Presented By O’Reilly and Cloudera
Make Data Work
September 11, 2018: Training & Tutorials
September 12–13, 2018: Keynotes & Sessions
New York, NY

Diversification in recommender systems: Using topical variety to increase user satisfaction

Ahsan Ashraf (Pinterest)
2:05pm–2:45pm Wednesday, 09/12/2018
Data science and machine learning
Location: 1A 06/07 Level: Intermediate
Secondary topics:  Media, Marketing, Advertising, Recommendation Systems

Who is this presentation for?

  • Data scientists, data analysts, and machine learning engineers

Prerequisite knowledge

  • Familiarity with basic math concepts
  • Experience with machine learning and recommender systems (useful but not required)

What you'll learn

  • Understand how measuring and evaluating diversity can lead to increased user satisfaction


Recommender systems that rely heavily on engagement features can often overoptimize for a user’s interests or preferences. A user that engages heavily with content relating to food may be served with food-related content disproportionately compared to other types of content. This leads to the unintentional consequence of optimizing for short-term gains in engagement at the cost of providing a variety of content to the user. Over time, this may lead to a staleness of the type of content the user sees. Furthermore, this can also cause the overdistribution of a certain type of content that has higher engagement rates in general due to some characteristics of the content. For example, a typical user may be more likely to click on results relating to food than on results for which the user is satisfied without clicking, such as an image result of a webcomic.

Introducing diversification into recommender systems to counter the problem of overoptimization has been of great interest in the field. Measuring and evaluating the impact of diversification is difficult, but when done appropriately, can lead to measurable increases in user satisfaction. The first step in understanding diversity is to define diversity. The definition of diversity can play an incredibly important role in the effectiveness of the implementation of diversity in the recommender system. For example, if diversity is defined too broadly or too narrowly, it could either be ineffectual or even detrimental. Although the importance of diversity in recommender systems has been shown to be important by several research groups in the past, it has been challenging to establish a preferred measurement.

Ahsan Ashraf shares approaches for defining and evaluating diversity at Pinterest and discusses the advantages and disadvantages of these approaches. Ahsan then outlines an experiment that implemented the established definition and drove an increase of ~2–3% in impressions of images and an increase of ~1% in time spent on Pinterest.

Photo of Ahsan Ashraf

Ahsan Ashraf


Ahsan Ashraf is a data scientist at Pinterest focusing on recommendations and ranking for the discovery team. Previously, Ahsan worked with personal finance startup as part of an Insight Data Science Fellowship, where he designed and built a recommender system that drew insights into users’ spending habits from their transaction histories. Ahsan holds a PhD in condensed/soft matter physics.