Data Science minor final project

Link to Shiny application zip. Includes:

• app.R - application
• w2v_90m - word2vec model trained on full dataset (3 files due to model size)
• artistsChoice.csv - list of artists (for app's input)

• models - word2vec models; (includes only the one trained on 10M)
• notebooks - notebooks
• slides - presentations
• other - code snippets and other stuff
• app - shiny application 🚧 --> https://vk.cc/9v83UO

Data: song playlists from russian SNs, including ~950K users and 90M user-item pairs in total.
Example:

Method: word2vec skip-gram
Idea: Each user's playlist is represented as a sentence, and if artists appear in the same playlists, they are similar and belong to the same context. The model takes artists (at least one) as an input, and recommends n artists.

The final model was trained on full dataset: approximately 90 000 000 user-item, ~ 950 000 users, time: 9 hours.

Case: something epic for the one who loves Game of Thrones

model_w2v.wv.most_similar('ramin djawadi', topn=10)

Model recommends other authors of soundtracks. Interesting case: soundtrack to 'the witcher 3 wild hunt' - 'The Trail' is quite similar to the TV series main song.

[('drake',	0.6897857189178467),
('hans zimmer',	0.774951577186584),
('ramin djawadi',	0.761010468006134),
('westworld',	0.747692465782166),
('the witcher 3 wild hunt',	0.722944676876068),
('daniel pemberton',	0.721352934837341),
('howard shore',	0.719944596290588),
('jeremy soule',	0.715254724025726),
('two steps from hell',	0.711450159549713),
('hans zimmer',	0.709886014461517),
('akihiro honda',	0.703315019607544)]

Case: the one who listens to placebo and radiohead and is not in a good mood

placebo_sim = [a[0].strip() for a in model_w2v.wv.most_similar(['placebo'], topn=15)]
placebo_sim

['iamx',
 'arctic monkeys',
 'blue october',
 'franz ferdinand',
 'radiohead',
 'the smiths',
 'the killers',
 'him',
 'the pretty reckless',
 '30 seconds to mars',
 'hypnogaja',
 'my chemical romance',
 'sea wolf',
 'she wants revenge',
 'stereophonics']

And who does not match there?
human prediction: 'my chemical romance' or 'franz ferdinand'
model:

print(model_w2v.wv.doesnt_match(placebo_sim))

my chemical romance