<?xml version="1.0" encoding="utf-8" ?><rss version="2.0"><channel><title><![CDATA[casmaster的专栏]]></title><description><![CDATA[]]></description><link>https://blog.youkuaiyun.com/casmaster</link><language>zh-cn</language><generator>https://blog.youkuaiyun.com/</generator><copyright><![CDATA[Copyright &copy; casmaster]]></copyright><item><title><![CDATA[[论文笔记]Modeling Interestingness with Deep Neural Networks]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/76039047</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/76039047</guid><author>casmaster</author><pubDate>Mon, 24 Jul 2017 22:33:18 +0800</pubDate><description><![CDATA[Abstract目的:根据用户正在阅读的doc,推荐相近的doc 
数据:浏览器数据 
训练:将source和target doc投影到隐含空间,使得他们的距离最小 
展示有效性的两个兴趣度任务:自动高亮,上下文实体搜索1.Introduction兴趣度很重要的两个任务:]]></description><category></category></item><item><title><![CDATA[[论文笔记]Learning Deep Structured Semantic Models for Web Search using Clickthrough Data]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/76037883</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/76037883</guid><author>casmaster</author><pubDate>Mon, 24 Jul 2017 22:16:04 +0800</pubDate><description><![CDATA[[论文笔记]Learning Deep Structured Semantic Models for Web Search using Clickthrough DataAbstract
DSSM是一个判别模型.
训练方式:极大化给定query条件下,被点击的文档的条件似然(样本是clickthrough data).
为了能够用于大规模搜索应用,使用了word hash技术
Introductio]]></description><category></category></item><item><title><![CDATA[[阅读笔记]Introduction to Boosted Trees]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/76020283</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/76020283</guid><author>casmaster</author><pubDate>Mon, 24 Jul 2017 15:06:53 +0800</pubDate><description><![CDATA[概念(要学什么)回归树(CART)回归树emsemble模型$$y_i = \sum_{k=1}^{K}f_k(x_i) , f \in F$ $参数GB(如何学习)additve training(boosting)L(t)=∑i=1nl(yi,ŷ (t−1)i+ft(xi))+Ω(ft)L^{(t)}=\sum_{i=1}^n l(y_i, \hat{y}_{i}^{(t-1)} + f_t(]]></description><category></category></item><item><title><![CDATA[[阅读笔记]AutoRec: Autoencoders Meet Collaborative Filtering]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/76014619</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/76014619</guid><author>casmaster</author><pubDate>Mon, 24 Jul 2017 12:10:47 +0800</pubDate><description><![CDATA[introductionAE之前成功应用在图像和语音上,现在迁移到推荐上模型传统的AE 
minθ∑r||r−h(r;θ)||22\min_{\theta}\sum_{r}||r-h(r;\theta)||_2^2AutoRec在传统AE的基础上做了如下变化 
- 损失函数只和观察到的元素有关 
- 加上正则化项AutoRec和RBM-CF的区别 
- RBM是生成式模型 
- RBM通过极大化对数]]></description><category></category></item><item><title><![CDATA[[公开课笔记]线性autoencoder和FM的关系]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75949783</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75949783</guid><author>casmaster</author><pubDate>Mon, 24 Jul 2017 01:13:35 +0800</pubDate><description><![CDATA[autoencoder概念weights:是一种特征变换.也就是”编码” 
好的weights:information preserving encoding (保留原始信息的编码) 
information preserving: 编码后,可以精确的解码 
结论:pre-traing目标要是information preserving encoding基本的autoencoder:  d−d′−d]]></description><category></category></item><item><title><![CDATA[[读书笔记]机器学习-11章-特征选择和稀疏学习]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75666326</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75666326</guid><author>casmaster</author><pubDate>Mon, 24 Jul 2017 00:33:14 +0800</pubDate><description><![CDATA[11 特征选择和稀疏学习11.1 子集搜索和评价定义:特征很多,选出相关的特征的过程 
原因: 
- 维度灾难(和10章讲的降维有关系) 
- 降低学习难度问题的难点:组合爆炸. 
解决的两个环节:如何评价子集好坏一个特征划分样本和正本真实划分约接近,特征效果越好. 
书中的例子是使用信息熵增益 
gain(A)=Ent(D)−∑|Di||D|Ent(Di)gain(A)=Ent(D) - \sum]]></description><category></category></item><item><title><![CDATA[[论文笔记]Adaptive Subgradient Methods for Online Learning and Stochastic Optimization]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75660198</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75660198</guid><author>casmaster</author><pubDate>Fri, 21 Jul 2017 16:28:55 +0800</pubDate><description><![CDATA[Abstract作者提出一个新的次梯度方法家族,可以动态地吸收之前的看过的数据的信息,来进行更informative的基于梯度的学习.它可以找到很有信息的特征,即使非常罕见.1.Introduction背景:在很多在线学习中,样本维度很高,但是只有少量非零特征.罕见的特征经常很有信息量.前人基于这点,做了很多工作,比如TF-IDF. 
直觉:高频学习率低,低频学习率高.高速模型,低频时间要”注意”.]]></description><category></category></item><item><title><![CDATA[virtualenv中使用jupyter的问题]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75578511</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75578511</guid><author>casmaster</author><pubDate>Thu, 20 Jul 2017 21:50:57 +0800</pubDate><description><![CDATA[把这个帖子中,对sys.path的操作注释掉就可以]]></description><category></category></item><item><title><![CDATA[[阅读笔记]An overview of gradient descent optimization algorithms]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75578506</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75578506</guid><author>casmaster</author><pubDate>Thu, 20 Jul 2017 21:50:06 +0800</pubDate><description><![CDATA[introdction个人理解,梯度下降的元素 
- 梯度方向 
- 步长梯度下降的种类根据计算梯度时候,用的数据量的不同,分为三类.数据量不同,决定了准确度和耗时的trade-off批量gd随机gdminibatch gd吸取了两者的优点挑战mini-batch gd 不能保证收敛,而且提出了一些挑战
很难选择学习率
第二点没懂
所有参数的学习率都相同.如果数据很稀疏,并且特征出现频率差异很大,我]]></description><category></category></item><item><title><![CDATA[[阅读笔记]Programming Models for Deep Learning]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75575086</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75575086</guid><author>casmaster</author><pubDate>Thu, 20 Jul 2017 19:22:00 +0800</pubDate><description><![CDATA[原文:Programming Models for Deep Learning 
这篇文章的主题是深度学习的编程模型. 
现在有很多深度学习库.这篇文章编程模型来分析的优缺点.符号式 vs 命令式 编程不同点在于符号式编程,C=B∗AC = B * A执行的时候,并没有真正计算,而是生成一个计算图命令式编程更加灵活本事原因在于,命令式编程更加native符号式编程效率更高效率包括内存和运行效率 
命]]></description><category></category></item><item><title><![CDATA[difacto源码分析]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75574905</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75574905</guid><author>casmaster</author><pubDate>Thu, 20 Jul 2017 18:55:28 +0800</pubDate><description><![CDATA[async_sgd.hclass AsyncScheduler  : public solver::MinibatchSchedulerstruct ISGDHandlestruct AdaGradEntrystruct AdaGradHandle : public ISGDHandleclass AsyncServer : public solver::MinibatchServer包含ps::K]]></description><category></category></item><item><title><![CDATA[逻辑回归系数的意义]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75449035</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75449035</guid><author>casmaster</author><pubDate>Wed, 19 Jul 2017 20:33:30 +0800</pubDate><description><![CDATA[基础概念odd发生比 
odd=p(1−p)odd=\frac p {(1-p)} 
和概率的关系 
p=odd1+oddp=\frac {odd} {1+odd} 
logit: 
logit=log(odd)logit=log(odd) 
对于逻辑回归来说: 
logit=w∗xlogit=w*x逻辑回归系数的意义一个简单的例子.比如预测男女喜欢打篮球的概率,唯一的特征是gender,取值0或1]]></description><category></category></item><item><title><![CDATA[11 条件随机场]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75330841</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75330841</guid><author>casmaster</author><pubDate>Tue, 18 Jul 2017 21:20:34 +0800</pubDate><description><![CDATA[概率无向图模型(马尔科夫随机场)又叫马尔科夫随机场,是一种概率无向图模型. 
HMM和马尔科夫随机场都是对联合分布进行建模,所以都是生成式模型.条件随机场对条件概率建模,所以是判别模型.基本概念团:无向图中节点的一个子集,如果其中任意两个节点都有边连接,那就成为一个团 
极大团:不能被其他团包含的团.每个节点都至少在一个极大团中. 
分离集:如果节点集A和节点集B被节点集C分离,那么C被称为分离集.]]></description><category></category></item><item><title><![CDATA[统计学习方法_读书笔记_10章_HMM]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75094196</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75094196</guid><author>casmaster</author><pubDate>Fri, 14 Jul 2017 15:11:53 +0800</pubDate><description><![CDATA[HMM是一种生成模型.描述了隐含的马尔科夫链随机生成观测序列的过程. 
所以HMM=隐含的马尔科夫链->随机生成的观测序列. 
更具体的说,HMM=隐含的马尔科夫链->随机生成状态序列->随机生成观测序列基本概念定义λ=(A,B,π)\lambda=(A,B,\pi) 
我理解,状态序列和观测序列都是离散可枚举的. 
感觉中间也有隐变量.那它和EM算法可以找到什么区别呢?其中一个是,HMM是和顺序有]]></description><category></category></item><item><title><![CDATA[difacto中的loss.h]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75068924</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75068924</guid><author>casmaster</author><pubDate>Thu, 13 Jul 2017 15:28:31 +0800</pubDate><description><![CDATA[/**
 * \brief the loss function
 */
template <typename T>
class Loss {
 public:
  /**
   * create and init the loss function
   *
   * @param data X and Y
   * @param model w and V
   * @param model_si]]></description><category></category></item><item><title><![CDATA[dmlc中的data.h]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75051098</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75051098</guid><author>casmaster</author><pubDate>Thu, 13 Jul 2017 13:51:11 +0800</pubDate><description><![CDATA[/*!
 * \brief a block of data, containing several rows in sparse matrix
 *  This is useful for (streaming-sxtyle) algorithms that scans through rows of data
 *  examples include: SGD, GD, L-BFGS, kmean]]></description><category></category></item><item><title><![CDATA[最大熵模型的理解]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75043401</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75043401</guid><author>casmaster</author><pubDate>Thu, 13 Jul 2017 09:24:28 +0800</pubDate><description><![CDATA[最大熵模型是理解机器学习的重要一环.最大熵原理最大熵模型的内核最大熵原理.最大熵原理就是,在满足某些已知约束情况下,对于未知情况,不做任何假设,就是未知那部分事情,都是等可能的.因为不好量化”等可能”,所以用熵来衡量.最大熵模型]]></description><category></category></item><item><title><![CDATA[机器学习的损失函数]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75041707</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75041707</guid><author>casmaster</author><pubDate>Wed, 12 Jul 2017 22:41:36 +0800</pubDate><description><![CDATA[李航的统计机器学习方法说: 
统计学习方法三要素:模型,策略,算法 
模型是假设,策略就是验证这个假设好坏的方法,算法是知道了假设好坏程度,怎么让假设向好的方向演进.所以需要定量地知道模型的好坏程度.模型好坏的理解如何定义模型的好坏呢.比如拿来一堆真实数据(Xi,yi)(X_i,y_i),然后一个模型根据XiX_i预测yi^\hat{y_i},显然这个yi^\hat{y_i}不一定和真实的yiy_i]]></description><category></category></item><item><title><![CDATA[因式分解梳理]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75041392</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75041392</guid><author>casmaster</author><pubDate>Wed, 12 Jul 2017 21:30:27 +0800</pubDate><description><![CDATA[因式分解在推荐系统中广泛应用.最近看了一些因式分解的论文,梳理一下它的发展历程.关系SVD -> biasSVD -> Asymmetric-SVD -> SVD++ -> FM 
SVD+用户的隐式反馈 = SVD++
SVD
biasSVD
Asymmetric-SVD
论文
  Koren, Yehuda. “Factorization meets the neighborhood: a mu]]></description><category></category></item><item><title><![CDATA[信息论笔记]]></title><link>https://blog.youkuaiyun.com/casmaster/article/details/75008431</link><guid>https://blog.youkuaiyun.com/casmaster/article/details/75008431</guid><author>casmaster</author><pubDate>Wed, 12 Jul 2017 13:55:04 +0800</pubDate><description><![CDATA[与机器学习相关的信息论概念基本概念信息量(自信息)I(p)=−logb(p)I(p)=-log_{b}(p) 
当b=2,熵的单位是bit;b=e,熵的单位是nat;b=10,熵的单位是hart 
p=1,I(p)=0p=1,I(p)=0 
p→0,I(p)→∞p\to0 , I(p)\to\infty 
p=1/2,I(p)=1p=1/2, I(p) = 1 熵H(X)=EX[I(X)]=EX[−]]></description><category></category></item></channel></rss>