The Vixen Blog Awards, which took place a few of months ago, is pretty much the highlight of the year for most Norwegians. Or so you would think, anyway.
Lately, it seems not a day goes by without one or several of our treasured fitness, lifestyle, fashion or furniture bloggers being front and center in some newspaper or magazine with hot off the press social commentary.
Common topics include how to oppose the use of palm oil while still using palm oil, making food with zero calories, “janteloven”, flashing of various body parts, and of course how to not care about the looks of said body parts because you’re pretty just the way you are.
The annual Vixen Blog Awards is an opportunity to reap the rewards from these endeavours from the last twelve months. It’s basically like the Oscars, but for blogs, and without the Leonardo DiCaprio-memes.
The truth is I actually think blogs are pretty cool (especially if they’re about food), and that blogging is an extremely interesting medium. And I do have my own blog, after all, so who am I to make fun?
I’m just a little jealous because I’ll never qualify for the Vixen Blog Awards myself. My chances of that pretty much went down the toilet when I stopped blogging about nutrition and exercise several years back.
Still, what if I wanted to get back into that kind of blogging? I can’t help but wonder which blogger I have the most in common with now and which blog award I’ve got the biggest chance of winning next year if I stop writing about tech stuff and start writing about everything else.
What, don’t tell me you’ve never thought about which “rosablogger” you are, or how many likes, shares and <3’s you would get if you started writing about about mundane, everyday stuff on the internet. It’s right up there with the meaning of life!
Unfortunately, machine learning cannot tell us the meaning of life. Well, not yet anyway. It can, however, tell you which blogger you have the most in common with based on what you write, and by extension which blog award you’re likely to win next year if you were to start blogging today.
That’s right – it’s time to use our powers for evil again, and this time it’s multi-class classification on eleven different classes.
Dude, I don’t care about any of this rocket science stuff, I just want to put all my friend’s Facebook posts into a web service and make fun of them.
Go to the Vixen Blog Awards, Machine Learning Edition web service, enter any NORWEGIAN block of text of any size, and press the submit button. You might need to give the page 10-15 seconds to work, especially if the VM running the web service needs to boot up.
After this short wait, what you will get is a set of probabilities, each one representing the probability of your text being something a particular blogger would write, and by extension which blog award this piece of text is worthy of. Neat, huh?
Things you can try out:
- Entire blog posts, paragraphs or excerpts
- News articles
- Facebook posts and tweets
- Government reports
- Legal texts
- Bible passages
- Random quotes
I put in a few of my recent Facebook posts and this is what I got:
I guess I’m still a fitness blogger whether I like it or not …
Like I said earlier, the model is for Norwegian text. It is not built for any other language, so don’t expect reliable or meaningful results on text written in English, Spanish or anything else that is not Norwegian.
My good sir, I care neither about any of this rocket science stuff or any of this stupid stuff. I want to know how to know how my business can make or save money on this stuff!
Fine, Uncle Scrooge. Scroll all the way to the very bottom of the post and look for the headline “An actual business case.”
If you want to know what goes on behind the scenes and how you might do something like this yourself, read on. In the end I’ll also teach you how to manipulate the model and bend the results to your will so you can troll your friends!
So how does this work, and how do I create a Blog Classifier in Azure Machine Learning?
You need three things to create your own blog classifier the same way I did it:
If you want to operationalize your service in the cloud for others to access and use, you also need a Microsoft Azure subscription – either trial, MSDN or paid.
Obviously, you don’t have to do any of this with blogs. This is a generalized approach that you can use for any kind of text mining from web sources, be it tabloids, news magazines or scientific studies.
You don’t have to use the data for classification either. How about predicting the number of clicks, likes, shares or comments that any block of text is likely to get, for example?
1. Gather training data
If you want to scrape data from the web or crawl sites for the information you need, import.io has got you covered.
I used this awesome service to create API’s on top of the 11 winning blogs and extract more or less all their blog posts from 2015 in tabular format.
Basically, this means that I turn the web pages containing the blog posts into tables with columns containing the post title, post content, post timestamp, comments, likes etc. This is exactly what you want the data to look like for what we’re going to do.
I saved the tables as CSV’s – one file for each blog – and created connections to all the 11 files in Excel using Power Query.
Oh, and did I mention that import.io rocks? (They’re really not paying me for this – I swear!)
2. Prepare training and test data
Using Power Query, I chose the columns I wanted from the CSV’s – most importantly the actual blog post contents, as well as the post titles and dates for traceability purposes. I then labeled the posts with the correct author based on the name of the source file (one file for each blogger, remember?) and appended them together into one giant query.
Finally, I split up the post contents column into multiple columns using periods as a delimiter, unpivoted the resulting columns to make a nice mix of sentences and paragraphs to use as training and testing samples for the model, dumped the results into Excel and saved them as a UTF-8 TSV (to keep those pesky Norwegian special characters).
Using smaller blocks of text as the unit if analysis makes the classification problem a lot more difficult. However, it’s more appropriate, given that the algorithm will most likely be analyzing sentences and paragraphs rather than whole posts when it’s operationalized.
If I’d made the model with whole posts, I’m quite certain that it would’ve been freakishly accurate from the get-go. And where’s the fun in that?
3. Text preprocessing in Azure ML
I uploaded the TSV into Azure ML Studio to start creating the training experiment.
Remember that it’s very important to do all the text preprocessing in Azure ML for an experiment such as this if you’re going to operationalize the model online or in your business.
If you use Power Query, R or other tools locally for anything other than combining data sources and tweaking your source material to reflect the unit of analysis that your operationalized model will be working with in its “natural” environment, your model is not going to be very reliable.
Simply put: the “real” data, meaning the blocks of text your model will be analyzing in “real life,” needs to go through the exact same preprocessing that your training and testing data does. That means it should be done in Azure ML, unless you’d rather automate the preprocessing using another tool or service.
First, you’ll probably have to replace any specific foreign characters with their internationally friendly alternatives. Because even though you can get these characters into Azure ML Studio using an UTF-8 encoded file, you lose them when you try to put them into an R data frame running on a computer with English settings – which is apparently the case with the computers who run R for Azure ML.
Then comes standard text preprocessing like the removal of special characters (like punctuation, parentheses, brackets, hyphens etc.), conversion to lower case, removal of stop words and word stemming.
Azure ML has example experiments featuring text preprocessing, as well as a readymade text preprocessing script for R which can easily be “hacked” to work with any language even with minimal knowledge of R. I’ll leave this challenge to you for now, but I might do a short writeup later on how to do this.
4. Feature extraction
Now comes the fun part – it’s time to turn the text into features for training our models. There are several ways both to look at text for this purpose and to actually turn it into features.
A simple and superficial way to look at text is as sequences of words. A sequence might consist of just one word – a unigram – or any other number of words, where a sequence of two words is a bigram, a sequence of three words is a trigram, a sequence of four words is a quadgram and so on.
We won’t be using anything longer than trigrams for this case, which I understand is pretty common for many problems. This might be because of scaling issues, but one might also intuitively assume that by going up to trigrams, you ensure that each word can be associated with both its preceeding and subsequent word in a feature, and that trigrams therefore represent a natural threshold.
The easiest way to make features from these so-called n-grams is to turn them into hash keys of limited size, and then count their occurences in the text. By specifiying the hasing bitsize, you control how much information is retained and by extension how many columns you end up with. This process is called feature hashing, or simply the “hashing trick”.
The resulting columns from feature hashing are then usually reduced into principal components, but they can also be used directly in the models.
The feature hashing approach has the advantage of being very fast (the PCA might take some time, though) while allowing us to keep a limited amount of information from word doubles and triples. It also lets the model work with previously unseen n-grams, and allows you to quickly establish a great benchmark for any text classification case early on.
The disadvantage is that we lose traceability and we’ll have a very hard time figuring out exactly which aspect of the text our models are using to make their predictions. We’ll also lose some information that might help us distinguish the least represented bloggers and fine-tune the model later on. You might choose to look at feature hashing + PCA as the quick and dirty way to do text classification.
Another way of creating features is to create a dictionary of the n-grams in the text, calculate their relative frequencies and use the unigrams, bigrams and trigrams as features directly in the models.
The dictionary approach has the advantage of being very traceable, and it provides a high level of detail. However, it does not scale well beyond single words and might produce hundreds of thousands of columns when you use it to create bigrams or trigrams – not to mention quadgrams – depending on the amount of text.
While some machine learning models can handle such a large amount of features, it’s hardly a very cost- and time-efficient way to do things.
My preferred, general approach to get the best of both worlds is the following:
- Generate as many hashing features as is feasible to run a PCA on, and reduce them to 10-20 principal components (look for when the standard deviation starts stabilizing).This allows me to keep a reasonable amount of information and variance from unigrams, bigrams and trigrams, as well as enabling the models to work with unseen n-grams and giving them some robust, dense features to use for for training.It also lets me quickly establish a benchmark and compare algorithms on an even playing field early on, since training and testing some of the more complex models with a very sparse feature vector containing thousands of features takes a lot of time.
- Supplement these features with a deliberate selection of n-gram frequencies.This gives me some traceability and allows me to do some more meaningful exploratory analysis if I want to. This is the part that enables me to tell you that if the model tells you that you’re Funkygine – like I am, apparently – it might be because you’re writing about amino acids, “starting positions” or Amsterdam. To do this with feature hashing + PCA, you would need to do some heavy simulations.
In Azure ML, this whole shebang looks like this:
Yeah, I love creating sexy workflows in the ML studio. Sue me.
Note that I actually preprocess the text slightly differently for creating the bigram and trigram dictionary. Specifically, this has to do with the treatment of stopwords, which I remove before creating bigram and trigram dictionaries.
Why? Well, I’m definitely interested in analyzing the frenquency of single stopwords for the purpose of telling the bloggers apart. However, I’m less interested in analyzing those stopwords in combination with their immediately surrounding words – simply because this causes the bigram and trigram dictionaries to explode in size.
On another note: the pre-packaged R script bundle for text preprocessing in the studio is not configured to let you extract anything more than unigrams. I had to “hack” it to enable the extraction of bigrams and trigrams.
I might cover this in a later post, as it was a little bit more cumbersome to do than changing the language.
5. Feature selection
Like I mentioned earlier, the dictionary approach does not scale very well when it comes to the actual modeling, so I wanted to do some deliberate feature selection before releasing the hounds (the algorithms, obviously).
I dug up some research on selection of features for text classification, and ended up with the following approach:
1. Filter out the rarest terms during the dictionary creation process
I set a frequency threshold of 5 for unigrams, 6 for bigrams and 3 for trigrams. If a term does not appear at least that many times in the whole dataset, it is dropped. I chose these specific numbers because they gave me a manageable total dictionary of ~24 000 features to continue selecting from.
One drawback to this approach is that it is a bit biased in favor of the dominant classes in the dataset, meaning the bloggers who have produced the most text in 2015.
As seen above, my labels are quite imbalanced, and we’re more likely to lose some information relevant to the minority classes if we remove features based on unadjusted frequencies. The next step does not really reverse or mitigate this information loss in any way, but it does not fuel the bias in favor of the dominant classes and puts equal emphasis on all of them.
2. Do seperate feature selections for each class
For a rather complex and quite imbalanced multi-class problem such as this, a simultaneous feature selection might ignore the terms that can be used to distinguish the least represented classes from the others.
Let’s say that Jenny Skavlan quite often writes “parrots eat seeds” in her blog posts. None of the other bloggers care much about parrots eating seeds, so writing something like this is pretty distinctive of Jenny Skavlan.
In practice, this means that a machine learning model can make good use of a feature counting the frequency of the trigram “parrots eat seeds” to classify any piece of text as being either Jenny Skavlan or not being Jenny Skavlan.
But the same feature is pretty useless for distinguishing Sophie Elise from Caroline Berg Eriksen, for example, because neither of them write anything about parrots eating seeds at all. This wouldn’t necessarily be a problem, had it not been for the fact that Sophie Elise and Caroline Berg Eriksen – together with all the other bloggers – make up a much bigger part of the dataset than Jenny Skavlan does.
So since Jenny Skavlan only has a pretty small amount of the text in the data set, and the term in question is only really useful for distinguishing this small amount text from a much larger amount of text, traditional feature selection on the entire dataset at once is not likely to place much importance on the “parrots eat seeds” feature.
The way to deal with this is to emulate the situation I described earlier, in which the feature “parrots eat seeds” is used to classify a piece of text as being either Jenny Skavlan or not Jenny Skavlan. This is a very different situation than using it to classify a piece of text as belonging to any of the 11 bloggers.
In practice, what I did was create 11 complete versions of the dataset, one for each blogger. I then replaced the labels to either represent that particular blogger (“true”) or any of the other bloggers (“false”), making it a binary classification problem for the purposes of feature selection.
I then used the chi-squared test, which is the generally preferred feature selection method for text classification, to pick the same number of features from each of the 11 datasets. I finally merged all the feature sets together, dropping duplicate features.
The result is a limited set of all my initial features selected on a class-by-class basis, balanced for distinguishing all of the 11 classes from each other and not just for the dominant ones.
What the hell does that look like in Azure ML, you say? I’m so glad you asked!
Of course, this could’ve been solved much more elegantly with a single R script module – probably with only a few lines of code – but where’s the fun in that?
When it comes to the number of features to keep for each category, there’s really no right or wrong answer. You should experiment with different amounts. It’s most likely going to be a decision based on some of the following factors:
- How much time you have
- How general you want your model to be
- Which models you’re going to try
Another benefit of this type of feature selection is that it allows me to easily discover predictive features that shouldn’t be in the model, even for the least represented classes.
6. Model training and evaluation
Alright, now comes the REALLY fun part. It’s time to release the hounds.
For feature-rich text classification problems such as this where you probably want to take advantage of the large feature space, logistic regression (LR) and support vector machines (SVMs) might be the best way to go. Neural networks are also commonly used for text analysis, but to take full advantage of their capabilities in this case you’ll probably want to do feature preparation differently and use a specialized network architecture. Maybe I’ll try this next time!
Still, you shouldn’t make these decisions before you’ve tried. Therefore, I started out with a smaller number of features to test out boosted decision trees, decision forests and neural networks, together with logistic regression and SVMs.
Since I’ve spent so much energy on an unbiased feature selection, I’m going to use macro-averaged precision and recall as my chosen performance metrics. This should allow me to measure the benefits of my deliberate class-by-class feature selection directly.
If I didn’t place equal value on classifying each of the bloggers and cared more about getting the best represented classes right, I’d use micro-averaged performance metrics instead. The choice of performance metrics is always a “business” or subject matter decision, rather than a statistical or technical one.
For computing the evaluation metrics, I use 10-fold cross-validation.
Let’s get modeling!
~1 000 features
I got pretty decent results from most of the models even when using only 1 000 features “carefully” chosen on a class-by-class basis using the approach described in the last section. The results seemed to be a lot better with this method than they were with simultaneously selected features.
Below is the confusion matrix and macro-averaged evaluation metrics for Microsoft’s legendary boosted decision tree (BDT) model with features selected simultaneously for all classes.
Macro-averaged precision: 0.4788 Macro-averaged recall: 0.4271
Compare that to the results from the class-by-class feature selection.
Macro-averaged precision: 0.5250 (+0.0462 from feature selection) Macro-averaged recall: 0.4631 (+0.0360 from feature selection)
I’d say that’s a pretty major difference for using exactly the same amount of features!
It should be noted that I only saw this improvement on models who produced decent results in the first place. Decision forests and jungles generally performed poorly for this problem, and feature selection didn’t seem to make much of a difference for these models.
Still, the results lead me to believe that this form of careful and deliberate feature selection can in fact be expected to produce more balanced results with a smaller amount of features for imbalanced text classification problems using the appropriate models. Not necessarily surprising, but still interesting to see that it actually works very well in practice.
~3 000 features
As I gradually increased the number of features, logistic regression started pulling ahead slightly while the others couldn’t quite keep up. Also, the models based on class-by-class feature selection continued to outperform the one based on simultaneous feature selection.
Below is a BDT with three times as many features as above.
Macro-averaged precision: 0.5780 (+0.0530 from more features) Macro-averaged recall: 0.5103 (+0.0472 from more features)
Compare that to the results from logistic regression below.
Macro-averaged precision: 0.6425 (+0.0645 compared to BDT) Macro-averaged recall: 0.5251 (+0.0148 compared to BDT)
Now we’re talking!
Neural networks also continuted to do fairly well, but they also start becoming slow to train when the number of features increases into the thousands.
None of the other algorithms could even come close to the speed of logistic regression at this point, and when it also produces both the most accurate and the most balanced results, it’s pretty much a no-brainer.
The icing on the cake is that logistic regression has much better transparency than a neural network or any of the tree ensemble models. All you have to do if you want the reasoning behind a particular prediction is to look up the relevant feature weights in the formula!
~6 000+ features
I decided to double the number of features and, unsurprisingly, the model performed even better across the board.
Macro-averaged precision: 0.6554 (+0.0129 from more features) Macro-averaged recall: 0.5523 (+0.0272 from more features)
Of course, increasing the number of features to these amounts gradually reduces the usefulness of our deliberate feature selection, making it sound like we wasted a lot of time and energy in data preparation.
But here’s the kicker: in my case, a model based on 6 000 features selected on a class-by-class basis performs more or less exactly the same as a model based on 9 000 simultaneously selected features. And when using the same amount of features, the model based on features selected on a class-by-class basis ALWAYS seems to outperform the one with the simultaneously selected features.
Below is the model with 9 000 simultaneously selected features.
Macro-averaged precision: 0.6475 (-0.0079 from feature selection) Macro-averaged recall: 0.5550 (+0.0027 from feature selection)
If I have the choice between a 6 000 feature model and a 9 000 feature model with more or less the same performance, I’ll go with the first one.
I decided to stop at 6 000 features, although I probably could’ve kept increasing the number of features since LR is blazingly fast even with a huge feature space. Still, with 6 000 features I’ve got a 1:6 ratio of features to training samples, which does sound like it’s about to become a little excessive. More importantly, the improvements to my evaluation metrics from adding more features seemed to become marginal from this point onwards.
Why not use even fewer features since I went to such lengths to select the best ones? Well, in the end I figured that having many features would make the operationalized model more fun to play with, because you can expect it to take advantage of more of the terms you feed it and change its predicted probabilities accordingly. Since this is just for fun and learning and not exactly business critical, I’m not really that concerned with generalization.
So which “rosablogger” are you?
Being done with the model, I published it as a web service using Azure ML and Azure App Service for all to enjoy!
And like I said in the start of my post: all you have to do is go to this website, paste or write any Norwegian block of text, and the model will score that text with probabilities for being written by each of the 11 bloggers and by extension which award you’re most likely win next year.
What is actually happening to the text behind the scenes? It’s processed through the finalized predictive experiment workflow below.
Exploring the final model, there are some interesting takeaways that we can expect to manifest “in production.” Look out for these things – in Norwegian, even though I’ve translated everything here – when you’re trying it out!
- Generally, the model still favors the best represented classes. It is prone to think you are either Sophie Elise or Casa Kaos, especially if it has little meaningful information to go on.
- Distinguishing the writings of “man-blogger” Ørjan Burøe from the other, exclusively female bloggers is not as easy as you might think. In fact, the model believes he is Sophie Elise 18.1 % of the time …
- Writing about kids makes you either Ørjan Burøe or Casa Kaos, and it makes you extremely unlikely to be Sophie Elise. However, writing about dads makes you very likely to be Ørjan Burøe and very unlikely to be Casa Kaos.
- Another thing that makes you very unlikely to be Sophie Elise is using the words “if” and “of course” – “hvis” and “selvfølgelig” in Norwegian. Apparently, there is no if for the blogger of the year, and she takes nothing for granted.
- Writing about “Nelia” and “Lars-Kristian” will most definitely make you Caroline Berg Eriksen.
- Writing in English is likely to make you Camilla Pihl. But again; the model is not really built for English text at all, so don’t expect any reliable or meaningful results on anything other than Norwegian text.
- Laughing (“haha”) makes you Sophie Elise, Caroline Berg Eriksen or Nette Nestea. Interestingly, it also makes you a lot less likely to be Camilla Pihl, Ørjan Burøe and Mat på bordet, and especially Casa Kaos. Those blogs are apparently no fun.
- Hviit is difficult to classify with few strong predictors. However, you can try writing about styling, posters, design and Copenhagen, and avoid the word “dere” like the plague. Hviit apparently has a strict policy of not talking directly to her readers!
- Mat på bordet is the easiest blog to classify, with a recall of 81.8 %. I guess writing about food and recipes does make for a pretty unique blog after all!
- Writing about training will obviously make you Funkygine, but it might also make you Caroline Berg Eriksen. It is very unlikely to make you Camilla Pihl, who apparently writes even less about training than Mat på bordet.
- Amino acids is another thing that’s likely to make you Funkygine, as is Amsterdam and “initial position,” whatever that is.
- Jenny Skavlan is the hardest blogger to distinguish from the others, with a recall of only 32.4 %. However, you can try writing about Fretex and see what happens.
- Using passive formulations such as “man” or writing about Harstad and someone named Robin is likely to make you Sophie Elise. Interestingly, passive voice makes you extremely unlike to be Caroline Berg Eriksen. It makes you a little more likely to be Jenny Skavlan.
- Misspelling “i dag” and “i kveld” as “idag” and “ikveld,” or misspelling “spesielt” as “spessielt,” also makes you Sophie Elise.
- Writing about dough and bread will most definitely make you Mat på bordet, as does writing about cheese. Interestingly, writing about dough should also make you a lot less lilkely to be Casa Kaos.
- Makeup stuff will make you Agnes Lovise, especially nail polish and eye shadow.
An actual business use case
So we’ve just used this methodology and technology for something that is pretty fun, but also pretty useless.
But let’s say that you’ve got thousands of documents or other blocks of text at work. Maybe it’s application forms, maybe it’s free text fields in feedback forms, maybe it’s something completely different. Or maybe it’s just old text files with information you’re unable to take advantage of because the documents are completely disorganized.
If you’re in the latter situation, a good place to start would be to simply categorize and tag these documents so that you can organize and use them. Instead of going through each and every document manually, you can:
- Select a small but presumably representative sample of documents. These will be your training documents.
- Manually classify, tag or otherwise label this sample with the information you’re interested in extracting from all your documents, for example categories, sentiment etc.
- Process all your documents – not just the training documents – in Azure Machine Learning like we did with the blog posts, and train models to predict your labels on your sample.
- Use your models to predict the labels for the rest of your documents that you didn’t manually label earlier.
Now, there are of course dedicated software suites which can do this and much, much more. And in some – perhaps many – cases, you’ll probably want to use those.
But what I think is so great about tools like Azure Machine Learning in combination with R is that you can learn to build solutions like this all by yourself and customize them exactly how you want.
By getting familiar with machine learning technologies and data science methodology, you’ll be much better equipped to deal with the next potential use case for machine learning when it pops up. This skillset is a lot more valuable than any particular software.
That’s all for now! Thanks for sticking with me through this monster blog post, and don’t forget to check back later for more ridiculously stupid use cases for Azure Machine Learning.