tushushu / imylu Goto Github PK

View Code? Open in Web Editor NEW

443.0 17.0 258.0 8.38 MB

Pure Python implementation of machine learning algorithms

License: Apache License 2.0

Python 100.00%

imylu's Introduction

What' new?

Unlike Imylu 0.1, Imylu 0.2 will be based on numpy and scipy to make the code simple and efficient.

What does imylu do?

Most of popular machine learning algorithms have been implemented by pure Python code in the project, which is specially recommended for people who would like to learn the algorithm details by reading Python code rather than lots of mathematical formulas. However, the necessary mathematical formulas and derivations are included in the code comments.

Imylu is compatible with: Python 3.6-3.7.

Why this name, imylu?

Chinese used to pronounce ML as [aimu'ailu], and the name imylu was inspired by this.

Folders guide

imylu's People

Contributors

Stargazers

Watchers

Forkers

arfu2016 joan-guo pandasxx gavinju dukeyuan sportzhang maliang668 uid00000000 411324590 lijia-xing pigts chrisliu007 dingxg lilitom lu839684437 concenterate xrzlizheng lzjtt2017 yuebin1009 wangshubing1 zain1iu peterboyyy dago0 846401882 laurelhhq 5up3rc futurebody onlyprogram fengzhi19940518 chaoyue729 jrxgugi chenxushu1025 mao-tool da-boy lijinchao-fork jasonwu1357002 muniuliuma xingyingao githappyboy alvinwangjunqiang helianghit jenniferxq 726663676 cornercao leerelive saki-xiong 10444005404 simonmenn zwleagle awalkinclouds zhuya1996 foosynaptic delaiahz mengyixin oceanlight cht649765744 lufenggui lovelyanyan qiaojuliu qianrenjian allenmien goldenbinxin zhang45 lukealee lushers huang027 ktt0522 hello-zzy zhangziquan chookg xinihe ant-dong newsong12 liu1695347 kepmov wangyn137 cowcomic jxliao13 carpdiemliang wurentidai sunjiamin chuangzhang tanmaohong modelerguanxusu liujun26 xyang65 watkyns duyining 2322349031 apple006 jamesdu0504 sanmisanfan liuriver123 mazelinux sarah-wy jacktank carry-ji a957994193 2015tdlp chenyang233

imylu's Issues

There seems to be no missing value process

Thank you for your sharing.
However, when in the true development environments, there are always having missing value.
It seems no missing value process for each algorithm.

_get_split_mse 函数计算

_get_split_mse 中计算MSE的话 split_sqr_sum[0] - split_sum[0] * split_avg[0] 是不是应该为split_sqr_sum[0] - split_avg[0] * split_avg[0] 因为你是Y的实际值-Y的均值的平方

矩阵分解, 相乘出现负值问题探讨

博主你好，谢谢你分享代码。代码特别适合学习，书写整洁。
我是通过ALS搜索到您的代码，借此我想请教你一个问题。

ALS分解出来的两个矩阵，相乘之后，出现了很多负值，这样的负值怎么理解呢。举例如下：

userId,movieId,rating,timestamp
1,2,4.5,964982703
1,3,2.0,964981247
2,1,4.0,964982224
2,3,3.5,964983815
3,2,5.0,964982931
3,4,2.0,964982400
4,2,3.5,964980868
4,3,4.0,964982176
4,4,1.0,964984041

对应于评分矩阵
{{0,   4.5,  2,   0}, 
 {4.0, 0,  3.5,  0},
 {0,   5.0,  0,   2.0}, 
 {0,   3.5, 4.0, 1.0}}

通过你的代码计算结果如下

model = ALS()
model.fit(X, k=2, max_iter=10)

## ===>

user_matrix = 
   [[0.7570282336382094, 0.03844973056986965, 0.8341276635923996, 0.7159946957782793],
    [0.2987763898981603, 1.2236650397020115, -0.09214284357731845, 0.661286003351013]]

item_matrix = 
   [[-0.9352925817978953, 5.779445419172115, 1.301780265376408, 1.4251036233547738], 
    [2.7179214250950294, -0.33085416701657533, 3.272648111651669, -0.2344265756021138]]

user_matrix.transpose x item_matrix = 
[[0.10400786028337483, 4.276351943480331, 1.9632744030892986, 1.0088025527850888],
 [3.2898636807717296, -0.1826365582074784, 4.054678181939848, -0.23206475458923115],
 [-1.0305904247583555, 4.851281148112146, 0.7842998282335908, 1.2103190870120526], 
 [1.1276588690551161, 3.919263034868889, 3.0962241552067207, 0.865343621997239]]

我通过Mathematica编程也计算出同样的结果

R = {{0, 4.5, 2, 0}, {4.0, 0, 3.5, 0}, {0, 5.0, 0, 2.0}, {0, 3.5, 4.0,1.0}};
X = RandomReal[1, {4, 2}]
(* {{0.242376,0.511595}, {0.661005,0.105123}, {0.00835057,0.137917}, {0.980405,0.334949}} *)
Do[
   (* compute Y *)
   Xt = Transpose[X];
   Y = Transpose[LinearSolve[Xt.X, Xt.R]];
   (* compute X *)
   Yt = Transpose[Y];
   X = Transpose[LinearSolve[Yt.Y, Yt.Transpose[R]]],
   {i, 10}
]

很容易看到两个矩阵相乘出现了负值, 怎么理解这样的情况？谢谢

ALS get_rmse 函数计算问题

n_elements = sum(map(len, ratings.values()))
elements 数目偏大，在你的矩阵计算过程中被赋予的0 的默认值
导致ratings 的规模是m*n ,使得RMSE 的值偏小

imylu/imylu/utils/load_data.py/

def gen_data(low, high, n_rows, n_cols=None):

Returns:       
        list -- 1d or 2d list with int

but in imylu/imylu/utils/preprocessing.py/def min_max_scale(X): X is ndarray .

and you can see imylu/examples/kd_tree_example.py/ 's line 51~line 53

X = gen_data(low, high, n_rows, n_cols)        
y = gen_data(low, high, n_rows)        
Xi = gen_data(low, high, n_cols)

I think just add a line in def gen_data():

def gen_data(low, high, n_rows, n_cols=None):
    """Generate dataset randomly.
    Arguments:
        low {int} -- The minimum value of element generated.
        high {int} -- The maximum value of element generated.
        n_rows {int} -- Number of rows.
        n_cols {int} -- Number of columns.
    Returns:
        list -- 1d or 2d list with int
    """
    if n_cols is None:
        ret = [randint(low, high) for _ in range(n_rows)]
    else:
        ret = [[randint(low, high) for _ in range(n_cols)]
               for _ in range(n_rows)]
    ret = array(ret)  # This is my add.
    return ret

Maybe I ignore something in somewhere, thanks!