Increasing 'For You' feed watch time

Sune, a video shopping app by QVC, redefines how younger generations discover products from remarkable brands through a personalized video feed. Through engaging and entertaining videos and livestreams, Sune is a unique way to discover new products, makers and brands.

Based in New York, Sune sells products from mission-driven brands through a personalized video feed. The platform is designed to bring authentic connections and joy back to the way consumers, content creators, and brands interact. The current storefronts that are offered within the app include Luxie, Multitasky, Little Words Project, Doviana, HiBar, Shapes Studio, Bundle x Joy, M2U, and Eleventh Hour. Products are sold directly in-app via brand storefronts.

Sune’s existing feed was logic-based. Logic-based ranking refers to a system where content or items are ordered based on predefined rules or heuristics rather than advanced machine learning models. This approach uses fixed criteria and manual configurations to determine the ranking of items, such as:

Rule-Based Criteria: Specific rules or conditions set by developers, such as prioritizing new items or trending content.
‍
Static Algorithms: Algorithms that follow a fixed formula or logic, such as sorting items by recency or popularity.
‍
Manual Adjustments: Ranking based on manually inputted factors or business decisions, such as promotions or seasonal content.

Sune’s goal was to build their app centered around a personalized ‘For You’ feed, with bespoke, dynamic content tailored for each user at the core of their product. Niche content appeals to user interests and leads to increased watch time, higher conversion, and ultimately more revenue.

Sune’s objectives were clear, they needed to improve the recommendation accuracy of user-to-video ranking i.e personalizing the 'for you' page based on user preferences,
video-to-video recommendations - suggesting related videos to maintain user engagement,
user-to-product ranking - tailoring product recommendations to individual user preferences and product-to-product recommendations - enhancing product discovery through related product suggestions.

The two technical fundamental problems Sune had front of mind were solving for real-time event ingestion and the cold start problem. They needed a sophisticated recommendation system. The only problem for Sune - they had no machine learning team to build it.

Building a personalized feed on a v-commerce app without a dedicated machine learning team is particularly difficult due to the need for sophisticated models to analyze and predict user preferences. Personalization requires handling vast amounts of data, including user interactions, purchase history, viewing patterns, and demographic information.

Even with machine learning expertise, developing a system to accurately process this data and generate relevant recommendations is a significant challenge. Additionally, fine-tuning models to adapt to changing user behaviors and preferences requires ongoing analysis and iteration, tasks that are beyond the capabilities of a standard development team, and highly time consuming for a specialized ML team.

Combining this ML resourcing scarcity with top down pressure to ship a product in a tight timeframe, the team at Sune needed a solution fast.

Instead of building a system from scratch, hiring a team with at least one data engineer ($130k), one machine learning engineer ($165k), and one data scientist ($145k), with timelines starting at months to years, the team at Sune integrated Shaped, and had a ‘for you’ feed up and running in less than one sprint.

With Shaped, Sune empowered their dev team with tooling to emulate an entire machine learning team, including:

Real-Time Ingestion: Shaped connected directly to Sune’s customer data platform enabling instant setup. This allowed streaming of end-user behavioral data (e.g., clicks, impressions, likes), allowing real-time updates of embeddings.

Declarative Stateful SQL Data Pipelines: Shaped provided a declarative DuckDB SQL interface - this eliminated the need for Sune to build custom data and ETL pipelines.

State-of-the-Art Embedding Encoders: Shaped offers a library of large multimodal machine-learning models that encode unstructured data. Shaped experimented with the best models for Sune to unlock the highest possible user engagement and watch time.

Visualization and Monitoring: Shaped provided a dashboard for viewing the state and health of data, ensuring features and embeddings were accurate. Shaped’s analytics pane helped Sune understand users and items from generated embeddings and AI-driven analytics.

End-to-End Recommendation and Retrieval API: Shaped offers end-to-end recommendation APIs, this allowed Sune’s dev team to quickly demonstrate value without needing to fine-tuning models or build end-to-end ML tasks.

Moving from a logic-based system to Shaped improved Sune’s watch time by double digits, leading to improved engagement, user retention and conversion. Users now spend more time on the platform, engaging more deeply with the content, and discovering vendors and products that are aligned with their preferences.

From Sunes initial interaction with Shaped to production, integrating an end-to-end recommendation system took a matter of days.