mirror of
https://github.com/kuhyx/WUT_Computer_Science.git
synced 2026-07-04 22:23:11 +02:00
Krzysztof kuhy Rudnicki
be962d0c35
git-subtree-dir: Programming/EARIN git-subtree-mainline:635e287095git-subtree-split:a09c96dd65
39 lines
1.5 KiB
Plaintext
39 lines
1.5 KiB
Plaintext
Parameters:
|
|
- datalimit (usable between 500k and max) [max = 109,224,747 ]
|
|
- seed (very important make sure it stays the same through all testing [maybe just 42?])
|
|
- metric (either cosine, mahalanobis or euclidean as in preliminary report)
|
|
- NN algorithm (either auto, ball_tree, kd_tree, brute)
|
|
- neighbors - number of nearest neigbors
|
|
- User threshold - minimal numbers of votes for user to be included in data
|
|
- Anime threshold - same for anime
|
|
|
|
|
|
These are 6 parameters that influence program behaviour and 1 parameter for seed
|
|
Probably would do simulations for 3 variants of each parameters (excluding seed), rest will be default
|
|
|
|
so in total 6 * 3 = 18 simulations
|
|
|
|
Default values:
|
|
Datalimit: all of data
|
|
Seed: 42
|
|
Metric: cosine
|
|
NN algorithm: brute
|
|
Neighbors: 5
|
|
User threshold: 500
|
|
Anime threshold: 200
|
|
|
|
Neighbors number count:
|
|
k = 3-5: default starting points for small-medium dataset
|
|
k = sqrt(n): rule of thumb, n is number of instances in dataset (balanced between underfitting and overfitting)
|
|
l = n / 2: look at half of dataset for each prediction
|
|
k = log(n): for very large datasets
|
|
k = n - 1: Use all data except one, will probably overgenarlize the model
|
|
|
|
|
|
Values spread:
|
|
Datalimit: [27306186, 54612373, 109224747] (max on the right, then halved and halved)
|
|
Metric: ["cosine", "mahalanobis", "euclidean"]
|
|
NN algorithm: ['auto', 'ball_tree', 'kd_tree', 'brute']
|
|
neighbors: [5, sqrt(n), n / 2, log(n), n - 1]
|
|
User threshold: [0, 500, 1000]
|
|
Anime threshold: [0, 200, 500] |