Today, I was thinking about how rhythmbox happens to be playing songs that I enjoy hearing, one after another. It for whatever reason, happens to know exactly the next song that I would like to hear. This got me thinking about the why this works. With these thoughts, I also got an idea how this processes could be improved to be even better.
When a new playlist is loaded, every song can be sorted on the following fields:
- Play count
- Last 10 time times
- Surrounding tracks at last rating
- Surrounding tracks at current rating
Play count is initialized to 0. Next, the user begins playing new tracks. This would increase the play count under the following conditions:
- The song was completely played, or
- The song was played atleast 75% and the next played track has a rating of atleast 3.
Upon updating the play count, the surrounding tracks are placed in the temporary surrounding tracks list. Later, this list becomes either the current rating or previous rating surrounding track list.
When a certain number of tracks have been seeded, the initialization of the random list can begin.
The first 10 tracks are sorted and the one with the most popular play time ( See most updating most popular play time ) closest to the current time is selected for play. This ensures that the first played song is a of high rating. Songs played at the same time, e.g. after work help build a profile of the listener. Perhaps, the individual begins preparing a meal when coming home from work at 6 pm. However, Friday he comes home at a different time and instead enjoys a different flavor of music. This sort of behavior requires that days are considered in the popular play time calculation.
Interesting things can begin to happen. The user at this time is always capable of switching to the next track. However, if the track is played through, note that its count and rating will be adjusted. This will continue for each succesively played track. Note that, occacionally, more frequent during first few minutes of playtime or number of songs played, the next selected track will be skewed from the current play list position. The exact offset is determined by a selective algorithm that considers the following information:
- Tracks in the surrounding track list of the current and previous rating
- Position at a random offset of the current. The exact offset is picked based on:
- Genre similarity of next track compared to current
- The above suggests that certain the next song should be as similiar to the current track as possilbe. It would be interesting to determine the beat patterns of the next songs in the list. It would also be beneficial to analyze and record a priority list of beat patterns which are most enjoyable to the current listener. This would be most ideal however suffers from a critical flaw in that different genres, even those in same, may have compeletly different patterns however they are appropriate to be played at the present time.
Now, certain songs are penelized in rating, more so, repelled from the list of surrounding tracks of current rating of the switching track. This may assist in composing distinctions between “moods” For instance, when I am listening to soft music and the next track isn’t desirable, I will change it. This behavior indicates that the two tracks are not related and instead, priority should be given to the track which was selected. This binding will enforce a playlist of desirable songs. Frequently, a song can place an individual into a specific mood. This is more strengthened when following songs further reinforce the mood. The listener is overcome with a feeling of “perfectly selected” tracks.
It would also be interesting to create a sorted list of authors based on rating. For instance, a list of favorite Do or Die songs. Note that at anytime that a song is played, a rating is increased. The author hopes that the listener is not consistently in random play and is continually refiniforcing behavioral patterns. However, this is far fetched. In reality, this cannot be guaranteed and the random play must exploring new songs on its own. It will therefore attemtp to select sogns over time which have never been visited. This process will create an organic playlist since new songs are selected. However, unattended play or attempt to find new songs can have negative effect since poisoning of the list can occur. This means that there must be an aggresive algorithm for 1. Retaining strong bounded songs and mood files 2. Be flexible in adjusting this list 3. Efficiently label songs as unwanted. The challange remains in that at one particular time a song was not friendly since it had bad timing (perhaps the mood algorithm was incorrect) it should still maintain its rating but analyses of whether the song should mainting its strong bound should be considered, perhaps simply weakened.
Allow people to queue next song – strong indicator of mood and acceptable next song – however, user may be experimenting and this should be accounted for. Allow users to quickly enable and disable the list. Also, allow users to easily remove songs from the queue.
Problems with iTunes Smart Playlist
The iTunes is apparently a successful manager of enormeous playlists. However, preminelary analysis shows critical falts including:
- Only supports organization based on limited criteria which do not nesesseraly reflect the mood of the listener and validity of next play song. Smart Playlist is limited to:
- Play count
- File type filtretation
- Conditions based on:
- Time, genre, etc which includes field values created by customized user input – especially mood
- In the mood is based on keywords associated with a song. Therefore, each song must be individually labelled with a particular mood.
Positive of Smart Playlist
There are qualities of Smart Playlist which make it a successful algorithm, including:
- Creation and composition of multiple playlists. N x N relationships can be created among playlists. In other words, subsets of playlists can be composed to create a new smart playlist.
Note that additional research is needed into the Smart Playlist aswell as alternative solutions in order to leverage and incorporate successful solutions.
Note that algorithm must be able to determine songs that are enjoyable in a series, for instance, I like to hear all Phil Collins track. It’s my Phil Collins mood.