From 8aa02590ee9e5d23fbbf4f6c347bb05a8d3e6952 Mon Sep 17 00:00:00 2001
From: shataowei <taowei.sha@ymm56.com>
Date: Fri, 29 Nov 2019 17:37:33 +0800
Subject: [PATCH] =?UTF-8?q?=E6=8E=A8=E8=8D=90=E9=83=A8=E5=88=86=E8=A1=A5?=
 =?UTF-8?q?=E5=85=85?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 .idea/workspace.xml                       | 159 +++++++++++++++++-----
 README.md                                 |  17 +++
 "\346\216\250\350\215\220/DeepFM.md"      |  25 ++++
 "\346\216\250\350\215\220/YouTubeNext.md" |  44 ++++++
 4 files changed, 210 insertions(+), 35 deletions(-)
 create mode 100644 "\346\216\250\350\215\220/DeepFM.md"
 create mode 100644 "\346\216\250\350\215\220/YouTubeNext.md"

diff --git a/.idea/workspace.xml b/.idea/workspace.xml
index 45968d9..7bd0319 100644
--- a/.idea/workspace.xml
+++ b/.idea/workspace.xml
@@ -2,8 +2,10 @@
 <project version="4">
   <component name="ChangeListManager">
     <list default="true" id="be5bf74e-ad1c-4197-8a1f-ac93566d8aa7" name="Default" comment="init">
+      <change type="NEW" beforePath="" afterPath="$PROJECT_DIR$/推荐/DeepFM.md" />
+      <change type="NEW" beforePath="" afterPath="$PROJECT_DIR$/推荐/YouTubeNext.md" />
+      <change type="MODIFICATION" beforePath="$PROJECT_DIR$/.idea/workspace.xml" afterPath="$PROJECT_DIR$/.idea/workspace.xml" />
       <change type="MODIFICATION" beforePath="$PROJECT_DIR$/README.md" afterPath="$PROJECT_DIR$/README.md" />
-      <change type="MODIFICATION" beforePath="$PROJECT_DIR$/深度学习/Attention.md" afterPath="$PROJECT_DIR$/深度学习/Attention.md" />
     </list>
     <ignored path="Reflection_Summary.iws" />
     <ignored path=".idea/workspace.xml" />
@@ -24,7 +26,7 @@
   </component>
   <component name="FileEditorManager">
     <leaf SIDE_TABS_SIZE_LIMIT_KEY="300">
-      <file leaf-file-name="README.md" pinned="false" current-in-tab="false">
+      <file leaf-file-name="README.md" pinned="false" current-in-tab="true">
         <entry file="file://$PROJECT_DIR$/README.md">
           <provider editor-type-id="split-provider[text-editor;MarkdownPreviewEditor]">
             <state split_layout="SPLIT">
@@ -39,8 +41,8 @@
           </provider>
           <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
             <state split_layout="SPLIT">
-              <first_editor relative-caret-position="264">
-                <caret line="297" column="36" selection-start-line="297" selection-start-column="36" selection-end-line="297" selection-end-column="36" />
+              <first_editor relative-caret-position="295">
+                <caret line="353" column="40" selection-start-line="353" selection-start-column="40" selection-end-line="353" selection-end-column="40" />
                 <folding />
               </first_editor>
               <second_editor />
@@ -48,7 +50,55 @@
           </provider>
         </entry>
       </file>
-      <file leaf-file-name="Attention.md" pinned="false" current-in-tab="true">
+      <file leaf-file-name="DeepFM.md" pinned="false" current-in-tab="false">
+        <entry file="file://$PROJECT_DIR$/推荐/DeepFM.md">
+          <provider editor-type-id="split-provider[text-editor;MarkdownPreviewEditor]">
+            <state split_layout="SPLIT">
+              <first_editor relative-caret-position="0">
+                <caret line="0" column="0" selection-start-line="0" selection-start-column="0" selection-end-line="0" selection-end-column="0" />
+                <folding />
+              </first_editor>
+              <second_editor>
+                <js_state />
+              </second_editor>
+            </state>
+          </provider>
+          <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
+            <state split_layout="SPLIT">
+              <first_editor relative-caret-position="375">
+                <caret line="25" column="0" selection-start-line="25" selection-start-column="0" selection-end-line="25" selection-end-column="0" />
+                <folding />
+              </first_editor>
+              <second_editor />
+            </state>
+          </provider>
+        </entry>
+      </file>
+      <file leaf-file-name="YouTubeNext.md" pinned="false" current-in-tab="false">
+        <entry file="file://$PROJECT_DIR$/推荐/YouTubeNext.md">
+          <provider editor-type-id="split-provider[text-editor;MarkdownPreviewEditor]">
+            <state split_layout="SPLIT">
+              <first_editor relative-caret-position="0">
+                <caret line="0" column="0" selection-start-line="0" selection-start-column="0" selection-end-line="0" selection-end-column="0" />
+                <folding />
+              </first_editor>
+              <second_editor>
+                <js_state />
+              </second_editor>
+            </state>
+          </provider>
+          <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
+            <state split_layout="SPLIT">
+              <first_editor relative-caret-position="195">
+                <caret line="13" column="22" selection-start-line="13" selection-start-column="22" selection-end-line="13" selection-end-column="22" />
+                <folding />
+              </first_editor>
+              <second_editor />
+            </state>
+          </provider>
+        </entry>
+      </file>
+      <file leaf-file-name="Attention.md" pinned="false" current-in-tab="false">
         <entry file="file://$PROJECT_DIR$/深度学习/Attention.md">
           <provider editor-type-id="split-provider[text-editor;MarkdownPreviewEditor]">
             <state split_layout="SPLIT">
@@ -63,8 +113,8 @@
           </provider>
           <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
             <state split_layout="SPLIT">
-              <first_editor relative-caret-position="540">
-                <caret line="36" column="0" selection-start-line="36" selection-start-column="0" selection-end-line="36" selection-end-column="0" />
+              <first_editor relative-caret-position="525">
+                <caret line="35" column="76" selection-start-line="35" selection-start-column="10" selection-end-line="35" selection-end-column="76" />
                 <folding />
               </first_editor>
               <second_editor />
@@ -120,8 +170,10 @@
         <option value="$PROJECT_DIR$/自然语言处理/LDA.md" />
         <option value="$PROJECT_DIR$/数学/牛顿法/牛顿迭代法求平方根.md" />
         <option value="$PROJECT_DIR$/自然语言处理/Word2Vec.md" />
-        <option value="$PROJECT_DIR$/README.md" />
         <option value="$PROJECT_DIR$/深度学习/Attention.md" />
+        <option value="$PROJECT_DIR$/推荐/YouTubeNext.md" />
+        <option value="$PROJECT_DIR$/推荐/DeepFM.md" />
+        <option value="$PROJECT_DIR$/README.md" />
       </list>
     </option>
   </component>
@@ -230,6 +282,20 @@
               <option name="myItemType" value="com.intellij.ide.projectView.impl.nodes.PsiDirectoryNode" />
             </PATH_ELEMENT>
           </PATH>
+          <PATH>
+            <PATH_ELEMENT>
+              <option name="myItemId" value="Reflection_Summary" />
+              <option name="myItemType" value="com.intellij.ide.projectView.impl.nodes.ProjectViewProjectNode" />
+            </PATH_ELEMENT>
+            <PATH_ELEMENT>
+              <option name="myItemId" value="Reflection_Summary" />
+              <option name="myItemType" value="com.intellij.ide.projectView.impl.nodes.PsiDirectoryNode" />
+            </PATH_ELEMENT>
+            <PATH_ELEMENT>
+              <option name="myItemId" value="推荐" />
+              <option name="myItemType" value="com.intellij.ide.projectView.impl.nodes.PsiDirectoryNode" />
+            </PATH_ELEMENT>
+          </PATH>
         </subPane>
       </pane>
       <pane id="Scope" />
@@ -601,27 +667,6 @@
     <watches-manager />
   </component>
   <component name="editorHistoryManager">
-    <entry file="file://$PROJECT_DIR$/数据预处理/特征提取/数据变换.md">
-      <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
-        <state split_layout="SPLIT">
-          <first_editor relative-caret-position="165">
-            <caret line="16" column="8" selection-start-line="16" selection-start-column="8" selection-end-line="16" selection-end-column="8" />
-          </first_editor>
-          <second_editor />
-        </state>
-      </provider>
-    </entry>
-    <entry file="file://$PROJECT_DIR$/README.md">
-      <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
-        <state split_layout="SPLIT">
-          <first_editor relative-caret-position="557">
-            <caret line="52" column="32" selection-start-line="52" selection-start-column="23" selection-end-line="52" selection-end-column="32" />
-            <folding />
-          </first_editor>
-          <second_editor />
-        </state>
-      </provider>
-    </entry>
     <entry file="file://$PROJECT_DIR$/README.md">
       <provider editor-type-id="split-provider[text-editor;markdown-preview-editor]">
         <state split_layout="SPLIT">
@@ -1192,7 +1237,7 @@
         </state>
       </provider>
     </entry>
-    <entry file="file://$PROJECT_DIR$/README.md">
+    <entry file="file://$PROJECT_DIR$/深度学习/Attention.md">
       <provider editor-type-id="split-provider[text-editor;MarkdownPreviewEditor]">
         <state split_layout="SPLIT">
           <first_editor relative-caret-position="0">
@@ -1206,15 +1251,59 @@
       </provider>
       <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
         <state split_layout="SPLIT">
-          <first_editor relative-caret-position="264">
-            <caret line="297" column="36" selection-start-line="297" selection-start-column="36" selection-end-line="297" selection-end-column="36" />
+          <first_editor relative-caret-position="525">
+            <caret line="35" column="76" selection-start-line="35" selection-start-column="10" selection-end-line="35" selection-end-column="76" />
             <folding />
           </first_editor>
           <second_editor />
         </state>
       </provider>
     </entry>
-    <entry file="file://$PROJECT_DIR$/深度学习/Attention.md">
+    <entry file="file://$PROJECT_DIR$/推荐/YouTubeNext.md">
+      <provider editor-type-id="split-provider[text-editor;MarkdownPreviewEditor]">
+        <state split_layout="SPLIT">
+          <first_editor relative-caret-position="0">
+            <caret line="0" column="0" selection-start-line="0" selection-start-column="0" selection-end-line="0" selection-end-column="0" />
+            <folding />
+          </first_editor>
+          <second_editor>
+            <js_state />
+          </second_editor>
+        </state>
+      </provider>
+      <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
+        <state split_layout="SPLIT">
+          <first_editor relative-caret-position="195">
+            <caret line="13" column="22" selection-start-line="13" selection-start-column="22" selection-end-line="13" selection-end-column="22" />
+            <folding />
+          </first_editor>
+          <second_editor />
+        </state>
+      </provider>
+    </entry>
+    <entry file="file://$PROJECT_DIR$/推荐/DeepFM.md">
+      <provider editor-type-id="split-provider[text-editor;MarkdownPreviewEditor]">
+        <state split_layout="SPLIT">
+          <first_editor relative-caret-position="0">
+            <caret line="0" column="0" selection-start-line="0" selection-start-column="0" selection-end-line="0" selection-end-column="0" />
+            <folding />
+          </first_editor>
+          <second_editor>
+            <js_state />
+          </second_editor>
+        </state>
+      </provider>
+      <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
+        <state split_layout="SPLIT">
+          <first_editor relative-caret-position="375">
+            <caret line="25" column="0" selection-start-line="25" selection-start-column="0" selection-end-line="25" selection-end-column="0" />
+            <folding />
+          </first_editor>
+          <second_editor />
+        </state>
+      </provider>
+    </entry>
+    <entry file="file://$PROJECT_DIR$/README.md">
       <provider editor-type-id="split-provider[text-editor;MarkdownPreviewEditor]">
         <state split_layout="SPLIT">
           <first_editor relative-caret-position="0">
@@ -1228,8 +1317,8 @@
       </provider>
       <provider selected="true" editor-type-id="split-provider[text-editor;markdown-preview-editor]">
         <state split_layout="SPLIT">
-          <first_editor relative-caret-position="540">
-            <caret line="36" column="0" selection-start-line="36" selection-start-column="0" selection-end-line="36" selection-end-column="0" />
+          <first_editor relative-caret-position="295">
+            <caret line="353" column="40" selection-start-line="353" selection-start-column="40" selection-end-line="353" selection-end-column="40" />
             <folding />
           </first_editor>
           <second_editor />
diff --git a/README.md b/README.md
index fe628ed..51da792 100644
--- a/README.md
+++ b/README.md
@@ -342,7 +342,24 @@
 # 推荐
 - DIN
 - DeepFM
+    - [DNN与DeepFM之间的区别](推荐/DeepFM.md#L164)
+    - [Wide&Deep与DeepFM之间的区别](推荐/DeepFM.md#L164)
+    - [你在使用deepFM的时候是如何处理欠拟合和过拟合问题的](推荐/DeepFM.md#L164)
+    - [DeepFM怎么优化的](推荐/DeepFM.md#L164)
+    - [不定长文本数据如何输入deepFM](推荐/DeepFM.md#L164)
+    - [deepfm的embedding初始化有什么值得注意的地方吗](推荐/DeepFM.md#L164)
+        - [Xavier初始化](推荐/DeepFM.md#L164)
+        - [He初始化](推荐/DeepFM.md#L164)
 - YoutubeNet
+    - [变长数据如何处理的](推荐/YoutubeNet.md#L164)
+    - [input是怎么构造的](推荐/YoutubeNet.md#L164)
+    - [最后一次点击实际如何处理的](推荐/YoutubeNet.md#L164)
+    - [output的是时候train和predict如何处理的](推荐/YoutubeNet.md#L164)
+    - [如何进行负采样的](推荐/YoutubeNet.md#L164)
+    - [item向量在softmax的时候你们怎么选择的](推荐/YoutubeNet.md#L164)
+    - [Example Age的理解](推荐/YoutubeNet.md#L164)
+    - [什么叫做不对称的共同浏览（asymmetric co-watch）问题](推荐/YoutubeNet.md#L164)
+    - [整个过程中有什么亮点？有哪些决定性的提升](推荐/YoutubeNet.md#L164)
 - Wide&Deep
 - MLR
 - Neural Network全家桶
diff --git "a/\346\216\250\350\215\220/DeepFM.md" "b/\346\216\250\350\215\220/DeepFM.md"
new file mode 100644
index 0000000..9abf2f4
--- /dev/null
+++ "b/\346\216\250\350\215\220/DeepFM.md"
@@ -0,0 +1,25 @@
+# DNN与DeepFM之间的区别?
+DNN是DeepFM中的一个部分，DeepFM多一次特征，多一个FM层的二次交叉特征
+
+# Wide&Deep与DeepFM之间的区别?
+DeepFM对Wide&Deep中的Wide层进行了优化，增加了交叉特征
+
+# 你在使用deepFM的时候是如何处理欠拟合和过拟合问题的？
+欠拟合：增加deep部分的层数，增加epoch的轮数，增加learning rate，增加正则化力度
+过拟合：在deep层直接增加dropout的率，减少epoch轮数，增加更多的数据，减少正则化力度，shuffle数据
+
+# DeepFM怎么优化的？
+- embedding向量可以通过FM初始化 
+- Deep层可以做优化
+    - NFM:把deep层的做NFM类型的处理，其实就是deep层在输入之前也做一个二阶特征的交叉处理和fm层一致
+- FM层可以变得交叉更多阶
+    - XDeepFM
+
+# 不定长文本数据如何输入deepFM？
+- 截断补齐
+- 结合文本id+文本长度，在做文本处理之前，先做不等长的sum_pooled的操作
+
+# deepfm的embedding初始化有什么值得注意的地方吗？
+- 常规的是Xavier，输出和输出可以保持正态分布且方差相近：np.random.rand(layer\[n-1],layer\[n])*np.sqrt(1/layer\[n-1])
+    - relu的情况下通常是HE，保证半数神经元失活的情况下对输出方差影响最小:：np.random.rand(layer\[n-1],layer\[n])*np.sqrt(2/layer\[n-1])
+- 文本项目上也可以用预训练好的特征
diff --git "a/\346\216\250\350\215\220/YouTubeNext.md" "b/\346\216\250\350\215\220/YouTubeNext.md"
new file mode 100644
index 0000000..43549a5
--- /dev/null
+++ "b/\346\216\250\350\215\220/YouTubeNext.md"
@@ -0,0 +1,44 @@
+# 变长数据如何处理的？
+- input数据中只拿了近20次的点击，部分用户是没有20次的历史行为的，所以我们记录了每一个用户实际点击的次数，在做embedding的时候，我们除以的是真实的history length
+    - 20次点击过去一周内的行为，曾经尝试扩大历史点击次数到40，60没有很明显的效果提升
+    - 点击行为是处理过的，一个session内只有一次点击的行为不会被记录
+    - 点击行为是处理过的，连续多次的重复点击会进行处理
+
+# input是怎么构造的
+- 最近历史20次点击商品id/文章id，如果不足不需要补充
+- 实际历史点击次数
+- 最后一次点击商品id/文章id
+- example age:商品/文章上线时长
+- user_info:age/gender/地理位置/注册时长
+- cross_info:最后一次点击距click时间，最后一次点击商品浏览次数
+- phone_info:设备信息，登录状态
+
+# 最后一次点击实际如何处理的？
+我们会以日进行切分，每日首次点击的lastclick会以\[unknow]进行替代，隔日的点击不会进行计算
+
+# output的是时候train和predict如何处理的
+- train的时候是进行负采样的
+- predict的时候是进行的all_embedding dot
+
+# 如何进行负采样的？
+- 该次点击时间之前所以的item或者article作为代选集
+- 负采样我们会进行剔除，把该次click下的同时show的样本进行剔除后采样
+- 均衡采样，不会根据其他样本show time进行加权
+    - 为了尽可能多的修正全量样本，尽快达到收敛
+    - 为了避免其他推荐产生的交叉影响
+
+# item向量在softmax的时候你们怎么选择的？
+是用初始化我们在进行history click embedding过程中使用的初始化的向量，没有在最后层重新构造一个item embedding的结果，实测效果翻倍的要好
+
+# Example Age的理解？
+实际使用的时候采取的是item上线时间-now的差值，article发表的时间-now的差值，并进行归一化。作用主要是为了平衡因为文章/商品的存量时间对当前选择的一个影响
+
+# 什么叫做不对称的共同浏览（asymmetric co-watch）问题？
+item对比nlp问题的时候，上下文信息更适用于文本等固定结果的探索问题（完形填空问题）；而在预测next one这种问题下，只通过上文进行预测，更加合适和合理。浏览信息大概率都是不对称的，而常规的i2i模型的假设都是对称的，上下文都可以用的
+
+# 整个过程中有什么亮点？有哪些决定性的提升？
+- 共享item embedding vector
+- 加入user feature和cross feature且对feature采取了归一化，归一化方式中均方法+log法对长尾数据有效的标准化了
+- 负样本的选取
+- attention所起到的作用并不是很明显，尝试了传统的DIN中的attention和传统的Transform中的attention，效果一般
+    - 这点其实违反直觉，可能原因是模型对负反馈没有很好的建模
\ No newline at end of file