Solving cold start for a new business

Overlap is a podcast app designed to enhance the discovery and consumption of podcast content, specifically tailored for business and tech enthusiasts. Overlap curates and delivers personalized podcast clips, making it easier for users to find and engage with content relevant to their interests. The app presents these clips in a format similar to TikTok, allowing for effortless vertical scrolling and quick consumption of high-quality podcast moments.

Overlap’s primary goal was to improve the discoverability of podcasts content with feed personalization. For a new startup, creating a curated feed at a user level is made difficult due to the fact they face the cold start problem for 100% of their users. Overlap approached Shaped, pre-production, pre-users, and therefore pre-interaction-data. They needed a solution that could personalize user’s feeds based on their actions in real time.

What is the cold start problem?
In recommendation systems, the cold start problem refers to the difficulty of making accurate recommendations when there is little to no historical data available. This occurs in several scenarios:

New user cold start - When a new user joins the system, there is no historical interaction data for that user, making it challenging to provide personalized recommendations. This was Overlap’s biggest challenge.

New item cold start - When a new item (in this case a podcast) is added to the system, there is no interaction data (e.g ratings, views) for that item, making it difficult to recommend it effectively.

Causes of the cold start problem:
‍
Lack of Interaction Data - Recommendations are typically based on past user interactions (such as ratings, clicks, purchases). Without this data, traditional recommendation algorithms like collaborative filtering struggle to make accurate suggestions.
‍
Data Scarcity - Insufficient data about new users or items leads to poor understanding and modeling of their preferences.

Shaped used several methods to help Overlap address the cold start problem for their podcast recommendation system. Here’s a breakdown of how Overlap leveraged Shaped to get their recommendation system up and running in 24 hours:

Real-time data ingestion
Shaped connected directly to Overlap’s CDP (Amplitude) allowing Overlap to stream real-time event data directly. This real-time data ingestion allowed the Shaped to immediately incorporate new user interactions and preferences into the recommendation model, instantly enhancing personalization and reducing latency.

Handling user and item attributes
To help Overlap’s integration, Shaped’s expert ML team consulted Overlap on best practice. Two crucial attributes to solve the cold start problem were:

1. User profiling - collecting user attributes, such as interests and demographic information, to build initial profiles. This data was crucial for making initial recommendations before sufficient interaction data was available.
‍
2. Item metadata - leveraging item attributes like video descriptions, channels, and publisher IDs to enhance content-based filtering. These attributes helped in generating recommendations based on the content characteristics rather than user interactions.

Continuous model development
Shaped adopted an evolving approach to model improvement by continuously updating the recommendation models based on new data and feedback. Working closely with Overlap to debug issues, and ensuring that the models further improved metrics over time.

Active user engagement
Overlap gathered explicit feedback from users through ratings and likes. This engagement helped in collecting valuable data points quickly, which were then used to fine-tune the recommendation models.

When companies start with no event data, they often employ rules-based approaches for ranking, for example ordering chronologically. Overlap, as a pre-launch podcast app, implemented Shaped to tackle the cold start problem and avoid biases often associated with low data regimes. By partnering with Shaped, Overlap was able to provide more personalized and diverse content recommendations, promoting visibility for new content and enhancing user engagement from the outset.

From end to end integrating Shaped took less than 24 hours.