2020-07-23

<section style="display: none;" data-tools="新媒体管家" data-label="powered by xmt.cn">

</section>

<img class="rich_pages js_insertlocalimg" data-ratio="0.5" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_gif/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Z942a023MFvicQ2Kd8oXNic4ibRRZyIMwH031U0t3v1zDiaeGPhbQyYc61g/?wx_fmt=gif" data-type="gif" data-w="800" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_gif/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Z942a023MFvicQ2Kd8oXNic4ibRRZyIMwH031U0t3v1zDiaeGPhbQyYc61g/?wx_fmt=gif">





<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">
相信大家都在淘宝购买过相关的商品。<mpchecktext contenteditable="false" id="1595491521532_0.6255523734265047">
</mpchecktext>比如某一天用户购买了一台台式主机后，淘宝会立刻给该用户推送与电脑相关的商品，例如电脑显示屏，键盘和鼠标等等。<mpchecktext contenteditable="false" id="1595491521533_0.27913958553839135">
</mpchecktext>这些推荐的物品完全是淘宝根据大量的点击数据去学习到的。<mpchecktext contenteditable="false" id="1595491521534_0.7241144618281035">
</mpchecktext>
</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">


</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">
比如，在100个人中，有99个人在买了电脑之后，还会去购买鼠标，键盘，无线网卡等等。<mpchecktext contenteditable="false" id="1595491521535_0.09542750114577325">
</mpchecktext>也就是说，用户最近是否买了电脑是一个非常强的信号，暗示着如果用户最近买了电脑，那么用户再次购买电脑显示屏，键盘等物品时的概率非常高，那么这个时候推荐系统及时的去推荐这些物品给用户，用户可能很可能会购买这些物品，这些商铺也能提高自身的销量。<mpchecktext contenteditable="false" id="1595491521536_0.35301935459817524">
</mpchecktext>因此，可以通过特征工程生成很多个这样强关联的信号，提升推荐系统的性能。<mpchecktext contenteditable="false" id="1595491521537_0.32229531588976856">
</mpchecktext>
</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">


</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">
除了电商领域，推荐系统还广泛的应用在电影推荐，兴趣推荐等等。<mpchecktext contenteditable="false" id="1595491521538_0.9425321787553453">
</mpchecktext>相信最近大家都看过大火的<隐秘的角落>，可以堪称是今年最好看的国产剧。<mpchecktext contenteditable="false" id="1595491521539_0.8506786199549834">
</mpchecktext>在豆瓣上的评分如下：<mpchecktext contenteditable="false" id="1595491521540_0.2215451077661421">
</mpchecktext>
</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">


</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
<shape type="#_x0000_t75" filled="f" style="text-indent:0;left:0;width:415.3pt;height:228.65pt;">
<imagedata title="image1">
</imagedata>
</shape><img data-ratio="0.550641940085592" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZGXzqEwKp0icjxTxicibs4o0spWXibRmZR5b4koibbZ1hKNmDhVJMs0wGAgw/?wx_fmt=png" data-type="png" data-w="701" style="width:553.72998046875px;height:304.8699951171875px;" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZGXzqEwKp0icjxTxicibs4o0spWXibRmZR5b4koibbZ1hKNmDhVJMs0wGAgw/?wx_fmt=png">
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">

</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
看完这部剧之后，还觉得不过瘾怎么办？<mpchecktext contenteditable="false" id="1595491521541_0.7292352150514376">
</mpchecktext>往下翻，你就会惊喜的发现一系列的推荐：<mpchecktext contenteditable="false" id="1595491521542_0.7290884520585745">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">


</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
<img data-ratio="0.6638537271448663" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZdDtibXxF2dgnKPyzz18hsDISeB7h4E1KawZaSNuJegEicZOKnj6Jibylw/?wx_fmt=png" data-type="png" data-w="711" style="width:553.72998046875px;height:367.5299987792969px;" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZdDtibXxF2dgnKPyzz18hsDISeB7h4E1KawZaSNuJegEicZOKnj6Jibylw/?wx_fmt=png">
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
是不是感觉推荐的电影都很符合自己的口味？<mpchecktext contenteditable="false" id="1595491521543_0.9954835170986134">
</mpchecktext><隐秘的角落>属于悬疑类的连续剧，而下面推荐的相关剧集也大多属于这个风格。<mpchecktext contenteditable="false" id="1595491521544_0.25257706867214935">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">


</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
这个例子也同样引出了推荐系统中一个重要的分支——协同过滤<mpchecktext contenteditable="false" id="1595491521545_0.33400725384448315">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">

</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
它主要可以分为基于用户的，基于物品的和基于模型的协同过滤算法。<mpchecktext contenteditable="false" id="1595491521546_0.9858037128364041">
</mpchecktext>基于用户的协同过滤算法的核心思想是：<mpchecktext contenteditable="false" id="1595491521547_0.5885428485748083">
</mpchecktext>给当前用户推荐相似的用户看过的物品。<mpchecktext contenteditable="false" id="1595491521548_0.6274684417711525">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">


</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;text-indent: 0em;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">
例如，用户A看过了<白夜追凶>和<无证之罪>，而资深悬疑剧爱好者不仅看过这两部剧，还看过了<非自然死亡>和<我们与恶的距离>。<mpchecktext contenteditable="false" id="1595491521549_0.068484443944403">
</mpchecktext>那么根据一定的相似度算法，在所有的用户中找到和用户A最相似的用户，比如我们找到了用户B，将用户B看过的另外两部剧推荐给了用户A，这样一次推荐过程就完成了。<mpchecktext contenteditable="false" id="1595491521550_0.6431926589028292">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">

</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
推荐系统的另一个典型场景是广告点击率预估。<mpchecktext contenteditable="false" id="1595491521551_0.9157129969899271">
</mpchecktext>在我们浏览手机的应用商店时，经常可以看到商店会给我们推荐某些应用，某些推荐的应用往往是广告主通过付费来进行推广。<mpchecktext contenteditable="false" id="1595491521552_0.15147500113263201">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">


</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
为了提高广告投放的准确度，广告点击率预估模型需要评价用户点击某些应用的概率，将用户最可能点击的应用进行推送，达到准确的投放广告的目的。<mpchecktext contenteditable="false" id="1595491521553_0.216891167302391">
</mpchecktext>一旦我们点击了其中的一个应用，商店就可以成功的向广告投放商进行收费，而广告主们也达到了应用推广的目的。<mpchecktext contenteditable="false" id="1595491521554_0.01428938792615253">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">


</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;text-align: center;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">
<img data-ratio="1.7777777777777777" src="https://mmbiz.qpic.cn/sz_mmbiz_jpg/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Z9k0vhn6HibXib7lV9ypcohVROorNYYlDFsgAVjlnG7icAXTRM3lAq7O5w/?wx_fmt=jpeg" data-type="jpeg" data-w="432" style="width:35%;height:auto;" _src="https://mmbiz.qpic.cn/sz_mmbiz_jpg/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Z9k0vhn6HibXib7lV9ypcohVROorNYYlDFsgAVjlnG7icAXTRM3lAq7O5w/?wx_fmt=jpeg">
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;text-align: center;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">


</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
除了上述领域，推荐系统还广泛的应用在生活中的方方面面，可以说是机器学习在工业界中最成功的落地场景之一了。<mpchecktext contenteditable="false" id="1595491521555_0.9139107306696308">
</mpchecktext>美国著名的电影和电视节目提供商Netflix曾经发起了奖金为百万美元的推荐系统比赛，旨在提升推荐系统的准确度。<mpchecktext contenteditable="false" id="1595491521556_0.9963634113367776">
</mpchecktext>在广告点击率预估的场景，性能提高了1%的模型往往可以为公司带来巨大的收入。<mpchecktext contenteditable="false" id="1595491521557_0.4460567705526448">
</mpchecktext>
</section>
<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">


<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: center;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;text-indent: 21pt;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
<img data-ratio="1" src="https://mmbiz.qpic.cn/sz_mmbiz_gif/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZToYFCfJDJAibsGEiatEY1ItKbJicRJdJ22HOvCoicluomwebWzppqAk3Ng/?wx_fmt=gif" data-type="gif" data-w="240" style="width:20%;height:auto;" _src="https://mmbiz.qpic.cn/sz_mmbiz_gif/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZToYFCfJDJAibsGEiatEY1ItKbJicRJdJ22HOvCoicluomwebWzppqAk3Ng/?wx_fmt=gif">

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;text-indent: 21pt;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">




<img class="rich_pages js_insertlocalimg" data-ratio="0.153125" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZpwGtXYFpb631hTrk0wDVuiauvLInJoe8ZTSEtqz3NYxic9RDV4ckiahhw/?wx_fmt=png" data-type="png" data-w="1280" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZpwGtXYFpb631hTrk0wDVuiauvLInJoe8ZTSEtqz3NYxic9RDV4ckiahhw/?wx_fmt=png">

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">
传统的推荐系统以协同过滤为代表，而随着深度学习的快速发展，越来越多的研究者尝试应用深度学习技术到推荐系统中。<mpchecktext contenteditable="false" id="1595491521558_0.41343210868856506">
</mpchecktext>在此我们主要介绍其中的Wide&Deep网络。<mpchecktext contenteditable="false" id="1595491521559_0.41433554234299685">
</mpchecktext>


<h3 style="margin-top: 13pt;margin-bottom: 13pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;break-after: avoid;font-size: 16pt;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 1px;">
总体流程<mpchecktext contenteditable="false" id="1595491521560_0.5873566370470085">
</mpchecktext>
</h3>
<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">
在此我们以APP商店中的推荐系统为例，其整体流程可以如下图所示<mpchecktext contenteditable="false" id="1595491521561_0.5709115780989085">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">


<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: normal;">
<img data-ratio="0.48764629388816644" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Z3YfgSJNicCibt3Wra3NCMdS3qqBiaaAO3ukkaXloCjH1Jk6PXvIIialgXw/?wx_fmt=png" data-type="png" data-w="769" style="height: 270px;font-family: Calibri;font-size: 16px;letter-spacing: 0.5px;white-space: normal;width: 553.73px;" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Z3YfgSJNicCibt3Wra3NCMdS3qqBiaaAO3ukkaXloCjH1Jk6PXvIIialgXw/?wx_fmt=png">
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: center;margin-left: 0pt;font-size: 10.5pt;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0px;line-height: normal;">
推荐系统流程图<mpchecktext contenteditable="false" id="1595491521562_0.3757896890338168">
</mpchecktext>
</section>
<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">


<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">
给定一个查询，这个查询可能是用户相关的特征，推荐系统首先会从数据库中检索到查询相关的APP，由于APP的数量非常巨大，因此我们可以取最相关的100个检索结果作为候选APP，这一过程通常叫作
粗排
。<mpchecktext contenteditable="false" id="1595491521563_0.6395355843349195">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">

然后将候选的100个APP送入排序模型中，此处的排序模型就是我们下面将要介绍的
Wide&Deep模型
，这一过程也被叫作
精排
。<mpchecktext contenteditable="false" id="1595491521564_0.23529303060617557">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">

排序完成后，我们可以将点击概率最高的APP放置于用户最容易注意到的地方。<mpchecktext contenteditable="false" id="1595491521565_0.5565327489677165">
</mpchecktext>无论用户是否点击了我们的推荐结果，我们都可以构造一个新的日志文件。<mpchecktext contenteditable="false" id="1595491521566_0.9512780018893234">
</mpchecktext>
在累积了一定数量的日志文件后，就可以继续微调排序模型，提高模型的准确程度
。<mpchecktext contenteditable="false" id="1595491521567_0.014898001300925001">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;break-after: avoid;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">


<p style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;text-indent: 0pt;font-size: 10pt;font-weight: normal;text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">



<img class="rich_pages js_insertlocalimg" data-ratio="0.153125" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Z50ekeDd6ME3ooXyTurzx9z8JZQL077sY1H1icH4eukqXlwT3kx06GNQ/?wx_fmt=png" data-type="png" data-w="1280" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Z50ekeDd6ME3ooXyTurzx9z8JZQL077sY1H1icH4eukqXlwT3kx06GNQ/?wx_fmt=png">

<p style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;text-indent: 0em;white-space: normal;font-size: 10.5pt;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
Wide&Deep模型自从被提出后就引起来非常大的关注。<mpchecktext contenteditable="false" id="1595491521568_0.6166093859588566">
</mpchecktext>正如其名，
此模型主要可以分成两个部分，Wide部分和Deep部分
。<mpchecktext contenteditable="false" id="1595491521569_0.4982185221764919">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;text-indent: 0em;white-space: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">


<h3 style="margin-top: 13pt;margin-bottom: 13pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;break-after: avoid;font-size: 16pt;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 1px;">
Wide模型<mpchecktext contenteditable="false" id="1595491521570_0.7791825705968476">
</mpchecktext>
</h3>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: normal;">
Wide部分就是一个线性网络，和我们在初中学的多元线性方程是一样的，只不过其中的参数求解采用梯度下降的方式。<mpchecktext contenteditable="false" id="1595491521571_0.010065523151344902">
</mpchecktext>例如下式<mpchecktext contenteditable="false" id="1595491521572_0.30357879789528996">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: normal;">


</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;font-size: 10.5pt;text-align: center;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;text-indent: 0em;line-height: normal;">
y=wx<mpchecktext contenteditable="false" id="1595491521573_0.32825578524402355">
</mpchecktext>
</section>
<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">



<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">
它的设计目的是
为了记住数据中特定的特征组合方式
。<mpchecktext contenteditable="false" id="1595491521574_0.7815469287683945">
</mpchecktext>例如前面介绍购买电脑的场景，购买了电脑主机的用户，再次购买显示器，键盘等物件的概率特别高。<mpchecktext contenteditable="false" id="1595491521575_0.5283978112833949">
</mpchecktext>因此可以将用户最近是否购买了电脑作为输入模型的输入，也就是特征。<mpchecktext contenteditable="false" id="1595491521576_0.034765396121916314">
</mpchecktext>






假设
当前wide网络只有一个特征
，当该特征取1时，y=wx可以得到y=0.9(假设w的值是0.9)，当该特征取0时，y=0。<mpchecktext contenteditable="false" id="1595491521577_0.8589743768217186">
</mpchecktext>y输出值越大，会增大模型对应用户点击概率的估计。<mpchecktext contenteditable="false" id="1595491521578_0.2556389060086466">
</mpchecktext>





回到我们APP应用推荐的场景中，可以看到图右侧的Cross Product Transformation，就是我们的wide部分的输入。<mpchecktext contenteditable="false" id="1595491521579_0.35153278867519555">
</mpchecktext>Cross Product Transformation是指特征交叉，即将User Installed App和Impression APP进行组合。<mpchecktext contenteditable="false" id="1595491521580_0.22830166062060675">
</mpchecktext>





例如用户手机已经安装了微信，且当前的待估计的APP为QQ，那么这个组合特征就是(User Installed App=’微信’, Impression APP=’QQ’)。<mpchecktext contenteditable="false" id="1595491521581_0.4814490195421415">
</mpchecktext>





<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">


<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 1px;">
Deep模型<mpchecktext contenteditable="false" id="1595491521582_0.48454028653919434">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">



<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">
介绍完Wide部分，我们重新回来再看下Deep部分。<mpchecktext contenteditable="false" id="1595491521583_0.5407387402111272">
</mpchecktext>
Deep部分的设计是为了模型具有较好的泛化能力
，在输入的数据没有在训练集中出现时，它依然能够保持相关性较好的输出
。<mpchecktext contenteditable="false" id="1595491521584_0.518570595723747">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">



<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">
在下图中，
Deep模型输入都是一些含义不是非常明显的特征
，例如设备类型，用户统计数据等类别特征(Categorical Features)。<mpchecktext contenteditable="false" id="1595491521585_0.9130945091243985">
</mpchecktext>类别特征一般属于高维特征。<mpchecktext contenteditable="false" id="1595491521586_0.17463496406949064">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">


<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">
例如手机的种类可能存在成千上万个，因此我们通常把这些类别特征通过
嵌入(Embedding)
的方式，
映射成低维空间的参数向量
。<mpchecktext contenteditable="false" id="1595491521587_0.3321784215471044">
</mpchecktext>这个向量可以被认为表示了原先这个类别特征的信息。<mpchecktext contenteditable="false" id="1595491521588_0.45890689491276815">
</mpchecktext>对于连续特征，其数值本身就具备一定的含义，因此可以直接将和其他嵌入向量进行拼接。<mpchecktext contenteditable="false" id="1595491521589_0.03564824598179883">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">


<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;break-after: avoid;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: normal;letter-spacing: 0px;">
<img data-ratio="0.48567870485678705" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZaZwn1a1OyEOJUxZQ8qKfVY3L5Eia3Jx2ibFQCNBlKX7TuthtU7MryJicw/?wx_fmt=png" data-type="png" data-w="803" style="width:553.72998046875px;height:268.92999267578125px;" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZaZwn1a1OyEOJUxZQ8qKfVY3L5Eia3Jx2ibFQCNBlKX7TuthtU7MryJicw/?wx_fmt=png">
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;text-indent: 0pt;font-size: 10pt;font-weight: normal;text-align: center;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: normal;letter-spacing: 0px;">
2Wide&Deep模型图<mpchecktext contenteditable="false" id="1595491521590_0.21435550354895083">
</mpchecktext>
</section>




在拼接完成后，可以得到大致为1200维度的向量
。<mpchecktext contenteditable="false" id="1595491521591_0.20305824849402532">
</mpchecktext>将其作为三层全连接层网络的输入，并且选择Relu作为激活函数，其中每层的输出维度分别为[1024, 512, 256]。<mpchecktext contenteditable="false" id="1595491521592_0.40081362589281455">
</mpchecktext>









<img class="rich_pages js_insertlocalimg" data-ratio="0.153125" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZicJbmxuv8vbtS0IUWLZsLsWLmY0mSSgzLmk8Tv6Ql2HTZS2kntPO3ew/?wx_fmt=png" data-type="png" data-w="1280" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZicJbmxuv8vbtS0IUWLZsLsWLmY0mSSgzLmk8Tv6Ql2HTZS2kntPO3ew/?wx_fmt=png">


在广告点击率预估的场景中，
模型的输出是一个0~1之间的值
，
表示当前候选APP被点击的概率
。<mpchecktext contenteditable="false" id="1595491521593_0.7811395612533354">
</mpchecktext>因此可以采用逻辑回归函数，将Wide&Deep部分的输出压缩到0~1在之间。<mpchecktext contenteditable="false" id="1595491521594_0.6454734822599122">
</mpchecktext>





首先，
Deep部分的输出是一个256维度的向量
，可以通过一个线性变换将其映射为维度为1的值，然后
和Wide部分的输出进行求和，将求和后的结果输入到逻辑回归函数中。<mpchecktext contenteditable="false" id="1595491521595_0.5259138217215569">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">



<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">




<img class="rich_pages js_insertlocalimg" data-ratio="0.153125" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Zc0Fk07Wy7jZy9zFJRoEbDiaeomx47dePavD6NQPxHBxZZHeLupibhTYg/?wx_fmt=png" data-type="png" data-w="1280" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Zc0Fk07Wy7jZy9zFJRoEbDiaeomx47dePavD6NQPxHBxZZHeLupibhTYg/?wx_fmt=png">


在介绍完模型之后，现在开始动手实践了，由于谷歌公司并没有将其在论文中使用的APP商店数据集进行公开，因此我们现在
采用Criteo公司发布在Kaggle的广告点击率预估数据集Criteo
。<mpchecktext contenteditable="false" id="1595491521596_0.0345083676694633">
</mpchecktext>




<section style="letter-spacing: 0.5px;line-height: normal;">
对应的Mindpsore代码可以在此处找到：<mpchecktext contenteditable="false" id="1595491521597_0.4003900177930393">
</mpchecktext>
</section>
<section style="letter-spacing: 0px;line-height: normal;">
https://gitee.com/mindspore/mindspore/tree/master/model_zoo/wide_and_deep<mpchecktext contenteditable="false" id="1595491521598_0.5888733828509147">
</mpchecktext>
</section>
<section style="letter-spacing: 0.5px;line-height: 1.5em;">
其中包含了模型的定义、训练以及数据的预处理过程。<mpchecktext contenteditable="false" id="1595491521599_0.8297184877292527">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">

</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">

</section>
<section style="text-align: center;letter-spacing: 0.5px;line-height: 1.5em;">
<img class="rich_pages js_insertlocalimg" data-ratio="0.153125" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZNWFfEd47icZXMS8NUnaSTvgEPGLwJsgqztNGLTzibD3RITRhGmTAgbibw/?wx_fmt=png" data-type="png" data-w="1280" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZNWFfEd47icZXMS8NUnaSTvgEPGLwJsgqztNGLTzibD3RITRhGmTAgbibw/?wx_fmt=png">
</section>

在模型定义前，我们需要对数据进行预处理。<mpchecktext contenteditable="false" id="1595491521600_0.31370116160281447">
</mpchecktext>
数据处理的目的是将数据集中的特征取值映射为数值id，并且去除一些出现次数过少的特征值
，避免特征值出现次数过少，导致当特征值对应的参数向量的更新次数过少，影响模型的精度。<mpchecktext contenteditable="false" id="1595491521601_0.1400535888898924">
</mpchecktext>Criteo数据集由13类连续特征和26列类别特征，已经通过哈希方式映射为了32位数值。<mpchecktext contenteditable="false" id="1595491521602_0.027469570101304797">
</mpchecktext>对应的标签(label)的取值为0和1。<mpchecktext contenteditable="false" id="1595491521603_0.33448403395981896">
</mpchecktext>





<section style="letter-spacing: 0.5px;line-height: 1.5em;">
数据处理的核心思路如下图所示，针对类别特征，建立一个词表，里面记录着每次出现的特征取值的编号。<mpchecktext contenteditable="false" id="1595491521604_0.9503327169834963">
</mpchecktext>然后再遍历每一列的特征取值，将原始的特征值根据词表映射为对应的id。<mpchecktext contenteditable="false" id="1595491521605_0.4568711305554223">
</mpchecktext>
</section>
<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">


<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
<img data-ratio="0.4322396576319544" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZE2MTjVMfSQRic5pJ9HbI1Drrw1iaV7KppZfK6bEichaicKsicbp3986DnfQ/?wx_fmt=png" data-type="png" data-w="1402" style="width:553.72998046875px;height:239.3300018310547px;" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZE2MTjVMfSQRic5pJ9HbI1Drrw1iaV7KppZfK6bEichaicKsicbp3986DnfQ/?wx_fmt=png">

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
还记得之前提到过的低维嵌入过程吗？<mpchecktext contenteditable="false" id="1595491521606_0.22387692283379823">
</mpchecktext>映射操作是为了模型中的嵌入向量查找而准备的。<mpchecktext contenteditable="false" id="1595491521607_0.3794396131282576">
</mpchecktext>例如，输入数据中的A，B，C被映射为了0,1,2。<mpchecktext contenteditable="false" id="1595491521608_0.8596751114792767">
</mpchecktext>而0,1,2分别表示在嵌入矩阵中(Embedding Table)的第0行，第1行和第2行。<mpchecktext contenteditable="false" id="1595491521609_0.07472105979453136">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">



<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
将Criteo数据集下载后解压
，可以看到train.txt和test.txt文件。<mpchecktext contenteditable="false" id="1595491521610_0.9955548280550606">
</mpchecktext>查看train.txt中的文件，
可以看到其中某些行存在缺失
。<mpchecktext contenteditable="false" id="1595491521611_0.7292858768747217">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">



<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
<img data-ratio="0.13284804367606914" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Zcgz9XPz046HSicl60h3XiciaWj89CXuZmtibfwngAlJIqA0pzlqmNsxlbA/?wx_fmt=png" data-type="png" data-w="1099" style="width:553.72998046875px;height:73.5199966430664px;" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Zcgz9XPz046HSicl60h3XiciaWj89CXuZmtibfwngAlJIqA0pzlqmNsxlbA/?wx_fmt=png">

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">



<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
我们可以调用<mpchecktext contenteditable="false" id="1595491521612_0.418118260522949">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">



<pre class="code-snippet__js code-snippet code-snippet_nowrap">
<code>model_zoo/wide_and_deep/src/preprocess_data.py<mpchecktext contenteditable="false" id="1595491521613_0.05205307756937949">
</mpchecktext></code>
</pre>
<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">


<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
进行数据的下载和处理，
对于缺失值，可以将其标记为OOV(Out Of Vocabulary)对应的id
。<mpchecktext contenteditable="false" id="1595491521614_0.2193681441461055">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">


<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">




<img class="rich_pages js_insertlocalimg" data-ratio="0.153125" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Zrp1W7cEcwlZficyWEpEwHtS7qx9pJV6nvkOW60oP9NNl3Em4llnicIUg/?wx_fmt=png" data-type="png" data-w="1280" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Zrp1W7cEcwlZficyWEpEwHtS7qx9pJV6nvkOW60oP9NNl3Em4llnicIUg/?wx_fmt=png">


介绍完数据处理后，我们在此开始
定义模型
。<mpchecktext contenteditable="false" id="1595491521615_0.8887462425029617">
</mpchecktext>模型的核心代码在mindspore仓库下<mpchecktext contenteditable="false" id="1595491521616_0.5841588145838044">
</mpchecktext>





<pre class="code-snippet__js code-snippet code-snippet_nowrap">
<code>model_zoo/wide_and_deep/src/wide_and_deep.py<mpchecktext contenteditable="false" id="1595491521617_0.784860819060053">
</mpchecktext></code>
</pre>





在MindSpore中，网络的定义方式和Pytorch比较接近，先定义定义的操作，然后再construct函数中对调用对应的操作对输入进行处理。<mpchecktext contenteditable="false" id="1595491521618_0.32974386636848463">
</mpchecktext>





首先我们再回忆下
Wide&Deep网络
，它由一个Wide部分和Deep部分。<mpchecktext contenteditable="false" id="1595491521619_0.7208582330729274">
</mpchecktext>其中
Wide部分是一个线性网络
。<mpchecktext contenteditable="false" id="1595491521620_0.7959959689289937">
</mpchecktext>



在实现上，我们将线性网络中的权重视为维度为1，通过嵌入矩阵查找的方式即可获得输入x对应的权重，然后将其和输入的mask相乘，将结果求和。<mpchecktext contenteditable="false" id="1595491521621_0.191038163225294">
</mpchecktext>






值得注意的是，
我们将连续特征和类别特征进行等同处理，因此这里的mask是为了将连续特征和类别特征进行区而设计的
。<mpchecktext contenteditable="false" id="1595491521622_0.30369802872677254">
</mpchecktext>
连续特征mask中的值即为连续特征值，类别特征mask中的值为1
。<mpchecktext contenteditable="false" id="1595491521623_0.596134292785284">
</mpchecktext>Wide部分的核心代码如下所示，我们定义个名为self.wide_w的权重，它的形状为[词表大小,1]。<mpchecktext contenteditable="false" id="1595491521624_0.6899780507690909">
</mpchecktext>




<pre class="code-snippet__js code-snippet code-snippet_nowrap" data-lang="ruby">
<code>class WideDeepModel(nn.Cell):</code>
<code> def init(self, config):</code>
<code> super(WideDeepModel, self).init()</code>
<code> …</code>
<code> init_acts = [('Wide_w', [self.vocab_size, 1], self.emb_init),</code>
<code> ('V_l2', [self.vocab_size, self.emb_dim], self.emb_init),</code>
<code> ('Wide_b', [1], self.emb_init)]</code>
<code> var_map = init_var_dict(self.init_args, init_acts)</code>
<code> self.wide_w = var_map["Wide_w"]</code>
<code> self.wide_b = var_map["Wide_b"]</code>
<code> self.embeddinglookup = nn.EmbeddingLookup()</code>
<code> self.mul = P.Mul()</code>
<code> self.reduce_sum = P.ReduceSum(keep_dims=False)</code>
<code> self.reshape = P.Reshape()</code>
<code> self.square = P.Square()<mpchecktext contenteditable="false" id="1595491521625_0.533022001520425">
</mpchecktext></code>
<code></code>
<code> def construct(self, id_hldr, wt_hldr):</code>
<code> mask = self.reshape(wt_hldr, (self.batch_size, self.field_size, 1))</code>
<code> # Wide layer</code>
<code> wide_id_weight = self.embeddinglookup(self.wide_w, id_hldr, 0)</code>
<code> wx = self.mul(wide_id_weight, mask)</code>
<code> wide_out = self.reshape(self.reduce_sum(wx, 1) + self.wide_b, (-1, 1))</code>
<code> out = wide_out<mpchecktext contenteditable="false" id="1595491521626_0.8040137239832756">
</mpchecktext></code>
</pre>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">
 <mpchecktext contenteditable="false" id="1595491521627_0.45306220201784786">
</mpchecktext>
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;letter-spacing: 0.5px;line-height: 1.5em;">
现在我们开始定义Deep部分。<mpchecktext contenteditable="false" id="1595491521628_0.09296508594142838">
</mpchecktext>
Deep部分同样有一个嵌入矩阵查找
，以及5层的全连接层构成。<mpchecktext contenteditable="false" id="1595491521629_0.8505185533081938">
</mpchecktext>
</section>
<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
 <mpchecktext contenteditable="false" id="1595491521630_0.40822084572269834">
</mpchecktext>

<pre class="code-snippet__js code-snippet code-snippet_nowrap" data-lang="ruby">
<code>class WideDeepModel(nn.Cell):</code>
<code> def init(self, config):</code>
<code> super(WideDeepModel, self).init()</code>
<code> …</code>
<code> init_acts = [('Wide_w', [self.vocab_size, 1], self.emb_init),</code>
<code> ('V_l2', [self.vocab_size, self.emb_dim], self.emb_init),</code>
<code> ('Wide_b', [1], self.emb_init)]</code>
<code> var_map = init_var_dict(self.init_args, init_acts)</code>
<code> self.wide_w = var_map["Wide_w"]</code>
<code> self.wide_b = var_map["Wide_b"]</code>
<code> self.embedding_table = var_map["V_l2"]</code>
<code> self.dense_layer_1 = DenseLayer(self.all_dim_list[0],</code>
<code> self.all_dim_list[1],</code>
<code> self.weight_bias_init,</code>
<code> self.deep_layer_act,</code>
<code> convert_dtype=True, drop_out=config.dropout_flag)</code>
<code> self.dense_layer_2 = DenseLayer(self.all_dim_list[1],</code>
<code> self.all_dim_list[2],</code>
<code> self.weight_bias_init,</code>
<code> self.deep_layer_act,</code>
<code> convert_dtype=True, drop_out=config.dropout_flag)</code>
<code> self.dense_layer_3 = DenseLayer(self.all_dim_list[2],</code>
<code> self.all_dim_list[3],</code>
<code> self.weight_bias_init,</code>
<code> self.deep_layer_act,</code>
<code> convert_dtype=True, drop_out=config.dropout_flag)</code>
<code> self.dense_layer_4 = DenseLayer(self.all_dim_list[3],</code>
<code> self.all_dim_list[4],</code>
<code> self.weight_bias_init,</code>
<code> self.deep_layer_act,</code>
<code> convert_dtype=True, drop_out=config.dropout_flag)</code>
<code> self.dense_layer_5 = DenseLayer(self.all_dim_list[4],</code>
<code> self.all_dim_list[5],</code>
<code> self.weight_bias_init,</code>
<code> self.deep_layer_act,</code>
<code> use_activation=False, convert_dtype=True, drop_out=config.dropout_flag)<mpchecktext contenteditable="false" id="1595491521631_0.31540606986635167">
</mpchecktext></code>
<code></code>
<code> self.embeddinglookup = nn.EmbeddingLookup()</code>
<code> self.mul = P.Mul()</code>
<code> self.reduce_sum = P.ReduceSum(keep_dims=False)</code>
<code> self.reshape = P.Reshape()</code>
<code> self.square = P.Square()</code>
<code> self.shape = P.Shape()</code>
<code> self.tile = P.Tile()</code>
<code> self.concat = P.Concat(axis=1)</code>
<code> self.cast = P.Cast()<mpchecktext contenteditable="false" id="1595491521632_0.4011408530903695">
</mpchecktext></code>
<code></code>
<code> def construct(self, id_hldr, wt_hldr):</code>
<code> """</code>
<code> Args:</code>
<code> id_hldr: batch ids;</code>
<code> wt_hldr: batch weights;</code>
<code> """</code>
<code> mask = self.reshape(wt_hldr, (self.batch_size, self.field_size, 1))</code>
<code> # Wide layer</code>
<code> wide_id_weight = self.embeddinglookup(self.wide_w, id_hldr, 0)</code>
<code> wx = self.mul(wide_id_weight, mask)</code>
<code> wide_out = self.reshape(self.reduce_sum(wx, 1) + self.wide_b, (-1, 1))</code>
<code> # Deep layer</code>
<code> deep_id_embs = self.embeddinglookup(self.embedding_table, id_hldr, 0)</code>
<code> vx = self.mul(deep_id_embs, mask)</code>
<code> deep_in = self.reshape(vx, (-1, self.field_size * self.emb_dim))</code>
<code> deep_in = self.dense_layer_1(deep_in)</code>
<code> deep_in = self.dense_layer_2(deep_in)</code>
<code> deep_in = self.dense_layer_3(deep_in)</code>
<code> deep_in = self.dense_layer_4(deep_in)</code>
<code> deep_out = self.dense_layer_5(deep_in)</code>
<code> out = wide_out + deep_out</code>
<code> return out, self.embedding_table<mpchecktext contenteditable="false" id="1595491521633_0.04563167580958871">
</mpchecktext></code>
</pre>
<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">



<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">
我们采用交叉熵作为损失函数，Wide部分采用FTRL作为优化器
。<mpchecktext contenteditable="false" id="1595491521634_0.44376660721383443">
</mpchecktext>FTRL可以产生较好的稀疏权重，可以帮助筛选有价值的特征，并且可以压缩模型权重。<mpchecktext contenteditable="false" id="1595491521635_0.4155254440482379">
</mpchecktext>Deep采用Adam优化器。<mpchecktext contenteditable="false" id="1595491521636_0.37079385163273604">
</mpchecktext>
</section>
<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">



<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">




<img class="rich_pages js_insertlocalimg" data-ratio="0.153125" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZQSqk5ZYMWIicJ0iarUJrcD0gRqLbty6LeTfcO9EgfkPRgSdk6wpwggvQ/?wx_fmt=png" data-type="png" data-w="1280" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZQSqk5ZYMWIicJ0iarUJrcD0gRqLbty6LeTfcO9EgfkPRgSdk6wpwggvQ/?wx_fmt=png">

<section style="text-align: justify;line-height: 1.5em;letter-spacing: 0.5px;">
在train_and_eval.py中的定义了数据，模型的初始化以及训练过程，在定义完模型中，可以将初始化的网络送入Model类中，这个类和Tensorflow中Estimator比较接近，可以通过简单的接口实现网络的训练(model.train)和评估(model.eval)。<mpchecktext contenteditable="false" id="1595491521637_0.32891444885193355">
</mpchecktext>
</section>
<pre class="code-snippet__js code-snippet code-snippet_nowrap">
<code>model_zoo/wide_and_deep/train_and_eval.py<mpchecktext contenteditable="false" id="1595491521638_0.8007241218067571">
</mpchecktext></code>
</pre>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">

</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">

</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
另外，此函数中定义了一些回调函数<mpchecktext contenteditable="false" id="1595491521639_0.4561272766649207">
</mpchecktext>
</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
https://www.mindspore.cn/tutorial/-zhCN/master/advanced_use/customized_debugging_information.html<mpchecktext contenteditable="false" id="1595491521640_0.06552131892906488">
</mpchecktext>
</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
可以打印损失值等相关信息。<mpchecktext contenteditable="false" id="1595491521641_0.25631691777208765">
</mpchecktext>

</section>
<section style="text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">

</section>
<pre class="code-snippet__js code-snippet code-snippet_nowrap" data-lang="python">
<code>def test_train_eval(config):</code>
<code> """</code>
<code> test_train_eval</code>
<code> """</code>
<code> data_path = config.data_path</code>
<code> batch_size = config.batch_size</code>
<code> epochs = config.epochs</code>
<code> ds_train = create_dataset(data_path, train_mode=True, epochs=epochs, batch_size=batch_size)</code>
<code> ds_eval = create_dataset(data_path, train_mode=False, epochs=epochs + 1, batch_size=batch_size)</code>
<code> print("ds_train.size: {}".format(ds_train.get_dataset_size()))</code>
<code> print("ds_eval.size: {}".format(ds_eval.get_dataset_size()))<mpchecktext contenteditable="false" id="1595491521642_0.30555821926604376">
</mpchecktext></code>
<code></code>
<code> net_builder = ModelBuilder()<mpchecktext contenteditable="false" id="1595491521643_0.9736738162578458">
</mpchecktext></code>
<code></code>
<code> train_net, eval_net = net_builder.get_net(config)</code>
<code> train_net.set_train()</code>
<code> auc_metric = AUCMetric()<mpchecktext contenteditable="false" id="1595491521644_0.5805876421991347">
</mpchecktext></code>
<code></code>
<code> model = Model(train_net, eval_network=eval_net, metrics={"auc": auc_metric})<mpchecktext contenteditable="false" id="1595491521645_0.3917814538148676">
</mpchecktext></code>
<code></code>
<code> eval_callback = EvalCallBack(model, ds_eval, auc_metric, config)<mpchecktext contenteditable="false" id="1595491521646_0.5703433659749202">
</mpchecktext></code>
<code></code>
<code> callback = LossCallBack(config=config)</code>
<code> ckptconfig = CheckpointConfig(save_checkpoint_steps=ds_train.get_dataset_size(), keep_checkpoint_max=5)</code>
<code> ckpoint_cb = ModelCheckpoint(prefix='widedeep_train', directory=config.ckpt_path, config=ckptconfig)<mpchecktext contenteditable="false" id="1595491521647_0.5985724627518798">
</mpchecktext></code>
<code></code>
<code> out = model.eval(ds_eval)</code>
<code> print("=====" * 5 + "model.eval() initialized: {}".format(out))</code>
<code> model.train(epochs, ds_train,</code>
<code> callbacks=[TimeMonitor(ds_train.get_dataset_size()), eval_callback, callback, ckpoint_cb])<mpchecktext contenteditable="false" id="1595491521648_0.273273385499504">
</mpchecktext></code>
</pre>







<img class="rich_pages js_insertlocalimg" data-ratio="0.153125" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZujUZQpeQpMVfP8ZQLjpUmIfPeC2crwGW7B8FfKrlFgfayFl92wIWiaw/?wx_fmt=png" data-type="png" data-w="1280" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZujUZQpeQpMVfP8ZQLjpUmIfPeC2crwGW7B8FfKrlFgfayFl92wIWiaw/?wx_fmt=png">


一旦训练完成后
，就可以装载模型参数进行评估。<mpchecktext contenteditable="false" id="1595491521649_0.57825032284277">
</mpchecktext>
评估网络和训练网络类似，只不过输出经过了一个Sigmoid层。<mpchecktext contenteditable="false" id="1595491521650_0.39571448633964024">
</mpchecktext>

<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;text-indent: 0em;">

</section>
<pre class="code-snippet__js code-snippet code-snippet_nowrap" data-lang="ruby">
<code>class PredictWithSigmoid(nn.Cell):</code>
<code> def init(self, network):</code>
<code> super(PredictWithSigmoid, self).init()</code>
<code> self.network = network</code>
<code> self.sigmoid = P.Sigmoid()<mpchecktext contenteditable="false" id="1595491521651_0.8135663424621313">
</mpchecktext></code>
<code></code>
<code> def construct(self, batch_ids, batch_wts, labels):</code>
<code> logits, _, _, = self.network(batch_ids, batch_wts)</code>
<code> pred_probs = self.sigmoid(logits)</code>
<code> return logits, pred_probs, labels<mpchecktext contenteditable="false" id="1595491521652_0.4314559528062756">
</mpchecktext></code>
</pre>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;text-align: justify;margin-left: 5%;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;">

</section>
<p style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;text-indent: 0em;line-height: 1.5em;letter-spacing: 0.5px;">
采用AUC作为评价指标
。<mpchecktext contenteditable="false" id="1595491521653_0.5075784026538523">
</mpchecktext>AUC广泛的应用在分类模型的评估中，
可以较好的反映模型学习的好坏，其值在0~1之间，值越高，模型的性能越好
。<mpchecktext contenteditable="false" id="1595491521654_0.2361147198557314">
</mpchecktext>


<section style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;text-align: justify;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: 1.5em;text-indent: 0em;">

</section>
<p style="margin-top: 0pt;margin-bottom: 0pt;margin-left: 0pt;font-size: 10.5pt;font-weight: normal;text-align: justify;text-indent: 21pt;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">



<img class="rich_pages js_insertlocalimg" data-ratio="0.153125" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Zr3nQGLHgqFQduOC7x3KdM1xnUp9bscZJHTKU0H0OWsYRqQSqcqSviaw/?wx_fmt=png" data-type="png" data-w="1280" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5Zr3nQGLHgqFQduOC7x3KdM1xnUp9bscZJHTKU0H0OWsYRqQSqcqSviaw/?wx_fmt=png">


本文内容介绍了推荐系统的原理和实践代码。<mpchecktext contenteditable="false" id="1595491521655_0.3108364278742566">
</mpchecktext>首先讲述了推荐系统在我们生活中的应用场景，并且介绍了推荐系统的核心原理。<mpchecktext contenteditable="false" id="1595491521656_0.5460435115551427">
</mpchecktext>然后详细介绍了Wide&Deep网络以及相关的代码实践，期望可以帮助大家入门<mpchecktext contenteditable="false" id="1595491521657_0.9002820412351322">
</mpchecktext><img data-ratio="1" src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZnyWkry1l4tmS5tzibBMichlJtejy3wGicxPPmkOBdpso0kHv4B1PtLy7w/?wx_fmt=png" data-type="png" data-w="19" style="display:inline-block;width:20px;vertical-align:text-bottom;" _src="https://mmbiz.qpic.cn/sz_mmbiz_png/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZnyWkry1l4tmS5tzibBMichlJtejy3wGicxPPmkOBdpso0kHv4B1PtLy7w/?wx_fmt=png">





<h2 style="margin-top: 13pt;margin-bottom: 13pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;break-after: avoid;font-size: 16pt;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">
</h2>




<hr style="border-style: solid;border-width: 1px 0px 0px;border-color: rgba(0, 0, 0, 0.1);transform-origin: 0px 0px;transform: scale(1, 0.5);">
参考文献
<mpchecktext contenteditable="false" id="1595491521658_0.2537673922628516">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: normal;letter-spacing: 0px;">
[1] Cheng H T, Koc L, Harmsen J, et al. Wide & deep learning for recommender systems[C]//Proceedings of the 1st workshop on deep learning for recommender systems. 2016: 7-10.<mpchecktext contenteditable="false" id="1595491521659_0.9547549872159344">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: normal;letter-spacing: 0px;">
[2]<span style="caret-color: rgb(25, 67, 156);color: rgb(25, 67, 156);font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;font-size: 12px;letter-spacing: 0.5px;text-decoration: underline;">https://www.mindspore.cn/tutorial/-zhCN/master/advanced_use/customized_debugging_information.html<mpchecktext contenteditable="false" id="1595491521660_0.14200408047701996">
</mpchecktext>

<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: normal;letter-spacing: 0px;">


<p style="margin-top: 0pt;margin-bottom: 0pt;text-align: justify;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;line-height: 1.5em;letter-spacing: 0.5px;">



<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: center;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-weight: normal;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0.5px;line-height: normal;">
<img data-ratio="1" src="https://mmbiz.qpic.cn/sz_mmbiz_gif/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZIsAd6NxHBtQU88StTygUhdbPicYg90wcJL7tEqToWdlo5n6hLFiavf9Q/?wx_fmt=gif" data-type="gif" data-w="240" style="width:15%;height:auto;" _src="https://mmbiz.qpic.cn/sz_mmbiz_gif/xfBicxTBmNWWYhNwiaTTcgiapMmtWvY9c5ZIsAd6NxHBtQU88StTygUhdbPicYg90wcJL7tEqToWdlo5n6hLFiavf9Q/?wx_fmt=gif">
</section>
<section style="margin-top: 0pt;margin-bottom: 0pt;text-align: center;margin-left: 0pt;text-indent: 0pt;font-size: 10.5pt;font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif;letter-spacing: 0px;line-height: normal;">
长按下方二维码关注↓<mpchecktext contenteditable="false" id="1595491521661_0.560668513046289">
</mpchecktext>
</section>

<img class="rich_pages" data-ratio="0.3975" data-s="300,640" src="https://mmbiz.qpic.cn/sz_mmbiz_jpg/xfBicxTBmNWX6gdFbo3zCqEBzN5bMufIGKmPiaUuGS2EvbhZs1icKJSl1LJ5m667QD05Zwks7pZ1MHctdprj2S5aQ/?wx_fmt=jpeg" data-type="jpeg" data-w="800" style="" _src="https://mmbiz.qpic.cn/sz_mmbiz_jpg/xfBicxTBmNWX6gdFbo3zCqEBzN5bMufIGKmPiaUuGS2EvbhZs1icKJSl1LJ5m667QD05Zwks7pZ1MHctdprj2S5aQ/?wx_fmt=jpeg">

<section style="margin-top: 0pt; margin-bottom: 0pt; text-align: justify; margin-left: 0pt; text-indent: 0pt; font-size: 10.5pt; font-weight: normal; font-family: "Helvetica Neue", Helvetica, "Hiragino Sans GB", "Microsoft YaHei", Arial, sans-serif; letter-spacing: 0.5px; line-height: 1.5em;">

</section>