Increasing diversity by 382%

Breakr connects artists and creators to permeate new music through culture. Creators can set up profiles, update their music preferences, and create digital campaigns available across all platforms. Breakr offers tools like a smart matching engine, AI-driven campaign optimization, and automated relationship management. Customers include Def Jam, Interscope Records, Meta and Capital Records.

Connecting tens of thousands of creators, artists, and record labels comes with a challenge. Without an advanced recommendation system, popular profiles get more popular, while new accounts are ‘penalized’, unfairly receiving little engagement. As a platform where one of the key goals is to surface the next hot and upcoming artists, Breakr needs to help them grow, the first step is through visibility.

In recommendation systems, diversity refers to the inclusion of a wide range of different items or recommendations within the result set presented to the user. The goal of promoting result diversity is to avoid presenting a monotonous or overly similar set of recommendations, ultimately improving user satisfaction and engagement. 

Result diversity is important for a number of reasons:

1. User Satisfaction - Diverse recommendations cater to different aspects of a user's preferences and needs, making the overall experience more enjoyable and engaging.

2. Exploration - Encouraging users to explore a wider variety of items leads to the discovery of new interests or products they may not have otherwise considered.

3. Mitigating Redundancy - A diverse set of recommendations reduces redundancy and prevents the system from suggesting too many similar items, which can be perceived as repetitive or boring.

4. Addressing Serendipity - Diversity enhances the chances of serendipitous findings, where users come across unexpectedly relevant or interesting items.

5. Avoiding Filter Bubbles - By introducing variety, result diversity helps prevent the reinforcement of existing preferences and biases, which can lead to filter bubbles where users only see a narrow slice of available content.

To improve result diversity while maintaining relevance, Shaped uses the Maximal Marginal Relevance (MMR) algorithm. MMR is a classic method that parameterizes the trade-off between diversity and relevance, and in this case, allowed Breakr to optimize the number of products that were being showed, while continuing to show highly relevant content.  For a deeper understanding of MRR see Goldstein, J., & Carbonell, J. (1998). The use of MMR, diversity-based reranking for reordering documents and producing summaries.

At a high level, Shaped works in 4 steps:

1. Retrieve candidate items from vector store (vectors are built from user and item embeddings which are created from several techniques).

2. Filter out candidate items that were already seen by the user.

3. Score candidate items based on pointwise binary classification model estimating probability of engagement/conversions

4. Order results by their estimated score, and penalize scores based on diversity (so that we optimize for both engagement and diversity fairly)

MMR provides:

‍An Improved User Experience - By presenting a diverse set of results, MMR helps users find information that covers different facets of their query, leading to a more satisfying search experience.

‍Coverage - Ensures that multiple aspects or subtopics related to the query are represented in the results, which is particularly useful in exploratory scenarios like Breakr.

‍Reduced Redundancy - Avoids presenting multiple similar results ir matches, making the best use of the limited space in search result pages or recommendation lists.

Improving diversity by almost 4x led to improved match rate, user retention and engagement. Users now spend more time on the platform, engaging more deeply with the content, and discovering artists and creators that are more aligned with their preferences. From ideation to production, integrating Shaped with Breakr took a matter of days.