of the recommendations and make necessary improvements.

It is expensive to build ML models and create an infrastructure to test the idea.

Teams try to test ideas through experiments done with the minimum amount of development. Therefore, Building a minimum viable product (MVP) is better, which can be first tested for a CUG.

One of the most common practices being used today is A/B Testing. Once any new model is made, it is not obvious that we should roll it out for all the customers, immediately. This is where A/B testing comes into the picture. Any new model must be A/B tested with the conventionally used model, with a pool of randomly selected users using the target model to understand using analysis or hypothesis testing which model works best for the users, and eventually for the business. Selecting the right KPIs for A/B testing is important for optimal results.

However,  it is considered best practice to release the new model to only 98-99% of users, with the remaining 1-2% being served by the control/conventional model. This set of users is known as the holdout set, which serves to identify potential issues with the new model. So, even if there is any issue in the future with the Recs system, the team could identify if it is impacting all the models or the newly deployed model.

The Dark Side of Personalization: The Pitfalls of Recommendation Systems: With the advantages of RS comes a fair share of challenges. These are some of the most important pitfalls which you as a PM need to address:

1. Cold Start Problem – RS relies heavily on user data; however, this also has its downside, especially when the customer is either new or has very little data available, or a new item is added to the catalog. The recs engine here does not understand what the taste of the new user is, or what the rate/reviews of the newly added product are, which can lead to less accurate results. 
2. Privacy concerns: As recommendation systems collect and use personal data to create personalized experiences, there is a risk of violating user privacy and data protection laws. Businesses must ensure that they are transparent about the data they collect and how they use it and that they obtain user consent where necessary. 

Implement robust security measures to protect sensitive information and obtain necessary consent from users. Avoid biases and ensure fairness in the recommendation algorithms to provide equal opportunities for all users.

3. Data quality: Recommendation systems rely heavily on data, and the quality of the data can significantly impact the accuracy and effectiveness of the recommendations. Poor quality data, such as incomplete or inaccurate user data, can lead to irrelevant or even incorrect recommendations. 

Just like every coin has two sides, RS comes with its positives and negatives. However, if built with proper research, the positives will supersede the negatives. From Netflix to Amazon, recommendation engines are everywhere, quietly working behind the scenes to make our lives better. As an AI Product Manager, these next Generation products can be a game changer for your business, if implemented with property strategy. With these engines constantly evolving and improving, we can expect to see even more amazing and personalized experiences in the future, and who knows? Maybe one day we’ll look back and wonder how we ever managed without them.

Learn more about machine learning and data on Mind the Product

• Data as a product – Omar Sallam on The Product Experience
• Product management in the era of AI and Machine Learning
• Understanding AI-based data and insights led digital product management