Measuring popularity in online music: social media, maths & the influence of fans


Photo by raygunb

I’ve just been in Groningen in the Netherlands to brainstorm and research Tribemonitor – an online information service to artists and record labels, created by New Music Labs.

The purpose of Tribemonitor is to measure the popularity of music artists based on social media buzz across a range of platforms, rather than on sales or radio airplay.

Measuring online buzz is not a simple thing to do, however. There are some scrapable and publicly accessible pieces of information such as Last.FM plays or numbers of MySpace friends that are obvious and countable. These simple statistical measures that make a good starting point that can act as a basis for artist consultancy (or reassurance): number of MySpace plays, number of artist followers on Twitter, number of YouTube views, etc.

But these metrics only measure what could be described as fan activity, rather than a useful and measurable social score, which would have more to do with the extent to which that artist is being discussed outside of their own sphere of influence. And this is the reason for this intervention.

Social capital and the popular music artist
Popularity is the basis on which commercial music derives income from recordings and performances. But popularity is not the same as CD sales or gig attendances. Instead, those are mechanisms of commercial activity based on the social capital afforded by popularity. In other words, popularity is the engine of commercial music success, and not simply its measure.

Thus, gauging that social currency allows for commercial approaches that understand and maximise popularity. Being interesting, noteworthy or remarkable is not a business strategy – but it is a platform on which a business strategy can be built. Monitoring and tracking the social capital of popular music artists offers up important market information for analysis and interpretation.

The first step is to be countable and comparable – both against competing artists, and longitudinally with respect to one’s own previous performance. By comparing data over time, it’s possible to get a sense of ‘what’s working’ and ‘what’s not’. It’s a blunt tool, but does flag up when and where there are things of interest going on. On that basis, targeted content analysis can be indicated and recommended.

Seeking a simple but practical approach
Leaving aside sentiment analysis (whether the mention is positive or negative, or to what degree), there is challenge enough in simply weighing that media data – let alone analysing it for mood. However, for the sake of simplicity, and on the pretext that ‘any publicity is good publicity’, merely identifying messages and their capturing their echoes through the social media environment provides the basis for a useful indicator of online ‘buzz’.

A good place to start is the area of mentions in social media contexts. While the demographics of different services are quite different (users of Twitter tend to be older and of higher socio-economic standing than users of the popular Dutch social network Hyves, for instance), it’s difficult to take the temperature of social media buzz in closed and largely private systems such as Facebook. A predominantly public and externally measurable system like Twitter gives a good leaping-off point in terms of measuring online social currency, and how that changes over time.

At the very least, it provides a starting point for an exploration of the complexities of this sort of social media data.

Simply measuring mentions of the artist is not sufficient. There is a difference in impact between a mention and a reply – but an algorithm that worked by identifying the artist’s Twitter user name would not necessarily distinguish between the two:

Mention:
I went and saw @thisiskrause perform last night. She was amazing!

Reply:
@thisiskrause I liked your show.

Twitter distinguishes between those two types of directed message – and so as far as ‘buzz’ is concerned (certainly from a promotional culture perspective), a mention is ‘worth more’ than a reply. And a mention may equally be the artist name (‘Krause’) or their Twitter handle (‘@thisiskrause’).

Now, if I tweeted those messages above, all of my followers would see the ‘mention’ – but only those people that followed both me and @thisiskrause (as well as Krause herself) would see the ‘reply’.

Next there are variables to be considered:

1) How many followers do I have?
2) How engaged/interested are those followers?
3) How influential are my followers on average?
4) How many people responded to or retweeted that particular message?

In order to factor those in, it’s necessary to come up with a calculation that accounts for each, and then arrive at a score that can be additive, so that a total figure across all mentions within a particular period can be arrived at. That score can be monitored over time for that one artist, or compared across the board with other artists.

Doing the maths
While the numerical value of ‘social media score’ is essentially an arbitrary figure, as it does not count a specific measurable object, when applied across the board it does provide a meaningful and (most importantly) a comparable index.

The first variable – number of followers (F) – is easy to count. But it’s perhaps not the most important thing. If I have 1000 followers, but they’re not really paying attention to what I have to say, then my tweets will have less impact than someone who has the same number of followers, but whose followers actively engage that person in conversation.

So we came up with a ‘Social Score Multiplier’ (M), which is simply a means by which it is possible to arrive at a weighted figure that is based on the average number of replies that the person receives each day. In other words, the degree to which the tweeter’s followers are ‘engaged’.

After playing around with a number of formulae to come up with a figure that would make what we considered to be a reasonable adjustment with respect to the level of attention and interest that person receives online (E), we decided on an algorithm that would return a figure that started at a multiplier of 1 (zero average replies a day), and increased by 0.1 (to a multiplier of 1.1) for each 1 daily reply, averaged from a year’s worth of data.

So the social score multiplier was 1 + (0.1 x (Replies in the past year ÷ 365))

M = 1 + (0.1 x (E ÷ 365))

Then we factored in the small impact of the overall influence of the followers of the person who has tweeted about the artist, by including the AVERAGE number of followers that the tweeter’s followers have (A), as well as an adjustment for the number of retweets and responses that individual tweet inspired (N).

In so doing, we arrive at a weighted ‘Buzz’ score (B) for each tweet:

B = ((F x M) + ((1+N) x A))) ÷ 100

To explain – the buzz score of a certain tweet is measured by the number of followers, adjusted by the social score multiplier (to account for how engaged that user is), PLUS the number of retweets multiplied by the average number of followers (plus one, to avoid a zero result where the tweet is not retweeted at all), divided by 100 – to give a usable and comparable score.

So… if a person with 1000 followers, who has had 1825 replies in the past year (an average of 5 replies per day), tweets about Krause, their social score multiplier is 1.5. If their tweet about Krause is not retweeted, and their followers collectively have an average following of 150, then the sum is as follows:

((1000 x 1.5) + ((1 + 0) x 150))) ÷ 100 = 16.5

The overall social score of that one tweet is measured at 16.5

Alternatively, a tweet from a person with 2000 followers, who had 1200 replies in the past year have a social score multiplier of 1.33 (ie: 1 + (0.1 x (1200 ÷ 365))). Let’s assume that their followers also have an average following of 150, but that they were retweeted once. The social score of their similar tweet will be as follows:

((2000 x 1.33) + ((1 + 1) x 150))) ÷ 100 = 29.58

And finally, that one person who retweeted our last example has the following impact, given that nobody retweets them, they have fewer followers (say, 100), their followers are less engaged with them (20 replies in the past year) and their followers have, on average, fewer followers themselves (60).

((100 x 1.01) + ((1 + 0) x 60))) ÷ 100 = 2.21

Now, in order to arrive at an overall social score for Krause for this week (day, month…), we simply add up the scores of all of the mentions in that period.

16.5 + 29.58 + 2.21 = 53.3

Now we have a number that we can compare to previous periods, and to other artists.

This is not a “value”
It’s important to bear in mind that there are plenty of other variables that could potentially influence this social score, if they were factored into the calculations. For instance, if a person who tweets a link to a YouTube video is then retweeted by two other people, there are some very simple calculations that would follow above. However, those calculations do not factor in the extent to which an audience overlaps. If I have 1000 followers, and I send out a message – and two people each with 1000 people retweet that message, we come to a conclusion as if there is no overlap between those three audiences of 1000 different people. However, it is conceivable that they are the SAME 1000 people (or at least, significantly overlapping), so that rather than reach 3000 people, you have reached 1000 people three times.

And there are plenty of other variables besides.

However, the social score is not a measurement as much as it is an indicative figure. It would be possible to complicate the algorithm with this, and any number of other variables, but there would be diminishing returns in terms of factoring those variables in – and counteracting factors such as the principle of reinforcement. Arguably, a message heard three times has more than three times the psychological influence than that message heard once – so what may appear on the surface to be a diminished impact may be more influential through repetition.

Without getting too deeply into cognitive science (and, for that matter, deep maths), it is possible to arrive at a figure that while neither pinpoint accurate nor comprehensive in terms of what it represents, can still be a meaningful and useful figure. While a broad and necessarily imprecise statistic, the social score can provide a consistent and comparable guide that factors in some of the main influencing factors within a social media framework.

This is not a “cause”
However, I’d caution further that it’s not possible to ascribe causal factors to that data.

For instance, while it is meaningful to assert that an artist with a score of 100 is more “interesting” in the online social sphere than an artist with a score of 10, one cannot draw correlations between those scores and sales of records or attendances at performances. An artist who has not released an album for ten years, but has recently died may draw more ‘buzz’ online, but will not attract any more concertgoers as a result.

And while it’s possible that the interest in this newly-deceased artist may result in an increase in sales, that would only be true if their record is in the store available in shops or online.

In other words, the social score gives an indication ONLY of how “interesting” the artist is at a given point. But that level of interest provides useful market information to artists and labels that might indicate where opportunities may exist.

It is neither the role of Twitter nor its effect to cause consumers to act in a particular way – nor is it connected with the extent to which businesses are supplying a market. The social score is a broad marker of discussion – and not a reason that one artist may be more successful than another in financial terms. Some moderately successful artists are entirely invisible in the online sphere, and some lesser known acts are very active with large and engaged fan communities online.

This is an instructive and indicative data set
There are three main ways in which this data can be used:

A snapshot figure – a single score looked at in a moment in time. Useful as a comparison amongst other, similar artists. (“How interesting am I, comparatively speaking?”

Over time – whether or not that number increases or decreases from week to week. (“Am I doing better or worse online? Did my publicity stunt draw more people?”)

Rate of change – the acceleration (or deceleration) of buzz. If the score goes from 1 to 10 to 100 to 1000 week by week, that is an exponential rate of change – as opposed to a score that goes from 10 to 20 to 30 to 40 – which has a linear growth. (“Is there something interesting going on with this artist that we can capitalise on?”)

Providing commercial clients with these statistics, which would be difficult to collect, analyse and interpret themselves, provides an opportunity to discuss ways in which the social media environment could be captalised on and changes in interest can be used as an occasion for marketing.

Moreover, it would be possible to identify and engage with key online influencers – in order to incentivise and reward those fans and opinion leaders who contribute the most to this social media process.

We don’t think the formula is correct
While the algorithm we’ve arrived at is simply a tool to arrive at a comparable and trackable social score, we think it will be very possible to improve upon its methods and become more sophisticated with time.

The second part of this research is to invite people to suggest alterations to the formula, integrate other factors you think may be important, and correct the ways in which the design of the formula calculates and communicates what we are trying to capture and present.

While Tribemonitor is a commercial service, which includes further analysis and interpretation of the data, we believe the means by which some of the data is arrived at would benefit from an open source approach. Mathematicians, statisticians, psychologists, social scientists – as well as musicians – will have opinions as to how this algorithm could be developed further.

As a piece of research, we are very keen to hear what effects could be incorporated to represent important social impacts; and ways in which the numbers could be more meaningful.

For instance, we were interested in ways in which the social score multiplier (M) could be expressed as a ratio of the number of followers (F). Having 1 reply per day, on average, could be considered not very engaged for a Twitter user with thousands of followers – but very engaged indeed for a Twitter user with only ten followers.

Likewise, adjustments could conceivably be made to allow for differences between occasional tweeters and verbose Twitter users. It’s certainly possible that a single tweet could be buried under an avalanche of tweets if the person is a heavy twitter user. To what extent is this significant, and how could this be represented in the formula?

We look forward to hearing (and sharing) your thoughts.

Leave a Reply