Launch HN: ClearBrain (YC W18) – predictive models for app conversions and churn
Our founding team comes from Optimizely & Google where we built similar predictive tools for our marketing teams. At each company, we kept building the same components of a predictive pipeline - javascript snippets to collect data, ETL jobs to transform that data, and cron jobs to run a regression. We were spending hours a week maintaining these pipelines, but the time-consuming part wasn’t the algorithms (as they’re open sourced) it was the transformations.
So with ClearBrain we decided to automate the data transformation steps. We built our system in Spark ML (scala), Data Pipeline, and Go. Instead of instrumenting yet another Javascript snippet, we use existing data in Segment (YC S11) and Heap (YC W13) through standard integrations. And because every Segment/Heap dataset has the same schema, our system can process it with the same transformations into a machine-readable feature matrix. When a customer selects a user action tracked in Segment/Heap to predict, our transformed matrix is run through a logistic regression via Spark ML, and outputs a probabilistic score for each user to perform that action based on users who performed it in the past.
This distills the predictive modeling process to a simple UI to identify high-probability users in minutes. We’ve built the tool with marketers in mind, to help them identify which users may convert or churn, and export those users to marketing tools like Facebook Ads, Hubspot, etc. We’ve also found good reception from startups that have marketing objectives but lack the resources to deploy ML-driven campaigns themselves.
We look forward to feedback from the HN community! :)
Bilal
20 comments
[ 4.5 ms ] story [ 52.5 ms ] threadShoot me an email at bilal@clearbrain.com and I can help you get started. You can also get set up immediately at https://www.clearbrain.com
Hi and congrats on ClearBrain! Hope all is well, super excited to see where this goes.
Main question...is this a domain where ML really adds a lot of value beyond the basic concept of High Interaction (lots of time spent and clicks/pages visited) -> High conversion Probability?
But the propensity models built in ClearBrain tend to be more specific. The target variable can be defined as any client or server-side event you've tracked in Segment, or any trait/attribute of your user. As such, common use cases tend to be around predicting conversion events to discrete stages of a user journey - separate models for whom will move from plan type A --> B --> C, etc. So even if a generic engagement metric shows up as highly correlated for these discrete stages, the benchmark in engagement would be different and hence still intuitively helpful to diffrentiate groups of users.
Any plans to connect to Amplitude? and/or if we feed events directly to you via a JS API? (we don't use heap or segment, that's why I'm asking)
Amplitude is definitely on our roadmap as one of the next integrations we're looking to support in 2018!
A random forest automatically gives you a "variable importance plot".
With respect to the points raised in the article - ClearBrain actually does not use deep learning techniques as a basis for our predictive models. Predictions in ClearBrain are based either on logistic regression or decision tree paradigms.
From the beginning of when we approached ClearBrain as well, we wanted to make sure we provided a service that wasn't merely a blackbox. We wanted to provide insight into how the models are performing, so we expose analyses such as feature importance, attribute benchmarks, and indications of which actions are informing the models.
This helps with both interpretability and actionability in customer workflows, but also in some of the GDPR issues noted.