Ian, Kona is wondering where you are…

So goes the title of the second such email I’ve received this year from United.

I'm right here, Kona

The thing is, I have a trip booked to Kona already…on United even.

Last time I got an email with this subject line, I had two trips booked to Kona on United (one for the Hawaii 70.3 race back in May, and one for Kona in October), and they were in fact the only upcoming trips I had at the time.

Somewhere at United, the left hand doesn’t know what the right hand is doing.

As ridiculous as this scenario seems, it happens all the time, and the problem isn’t unique to United Airlines. Marketing automation systems attempt to do personalized campaigns in order to get a better response rate. But they lack two central ingredients:

  1. Current situational awareness from other transactional systems. United’s reservation system and all ancillary systems, such as checkin systems, frequent flier account, etc, aren’t one monolithic database, but rather a series of systems that need to update one another. This doesn’t happen instantaneously (this is why your frequent flier account doesn’t have your most recent flight on it as soon as you land); if every system updated every other system instantly, everything would bog down and grind to a halt – at least in the IT infrastructure on which a lot of these systems are currently built. Everyone wants “real time,” but most often the cross-system updates are still scheduled in large offline batch operations. (A good lesson on the IT complexity of airlines is talked about here and other places with respect to the merger of United and Continental.)
  2. Intelligent, adaptive systems that take as much current customer context into account as possible. This is the holy grail of personalized marketing, but more often than not, campaign management systems are rule or model driven, and they only look at a few factors, such as when I made trips in the past and to where. The timing of my Kona emails is no surprise when you think about my past history: I’ve done the Hawaii 70.3 race in May a grand total of seven times since 2006, and this is my fourth Hawaii Ironman in October. The fact that I’ve already got reservations is a comical illustration of point 1 above, but anyone who knows me, either personally or through my various social networks, would know which races involving travel I’ve already signed up for, and combining that intelligence with my current transactional status would prevent emails like this that make it seem like the airline doesn’t know me, when in fact I’ve flown well over 2 million miles on United. They should know me very well by now!

In fact, if they were smart, they’d be making me offers for business class fare to Klagenfurt, Austria next summer, seeing as (a) I’ve signed up for Ironman Austria, and (b) I likey my business class as long as you give me a reasonable deal. True personalization like this is something we’ve worked very hard on a Saffron with our autonomous learning platform, and we have some very good proof points in accuracy of individualized product prediction. We co-presented on this with our customer USAA at Wharton back in April. “Cluster of one” personalization is just in its infancy, but with the IT industry on its way to adopting technologies that solve the two challenges above – low-latency updates from transactional systems and intelligent systems that can treat all of us like the individuals we are – we shouldn’t be surprised to see the marketing of the future become “scary good.”

Meanwhile, back in the present, two days after I booked my trip to Phoenix for this November’s Ironman Arizona race, I got this:

Phoenix, I'm right here!

<sigh> Really, United?

Aggressive Drivers: Time for a “Page of Shame”?

The San Francisco Peninsula is a great place to live and train; the cycling in the Santa Cruz Mountains and along the coast is hard to beat (especially if you like climbing). But the one fly in the ointment in training paradise has been the few drivers and motorcyclists who make cycling not just unpleasant, but downright dangerous — and sometimes fatal.

Today was a running day, and as I was heading home from Woodside Elementary School (where we park to go running in Huddart Park), a Mini Cooper turned onto Woodside Rd. from Mountain Home Rd. without stopping and proceeded to tailgate the SUV directly in front of him. The SUV had another car in front of him and was going the speed limit. Nothing really out of the ordinary, though, until the Mini driver started “slaloming” in the line, alternately cross the centerline and going into the bike lane. He almost took out a cyclist.

Then he continued weaving, even trying to pass the two cars ahead by going right into the bike lane. Luckily there were no cyclists there at that moment, but he got stopped at the light at I-280, where Woodside Rd. splits from one lane in each direction to two. I ended up right behind him and snapped his photo:

2014-02-16 11.16.48

He wasn’t pleased — he got out of his car (as did his passenger) and started threatening me. Luckily, I had witnesses all around, so he thought better of escalating things any further. I would have been happy to discuss things with a county cop present.

The point of this is not to single out this one guy (though he’s definitely one of the most aggressive drivers I’ve ever encountered, especially in a town like Woodside, which is usually crawling with cops on weekends); the larger question is how to use the power of social media to rally cyclists around identifying these dangerous drivers and doing something about them. They are literally life threatening.

I’m reminded of a great early website from the mid 90s: The Highway 17 Page of Shame. It was entertaining reading back then, and all of us have experiences like this at one point or another. But with all of the technology available today — camera phones and small video cameras with amazing resolution, plus the social networks to distribute the content to a wide audience — maybe it’s time to put that to use to ferret out not simply those drivers that annoy us but rather those who recklessly put our lives at risk.

Can social media predict tonight’s winner of “The Voice”?

Earlier on in the season, I’d have said an emphatic “no” based on the correlation (or lack thereof) between buzz on Twitter, positive buzz on Twitter, number of followers on Twitter, week-to-week growth in followers on Twitter, Klout score — and just about anything else I could measure — and the actual results. The past couple of weeks, however, have gotten slightly better results, and that could be for several reasons:

  • Better odds in general as the number of contestants declines.
  • More focused social media discussion of the remaining contestants — i.e., less “noise” — overlapping better with voting behavior.
  • More audience engagement as we hit the semi-finals and finals, similar to what we see with big-time sports (e.g., Super Bowl, March Madness, NBA finals, Stanley Cup raise the volume of discussion over what you see in the regular season).

Remember, too, that we are using Twitter buzz only as a proxy for voting behavior — tweets don’t actually count in the voting. What does count is telephone voting, SMS voting (but only if you’re on Sprint), Facebook voting and iTunes downloads, this last category having a disproportionately large effect with sufficient volume. We do not have access to any of the metrics for these sources, so the experiment has been to see how well buzz on Twitter correlates with the outcomes each week. So this is the final week of the experiment for this season.

So let’s look at overall Twitter buzz:

voice all 2013-06-18 at 2.22.52 PM

And now restrict the view to only “positive” buzz:

voice positive 2013-06-18 at 2.25.52 PM

This does change the order between Michelle and Danielle if we look at the Top Entities view; the Swon Brothers are a distant third in both sets of metrics. We could posit that the Swons’ fan base doesn’t tweet, but that actually runs counter to generally observed engagement levels — the brothers are very active on Twitter compared with the other two contestants, not just with posting but also with replies and retweets. Perhaps two is better than one for keeping the chatter high.

Another interesting phenomenon became apparent in the buzz charts if we look at Top Hashtags: Michelle has a hashtag, #4eyesontheprize, that didn’t exist when I set the topic up, so this tag is not being counted in her totals under Top Entities. This means that her numbers are potentially higher than the Entity chart shows, though only for tweets that contain that hashtag and no other reference to her. So we can’t simply add the two numbers together. There’s also another telling phenomenon in that in overall buzz, #teamusher beats #teamblake, even though Team Blake has two of the three finalists. However, those numbers look vastly different if we just look at positive buzz. The thing about the sentiment coding here, though, is that it favors precision over recall, which is a somewhat technical way of saying it undercounts rather than overcounts. So when push comes to shove, I’m going with the old Oscar Wilde adage: “the only thing worse than being talked about is not being talked about.” — any buzz at all trumps positive buzz.

In that vein, however, we do have to remember that Danielle has consistently out-buzzed Michelle in many of the other weeks, and also that Danielle has more than 40% more followers than Michelle. On the other hand, follower counts have not correlated at all well with actual results, neither for “The Voice” nor for “American Idol”. Also, this week’s overall volume is significantly higher than in those previous weeks, so we’ll assume it follows the engagement trend of other finales, in which overall votes or other forms of engagement tend to correlate with viewing audience.

Anyway, it would appear the numbers are telling us that the Swon Brothers get third, and that there’s a somewhat conflicted picture of whether Danielle or Michelle wins, but I’m going with the overall buzz numbers and calling it for Team Usher’s Michelle Chamuel.

One last side note: it appears that “Voice” executives finally ponied up to Twitter, as all finalists’ Twitter accounts are now verified. In Twitter-speak, that means they now count as celebrities. 🙂

Social media numbers for “The Voice” semi-finals

Back from a hiatus on the Big Island to race and relax, it’s time once again to review the social numbers for this week’s “Voice”.” Past weeks have shown a relatively poor correlation between buzz on Twitter and actual results, but as the pool of contestants has decreased, the correlation has improved. This can of course be due to the increased probability of being correct simply as a function of smaller potential outcomes, but let’s see what the numbers say regardless.

For the period spanning last night’s live performances on the East Coast through midday today, we see the following:

the voice overallCompare this with only positive mentions, and we see a similar distribution:

the voice positive

The good news is that there’s a clear separation between the top two and the remaining three, so it appears as though Amber and Danielle are clearly safe. Two will be sent home tonight, though, and while it appears as though Sasha will be one of them, the difference between Michelle and the Swon Brothers is not very large. Both also seem to have a very loyal following that has kept them on the show even when the buzz numbers were relatively low, so at this point it’s just a guess as to which of the two will remain after tonight’s show.

Here I’m going with the Swon Brothers, purely because the novelty of a duo going to the finals vs three female soloists might have gotten the vote out better.

As this season has shown, though, the social numbers have been wrong before, so don’t take it to Vegas. 🙂

Another Week, Another Set of Live Rounds of “The Voice”

This week on “The Voice,” we’re down to 10 contestants, another two of whom go home tonight, so if nothing else our odds of predicting who’s going home are naturally improving week to week. But let’s look at the social media numbers again. First, overall “share of voice”:

thevoice-overall 2013-05-21 at 10.07.39 AM

Then, positive “share of voice”:

thevoice-positive 2013-05-21 at 10.07.19 AM

For the first week ever, both rankings produce the same bottom two, Amber and Kris.

If we look at week-to-week follower count growth, however, we get a different ranking:

Twitter handle Followers 5/14 Followers 5/21 Delta Percentage
@theswonbrothers 14750 19588 4838 33%
@michellechamuel 21305 27370 6065 28%
@josiahhawley 40894 50517 9623 24%
@dbradbery 45723 55430 9707 21%
@hollytmusic 13454 16273 2819 21%
@ambercarrington 14520 17149 2629 18%
@kristhomasmusik 12130 13966 1836 15%
@sarahsimmusic 25314 28828 3514 14%
@sashaallenmusic 16674 18985 2311 14%
@judith_hill 39953 43041 3088 8%

So the question is whether buzz or follower count (or neither) correlates better with actual results. So far, follower count has correlated a lot less well, so this week I’m going with buzz — especially since overall buzz and positive buzz line up this time.

But social numbers haven’t been right yet this season, which I’m sure make Kris and Amber happy to no end. 🙂

What Did We Learn from This Week’s “The Voice”?

Yesterday I looked at a number of different cuts of social media data and found a number of candidates for the Bottom 2 on “The Voice.” The interesting thing about the actual results was that none of the data I had access to correlated with the outcome, at least in a discernable pattern that you could build any sort of rule or model off of. I could try using a stats tool like R to find non-intuitive features that better predict an outcome, but that is probably like using a bulldozer to pound in a nail — there aren’t enough samples to make that approach meaningful. So I’m left with hypotheses:

  • Twitter isn’t a good proxy for voting behavior in the case of this show due to either a demographic skew or to the fact that the audience gravitates in general to other channels of interaction.
  • I haven’t found the right features yet. I used growth in Twitter followers as one potential proxy for “momentum,” and that correlated with one of the results — Vedo had by far the lowest percentage growth in followers week to week (he started of course with the highest number, so it gets harder to maintain the same growth rate), but in the case of the other ousted contestant, Garrett Gardner, we saw the highest percentage growth in followers, so that piece of data doesn’t correlate at all. Garrett, in fact, by any of the numbers I had access to or derived, should still be on the show. (Though I personally found his vocal performance on Monday weak, despite an interesting arrangement.)
  • The vote tallies among the bottom half of the contestants are actually fairly close (which the counts I had showed as well), so until we get to a point where we see either much greater volume or much greater differences between contestants, the best we’re going to do is random guesses over a larger pool of contestants with similar numbers. In other words, absolute rankings don’t work until you start to see larger gaps.

One thing is for sure — pure follower count is meaningless in this show (as it was in last week’s “American Idol”). A lesson for all who use follower count as a proxy for influence.

Social Media Numbers for “The Voice” Top 12

Last night’s “Voice” episode featured the Top 12, and for the first time this season, audience votes and downloads are the only determining factor for who goes home tonight. Like last week, I’ll be looking at Twitter numbers as a proxy for voting behavior so that we can see the extent to which opinions expressed there correlate with results. The voting methods as explained on the show were phone, text, Facebook and iTunes download, so Twitter is an independent factor in this equation and has no direct bearing on the outcome.

Last week, the Twitter-derived numbers correlated reasonably well. Not perfectly, though. For Team Adam, I hadn’t set up the topic in time, so had about 1/5 of the numbers I had for the Team Blake and Team Shakira contestants. So there my numbers were just plain off. The other factor was the judge’s discretion, and Blake elected to save the Swon Brothers rather than Justin Rivers. Welcome to the big leagues!

I personally enjoyed many of last night’s performances, but what’s important is what those who might vote thought. I took a couple of different cuts at the data. The first is simply looking at “share of voice” from the period of the airing of the show (starting in the Eastern time zone) through late this morning Pacific time. Here’s what we get:


Simply taking the ranking from most mentions to least during this time period, we end up with a Bottom 2 of Kris Thomas and Michelle Chamuel. We could look at the data another way, though: if we filter by positive sentiment — the idea being that people who express positive opinions are more likely to vote (or perhaps “you vote FOR someone, not AGAINST someone else”) — then the rankings change a little:



Now Amber Carrington and Holly Tucker make up the Bottom 2. However, note that the overall numbers aren’t very high when we apply this filter (sentiment analysis isn’t an exact science by any means, and the software used here is biased towards precision over recall, so is somewhat conservative.

Confounding the data further is iTunes, which doesn’t give precise data but does provide a “popularity” meter. If we look at each of the above sets of Bottom 2, we do see that Amber and Michelle’s performances did not max out the popularity meter, whereas Kris’ and Holly’s did.

One other data point we can look at is a week-to-week difference in Twitter followers. If we rank by percentage gain first, then by overall number, we get:

Twitter handle FollowersLastWeek FollowersThisWeek Delta Percentage
@garrettgardner2 16893 24426 7533 45%
@michellechamuel 14872 21305 6433 43%
@dbradbery 32352 45723 13371 41%
@josiahhawley 29100 40894 11794 41%
@theswonbrothers 10644 14750 4106 39%
@hollytmusic 10232 13454 3222 31%
@ambercarrington 11274 14520 3246 29%
@kristhomasmusik 9721 12130 2409 25%
@sarahsimmusic 20826 25314 4488 22%
@sashaallenmusic 13735 16674 2939 21%
@judith_hill 36078 39953 3875 11%
@vedothesinger 59217 61606 2389 4%

Based on this, it looks like Michelle picked up a lot of followers, so I’m going to have to go with Kris and Amber as the ones going home, with a possibility that Holly ends up somewhere in there. But I’m not taking it to Vegas — the differences aren’t large. Tough call this week!

Enjoy the show!