#!/usr/bin/env python
import urllib2
import numpy as np
from keras.models import Sequential
from keras.layers import Dense
from keras.wrappers.scikit_learn import KerasRegressor
from sklearn.model_selection import cross_val_score
from sklearn.model_selection import KFold
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
def load_data():
    X = []
    Y = []
    data_url = 'https://archive.ics.uci.edu/ml/machine-learning-databases/housing/housing.data'
    for line in urllib2.urlopen(data_url).readlines():
        line = map(float, line.split())
        X.append(line[0:13])
        Y.append(line[13])
    return X, Y
def basic_model():
    # create model
    model = Sequential()
    model.add(Dense(13, input_dim=13, kernel_initializer='normal', activation='relu'))
    model.add(Dense(1, kernel_initializer='normal'))
    # compile model
    model.compile(loss='mean_squared_error', optimizer='adam')
    return model
d

逾期天数（days past due, DPD）

已逾契约书约定缴款日的延滞天数，贷放型产品自缴款截止日（通常为次一关账日）后第一天开始计算；信用卡比较特别，虽然缴款截止日约为关账日后20天，但逾期天数亦是由次一关账日后起算。

逾期期数（bucket）

逾期1期称为M1，2期称为M2，3期称为M3……以此类推。信用卡缴款截止日与次一关账日之间虽然不计算逾期天数，但其bucket称为M0。
注意，因为每月天数不一定相同，所以各期的长短会有不同。

逾期阶段（stage）

依bucket可分为前期（front end）、中期（middle range）、后期（hot core）、转呆账（write-off）。
stage的划分方式并无硬性规定，可依各银行的催收策略、转呆账政策与产品特性决定。以信用卡为例，一般将M1列为前期，M2-M3列为中期，M4以上列为后期，若已转列呆账者则列入转呆账。

即期指标（coincidental）

即期指标 = 当期各bucket延滞金额 / 当期应收账款

即期指标是计算延滞率时常用的两种方法之一，其概念为分析当期应收帐款的质量结构。一般公开信息所显示的延滞率，若无特别注明，皆是以coincidental的概念计算的。

递延指标（lagged）

递延指标 = 当期各bucket延滞金额 / 各bucket对应的历史月份应收帐款

即期指标的分母一律是当期应收账款，不过其分子实际上是由之前的应收账款产生的；因为为了回溯逾期起源，递延指标将分母改成了相对应的之前月份的应收账款。

期末结算（cycle end）

期末结算为信用卡特有的结算方式。因为信用卡客群最为庞大，作业处理相当耗时，许多银行会将其客户划分至不同账务周期（cycle），因此信用卡产品下通常有多个关账日。
银行必须就各个cycle客户分别管理，尤其是账务及催收单位皆以cycle为作业周期。

月底结算（month end）

月底结算报表主要表达各月月底结算数据，适用于消费金融所有产品，尤其在跨产品并列分析时，为实现资料切点一致，多采用月底结算数据。

参考：《互联网金融时代：消费信贷评分建模与应用》

• Date functions

-- change date format
from_unixtime(unix_timestamp('20150101' ,'yyyyMMdd'), 'yyyy-MM-dd')
-- add n days
date_add('2015-11-01', 30)  -- will return '2015-12-01'
-- calculate date difference
datediff('2015-12-01', '2015-11-01') -- will return 30

• Generate row number

row_number() over (DISTRIBUTE BY... SORT BY... DESC)

• Get partition information

analyze table xxx.yyy partition(dt = '2015-12-11') compute statistics;
describe formated xxx.yyy partition (dt = '2015-12-11');

Euclidean distance

from sklearn.metrics.pairwise import euclidean_distances
euclidean_distances([[1,2,3], [100,200,300]])
# return:
# array([[  0.        , 370.42408129],
#        [370.42408129,   0.        ]])

Cosine similarity

from sklearn.metrics.pairwise import cosine_similarity
cosine_similarity([[1,2,3],[100,200,300]])
# return:
# array([[1., 1.],
#        [1., 1.]])

Pearson correlation

from scipy.stats.stats import pearsonr
pearsonr([1,2,3], [100,200,300])
# return ('1.0', 0.0) // (Pearson’s correlation coefficient, 2-tailed p-value)

Install

apt-get install python-pip
pip install shadowsocks

Setup

Create config file /etc/shadowsocks.json:

{
    "server":"your_ip_address",
    "server_port":8388,
    "local_address": "127.0.0.1",
    "local_port":1080,
    "password":"your_password",
    "timeout":300,
    "method":"aes-256-cfb",
    "fast_open": false
}

You can set multiple ports in the config file:

{
    "server": "your_ip_address",
    "local_address": "127.0.0.1",
    "local_port": "1080",
    "port_password": {
        "8381": "password_1",
        "8388": "password_2"
    },
    "timeout": 300,
    "method": "aes-256-cfb"
}

Start

ssserver -c /etc/shadowsocks.json
# run at background
ssserver -c /etc/shadowsocks.json -d start
ssserver -c /etc/shadowsocks.json -d stop

Start on boot

Edit /etc/rc.local:

/usr/local/bin/ssserver -c /etc/shadowsocks.json -d start
exit 0

#!/usr/bin/env python
# -*- coding: utf-8 -*-
import urllib2
from numpy import mat, ones, shape, exp, array, arange
import matplotlib.pyplot as plt
def createDataSet():
    features = []
    labels = []
    lines = urllib2.urlopen('https://raw.github.com/pbharrin/machinelearninginaction/master/Ch05/testSet.txt').readlines()
    for line in lines:
        line = line.strip().split()
        features.append([1.0, float(line[0]), float(line[1])]) # set x0 to 1.0
        labels.append(int(line[2]))
    return features, labels
def sigmoid(value):
    return 1.0 / (1 + exp(-value))
def gradAscent(features, labels, alpha=0.001, iterations=500):
    '''
    梯度上升算法：
    - 批处理算法：每次更新回归系数时都需要遍历整个数据集
    '''
    featureMatrix = mat(features)
    labelMatrix = mat(labels).transpose()
    m, n = shape(featureMatrix)
    weights = ones((n, 1))
    for k in range(iterations):
        h = sigmoid(featureMatrix*weights)
        error = (labelMatrix - h)
        weig

user-based collaborative filtering

1. for each user, find similar users by calculating similarity of the ratings (e.g. euclidean distance, pearson similarity)
2. for each item of the seleted users, calculate the weighted rating according to each user's similarity
3. select top n new items for this user

item-based collaborative filtering

1. for each item, calculate similarity of each other item
2. select top rating items of this user
3. for each selected item, find similar items and calculate the weighted rating according to each item's similarity
4. select top n new items for this user

user-based or item-based?

• item-based method needs to maintain the item similarity table
• for sparse dataset, item-based method is better
• for dense dataset, both methods have the similar performance

• Get absolute value

awk '{printf("%d",sqrt($1*$1))}' test.csv

• User defined function

echo "4 105" | awk 'function max(a,b){return a>b?a:b}{print max($1, $2)}'