Interpret clustering results

[xurxodiz]

After picking the clustering the cluster algorithm (#23) and implementing it (#24), but before hardcoding the transitions (#26), we need to interpret the results so we can assign them values that make sense. What does each cluster represent, according to their disparity in the values?

[xurxodiz]

For example, if we follow through with using EM as our clustering algorithm, we get three different clusters: 0, 1 and 2. Analyzing the means:

• Cluster 0 has:
  • the lowest: timesofDeathByGoomba, totalTimeLittleMode.
  • the highest: totalCoinBlocks, totalCoins, totalEnemies.
• Cluster 1 has:
  • the lowest: timeSpentRunning.
  • the highest: GoombasKilled, aimlessJumps, coinBlocksDestroyed, coinsCollected, completionTime, duckNumber, emptyBlocksDestroyed, enemyKillByFire, jumpsNumber, percentageBlocksDestroyed, percentageCoinBlocksDestroyed, percentageEmptyBlocksDestroyed, percentagePowerBlockDestroyed, timeRunningLeft, timeRunningRight, timeSpentDucking, timesPressedRun, timesSwitchingPower, totalEmptyBlocks, totalTime, totalTimeFireMode, totalTimeLargeMode, totalTimeLittleMode, totalPowerBlocks.
• Cluster 2 has:
  • the lowest: GoombasKilled, aimlessJumps, coinBlocksDestroyed, coinsCollected, completionTime, duckNumber, emptyBlocksDestroyed, enemyKillByFire, jumpsNumber, percentageBlocksDestroyed, percentageCoinBlocksDestroyed, percentageEmptyBlocksDestroyed, percentagePowerBlockDestroyed, timeRunningLeft, timeRunningRight, timeSpentDucking, timesPressedRun, timesSwitchingPower, totalCoinBlocks, totalCoins, totalEmptyBlocks, totalEnemies, totalTime, totalTimeFireMode, totalTimeLargeMode, totalPowerBlocks.
  • the highest: timeSpentRunning, timesofDeathByGoomba.

[xurxodiz]

I think it's safe to conclude, then, that cluster 1 can be nicknamed explorer (it prefers a slower playthrough and careful examination of all objects) and cluster 2 speeder (runs all the time, doesn't stop for anything). Cluster 0 would be a intermediate position.

The odd (or unexpected) higher values of totalCoins in the intermediate cluster versus the explorer can be explained by a higher standard deviation. Note that that can't be said of totalEnemies -- intermediate players do play more enemies than explorers.

Refer to commit 998da7d651235469153ce7a77da93f6b74adb735 for the full EM cluster results.

[xurxodiz]

For future reference and reflection: could we take advantage of an apparent spectrum ranging from speeders to explorers, and position all players at some point along that line?

[xurxodiz]

They are interpreted enough for the paper. Moving it to thesis milestone.

[xurxodiz]

I think I have a pretty clear idea what each profile likes. Closing,