“Why a Strong Online Presence is Key: A Comprehensive Guide to Diversifying Your Search and Recommendation Systems” | Pinterest Engineering Shares Expert Insights | May 2023.

Utilizing Visual Skin Tone Signal to Improve Diversity on Pinterest

Pinterest is a platform whose mission is to inspire users to create a life they love. In today’s interconnected world, online platforms need to reflect the diversity of their users to facilitate content discovery and meet their needs and preferences. Improving representation online can lead to increased engagement, retention, and trust in the platform. This post will show how Pinterest improved diversification on three surfaces: Search, Related Products, and New User Homefeed.

The End-to-End Diversification Process

Pinterest has developed and deployed scalable diversification mechanisms to support a wide range of skin tones in recommendations for fashion using a visual skin tone signal. The end-to-end diversification process consists of the following components:

Detection of requests that trigger diversification
Retrieval of diverse content from the large content corpus
Diversity-aware ranking to balance the diversity-utility trade-off when ranking content
Accommodation of diversification across several dimensions
Multi-Stage Diversification

Multi-stage diversification allows the mechanism to operate throughout the pipeline, from retrieval to ranking, to ensure that diverse content passes through all the stages of a recommender system. Large-scale recommender systems can broadly be categorized into retrieval and ranking stages, followed by additional business logic. The retrieval stage generates a set of candidates from a large corpus of items, while the ranking stage finds an ordering of the candidates that maximizes a combination of objectives.

Diversity in Recommendations

Diversification aims to ensure that the ranked list of items surfaced by the system is diverse with respect to a relevant diversity dimension, which could include explicit and implicit dimensions. For a given query, the top-k diversity of a ranking system is the fraction of queries where all groups under the diversity dimension are represented in the top k ranked results for which the diversity dimension is defined.

Triggering Logic

Upon receiving a request, the system needs to determine whether to trigger diversification according to the dimension of interest. The triggering logic needs to account for the diversity dimension, the application, the production surface, and the local context, such as country and language, and can be based on heuristics or ML models.

Ranking Stage

Pinterest leverages a diversity-aware ranking stage that takes as input a list of items with utility scores and their diversity dimensions and produces a ranking according to a combination of both objectives. The first approach used is a class of simple greedy rerankers, e.g. Round Robin. We construct |D| number of ordered sub-lists corresponding to each skin tone range and containing items that have a utility score above the threshold. Then, we re-build a ranked list by greedily selecting the top item of each sub-list.

Conclusion

Pinterest’s diversification mechanisms use a visual skin tone signal to support representation of a wide range of skin tones in recommendations, improving user experience and satisfaction. The end-to-end diversification process consists of several components, and multi-stage diversification operates throughout the pipeline to ensure diverse content passes through all stages of a recommender system. By utilizing a diversity-aware ranking stage, Pinterest is able to produce a ranking that balances both diversity and utility objectives.