Cracking the Code on What’s Trending

With audience attention becoming more elusive every day, data that can help shed light on what matters most to people is at a premium. Facebook sits at the center of this treasure trove of social data, with more than 1 billion users actively engaging and sharing relevant content in real time.

Perhaps that’s why people have been all the more surprised by recent revelations that Facebook uses human editors to manually edit its Trending Topics section. If Facebook has the data and the ability to analyze it, why not just give the public direct access to these insights?

The crux of this problem was recently summed up by The New York Times using Wrigley’s popular fruit flavored candy as an example – an algorithm may suddenly detect a rise in the word “Skittles,” but surfacing that topic without context appears to have dubious news value. Facebook also used the example of the topic “#lunch” spiking every day at noon – certainly not the type of insight they would want to surface on the homepage. According to The New York Times, about 40% of topics identified by Facebook’s algorithms represent this kind of noise. Looking at the problem another way, are the #Kardashians or #JustinBeiber ever not trending?

As a B2B focused technology company, we don’t face the challenge of being a public news source, but our recent efforts to surface our own data publicly as well as the public discourse about Facebook got us thinking about the complexity of analyzing trends and the evolution of our own real-time data set.

For our customers, the issue is less about media neutrality than accuracy at scale. We need to help them cut through the clutter and take action based on what their audiences cares about right now. In the end, however, the challenge is the same – we must figure out how to make the most of our algorithms with minimal human intervention. Here are five key lessons learned from more than 5 years of algorithmic trend spotting:


1) Incorporate data from many sources
Every platform contains its own inherent biases. For example, people may be more likely to use Facebook to engage with one type of content vs. another or certain demographics may be more likely to use Twitter than others. Also, it’s important to consider that direct comments on publisher websites can provide a vast amount of data and “hints” about what topics are truly trending and why they’re part of the conversation. In short, it’s critical to mine the full extent of the web to build a balanced view.


2) Leverage historical data
Part of setting a trend watching machine up for success is providing enough historical data to understand patterns (and what breaks the pattern). With years of data it becomes easier to separate the normal ebb and flow of Justin Bieber fan chatter from buzz driven by a new album release or breaking scandal.

In the trend snapshot below, we can see a never ending stream of chatter from Justin Bieber’s fan base but a clear spike centered on the Billboard Music Awards, indicating a significant trend.


3) Breadth matters
Looking beyond simple volume and velocity to understand demographics and sentiment can also help separate the signal from the noise. For example, rapidly shifting sentiment can indicate an important trend among a topic that typically generates a high volume of conversation. Demographic segmentation can also be critical for trend monitoring, especially for businesses looking to understand what matters to their customers. It’s more the exception than the rule that a trend will be universally relevant.

The importance of integrating sentiment can be seen below. We see a consistent stream of buzz around the topic of Hawaii, but a sudden shift toward negative sentiment helps to flag this as a unique and important event.


4) Have a strong validation process
Watching changes in volume, velocity, and sentiment can help signal that a topic is trending. But why is it trending? This is the Skittles problem. By scanning thousands of articles and videos for frequency and context around trending topics, we can better infer the most likely possible reason for the trend and surface it within our results.

In the example below we can see that our system detected a spike in the word “barista.” Surfaced alone, this isn’t particularly meaningful. But, placed in it’s most likely context, we can identify this as a celebrity trend.


5) Provide transparency
If you’ve read this far, hopefully you believe that we’re walking the talk in our early attempts to provide more visibility into how we uncover trends. Facebook wisely responded to their controversy by offering up detailed information about how trending topics are selected for their homepage. Ideally, this should be done in the product UI as well. The more people have visibility into how trends are being calculated, the better they can interpret the data, including the occasionally quirky results that may sometimes appear when using a strictly machine-driven approach.

Basic changes can a make a big difference in the way people interpret data. For example, we took the simple step to bold trending topics in our dashboard. Immediately, customers were better able to understand at a glance which topics were trending and why we were flagging them.


Dealing with massive volumes data and the complexities of human language can be challenging, but we believe it’s worth it to unlock the vast potential of online trends to better inform online advertising and audience insights. You can visit our Trend Pulse page right now to see the unedited, real-time trends detected by our system for specific audience segments.