These are exciting times for Eventbrite! Since our beginning Eventbrite has served a platform for event organizers. We allow organizers to market their events, message their attendees, and most importantly we take care of the hassles associated with selling tickets. But recently we have began to take a more active role of connecting potential attendees to the events that they will enjoy and in doing so, we are becoming the marketplace of events!
Personalized Discovery – the Center of the Marketplace
And at the heart of this transition to a marketplace of events is the notion of discovery – the ability for users of our site to serendipitously discovery events that they will enjoy. At Eventbrite, we think of discovery as occurring in two predominant modes: search and recommendation. In search, the user is actively telling us what type of event they are looking for, and it is our job to return results that are most relevant to their search. In recommendation, Eventbrite will provide passive users with recommendations for events that they might be interested in. Personalization is an important aspect of discovery as we hope to provide users not only with recommendations of popular events or search results that match their query, but we also want to tune our recommendations and search results to match individual user preference based upon their previous interactions with Eventbrite and the things that they have explicitly told us about themselves.
But this is no small undertaking, and Eventbrite has some interesting challenges to overcome. A chief concern is that, unlike movies or consumer goods, events are an unusually short-lived type of product. Netflix and Amazon use customer interactions to build rich recommendation models of their products through a technique called collaborative filtering. But when an event is published to Eventbrite there is no user interaction data and by the time the event is finished we may have only started to adequately understand how an event may be matched to our users. Though this situation will be challenging, we believe that is can be overcome via a hybrid approach – a combination of collaborative filtering and content-based recommendation and search. Read on…
Blueprints for Personalized Discovery
We are still in the early stages of building out our discovery platform, but we have some good ideas about where we want to go and how we plan to get there.
A major component of our strategy is collaborative filtering. By watching how users interact with the events in our inventory (for instance through clicks, purchases, shares, etc.) we can build a model of user-to-event affinity. Currently we are using this affinity model to send personalized event recommendations to users that receive our event follow-up emails. And right now we are rolling out a new “related events” feature on our event landing pages that is driven by this same user-to-event affinity model. The model is quite flexible, so in the future we may use it to recommend events customized for a group of users (think “night out on the town” with friends), or we might tune our related events feature so that it the related events are further personalized for the viewing user.
But, as alluded to above, one of the short-comings of collaborative filtering is that building a model requires quite a bit of user interaction data, and at the moment an event goes live there is no interaction data and thus no chance that the new event will be recommended. Because of this we will also rely upon content-based strategies for recommendation and search. Initially this will be based largely upon the organizer-specified categorization for the events and the text we find in the title and description fields (along with filters on things like location). But we are also looking at creating an event tagging system which will allow for much more finely tuned categorization of events. We are also building up user profile pages and we may eventually allow users to tag themselves with the same system being used for events. This would make content-based recommendation as simple as finding events that have the same tags as the user.
There are some problems though in having two competing recommendation systems. How do you combine the results provided by each system? Is it really a good idea to maintain two competing systems which probably have their own technology stacks? And do you really want to query both systems every time a recommendation is required? We think this seems silly. So, we’ve decided to roll both collaborative-filtering and content-based recommendation into one system and we’re doing this with search technologies. Think about it this way, search engines are natively good at content-based search and recommendation – you just store the text and categorization for each event in the search engine and then you query it as you normally would. Then, if we want to combine this with collaborative-filtering, we have to find a way of including a new field or making a new type of query that incorporates the information from the user-to-event affinity model. As it turns out, there are a few different ways to do this; but that’s a discussion for a different blog post! Once content and collaborative filtering information are in the search engine then the two sources of information can be balanced in much the same way that matches in the title and description field would be balanced in more mundane uses of search. Ultimately search and recommendation become one and the same thing and both are all the more relevant because they incorporate content-based and collaborative filtering strategies.
Will it Work?
Well, this is the most exciting question of all, because we don’t know yet if this will all work as we’ve planned. In a previous life as a search consultant I built a prototype that closely resembles the system we’re working right now, and the results were very encouraging. But this time we’re building for much larger scale. However I have seen search engines perform some amazing feats of scaling in the past and I have a strong suspicion that search technologies will not let us down here.
If you’d like to see how we’re progressing, then check back in to the Eventbrite engineering blog from time to time. In the coming months I plan to blog fairly regularly about our progress. So pull up a chair, you’ll have a front-row seat to all of the action.
P.S. In case you’re interested in the broader discussion of search relevance, then you should check out Taming Search, a book I’m writing with Doug Turnbull. The techniques I’m using here will be covered in depth in a later chapter of the book. If you decided to buy the book, then please be sure to use the code
mlturnbull to get 50% off the price of the book.