Building a Generalized Feed Ranker with Amazon SageMaker Inference: The Success Story of Meesho
Introduction:
Meesho, India’s rapidly growing ecommerce company, is focused on democratizing internet commerce and making it accessible to the next billion users in the country. To enhance the user experience, Meesho has developed a powerful generalized feed ranker (GFR) using AWS machine learning services. This ML-driven solution considers individual preferences and historical behavior to personalize each user’s shopping feed, boosting user engagement and conversion rates. The GFR sources diverse and relevant recommendations from multiple channels and utilizes features such as browsing patterns, interests, and item scores to create personalized product recommendations. The GFR architecture consists of two components: model training and model deployment. Meesho used Amazon EMR with Apache Spark for distributed training at scale, reducing training time and improving efficiency. Model artifacts were stored in Amazon S3 for convenient access. For model deployment, Meesho used SageMaker inference endpoints with auto scaling enabled, resulting in low-latency serving of models. A custom inference image was built and pushed to Amazon ECR for specific needs. Meesho leveraged A/B testing and monitoring to continuously improve the performance of the model. By implementing this advanced feed ranker, Meesho has significantly improved user engagement and conversion rates, contributing to its overall business growth. The utilization of AWS services has reduced ML lifecycle runtime and improved efficiency for the team. Meesho continues to deliver tailored shopping experiences, fulfilling its mission to democratize ecommerce for everyone.
Full News:
Optimizing the Shopping Experience with a Generalized Feed Ranker: A Meesho Case Study
Meesho, India’s fastest-growing ecommerce company, has revolutionized the online shopping experience for millions of customers across the country. With a mission to democratize internet commerce and make it accessible to every Indian user, Meesho has become a popular platform for micro, small, and medium businesses, as well as individual entrepreneurs. By providing access to millions of customers, a wide selection of products, and efficient logistics and support services, Meesho empowers businesses to thrive in the ecommerce space.
To enhance the user experience further, Meesho set out to develop a powerful generalized feed ranker (GFR). The goal was to create a personalized and tailored shopping experience for each customer by considering their individual preferences and historical behavior. By utilizing AWS machine learning (ML) services, such as Amazon SageMaker, Meesho was able to build an ML-driven solution that optimized the user’s feed, resulting in increased user engagement, conversion rates, and overall business growth.
The GFR utilizes extensive historical data to analyze user browsing patterns, interests, and various other factors. These valuable insights are used to construct ranking models that personalize the user’s feed in real time. Factors like geography, prior shopping patterns, acquisition channels, and more are considered to provide a unique and relevant shopping experience for each customer. Additionally, the GFR captures user interaction-based features, including affinity towards specific items, item categories, price, ratings, and discounts. These features, along with user-agnostic features and item-level scores, are inputted into the Learning to Rank (LTR) model, which predicts the Probability of Click (PCTR) and the Probability of Purchase (PCVR).
For diverse and relevant recommendations, the GFR sources candidate products from multiple channels, including known user preferences, novel and potentially interesting products, trending items, and the latest additions to the platform. This ensures that customers receive personalized recommendations that align with their preferences while also introducing them to new and exciting products. The GFR architecture, divided into model training and model deployment components, facilitates scalable and reliable performance within the Meesho ecosystem.
During model training, Meesho used Amazon EMR with Apache Spark to process massive amounts of data points. Distributed training using Dask on Amazon EMR enabled efficient and cost-effective scaling of training jobs, significantly reducing training time. To maintain a historical record of feature values used for model training, Meesho employed an offline feature store. Model artifacts were stored in Amazon S3, providing convenient access and version management. To evaluate model performance, various metrics were tracked, including area under the ROC curve and area under the precision-recall curve. Calibration of the model was also monitored to ensure accurate probability score predictions.
For model deployment, Meesho utilized SageMaker inference endpoints with auto scaling enabled, providing low-latency deployment support for trained models. A custom inference image was built specifically to cater to Meesho’s needs, deployed using Amazon ECR. Meesho also developed an in-house A/B testing platform to make data-driven decisions promptly and observed an approximate 3.5% enhancement in the platform’s conversion rate and an increase in app open frequency through A/B experiments. Drifts such as feature drift and prior drift were monitored multiple times a day post-deployment to maintain model performance. Automations and triggers were set up using AWS Lambda for model retraining, endpoint updates, and monitoring processes.
With the implementation of the generalized feed ranker, Meesho successfully improved user engagement, conversion rates, and overall business growth. By delivering highly personalized product recommendations based on individual preferences and historical behavior, Meesho continues to provide tailored shopping experiences to its customers. The utilization of AWS services significantly reduced the ML lifecycle runtime, leading to increased efficiency and productivity for the Meesho team.
This case study showcases Meesho’s commitment to customer satisfaction and its continuous efforts to enhance the ecommerce experience for every user. Through advanced ML-driven solutions, Meesho fulfills its mission to democratize ecommerce and provide value to its customers. With guidance from industry experts and the support of AWS services, Meesho remains at the forefront of innovation in the ever-evolving world of online shopping.
—
As a guest post co-written by Rama Badrinath, Divay Jindal, and Utkarsh Agrawal at Meesho, this article provides valuable insights into Meesho’s implementation of a generalized feed ranker and the impact it has had on their business. By leveraging AWS machine learning services and optimizing the shopping experience for their customers, Meesho has created a platform that offers personalized recommendations and increased engagement. As an ecommerce company, Meesho’s focus on improving user experience and providing value to their customers sets them apart in the competitive Indian market. With their mission to democratize internet commerce, Meesho continues to empower businesses and entrepreneurs across the country, making the online shopping experience accessible to all.
Conclusion:
India’s fastest growing ecommerce company, Meesho, has implemented a generalized feed ranker using Amazon SageMaker to provide highly personalized product recommendations. By analyzing user preferences and historical behavior, the ranker tailors the shopping experience for each customer, resulting in improved user engagement, conversion rates, and overall business growth. This ML-driven solution, developed using AWS machine learning services, has streamlined the ML lifecycle for Meesho, reducing runtime from months to weeks and increasing efficiency and productivity. With this advanced feed ranker, Meesho continues to deliver tailored shopping experiences and democratize ecommerce for everyone.
Frequently Asked Questions:
1. What is Meesho’s generalized feed ranker?
Meesho’s generalized feed ranker is a system developed using Amazon SageMaker inference. It enables Meesho to organize and display a personalized feed of products to its users based on their preferences and browsing history. This ranker uses machine learning algorithms to analyze user behavior, product attributes, and other relevant data to determine the most relevant products for each user.
2. How does Meesho’s generalized feed ranker work?
The generalized feed ranker uses data from various sources, such as user interactions, product metadata, and historical data, to train machine learning models. These models are then used to predict the relevance of products for individual users. The ranker takes into account factors like user preferences, search history, ratings, and product popularity to generate a personalized feed for each user in real-time.
3. What is the role of Amazon SageMaker inference in Meesho’s feed ranker?
Amazon SageMaker inference plays a crucial role in Meesho’s feed ranker by providing a scalable and efficient platform for deploying trained machine learning models. It allows Meesho to process large volumes of data and make accurate predictions in real-time. SageMaker inference also enables Meesho to easily update and retrain models as new data becomes available, ensuring continuous improvement in feed ranking accuracy.
4. How does Meesho ensure the relevance of the feed ranker?
Meesho constantly collects and analyzes user feedback, engagement metrics, and other relevant data to refine its feed ranker algorithms. The company regularly conducts A/B testing and incorporates user feedback to fine-tune the ranking models. Meesho’s data science team also works on improving the machine learning models to adapt to changing user preferences and market trends, ensuring the feed ranker remains relevant and personalized.
5. Does Meesho’s generalized feed ranker consider user preferences?
Yes, Meesho’s generalized feed ranker is designed to prioritize products based on user preferences. It takes into account factors such as previous purchases, browsing history, clicks, and interactions to understand user preferences and recommend relevant products. By analyzing and learning from user behavior, the ranker continuously improves its ability to suggest products that match individual user tastes and needs.
6. Can the feed ranker handle large volumes of data?
Yes, Meesho’s feed ranker powered by Amazon SageMaker inference is designed to handle large volumes of data efficiently. SageMaker provides scalable infrastructure that can process and analyze vast amounts of data in real-time. This scalability ensures that Meesho can deliver personalized feeds to its millions of users with minimal latency.
7. How frequently does Meesho update its feed ranking models?
Meesho updates its feed ranking models regularly to reflect changes in user behavior, preferences, and market trends. The frequency of updates depends on the availability of new data and the impact of changes on user experience. Meesho’s data science team continuously monitors the performance of the ranker and incorporates improvements through frequent retraining and model updates.
8. What are the benefits of using Amazon SageMaker inference?
Using Amazon SageMaker inference provides several benefits to Meesho. It allows for the deployment of machine learning models at scale, ensuring real-time predictions for large user bases. SageMaker inference also offers auto-scaling capabilities, optimizing resource allocation based on demand. Additionally, it enables Meesho to experiment with different model architectures and algorithms, enhancing accuracy and performance.
9. Can Meesho’s feed ranker adapt to changing user preferences?
Yes, Meesho’s feed ranker is designed to adapt to changing user preferences. The ranker uses machine learning algorithms that continuously analyze user interactions and feedback to update the models and improve recommendations. This adaptive approach enables Meesho to personalize the feed for each user based on their evolving preferences, ensuring a relevant and engaging user experience.
10. How does Meesho measure the performance of its feed ranker?
Meesho employs various metrics to measure the performance of its feed ranker, including click-through rates, conversion rates, customer satisfaction, and revenue generated. These metrics help determine the effectiveness of the ranker in delivering relevant and engaging product recommendations. Meesho also conducts user surveys and collects feedback to gain insights into user satisfaction and make further improvements to the feed ranking system.
