So What? Marketing Analytics and Insights Live
airs every Thursday at 1 pm EST.
In this week’s episode of So What? we focus on prepping your data for analysis. We walk through pulling your planning from the previous week into your data extraction and exploration. Catch the replay here:
In this episode you’ll learn:
- what to look for in your data set
- what questions you should be asking
- how to get it prepped for analysis
- Data prep for analysis part 3 – 2/11/2021
- Data prep for analysis part 4 – 2/18/2021
- How do you benchmark a website’s performance? – TBD
Have a question or topic you’d like to see us cover? Reach out here: https://www.trustinsights.ai/resources/so-what-the-marketing-analytics-and-insights-show/
Katie Robbert 0:20
Well, Hi, and welcome to so what the marketing analytics and insights live show.
This week we are covering part two of how do I clean and prep my data for analysis.
If you missed part one, which we covered last week, which are the basics of preparing for data analysis getting organized, you can find that on our YouTube channel at Trust insights.ai slash YouTube, it is in the sowhat playlist.
Today in part two, we are covering the data itself.
So we’re going to be walking through an actual data set that Chris has been working on and sort of going through what to look for in your data set, what kinds of questions you should be asking, and based on how you got organized before you even pulled the data, what additional organizational steps you should be taking in order to set yourself up for success.
So that’s what we’re covering today.
Chris, john, anything else before we dive into the data? Lake, let’s go
Unknown Speaker 1:17 take it away.
Christopher Penn 1:20
So as your call last week, one of the first steps in the process is figure out what our goals are.
Right? What What is it we’re trying to prove? So I was actually writing a blog post about someone was asking how we do social media strategy planning.
And the first thing came up to mind was okay, well, we kind of need to know, like, what’s working? So before we talk about Data Prep, and data collection, and some social media data, we got to ask ourselves, what should we be caring about? So a while ago, we ran some customer journey analyses, these are attribution models.
This is on my personal website, I wanted to compare 2020 versus 2021.
Looking at social media, Now, obviously, Google and email far and away the things that drive a lot of conversion on my websites, we can give that a mess.
But in 2020, Twitter is sort of, you know, the fourth most useful source file by medium than Facebook.
And then going down here, if YouTube down the middle and LinkedIn, this, by the way, is why I stopped doing a daily YouTube show.
Because when we look at 2021, it’s still, you know, generating almost no conversion.
So even though it’s fun, you know, it didn’t really move the needle for conversion.
What we saw in 2021, is, if you look, Twitter generated at nine conversions, in January of last year, and 11, this year, so even though it is proportionately a smaller piece of the pie, because email is doing so much better this year, it still is generating good conversions.
Meanwhile, Facebook and LinkedIn are down in the middle of the pack like, well, this isn’t really working for me.
So the question that I had, and the data want to dig into is, Well, okay, Twitter is at least doing something, right.
It’s, it’s in there, what? What can I be doing on Twitter? That’s better? How can I improve it? So, Katie, when you’re facing a situation like this, and you look at data like this, what comes to mind? How do you start thinking about framing out the problem of how do I make my Twitter better?
Katie Robbert 3:21
How do I make my Twitter better? I like the alliteration.
So the first thing I need to look at is what the heck am I currently doing so because you want to take a snapshot of the current state, because in order to measure what you’re doing better, you need to know what you’re changing.
And that gives you an opportunity to, you know, change one thing at a time, I think that’s one of the mistakes that you know, a lot of people make that we’ve made historically is trying to change too many things all at once, then you that how you measure kind of gets away from you, because you don’t know what exactly it is it’s working.
So if you’re looking specifically at your Twitter account, Chris, the first thing I would ask you is, how active Are you on Twitter? You know, how many times a day are you posting? Is it consistent? Is it random? Do you have a current plan that you could look at to say, Okay, this is what I’m currently doing.
Where in this? Could I be doing better? Have I measured my engagements? Am I growing consistently with followers? You know, what are people doing? What things Am I posting what specific posts are actually converting? So there’s a lot to dig into before you even try to start changing things.
Christopher Penn 4:30
And so, when we start talking about those questions, we’re really talking about requirements gathering, what kinds of data do we need to be able to answer these questions? So when we go into the data, let’s go ahead and pull the actual actual Pope Twitter first.
So when you go to your Twitter account, if you go to analytics twitter.com you can export all the data that you see here, and what you get out of that is a pretty Yeah, okay, somewhat helpful spreadsheet.
You have the idea of the tweet you Have the tweet link, you have what you said in the tweet.
And you can see this whole bunches of stuff in here, the time and the date, the number of impressions, which is a number by the way, it does not show up in other social media monitoring tools, you have to get it directly from Twitter itself, your engagements, your engagement rate, and then you have all Twitter’s appended data.
So the number of retweets that tweet replies likes user profile clicks, URL clicks, so people will click the link in the tweet itself, detail expands.
So if there’s a piece of media attached to the tweet, like a photo, or a video, someone clicks on it, to look at it in greater depth permalink clicks, app opens app installs, if you have app based tweets, emails, if someone’s email the tweet, if there’s a video attached, you can see the number of media views the number of media engagements on it.
And then if you’ve been paying for it, you get all the same data for paid.
Now, here’s the challenge I run into with this.
And this is where we start getting into a Data Prep, there’s a, there’s a bunch of things that you just talked about Katie that aren’t in here, right? So we have the tweets, but we don’t really have a change in followers per se.
We do have all the different types of engagement, we have the time we have the impressions.
But I want to know how to make Twitter convert better for me.
Right? So ultimately, the measure I’m gonna have to go off of is URL clicks, because if I tweeted something, like a link to my website, I wonder if you clicked on it.
So that’s sort of my objective.
So one of the first things we have to figure out, anytime we’re looking at a data set is what’s our response variable? Like? What’s the the outcome that we’re looking for? If the outcome is not in here? You can’t really do an analysis.
Katie Robbert 6:43
Well, and Chris, I know that one of the strategies that you employ is you to actually tweet out and you know, we all do a lot of third party content, so content that isn’t yours.
So as you’re going through this, would you need to exclude those tweets and only be looking at the tweets that go to your web properties?
Christopher Penn 7:07
It’d be nice to do that.
You can’t do that.
But with this data, at least in its current form, because as folks, you know, who use Twitter, Twitter runs everything through its own link shorteners.
So every link is going to be the T dot SEO links in here.
So that’s something that to keep in mind, stuff, we’d want to engineer.
But there’s other stuff that we’d want to create, as well, that, again, Twitter doesn’t include in the data set, we have to do what’s called feature engineering, you have to take existing data and turn it into other kinds of data.
So looking at the data, what things could we engineer that might help us understand? Yeah, actually, you know, that might be a factor that that could boost, you know, clicks versus clicks.
Katie Robbert 7:52
You know, one of the I think one of the, quote unquote, easier things to feature engineer.
And again, feature engineering is not something that should be scary and unapproachable.
It really just means, you know, taking your existing stuff and sort of picking it apart to smaller piles of information, at least in this context.
So I would probably start with some of the date and time data, you know, and break it down a little bit more granular, you know, because I can look at the dates, but then I also then have to look at a calendar and say, what day of the week was that? Does that make a difference? Or what time of day was that? Does that make a difference? And if that all of that data is contained in the strain, then I would want to pull that out into individual pieces of data.
Christopher Penn 8:34
All the things that’d be pretty straightforward.
How many handles haven’t mentioned like here, I’ve been at mentioning ringcentral.
I’ve mentioned our friend Jeremy Ouyang, has a chat, that Twitter chat that was happening.
Can I engineer out the number of hashtags used in the tweet Absolutely.
Just do raw counts how many hashtags were used is and that’s something that we might be able to say, does the number of hashtags used in a tweet helpfully predict the likelihood somebody clicks on on that tweet? If you look at the syntax, any tweet that begins with an ad symbol, is a reply.
So we could code a tweet as to whether it’s a reply or not, and be able to differentiate that.
And I think critically for a lot of social media managers who are struggling with their content strategy.
The big question is, what content are you sharing? What content should you be sharing? So if we look at this list here, and this is just one month of tweets, I tweet a lot.
There’s a lot of data to go to.
And you can see there’s different topics you know, we got, you know, top three podcasts for marketing majors AI algorithm using the heart rate, and motion to predict agent and gender promotional tweets for for Trust Insights on 2021 marketing plans.
There’s a lot of information in here, one of the things we want to do is we want to boil this down from text into into numbers.
So one of the steps that we would take to do this is to basically take all of the, the copy of all the tweets, and get a list of the number of words used, right? Which individual words have been used in those tweets.
So let’s show you an example of how that would look.
Let’s go here.
So let’s go ahead and take our tweets, we’re going to bring them all into a big data frame.
Now this is in the programming language art, you don’t need to use this, I use it because it’s convenient.
It’s it’s easier for me to process the data.
You could do this in Excel, and then with some third party, things on the web, but we’re gonna go here and take all those tweets, I’m just gonna make a frequency list.
And this frequency list is just the number the number of times any given word appears, right? So I use marketing 161 times in this pile of tweets, use Word data, 112 times insights, and obviously, the company name Trust Insights, isn’t there? analytics shows up 45 times, we see some of my us guy answer stuff, we have business.
So we have data science showing up podcasting.
And so this gives me a good starting point to try and figure out how many different topics are there that I’m tweeting about? And so what I’ve done is I’ve gone ahead and said, Okay, I want to count how many times I mentioned social media and all these different words in social media like Twitter, Facebook, Tiktok, Instagram, etc.
How many times they’ve mentioned SEO stuff, how many times they’ve mentioned data, data science, Trust, Insights, podcast, Google Analytics, and then engineering? How many? Was there a URL on their ad handle? Was it a reply? How many hashtags? And so this process is the is the data preparation process to go through and understand? What are the things that I might want to predict on one of the things that I might want to run an analysis on and just even doing something as simple as getting raw counts? Like, okay, I, you know, tweet about marketing way more than data.
If I did that, and did those counts, I could say, huh, that would have meant to do like, did I actually mean to be tweeting so much about, you know, marketing technology, when I’d like to be known for data science? Are there different things?
Katie Robbert 12:34
Let me ask you this question, Chris.
And maybe it’s just a matter of depending on what your goal is, but so you’re looking at ROC counts based on frequency of words used in the text of those tweets? Would you recommend that someone look at, you know, a different metric first, like number of engagements or number of retweets, pull that information, and then look at what the context of those tweets were.
Christopher Penn 13:04
So that goes back to the original data science lifecycle, the Data Prep lifecycle, where we have to know what the audit strategy is, what is what method Are we going to be using to judge this outcome? In this case, we’ve already selected URL clicks as our response variable.
And we know we’re going to be doing a regression analysis to say, Okay, what things how the most the strongest mathematical relationship to this outcome we care about? If we didn’t have that outcome, then yeah, you could actually do a whole bunch of mixing and matching and explore.
What are some other possibilities? It might turn out, for example, that in my if you were to do it, either with Oh, you could do with regression analysis.
engagements might matter a whole lot to URL clicks? We don’t know yet.
We haven’t we haven’t figured that out yet.
Because we haven’t run that part of the analysis.
We’re just trying to prepare the data so that you can can be used in analysis.
And that’s, that’s, sometimes it feels like it’s putting the cart before the horse.
But it’s one of those things, it’s very iterative.
Katie Robbert 14:10
Well, and I think that, you know, it goes back to, to your point of what we’re talking about last week with that lifecycle is the one of the first things Chris, you pointed out on this live stream was that I asked a bunch of questions that the data can’t answer because that data does not exist in that data set.
And so those are the types of things that you would want to sort of write down and figure out what data do I need in order to answer those questions? Because if you know, I say, Chris, I want to know, the number of followers month over month that we grew, and you hand me, you know, the export from that Twitter, I’d say I can’t answer that question.
And there’s, you know, that’s then just sort of wasted time and wasted resources.
You know, there might be a different kind of data to look at or different way to capture that information.
So it really does, you know, Go back to all of the planning upfront, which might feel daunting, is going to save you a lot of time, especially when you start getting into expensive resources who know how to use programming language languages, such as R.
That’s not an inexpensive ask.
Christopher Penn 15:15
It’s definitely not and the techniques that you need to do the analysis, in addition to the computational costs, you have the expertise costs, because in order to run this particular model, the way that I’ve set it up, you have to know, just a quick scroll down here shows you just how much more goes into this, before you actually get to the answer.
This is not a small amount of coding and programming to pull this off.
So there is absolutely a very large cost to not doing the prep work, and then building a model and then finding out the model sucks, and then having to redo it over and over and over again.
Katie Robbert 15:53
So john, you, you know, you don’t actually use Twitter a lot.
I know because I have access to your Twitter accounts.
Because as a CEO, that’s what I get to do is I get to see everything.
But if you were you know, just hypothetically for yourself, if you were to start tweeting more, or setting up some sort of a strategy, what kinds of things would you particular, want to be known for, aside from, you know, your partnership with Trust Insights?
John Wall 16:22
Yeah, I think, you know, the biggest we always joke about the cobblers kids have no shoes, you know, the the gap that I’m missing is there’s all this marketing over coffee content, which could just be recycled every week, you know, there’s no reason the featured episode from two years ago, just couldn’t get plugged in, fill that all up.
So yeah, you know, one of these days, I’ll get around to actually doing at least the baseline adequate marketing.
And then after that, I, you know, I don’t know the channel doesn’t seem to fit that well, because you know, LinkedIn and Facebook are a little bit better, because they give better previews of podcast or of links that you’re sharing.
So, you know, the content, it’s just a kind of falls to the bottom of the list.
The other thing, though, and this, I’d be interested in what you guys think about this, is that, you know, is it worth going to the effort of really doing a drive to get more followers to try and get in front of more people? Because that’s always been a question.
You know, if I were to actually do a full on pressing campaign to see how much action I could just get happening in that channel, and it would feed upon itself, you know, as you get more people you start to publish more times per day? Is that a kind of is would it be worthwhile to build that kind of flywheel?
Christopher Penn 17:35
Um, it depends.
It depends on a whole bunch of things.
We know, though, that social networks in particular, really do focus on engagement, their algorithms are tuned towards getting eyeballs to be sticky.
And so one of the challenges and this this is something that has come back to bite a lot of brands hard, is they spent enormous sums of money building huge followings.
And then those followings were not engaged.
And so it actually created their, their Facebook pages, Twitter very similarly, because you can’t, you can’t have 10,000 followers have one person engaged, you look terrible, versus having 10 followers, and one of them engaged, you know, 10% engagement rate.
And, you know, again, we don’t know the back end scoring mechanism.
And but we can see through empirical evidence that the less engagement you have, the less well you do.
So just on a campaign based campaign basis, I would say no, you probably shouldn’t be buying.
John Wall 18:34
And it was funny, because I noticed on Twitter, there was a tweet from Twitter that was being retweeted like crazy.
And it just said, four likes is a lot.
And when you think about it among billions of tweets, yeah, actually, if you get four likes like that you’re doing well.
Katie Robbert 18:52
That’s a really good point.
And I think, you know, that engagement piece, and that’s something that we’ll look into with this data is important, because it’s not just about people liking it.
But if someone asks a question, and it goes ignored, you know, are you do you have the time and availability, to then be responding to people on Twitter to keep that conversation going on one single tweet or one thread? And so engagement is more than just someone clicking the like button, like, you have to do work as well.
Christopher Penn 19:22
So we’re at a point now where we’ve prepared the data, I think, probably as well as as it’s going to get.
Because we’re now at a point where we don’t know what we don’t know, right? We know we’ve got this data we’ve engineered, I think, a decent number of features out of it.
But we’re kind of stuck now as to what should we be doing next.
So in terms of what to do next, if we go back to the lifecycle and try to pull this up here probably should have had this up and running up in the first place.
Let’s see where is my data science lifecycle here.
We’ve collected our data.
We’ve taken a look at it to see like, what’s in here, what dimensions of it what metrics, what data types, right.
And we’ve done some very initial analysis.
So now we’re kind of at the requirements gathering stage.
And we’ve actually had to do some feature engineering upfront, because we talked about it early on.
But we know that in the initial data, the some of the required features are, we’re simply not there.
So the next thing to do in the process is just do a quick quality check.
Like how much data is the data in good condition or not? So let’s go ahead and run this.
Katie Robbert 20:42
Well, and I think, a good PSA, if you’ve only tweeted three times in the past month, and you’re looking at trying to do this kind of analysis, you probably don’t have enough data you have to have, you know, there’s no magic number.
But if you’ve only tweeted three times, for example, like that’s really not going to tell you a whole lot.
John Wall 21:04
There’s no magic number, but it’s way bigger than four.
Katie Robbert 21:06
It’s bigger than that.
It’s not magic.
Here I am, I’ll be here all day.
You know, sample sizes.
Christopher Penn 21:16
Katie Robbert 21:19
This is why you hang out with us, Chris.
Christopher Penn 21:21
So this is a type of exploratory data analysis is just very simple univariate analysis to look at what’s in the box, right? So we’re taking our our data and ask him what’s in the box.
There are some tweets here, for example, that have occurred more than few times they’re on repeat, right? So we would want to know, how much repetition is there in the data set? There’s actually a fair amount here.
We have our our timestamps, we have our impressions, how many impressions, so we have a minimum number of four, we have a median of 1200, and a maximum of 14,000.
Right? pretty wide range in terms of impressions.
JOHN, to your question about like, you know, should you be focusing on building awareness, you might want to first just take a look at your data that exists and say, What is the range of reasonable impressions, on on the tweets that you’re sharing? Same thing engagements, right? minimum zero, no surprise there, the median, six, and a maximum of 343.
Now, bear in mind, this is on a Twitter account with 95,000 followers.
343 engagements is the maximum kinda sucks, right? Definitely goes show that just because you have a lot of followers doesn’t necessary mean it’s a good thing.
engagement rate, the median zero.
Katie Robbert 22:35
It makes all of us feel better, Chris?
Christopher Penn 22:37
retweets, replies and thinks about the likes.
And then we start getting into the stuff that we started engineering.
So looking at how many how often is it you know, talk about topic marking, most of these topics have zeros, right.
So they’re not frequent.
So there’s, there’s not a lot in here.
So that would make me start to question the predictive power of some of these topics, if there’s not enough to even be, you know, the median topic, then a, my Twitter feed is probably not as focused as it could be, would be, it might not be a lot of predictive power.
When we run an actual algorithm, we look at the number of times you use handles.
One treatment apparently had 19, handles stuffing them, that must have been a fun tweet, to read
Katie Robbert 23:24
through those young tense,
Christopher Penn 23:28
up to five hash tags, and so on, and so forth.
So even just this very simple univariate analysis, just looking across the board at the data set, tells us a lot about what’s in what’s in the data.
The things that are not in here that I’m glad to see aren’t in here, are there’s no missing data.
So looking at this, there’s no file, you know, big piles of anaise, there’s a if you look in this column here, valid, you got all 100 percents, which is looking good, makes me happy.
If you had, like, you know, 20%, missing, 30% missing, you probably can’t go on a whole lot, right? There’s you’re gonna get stuck, because you’ve got big chunks, missing data.
And at that point, you have to say, do we need to supplement the data? Do we need to do imputation to try and repair it? Do we exclude the missing data? What do we do with it? But in this case, because it’s a clean data source, it came out really nicely, well ordered.
Katie Robbert 24:27
Question for you, Chris.
Would so I often see a lot of people tweeting in different languages, they have keyboards that handle special characters that the standard, you know, English based keyboard doesn’t have would that throw an error into your, you know, the script that you just ran? Or are those things that you would have to account for ahead of time, if, let’s say you were fluent in a different language, and so you were constantly tweeting in both kinds of languages.
Christopher Penn 24:59
So that’s it.
Actually a data engineering question.
And the short answer, at least in the code that I’m writing is that everything gets transliterated down to the ASCII character set.
So you would want to change character sets for your tweets, if you use UTF, eight UTF 16, something that accommodate that, and then the code would then function exactly the same.
But you’d need to know that going into it.
So if you were doing this on, say, a customer service data set, where there’s the possibility, you don’t know what languages are in there, then you’d want to probably use something like UTF, eight UTF 16, to make it as maximally permissible as possible for all the different languages, one of which UTF eight is UTF.
Eight mp4 is one of the format’s that is an extended multibyte character set that accommodates emoji.
If you use standard UTF.
Eight emoji don’t work, they come in as these weird strings of characters, you have to use that other characters that.
And again, knowing you’re looking at social media data, that might be something you want to plug in here.
Now we’re looking at my tweets, I don’t use emoji and tweets, so we’re okay there.
Katie Robbert 26:02
But if I was looking at mentioning you, they might
Christopher Penn 26:05
exactly if I was looking at social media monitoring data from like Talkwalker, for example, then like, ah, I need to use a different characters that again, that’s an engineering question.
Katie Robbert 26:14
Well, it’s an engineering question.
But it also goes back to prepping your data for analysis, things that you should probably be aware of, if you’re looking at social media data, there is likely to be an emoji or two thrown in there.
Christopher Penn 26:29
So we’ve asked questions of the data set, we’ve obviously dug around looking at the raw data itself and seeing what what comes in the box.
And we even started doing some of the preparations for analysis, engineering, new fields, doing quality checks and things like that.
Katie, what’s coming up next week, in this topic?
Katie Robbert 26:51
Next week, we are actually, you know, looking at, is the data ready for more heavy duty machine learning? What do we need to do to get the data to that predictive space to, you know, do more of that sentiment analysis, for example, but you can do some of that out of the box.
But it’s not as good as if you’re doing, you know, more heavy duty machine learning? You know, Chris, you had mentioned predictive power, those kinds of things, you know, what is it that you need to do to get your data ready for those things?
Christopher Penn 27:30
One of the things that you have to do to know when to I don’t know if I’m phrasing that correctly or not, is you got to be able to at least run some quick tests on the data to see if what you what you’ve got to work with is is even useful.
And this is part of that requirements verification stage.
But it’s also part of modeling and insight.
So I guess I was working on this data for a blog post.
And I’ve got two different results here.
of basically, how good is the data at answering the question I wanted to ask answer.
And there’s two different fields here.
On this one is a one set of parameters for testing, this is another set.
And looking at the root mean squared error, which a lower number is better, it means that the data is less noisy.
And then the R squared error range, which again, a closer to one is better, because it means that the data fits better.
So far, just looking at these initial results from an initial test model run, they kind of suck AI root mean squared error is good.
That’s it’s a pretty low number.
But the R squared error is is way below one, which means that the data doesn’t really answer the question, right? I asked it What explains URL clicks, and of all the stuff that we’ve put in so far, none of it’s a particularly good fit.
Now, there’s two things we could do at this point.
Number one, is we have to question and this goes back to our goal and strategy, are we using the right algorithm? We may not be? And two, if this isn’t the case, if we are pretty confident in the algorithm, then we have to ask ourselves, are we missing some data? Is there some data that is not in here, like for example, you would point out earlier followers change, or a number of followers that would assist the model in in building a prediction about this right now, these numbers are so low, I would not put this in front of a client, I would say, you know, this, this needs some more work in the lab, because neither of these results, I would would to me would be acceptable enough to say yes, bet your strategy on these results.
John Wall 29:41
Well, the the R squared, if that gets above point five would you go with that are you needed to be all the way up to one
Christopher Penn 29:48
I would be looking for above point five.
More than anything, though.
Because one of the things, I can show you what the actual files look like there’s all these different variations.
So I’d run probably Like 35, or 40 different variations of all the ways to test the data, and then look at the cohort of them and find, okay, do any of these parameters, you know, show up as a solid test.
As you can see, I’ve done six unit tests.
So far, none of them are reaching the finish line yet.
Katie Robbert 30:22
And that’s all with the same dataset,
Christopher Penn 30:25
all the same data set, it’s at this point, it’s just it’s called called what’s called hyper parameter optimization, it’s like, easiest way to explain is you’ve got the same ingredients.
Now you’re tweaking all the dials on the oven to see like, you know, different temperature, how long you cook it for, and stuff like that, to try and make a good cook batch of cookies.
So far, I’ve made six batches of bricks.
Katie Robbert 30:44
Well, and this, this goes back to sort of the point of this particular episode is what questions that are, should you be asking when you have the data? And so the first question I’m going to ask you, Chris, is, are you looking at the right data? To answer the question, it sounds like you’re not?
Christopher Penn 31:00
Yeah, something’s missing, something’s missing.
Or it may turn out that URL clicks is just the wrong thing to be optimizing on now that from a logic and common sense perspective, there’s really only a couple things that you could have in Twitter data that you could optimize on, one of which would be, you know, the engagements, the impressions, or the clicks to essentially sort of the actions that people take to validate that, you’d actually have to set up a different experiment.
So you want to take the the Twitter data to your website, and then the different numbers inside Twitter, and do a regression analysis on that to say, Okay, what number most closely record looks like Twitter traffic, and then use that as a benchmark.
Knowing how I tweet and knowing what I’m doing a PI data, I’m fairly confident that, in fact, URL fix is the correct objective for this particular test.
But it could be something we could test and see if there’s a different possibility.
What I think is, the bigger problem is that most of my tweets that have URLs, go to the Trust Insights website, and stuff.
And so you know, that test would fail, essentially, because it’d be benchmarking off my personal website, when you should be using the Trust Insights website as the as the regression analysis.
Katie Robbert 32:14
Okay, so let me see if there’s a way to simplify this a little bit, because we’ve talked a lot about regression analysis, we’ve talked a lot about cleaning the data feature engineering, bring it into a system like our What if someone doesn’t have access to those things? Or doesn’t have that skill set? What are some of the things that they could do with? You know, obviously, anybody can export that data from analytics.twitter.com? What could they do with just that spreadsheet to try to answer some of these questions? Of course, you know, going into the analysis, knowing this is what I have available to me, we covered in last week’s episode, what tools do you have? And if the only tool you have is Excel? What can they do with that?
Christopher Penn 33:00
Some of the basic feature engineering, you can still you could do this inside of Excel, no.
So you can see here, I’m looking for specific words, I go back into our Excel spreadsheet and just add a column here that has a much bigger column I expected.
And then I could create, for example, let’s go to the top.
Let’s call this marketing.
I can do it equals count, if in this cell,
it’s going to give me a count of the number of times marketing appears in that cell.
So I could repeat this column by column for all these different topics and engineer the features in here.
And then I could do using excels built in correlation function, I could do it and sort of accurate version of that the multiple regression model in Excel.
I don’t know if I if you can do multivariate as easily as you can in ROI, but you could at least identify, you know, column by column.
Okay, yeah, this there’s at least something so you could do 60% of this in Excel.
I think it is way easier to use either R or platform like Watson Studio.
If you or SPSS or alteryx.
Any of the bigger, more expensive BI tools.
If you’ve got access to those even like the free public versions of them, you could certainly do that type of analysis like Watson Studio, you get 50 hours free a month.
So you don’t have to buy anything from IBM in order to be able to try it out on a data set like this.
Katie Robbert 34:39
Which I think is that’s really good news for people who are short on budget and resources.
But I think that it’s important.
You know, Chris, as you pointed out, you’ve run six different tests and the data set isn’t necessarily answering the question.
You know, so it sounds like you still need to do all of that prep upfront.
Before For you bring it to the machine learning models, because you could be wasting a lot of those free hours that you have on data.
That’s not going to answer the question.
Christopher Penn 35:10
So why should sign up for multiple accounts?
Katie Robbert 35:14
Well, let’s not talk about gaming the system, let’s stay focused.
Christopher Penn 35:21
Doing doing requirements gathering during that initial analysis, doing feature engineering, trying to come up with as many options as possible.
On the first run through is important, right, though, the more that you can do that, the easier it’s going to be to actually run those models, because you still have to do all the tuning and stuff in the model itself, you want to make sure the data going into the model is as good as it’s going to get.
You know, for example, you could look for the syntax around like, you know, is this a Twitter chat I was participating in yes or no, and then code that as well, maybe Twitter chats work, or they don’t.
But it’s important to be able to dig into those questions upfront, as much as you can and create as much of that engineer data as possible.
It’s not easy, but it is something that is, is important skill.
Katie Robbert 36:14
Well, and you know, to your point, if you’re being asked, what should we be doing with our social media this year, and the company is looking to hand over a lot of money to invest in social media, you want to make sure that you can answer that question with some data behind it.
Because come the end of the year, when nothing’s changed, nothing’s happened.
You know, you’re gonna be sLl.
Christopher Penn 36:36
So you have to do the analysis upfront.
John Wall 36:39
We do a lot of work on LinkedIn.
Christopher Penn 36:49
So coming up next week, we’re gonna be doing the more advanced feature engineering, if appropriate, and then taking a look at what it looks like to take out some of the features.
So we want to talk about cleaning and taking out stuff that doesn’t matter.
Because we didn’t cover that in today’s episode.
And there’s a lot of junk in here too, that needs to come out.
So that’s, that’ll be in next week’s episode.
Katie Robbert 37:11
Yep, that’ll be part three.
And then Part Four, we get to the meat of it.
The so what we’ve done all this work.
So what what do we do with it? What do we do with all this data? Did we answer the question? How do we put it all together? So stay tuned for those, you know, the following two Thursdays 1pm.
Eastern, same bat time, same bat channel.
Christopher Penn 37:32
We’ll talk to you next time.
Thanks for watching today.
Be sure to subscribe to our show wherever you’re watching it.
For more resources.
And to learn more.
Check out the Trust Insights podcast at Trust insights.ai slash t AI podcast, and a weekly email newsletter at Trust insights.ai slash newsletter.
got questions about what you saw on today’s episode.
Join our free analytics four markers slack group at Trust insights.ai slash analytics pro marketers.
See you next time.
Need help with your marketing data and analytics?
You might also enjoy:
Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, Data in the Headlights. Subscribe now for free; new issues every Wednesday!
Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new 10-minute or less episodes every week.