<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[schwt的博客]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/schwt</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; schwt]]></copyright><item><title><![CDATA[matplotlib 画折线图]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/103593435</link><guid>https://blog.youkuaiyun.com/schwt/article/details/103593435</guid><author>schwt</author><pubDate>Wed, 18 Dec 2019 10:44:28 +0800</pubDate><description><![CDATA[有三组数据，共用x轴。两组量纲相同，使用左侧纵轴刻度，第三组量纲不同，使用右侧Y轴刻度。
import numpy as np
from numpy import genfromtxt
import matplotlib.pyplot as plt

f_avg = 'loss_avg.txt'
f_los = 'loss_verify.txt'
f_err = 'loss_error_verif...]]></description><category></category></item><item><title><![CDATA[数据直方图脚本]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/80424837</link><guid>https://blog.youkuaiyun.com/schwt/article/details/80424837</guid><author>schwt</author><pubDate>Wed, 23 May 2018 19:30:09 +0800</pubDate><description><![CDATA[#!/usr/bin/python
#encoding:utf-8
import sys 
from numpy import genfromtxt
import matplotlib.pyplot as plt 

nbins = 100
file_in  = sys.argv[1]
file_out = sys.argv[2]

data = genfromtxt(file_in)
print...]]></description><category></category></item><item><title><![CDATA[bi-graph with Map-Reduce implementation]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/80268281</link><guid>https://blog.youkuaiyun.com/schwt/article/details/80268281</guid><author>schwt</author><pubDate>Thu, 10 May 2018 15:17:01 +0800</pubDate><description><![CDATA[formular:

Sij=1kρj∑uruirujkλud(tui,tuj)Sij=1kjρ∑uruirujkuλd(tui,tuj) S_{ij} = \frac 1 {k_j^\rho} \sum_u \frac {r_{ui} r_{uj}} {k_u^\lambda} d(t_{ui}, t_{uj}) 
d(t1,t2)=exp[−(t1−t2)22τ2]d(t1,t2)=exp⁡[...]]></description><category></category></item><item><title><![CDATA[ctr预估的负采样比率修正公式]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/79283556</link><guid>https://blog.youkuaiyun.com/schwt/article/details/79283556</guid><author>schwt</author><pubDate>Wed, 07 Feb 2018 19:32:53 +0800</pubDate><description><![CDATA[p=c1p′−1+cp=c1p′−1+c
p = \frac {c}{\frac 1 {p'} -1 +c}



c∼(0,1]c∼(0,1]c \sim (0,1]： 负样本采样比例。如果正负样本都采样，采样比分别为c1,c2c1,c2c_1, c_2, 则 c=c2/c1c=c2/c1c = c_2/c_1
p′p′p'：使用有采样的样本预估的ctr
ppp：修正ctr（理论真实值）


特...]]></description><category></category></item><item><title><![CDATA[常用值域变换公式]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/78662863</link><guid>https://blog.youkuaiyun.com/schwt/article/details/78662863</guid><author>schwt</author><pubDate>Wed, 29 Nov 2017 10:55:58 +0800</pubDate><description><![CDATA[小结，持续更新1. 线性伸展]]></description><category></category></item><item><title><![CDATA[word2vec 核心简记]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/73692004</link><guid>https://blog.youkuaiyun.com/schwt/article/details/73692004</guid><author>schwt</author><pubDate>Sat, 24 Jun 2017 14:59:46 +0800</pubDate><description><![CDATA[word2vec 核心间记
Huffman编码：
  用词频作为词权重，构造 Huffman 树生成的 Huffman 编码，满足语料库总编码长度的期望最短。（频率高的编码短，频率低的编码长）n-gram:
  (n-1)阶Markov假设：词概率只与它前面 n-1 个词有关。复杂度 O(Nn)O(N^n)。需要平滑（类似 Bayes) 
  n=1n=1即是 unigram2 Hierarchic]]></description><category></category></item><item><title><![CDATA[boosting 方法]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/49762945</link><guid>https://blog.youkuaiyun.com/schwt/article/details/49762945</guid><author>schwt</author><pubDate>Tue, 10 Nov 2015 17:41:21 +0800</pubDate><description><![CDATA[boosting方法]]></description><category></category></item><item><title><![CDATA[排序比较指标]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/49467083</link><guid>https://blog.youkuaiyun.com/schwt/article/details/49467083</guid><author>schwt</author><pubDate>Wed, 28 Oct 2015 12:34:06 +0800</pubDate><description><![CDATA[衡量两个次序的差异的指标



1. Fitness

F=1Z&#x2211;jwj&#x03B1;(|pi&#x2212;qj|+1)+(1&#x2212;&#x03B1;)pi" role="presentation">F=1Z∑jwjα(|pi−qj|+1)+(1−α)piF=1Z∑jwjα(|pi−qj|+1)+(1−α)pi F = \frac{1}{Z} \sum_j \frac{w]]></description><category></category></item><item><title><![CDATA[FTRL 笔记]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/49466939</link><guid>https://blog.youkuaiyun.com/schwt/article/details/49466939</guid><author>schwt</author><pubDate>Wed, 28 Oct 2015 12:05:50 +0800</pubDate><description><![CDATA[这篇笔记主要参考冯杨的五篇博客：在线最优化求解(Online Optimization)。因为对于在线学习方法，稀疏性问题需要特别关注：每次在线学习一个新 instance 的时候，优化方向并不一定是全局最优，不容易产生稀疏解，而简单截断又可能将从全局看不该稀疏掉的特征变为零。所以这里以 L1 正则为基础，比较几种在线学习算法。0，预备每个 instance 由特征向量和预测目标组成: (x,y)(]]></description><category></category></item><item><title><![CDATA[RBM (受限玻尔兹曼机)]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/49466879</link><guid>https://blog.youkuaiyun.com/schwt/article/details/49466879</guid><author>schwt</author><pubDate>Wed, 28 Oct 2015 11:59:11 +0800</pubDate><description><![CDATA[1. 玻尔兹曼分布：p(E)∼e−E/kTp(E) \thicksim e^{-E/kT} 2. RBM两层：隐层和]]></description><category></category></item><item><title><![CDATA[Logistic Regression]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/49466837</link><guid>https://blog.youkuaiyun.com/schwt/article/details/49466837</guid><author>schwt</author><pubDate>Wed, 28 Oct 2015 11:53:48 +0800</pubDate><description><![CDATA[一、说明一个最常见的分类算法.
mm: 训练集个数
nn: 特征个数 
训练集：    {(x(1),y(1)),...,(x(i),y(i)),...,(x(m),y(m))}\{(\mathbf x^{(1)}, y^{(1)}), ..., (\mathbf x^{(i)}, y^{(i)}), ..., (\mathbf x^{(m)}, y^{(m)})\}.
x(i)∈Rn,  y(i)]]></description><category></category></item><item><title><![CDATA[SVD 的理论基础]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/49466617</link><guid>https://blog.youkuaiyun.com/schwt/article/details/49466617</guid><author>schwt</author><pubDate>Wed, 28 Oct 2015 11:36:44 +0800</pubDate><description><![CDATA[本文从理论基础的角度梳理对SVD的理解，不涉及SVD的实现算法。]]></description><category></category></item><item><title><![CDATA[RankNet]]></title><link>https://blog.youkuaiyun.com/schwt/article/details/48794365</link><guid>https://blog.youkuaiyun.com/schwt/article/details/48794365</guid><author>schwt</author><pubDate>Tue, 29 Sep 2015 00:20:24 +0800</pubDate><description><![CDATA[RankNetRankNet 论文的笔记。原文： 
Learning to rank using gradient descent.]]></description><category></category></item></channel></rss>