/arena_melon

traces made for the 3rd kakao arena challenge (melon playlist continuation)

Primary LanguagePython

arena_melon

  • traces made for the 3rd kakao arena challenge (melon playlist continuation) [star]

  • could be updated to verify other hypotheses even after the end of the official event :)

  • the final model submitted is basically based on MF(ALS), and CB model for cold-start cases

    1. first of all, popularity-based item filterting was done (only 30% of total items was used)
    2. for cold-start cases, playlist titles were exploited to make comparison and calculate similarities among them
    3. for non-cold cases, MF(ALS) method was applied thanks to implicit library, with a little hyper-parameter tuning
    4. for tag prediction, some heuristic approach was made: calculate the tags per each song and again aggregate the tags of the songs in each playlist

  • regarding the task as a kind of 'next-item recommendation' problem, several deep learning based approaches were tried.. : but some showed worse performances than MF, while others were found to hard to optimize for now (will fix soon)

  • happy to have a tons of trials & errors in order to (come up with ideas) & (make it work)

  • and most of all, many thanks to kakao for this opportunity :)