Data Science at eHarmony

Graph Optimization

Graph optimization plays a crucial role in enhancing matchmaking algorithms at eHarmony. Jon Morra explains that the transition from a round-robin approach to graph flow optimization allowed for a more equitable distribution of matches, addressing customer complaints about match quality 1. This method ensures that all users have a fair chance at desirable matches, rather than favoring those at the top of a list.

By doing a graph flow optimization, we have a much more socialistic approach that says we're going to jointly optimize for everybody at the same time.

--- Jon Morra

The approach is akin to logistics solutions used by companies like FedEx, where computational efficiency is key 2.

   

Bandit Algorithms

Bandit algorithms are integral to eHarmony's data science workflow, particularly in A/B testing scenarios. Jon Morra describes how these algorithms allow for dynamic testing of user experiences by evaluating different actions and their outcomes, rather than sticking to traditional A/B testing methods 3. This approach provides more nuanced insights into user behavior and preferences.

It's proven exceptionally effective here, like I said, within product for A, B, testing various experiences.

--- Jon Morra

The flexibility of bandit algorithms helps balance business objectives with data-driven insights, fostering innovation while managing risk 4.

   

Deep Learning

Deep learning is poised to revolutionize eHarmony's computer vision projects, although its implementation is still on the horizon. Jon Morra acknowledges the limitations of current off-the-shelf classifiers, which fail to provide the specific insights needed for eHarmony's unique dataset 5. Instead of generic classifications, the goal is to identify contextual details within images, such as activities or settings.

The deep learning stuff has just proven so much more effective.

--- Jon Morra

This shift will require significant data collection and training efforts, potentially utilizing platforms like Mechanical Turk to gather the necessary labeled data 5.