Machine Learning: Pioneering Social Media Software Helps Businesses Find New Customers in the Mass Universe of Tweets
Machines can’t think like humans can. Nor can they learn the way we do. But thanks to a branch of artificial intelligence that HipLogiq is using, computers can improve at certain jobs on their own, without being explicitly programmed.
As a result, companies like ours can improve their customer service, which in HipLogiq’s case means helping companies better mine Twitter for customer prospects. Machine learning technology is what Netflix, Pandora and Amazon.com use to give customers timely, relevant offers.
The “machine learning” I’m describing involves computer programs that can get better at tasks when they’re exposed to new data, according to the reference web site WhatIs.com. The programs essentially look for patterns in information and then change their actions accordingly, WhatIs.com says.
At HipLogiq, we’re using a machine learning program called Joshua to parse out relevant tweets in certain geographic areas that companies like Papa John’s International and Dunkin’ Donuts can use to find new customers through Twitter. Joshua is what’s called an “algorithm,” meaning it solves a problem – mining for relevant tweets – through a limited number of steps.
The upshot is that Joshua helps us better find those needle-in-a-haystack tweets that companies want – the ones in which customers have expressed a strong desire or need for something that our clients can help with.
Like most machine learning programs, Joshua needs a little help from humans. HipLogiq lends him a hand first by running millions of tweets through filters that help ferret out ones that have a key word, such as “pizza.” We also toss out tweets with words we don’t want (such as “Papa John’s,” because that person is probably already a customer) and toss out anything with profanity.
Finally, we’re left with a relatively small pool of tweets that a human can go through to pick out the most relevant ones for, say, Papa John’s. Joshua will learn not only from which tweets we select – “I could really use a pizza” – but also from the ones we reject (“a neighbor kid called my son a ‘pizza face.’”)
Joshua essentially builds what we call a “characterization scale” to determine the relevancy of a given tweet. Within in as few as 140 tweets, Joshua’s ability to predict which tweets are relevant approaches 80 percent of its maximum.
After we’ve run Joshua, we recalibrate the software several times with more tweets that humans have vetted. In other words, we double-check to ensure it is making precise measurements – or, to put it another way, we see that the tweets it is flagging as relevant are, in fact, pertinent to what the client is seeking.
Our results speak for themselves. For one restaurant client, for instance, we’ve responded to roughly 580 Twitter conversations. That group, in turn, has produced nearly 1900 customers – a 3.25 to 1 relationship.
In other words, one relevant tweet generates 3.25 new customers, because the people to whom we’ve responded have told friends about the offers they’ve received.
Those results come from all of the techniques I’ve described – using keywords, geotargeting, filtering, machine learning and the rest. But the bottom line is that Joshua can make a marked difference in presenting relevant tweets to the customer.
And as time goes on, the technology will only get better at finding those needles in the Twitter haystack that our clients desire.
Social media marketing company HipLogiq pioneers machine learning to help businesses find new customers on Twitter.
[image error]
