Additional Updates: October 2023 Quarterly Product Release
AI-generated event descriptions now power 1.4M+ events in our system
Previously, less than 10% of attendance-based events in our system had descriptions. Using OpenAI (GPT-3.5), the technology behind ChatGPT, we have generated high-quality descriptions for all attended events between 2018 and 2028 with a PHQ rank greater than 50.
This is a major improvement for our customers who surface event descriptions directly to their users – it will provide more background details around attended events that are relevant to them, which will directly improve the customer experience for end users.
This update also strengthens our other models, including improving data such as PHQ labels. With new AI-generated descriptions, we’re now able to get labels in 80% of cases compared to 63% without. Here are a couple of examples of attended events updated with event descriptions:
Get better data quality with less duplicate event records
Event duplicates are copies of existing events. Event duplicates in our system can introduce misleading data and diminish data quality – for example, doubling the predicted attendance.
We’ve now upgraded our deduplication model to use large language models to now analyze similarities between event titles. This includes alternate spellings, combinations, acronyms, and variations of the event name. This release cuts the number of duplicates in our system by half – bringing the total number to under 1%.
Easily find the events you need with alternative title search
We now support searching by alternative titles in both our web application and our Events API. Alternative titles may include different spellings and variations of event titles, and even acronyms rather than the fully spelled-out version of event titles. This means if you don’t have the full name of the event handy, but you know the acronym, for example, the event will still appear in your search results.
Here are a few examples of the search picking up alternative titles (log in for access):
Our ML de-duplication improvements are a key element to this release, as it reduces the number of duplicate events in our system – especially those caused by alternative titles. Now, when an event is de-duplicated due to an alternative title, that title is added to the searchable list for that event.
Check out the October Quarterly Product Release webinar hosted by Director of Customer Success, JJ Mills – or log in to your PredictHQ account to explore the new updates for yourself.
