100 Warm Tunas 2016 Prediction Analysis

It’s been a long time since the Hottest 100 of 2016 was aired. Unfortunately, I never really got around to publishing some analysis I performed on the prediction results. Fortunately, I managed to find some time recently!

Looking from afar, the results don’t look fantastic (when you compare them to my results from 2015 at least). The prediction unfortunately predicted the top two places out of order, however did manage to predict the third place correctly.

Lets take a look at the Top 10 of Triple J’s list and match it up with 100 Warm Tunas:

Triple J Rank vs Tuna Rank

Looking at this we see most predictions we can find some learnings:

  • The average error for the top ten rank was 1.9 rank positions.
  • If 100 Warm Tunas ignored rank and simply guessed the top ten, it would have predicted 8 of the top 10 songs.
  • If 100 Warm Tunas ignored rank and simply guessed the top 3 songs to win, it would have predicted all 3 songs. Woo!

Lets dive into a chart that shows error for all ranks:

Rank error per position

From this chart, we can deduce that the further away from position 1 we become, the higher the error. This information alone isn’t very useful. We can get a better understanding of error by finding the average for each ranking group:

Average Rank error per group

As we get closer to rank 1, the results become more and more accurate, however they are not perfect. This is more obvious if we use a scatter plot to compare Triple J ranks against Warm Tunas predictions:

Triple J vs Tunas Scatter Plot

It’s clear now that as we get closer to rank 1, the 100 Warm Tunas prediction gets better and converges upon the actual rankings played out on the day. However, unfortunately this year the difference between rank 1 and rank 2 was way too close to call - just 0.67% of voting volume was separating the two. A difference that was not enough to provide an accurate prediction of the winner.

Overall, whilst 100 Warm Tunas 2016 did get the two top positions out of order, it’s understandable as to why this happened. Hopefully this year there is a greater difference between ranks, giving further ability to predict the winner in position #1.

Reverse Engineering a 433MHz Motorised Blind RF Protocol

I’ve been doing a fair bit of DIY home automation hacking lately across many different devices - mostly interested in adding DIY homekit integrations. A couple of months ago, my dad purchased a bulk order of RAEX 433MHz RF motorised blinds to install around the house, replacing our existing manual roller blinds.

RAEX Motorised Blind

Note: If you are based in Australia, you can purchase these in bulk or individually via www.raexaustralia.com (Full disclosure – my father runs the site).

The blinds are a fantastic addition to the house, and allow me to be super lazy opening/closing my windows, however in order to control them you need to purchase the RAEX brand remotes. RAEX manufacture many different types of remotes, of which, I have access to two of the types, depicted below:

R Type Remote

R Type Remote (YRL2016)

X Type Remote

X Type Remote (YR3144)

Having a remote in every room of the house isn’t feasible, since many channels would be unused on these remotes and thus a waste of $$$ purchasing all the remotes. Instead, multiple rooms are programmed onto the same remote. Unfortunately due to this, remotes are highly contended for.

An alternate solution to using the RAEX remotes is to use a piece of hardware called the RM Pro. This allows you to control the remotes via your smartphone using their app

RM Pro Home Screen
RM Pro Blind Control Screen

The app is slow, buggy and for me, doesn’t fit well into the home-automation ecosystem. I want my roller blinds to be accessible via Apple Homekit.

In order to control these blinds, I knew I’d need to either:

  1. Reverse engineer how the RM Pro App communicated with the RM Pro and piggy-back onto this
  2. Reverse engineer the RF protocol the remotes used to communicate with the blinds.

I attempted option 1 for a little while, but ruled it out as I was unable to intercept the traffic used to communicate between the iPhone and the hub. Therefore, I began my adventure to reverse engineer the RF protocol.

I purchased a 433MHz transmitter/receiver pair for Arduino on Ebay. In case that link stops working, try searching Ebay for 433Mhz RF transmitter receiver link kit for Arduino.

Initial Research

A handful of Google searches didn’t yield many results for finding a technical specification of the protocol RAEX were using.

  • I could not find any technical specification of the protocol via FCC or patent lookup
  • Emailed RM Pro to obtain technical specification; they did not understand my English.
  • Emailed RAEX to obtain technical specification; they would not release without confidentiality agreement.
  • I did find that RFXTRX was able to control the blind via their BlindsT4 mode, which appears to also work for Outlook Motion Blinds.
  • After opening one of the remotes and identifying the micro-controllers in use, I was unable to find any documentation explaining a generic RF encoding scheme being used.
  • It may have been possible to reverse engineer the firmware on a remote by taking an I2C dump of the ROM chip. It seems similar remotes allow dumping at any point after boot

Capturing the data

Once my package had arrived I hooked up the receiver to an Arduino and began searching for an Arduino sketch that could capture the data being transmitted. I tried many things that all failed, however eventually found one that appeared to capture the data.

Once I captured what I deemed to be enough data, I began analysing it. It was really difficult to make any sense of this data, and I didn’t even know if what had been captured was correct.

I did some further reading and read a few RF reverse engineering write-ups. A lot of them experimented with the idea of using Audacity to capture the signal via the receiver plugged into the microphone port of the computer. I thought, why not, and began working on this.

The RF capturing setup

Audacity capture

This captures a lot of data. I captured 4 different R type remotes, along with 2 different X type remotes, and to make things even more fun, 8 different devices pairings from the Broadlink RM Pro (B type).

From this, I was able to determine a few things

  1. The transmissions did not have a rolling code. Therefore, I could simply replay captured signals and make the blind do the exact same thing each time. This would be the worst-case scenario if I could not reverse engineer the protocol.
  2. The transmissions were repeated at least 3 times (changed depending on the remote type being used)

Zooming into the waveform, we can see the different parts of a captured transmission. This example below is the capture of Remote 1, Channel 1, for the pairing action:

R1, CH1 PAIR capture

Zooming in:

Zoomed R1, CH1 PAIR capture

In the zoomed image you can see that the transmission begins with a oscillating 0101 AGC pattern, followed by a further double width preamble pattern, followed by a longer header pattern, and then by data.

This preamble, header and data is repeated 3 times for R type remotes (The AGC pattern is only sent once at the beginning of transmission). This can be seen in the first image.

Looking at this data won’t be too useful. I need a way to turn it digital and analyse the bits and determine some patterns between different remotes, channels and actions.

Decoding the waveform.

We need to determine how the waveform is encoded. It’s very common for these kinds of hardware applications to use one of the following:

By doing some research, I was able to determine that the encoding used was most likely manchester encoding. Let’s keep this in mind for later.

Digitising the data

I began processing the data as the raw scheme outlined above (even though I believed it was manchester). The reason for this is that if it happened to not be manchester, I could try decode it again with another scheme. (Also writing out raw by hand was easier than doing manchester decoding in my head).

I wrote out each capture into a Google Sheets spreadsheet. It took about 5 minutes to write out each action for each channel, and there were 6 channels per remote. I began to think this would take a while to actually get enough data to analyse. (Considering I had 160 captures to digitise)

I stopped once I collected all actions from 8 different channels across 2 remotes. This gave me 32 captures to play with. From this much data, I was able to infer a few things about the raw bits:

  • Some bits changed per channel
  • Some bits changed per remote.
  • Some bits changed seemingly randomly for each channel/remote/action combination.
    • Could this be some sort of checksum?

I still needed more data, but I had way too many captures to decode by hand. In order to get anywhere with this, I needed a script to process WAV files I captured via Audacity. I wrote a script that detected headers and extracted data as its raw encoding equivalent (as I had been doing by hand). This script produced output in JSON so I could add additional metadata and cross-check the captures with the waveform:

    "filename": "/Users/nickw/Dropbox/RF_Blinds/Export_Audio2/tracks2/R1_CH1.wav",
    "captures": [
        "data": "01100101100110011001100101101001011010010110011010011010101010101010101010011001101010101010101010101010101",
        "header_pos": 15751,
        "preamble_pos": 15071
        "data": "01100101100110011001100101101001011010010110011010100110101010101001101010011001101010101010101010101010101",
        "header_pos": 46307,
        "preamble_pos": 45628
        "data": "01100101100110011001100101101001011010010110011010010110101010101010011010011001101010101010101010101010101",
        "header_pos": 73514,
        "preamble_pos": 72836
        "data": "01100101100110011001100101101001011010010110011010101010101010100101010101101001011010101010101010101010101",
        "header_pos": 103575,
        "preamble_pos": 102895

Once verified, I tabulated this data and inserted it into my spreadsheet for further processing. Unfortunately there was too many bits per capture to keep myself sane:

Raw captures inside a spreadsheet

I decided it would be best if I decoded this as manchester. To do this, I wrote a script that processes the raw capture data into manchester (or other encoding types). Migrating this data into my spreadsheet, it begins to make a lot more sense.

Manchester captures inside a spreadsheet

Looking at this data we can immediately see some relationship between the bits and their purpose:

  • 6 bits for channel (C)
  • 2 bits for action (A)
  • 6 bits for some checksum, appears to be a function of action and channel. F(A, C)
    • Changes when action changes
    • Changes when channel changes.
    • Cannot be certain it changes across remotes, since no channels are equal.
  • 1 bit appears to be a function of Action F(A)
  • 1 bit appears to be a function of F(A), thus, G(F(A)). It changes depending on F(A)’s value, sometimes 1-1 mapping, sometimes inverse mapping.

After some further investigation, I determined that for the same remote and channel, for each different action, the F(A, C) increased by 1. (if you consider the bits to be big-endian.).

Encoded value increasing per different action

Looking a bit more into this, I also determined that for adjacent channels, the bits associated with C (Channel) count upwards/backwards (X type remotes count upwards, R type remotes count backward). Additionally F(C) also increases/decreases together. Pay attention to the C column.

Encoded value increasing with adjacent channels

From this, I can confirm a relationship between F(A, C) and C, such that F(A, C) = F(PAIR, C0) == F(PAIR, C1) ± 1. After this discovery, I also determine that there’s another mathematical relationship between F(A, C) and A (Action).

Making More Data

From the information we’ve now gathered, it seems plausible that we can create new remotes by changing 6 bits of channel data, and mutating the checksum accordingly, following the mathematical relationship we found above. This means we can generate 64 channels from a single seed channel. This many channels is enough to control all the blinds in the house, however I really wanted to fully decode the checksum field and in turn, be able to generate an (almost) infinite amount of remotes.

I wrote a tool to output all channels for a seed capture:

./remote-gen generate 01000110110100100001010110111111111010101

My reasoning behind generating more data was that maybe we could determine how the checksum is formed if we can view different remotes on the same channel. I.e. R0CH0, R1CH0, X1CH0, etc…

Essentially what I wanted to do was solve the following equation’s function G:


However, looking at all Channel 0’s PAIR captures, the checksum still appeared to be totally jumbled/random:

Identical channels / action jumbled checksums

Whilst looking at this data, however, another pattern stands out. G(F(A)) sits an entire byte offset (8 bits) away from F(A). Additionally the first 2 bits of F(A, C) sit at the byte boundary and also align with A (Action). As Action increases, so does F(A, C). Lets line up all the bits at their byte boundaries and see what prevails:

Identified Boundaries
Colours denoting byte boundaries
Aligned byte boundaries
Aligned boundaries

From here, we need to determine some function that produces the known checksum based on the first 4 bytes. Initially I try to do XOR across the bytes:

Attempt to find checksum function via XOR

Not so successful. The output appears random and XOR’ing the output with the checksum does not produce a constant key. Therefore, I deduce the checksum isn’t produced via XOR. How about mathematical addition? We’ve already seen some addition/subtraction relationship above.

Attempt to find checksum function via addition

This appeared to be more promising - there was a constant difference between channels for identical type remotes. Could this constant be different across different type remotes because my generation program had a bug? Were we not wrapping the correct number of bits or using the wrong byte boundaries when mutating the channel or checksum?

It turns out that this was the reason 😑.

Solving the Checksum

Looking at the original captures, and performing the same modulo additions, we determine the checksum is computed by adding the leading 4 bytes and adding 3. I can’t determine why a 3 is used here, other than RAEX wanting to make decoding their checksum more difficult or to ensure a correct transmission pattern.

I refactored my application to handle the boundaries we had just identified:

type RemoteCode struct {
    LeadingBit uint // Single bit
    Channel    uint8
    Remote     uint16
    Action     uint8
    Checksum   uint8

Looking at the data like this began to make more sense. It turns out that F(A) wasn’t a function of A (Action), it was actually part of the action data being transmitted:

type BlindAction struct {
    Name  string
    Value uint8

var validActions = []BlindAction{
    BlindAction{Value: 127, Name: "PAIR"},
    BlindAction{Value: 252, Name: "DOWN"},
    BlindAction{Value: 253, Name: "STOP"},
    BlindAction{Value: 254, Name: "UP"},

Additionally, the fact there is a split between channel and remote probably isn’t necessary. Instead this could just be an arbitrary 24 bit integer, however it is easier to work with splitting it up as an 8 bit int and a 16 bit int. Based on this, I can deduce that the protocol has room for 2^24 remotes (~16.7 million)! That’s a lot of blinds!

I formally write out the checksum function:

func (r *RemoteCode) GuessChecksum() uint8 {
    return r.Channel + r.Remote.GetHigh() + r.Remote.GetLow() + r.Action.Value + 3

Additional Tooling

My remote-gen program was good for the purpose of generating codes using a seed remote (although, incorrect due to wrapping issues), however it now needed some additional functionality.

I needed a way to extract information from the captures and verify that all their checksums align with our rule-set for generating checksums. I wrote an info command:

./remote-gen info 00010001110001001101010111011111101010100 --validate
Channel:    196
Remote:     54673
Action:     STOP
Checksum:          42
Guessed Checksum:  42

Running with --validate exits with an error if the guessed checksum != checksum. Running this across all of our captures proved that our checksum function was correct.

Another piece of functionality the tool needed was the ability to generate arbitrary codes to create our own remotes:

./remote-gen create --channel=196 --remote=54654 --verbose
00010001101111110101010111111111010011001    Action: PAIR
00010001101111110101010110011111101101000    Action: DOWN
00010001101111110101010111011111111101000    Action: STOP
00010001101111110101010110111111100011000    Action: UP

I now can generate any remote I deem necessary using this tool.

Wrapping Up

There you have it, that’s how I reverse engineered an unknown protocol. I plan to follow up this post with some additional home-automation oriented blog posts in the future.

From here I’m going to need to build my transmitter to transmit my new, generated codes and build an interface into homekit for this via my homebridge program.

You can view all the work related to this project in the nickw444/homekit/blindkit repo.

As mentioned above, if you are based in Australia, you can purchase these blinds and associated accessories in bulk or individually via www.raexaustralia.com (Full disclosure – my father runs the site)

100 Warm Tunas 2016

Last year I predicted the top 3 results in order in Triple J’s hottest 100. This year I’m back at it again, however, now with a webpage and a Spotify playlist.

Results are collected, optimised, and processed multiple times per day. Instagram images tagged with #hottest100 and a few others are included for counting.

Happy voting!

You can read about the process last year here. However, vote collection is a fair bit more accurate this year.


Other Mentions:

Edit: Woohoo, the Spotify playlist now has just over 1200 followers, and the website has had over 30,000 hits! That’s massive, thanks everyone!

Understanding and Tweaking some GPX data

As a casual bike rider, I enjoy tracking my rides with Strava so I can take a look at how my ride went and how well I performed throughout.

However, very rarely the Strava tracking application randomly crashes, or gets killed by iOS on my phone, during the ride. This means that the data was never recorded between the point at which the app died and the point when I became aware the app had died.

If we plot this type of failure, it looks something like this:

Map with missing data

Fortunately in this case, there wasn’t too much missing data. However, I was still determined to learn about the GPX format and see if I could patch up the GPX file programatically.

In the specific case of the above map, I was riding north west, and at a point Strava crashed. Between this point and when I pulled out my phone to check my progress, no points were plotted. Google maps interprets this lack of data as a straight line between to the 2 points (as per GPX specification).

If we crack open the GPX file and take a look, we can see exactly what this looks like:

<trkpt lat="-33.9014420" lon="151.1066810">
<trkpt lat="-33.8802920" lon="151.0702190">

In it’s simplest form, a GPX file is an XML document that contains a sequence of GPS points (with associated metadata like elevation, and other depending on the tracker). This makes it reasonably simple for us to get our hands dirty and begin fixing the data set.

In order to add the missing data back into the GPX file, we need 3 things:

  • The last coordinate recorded before the app crashed
  • The coordinate when the app was revived
  • A list of points of the track we want to use for our data points.

Fortunately, I was able to obtain a list of coordinates for the missing data since I travelled the same path on the return journey (As can be seen on the map above).

The other 2 app state points of interest are reasonably easy to find - just find 2 data points that have a (reasonably) large time distance between them.

In order to process the data, I used a python library called gpxpy which provided some good utilities for reading and processing a GPX file.

With this library, I was able to find the crash point, the revival point, and the list of the points of the track. With this data, I interpolated the start/end times of the crash points onto the track data, and spliced it back into the dataset.

After exporting the data set, we achieve a map that looks like:

Map with resolved data

Quite clearly, this has a few limitations, for example, the calculated velocity through all of the data points is simply an average. However, this did provide me with an improved dataset which I could re-upload to Strava.

You can find all the source for this script on my github

Accurately Predicting Triple J's Hottest 100 of 2015

In 2014, a prediction was accurately made for the Hottest 100 of 2013. The results were posted on warmest100.com.au.

The author of the prediction in 2014 managed to acquire accurate results because Triple J featured a social share button on their voting page, which posted your votes to your Facebook in text form. The author scraped results from public Facebook posts and aggregated all the votes. They managed to obtain 1.3% (1779 entries) of the expected total vote.

Consequently, voting for the Hottest 100 2014 and 2015 did not contain such a feature. Fortunately, voters still felt the need to share these results with their friends, and taking a screen shot or a photo of their screen and posting to social media was a concrete alternative. Using these images posted to Instagram, I was able to accurately predict the results of Triple J’s Hottest 100 of 2015.

Some Cool Stats Before You Continue

  • Triple J Tallied 2094350 Votes (209435 Entries) for Hottest 100 2015
  • I collected a sample size of ~2.5% of all entries
    • 7191 images initially collected
    • I categorised 5529 images as votes
    • ~4900 images contained the words “vote/votes/voting”
  • My Top 3 Results were 100% accurate

You’ll probably find this article interesting, but if you’re super eager, you can Skip To The Results.

Taking Advantage of Social Media

I decided to only target votes that were posted to Instagram, since a high majority of the pictures hashtagged with #hottest100 were in fact votes, and there was a reasonably high volume of them, and most publicly accessible.

I required means to acquire all pictures that had been posted to Instagram. Instagram have an official API, however you are required to have your API app usage approved before it can interface with non-sandbox users. Additionally, Instagram impose a rate limit on non-approved apps, as well as approved apps. I did not have time to waste, and wanted results immediately, so I found an alternative.

Fortunately, Instagram exposes a non-public API through their website ajax loading when you browse to a hashtag. By imitating the web browser with a simple python script using the requests library I managed to download all images from the latest until a cut off date that I specified (the day voting opened).

After scraping the hashtag #hottest100, I expanded my search to #hottest1002015 and #triplejhottest100.

Processing Images

After downloading 7191 images from Instagram, I needed to find an accurate way to filter out the images that were not votes.

I’ve had previous experience with using PIL in Python, so using PIL, I wrote a simple script to sort the photos into 2 categories; photos that appeared white-ish, and photos that were not.

A good vote looked like this:

A Good Vote

Unfortunately, not every image ended up in the right folder, and I ended up with both false negatives and false positives, however I wasn’t too concerned about false positives, as my OCR processing step would exclude them. Instead, I was more concerned about false negatives.

As the image processing and sorting continued, I manually moved false negatives to the positives folder. I calculated about 5% of the non-matching photos were incorrectly classified, however this was due to them being pictures taken of computer screens, similar to the photo below:

A Bad Vote

Some image statistics:

  • 7191 images collected initially
  • 1662 images categorised as non-votes
  • 5529 images categorised as votes
  • ~4900 images contained the words vote/votes/voting

Improving OCR Performance

After experimenting on raw photos from Instagram, I found that OCR accuracy was not very accurate. To remediate this, I utilised Imagemagick to flatten image definition to improve text results.

An improved image

Bringing in Tesseract (OCR)

After weeding out the junk, I still needed to turn these images into readable text.

Using Google’s Tesseract library, I slowly processed all the images and extracted the text from them.

Unfortunately, due to the layout of the Hottest 100 voting website the two columns were broken up inconsistently over the results.

Some were processed as:

Line by Line processing
Flight Facilities
Hayden James
Major Lazer
Weeknd, The
ZHU x Skrillex x THEY.
Jarryd James
Kendrick Lamar
Heart Attack {FL Owl Eyes)
(Radio Edit)
Something About You
The Buzz (Ft. Malaya/Young
Lean On (Ft. Mé/DJ Snake}

And others processed as:

Song/Artist line by line
Lucky Luke 1 Day
Mosquito Coast Call My Name
Tn ka Right By You
Tuka L.D.T.E.
Half Moon Run Trust
Spring King City
Tame Impala Let It Happen
Saskwatch I‘ll Be Fine
Jungle Giants. T Kooky Eyes

And others just did not process at all, due to resolution, colour, skewing, or simply because they were a photo of a computer screen:

Bad Image
'VHotllne Bling
Regardless (Ft. Julia Stone)

Parsing the Results

I processed the results line by line, and call these “terms”. These such terms could contain a single song title, a single artist, an artist name with song name, or just junk overhang from a previous line. Initially there were 31062 uncategorised terms.

I processed each term and aggregated number of results for each. This worked really well for songs with short names that were less prone to error, such as Hoops, however did not correctly capture terms where artist name and song name occurred on the same line, or where the OCR library interpreted a few characters incorrectly.

OCR Inaccuracy & Levenshtein

Even with photo enhancements, the OCR accuracy was somewhat subpar for some votes. Some l’s were interpreted as t’s, i’s as l’s, etc. Additionally, the longer the name of the song, the more prone to error it was.

Fiesh Without Blood
L D R U Keepmo Score Fl Pavqe IV)
Yam: unpala The Les I Knew The Bauer
The Tlouble Wilh US

A technique that can be used to fix these spelling errors of single/multi character errors is the Levenshtein algorithm for edit distance. Using this algorithm, we can compare 2 strings and determine how many edits need to be made to make the strings equal each other.

In order to perform this kind of matching, we needed an accurate list of songs that were released this year, along with a list of artists that released music this year.

Using Spotify To Help

To acquire an accurate list of songs released this year, I used Spotify and crawled various playlists from 2015. These included Spotify Charts, Triple J Hitlist, and various other genre-alike playlists.

In the end I ended up with a songs list with 1781 songs, and an artists list with 1229 artists. After the Hottest 100 aired, I compared the results of the countdown to the songs found in my list, and only 6 songs that occurred in the hottest 100 were not in my “truth” list.

During list gathering, I made sure to convert all unicode characters to their ASCII counterparts, so that characters with accents and similar would be matched correctly.

Continuing Processing

Now carrying reasonably accurate artists and songs lists we continue categorisation and processing. The processing algorithm worked in the following way:

  1. Load all terms from every image’s .txt OCR result. Every line is a “term”.
  2. Clean all the terms by turning them into lowercase and stripping whitespace.
  3. Loop through each term:
    1. If term exists in our known songs list, move the term to the songs aggregation and count the votes.
    2. If term exists in our known artists list, move the term to the artists aggregation and count the votes.
    3. If couldn’t find it in either of those:
      1. Loop through all artists in our artist known artist list.
        1. Check if the term starts with the current artist. If it does split it into artist and unknown term. Add the votes to the artist aggregation.
        2. If matched artist, check if the new unknown term exists in the songs list, if it does, add it to the songs aggregation. If not, add it back to the unknown. break loop.
      2. If it didn’t have a prefixed artist, just add it back to the unknown terms.

At this stage, we have a reasonably accurate aggregation of results. We have not yet used Levenshtein string matching. We now have 27294 uncategorised terms, down from 31062 uncategorised terms. So far our results:

==       Results       ==
1   Hoops                          998
2   King Kunta                     765
3   Lean On                        750
4   The Buzz                       646
5   Like Soda                      568
6   Never Be                       484
7   Let It Happen                  476
8   Magnets                        465
9   Do You Remember                409
10  Ocean Drive                    405
==    853 unique terms    ==

==       Top Unknown Terms       ==
1   Your Hottest 100 Votes:        2279
2   Your Votes                     2127
3   }                              320
4   Hottest Io                     248
5   V                              231
6   Throne                         222
7   Triple J?                      209
8   D] Snake                       203
9   The Less | Know The Better     203
10  Asap Rocky                     199
==    27294 unique terms    ==

However, we still haven’t aggregated any votes that had spelling errors due to OCR inaccuracies.

Employing the Levenshtein algorithm, we continue to process the unknown terms. I configure matching to allow lenience based on the length of the term - the maximum edits that were allowed was 2/5 * length of term. The process continues:

  1. For all unknown terms:
    1. Check term length > 3. Break if <= 3. Can’t match a short string.
    2. Match Songs:
      1. Loop through all songs in known songs list:
        1. Compare current song to current term. Get edit distance.
        2. If edit distance == 1, move votes for this term to the guessed song in our songs aggregation, then continue to the next term.
        3. Add distance to a dictionary of value/distances
      2. Using our value/distances dictionary, find the closest match that satisfies our 2/5 * len(term) rule. If it matches, move the votes for this term to the guessed song in our songs aggregation, then continue to the next term.
    3. Match Artists using the same method.

Some of the results of string matching, providing some reasonably accurate re-matching.

[A] weekncl, the -> weeknd, the with distance 2
[A] mm m. -> ms mr with distance 2
[S] km; kunta -> king kunta with distance 3
[A] macklelllore ex ryan lewis -> macklemore & ryan lewis with distance 5
[A] eulsch duke) -> deutsch duke with distance 3
[A] bloc pany -> bloc party with distance 2
[S] nommg's forevev -> nothing's forever with distance 5
[S] t he hllns -> the hills with distance 3
[S] emocons -> emoticons with distance 2
[S] better off without you -> better with you with distance 7
[S] - the less | know the better -> the less i know the better with distance 3
[S] vancejoy fire and the fiood -> fire and the flood with distance 10
[S] too much me togglhu -> too much time together with distance 6
[A] of mons-us and m. -> of monsters and men with distance 5
[S] gmek tragedy -> greek tragedy with distance 2
[S] marks to prove 1t -> marks to prove it with distance 1
[A] rlighx facilities -> flight facilities with distance 2
[A] gang 01 youth: -> gang of youths with distance 3
[A] fka lwlgs -> fka twigs with distance 2
[S] hoine bling -> hotline bling with distance 2

After performing this additional processing, I ended up with 18509 uncategorised terms, down from 27294 uncategorised terms.

That means we were able to successfully categorize 8785 terms via the Levenshtein distance algorithm!

==       Results       ==
1   Hoops                          1011
2   King Kunta                     1008
3   Lean On                        793
4   The Buzz                       667
5   Let It Happen                  637
6   Like Soda                      617
7   The Less I Know The Better     602
8   Magnets                        521
9   Never Be                       520
10  The Trouble With Us            501
==    1143 unique terms    ==

==       Top Unknown Terms       ==
1   Your Hottest 100 Votes:        2279
2   }                              320
3   Hottest Io                     248
4   V                              231
5   Throne                         222
6   Triple J?                      209
7   Thanks For Voting!             174
8   Tapz)                          170
9   Suddenly                       155
10  Once                           140
==    18509 unique terms    ==

Quite an improvement, however still not great. Some of the terms there weren’t able to be categorised which caught my attention included:

9   Suddenly                       155
16  Big Jet Plane                  123
17  Heart Attack                   120
18  True Friends                   114
23  Rumour Mill                    107
35  The Less | Know The            76
63  & Chet Faker The Trouble With Us 46

Paying special attention to The Less | Know The, if I were to add it’s sum to our results, it would have placed 4th, however, the results we already have look reasonably accurate.

Final Results

==       Results       ==
1   Hoops                          1011
2   King Kunta                     1008
3   Lean On                        793
4   The Buzz                       667
5   Let It Happen                  637
6   Like Soda                      617
7   The Less I Know The Better     602
8   Magnets                        521
9   Never Be                       520
10  The Trouble With Us            501
11  Do You Remember                480
12  Ocean Drive                    463
13  Can'T Feel My Face             457
14  You Were Right                 444
15  Middle                         423
16  Magnolia                       381
17  Young                          380
18  The Hills                      369
19  Hotline Bling                  356
20  Keeping Score                  321
21  Embracing Me                   319
22  Mountain At My Gates           318
23  Loud Places                    300
24  Run                            298
25  I Know There'S Gonna Be        287
26  Some Minds                     287
27  Say My Name                    283
28  Fire And The Flood             280
29  Visions                        275
30  Greek Tragedy                  274
31  Long Loud Hours                272
32  Shine On                       254
33  Asleep In The Machine          249
34  Leave A Trace                  242
35  Like An Animal                 235
36  Something About You            224
37  Dynamite                       224
38  All My Friends                 218
39  Deception Bay                  217
40  Downtown                       210
41  Ghost                          200
42  Son                            196
43  Hold Me Down                   196
44  No One                         196
45  Kamikaze                       196
46  Puppet Theatre                 192
47  Vice Grip                      191
48  Forces                         185
49  Better                         185
50  Counting Sheep                 184
==    1143 unique terms    ==

Some Notes

  • Run appeared so high on the leaderboard because both Seth Sentry and Alison Wonderland released similar tracks titled RUN/Run. Since I lowercased all comparisons and removed special characters, these votes merged.

Improving the Analysis

After reviewing the method used for analysis, I have identified a few places for improvement that could possibly improve the results.

  1. Improved Levenshtein Algorithm. The Levenshtein algorithm is great for calculating edit distance, however I could not weigh edits of similar characters such as t’s, i’s and l’s less, thus improving matching due to OCR inaccuracies. I expect that string matching could have been significantly improved if this was explored.
  2. Songs that had long titles, such as The Less I Know The Better generally were split across multiple lines. This caused their aggregation to not sum correctly. It would be good if I could determine if a song was split across two lines.
  3. Songs that were in the format of artist song and were spelt incorrectly were most likely not picked up by string matching, as we only matched against songs and artists individually. In order to improve matching for this, an additional list for joined songs/artists could have been used and compared against for remaining terms.

Some Cool Stats

  • Triple J Tallied 2094350 Votes (209435 Entries)
  • I collected a sample size of ~2.5% of all entries
    • I collected 7191 images collected initially
    • I categorised 5529 images as votes
    • ~4900 images contained the words “vote/votes/voting”
  • My Top 3 Results were 100% accurate