Conclusions

​

After exploring and testing on the methods we proposed, word processing techniques such as Word2Vec and specialized language models like LSTM are effective in improving classification accuracy. More specifically, with top 3 generated predictive tags, we are very likely to capture the true story hidden behind the lyrics over 60% of the time for a completely new song given with such model.

 

Lyrics are useful in labeling songs with the most relevant themes, and can thus be used as a complement to song acoustic features in making playlist recommendations. More generally, with a more personalized playlist that tells your desired story, Spotify should be able to further increase its current users’ stickiness as well as attracting more users. Having a large market coverage is a huge business value by itself, since the company can thus have more advertisers reaching out driving up its revenue. Most importantly, as users for the music streaming platform ourselves, we should be able to have a much better enjoying experience with the new recommended list.

​

Future work

​

• As target tag processing now is still been done in a relatively naïve approach(using majority votes), if time permits, we would also consider grouping all distinct tags into larger batches, extract some themes out of the cluster and use them as the prediction response.

 

• With the current lyrics-based model we have now, another natural further step to take is incorporating the acoustic features when making playlist generations to make the model more versatile. More specifically, as lyrics are helpful and accurate in some specific themes or mood recommendation, and acoustic features may be good at finding songs that sound similar, we would like to combine the two aspects and thus provide the playlist that have both similar rhythm songs as well as songs telling a similar story without being too monotonic.