Angel推荐算法在游戏推荐中的应用

{"type":"doc","content":[{"type":"blockquote","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"导读：","attrs":{}},{"type":"text","text":"Angel是腾讯自研的分布式高性能的机器学习平台，支持机器学习、深度学习、图计算以及联邦学习等场景。Angel的深度学习平台已应用在腾讯的很多个场景中。本次分享为大家介绍Angel推荐算法在游戏推荐中的应用。主要内容包括：游戏平台上的游戏推荐、Tesla平台上的推荐算法、经典算法的线性特点、DeepFM算法的非线性特点、DeepFM应用过程。","attrs":{}}]}],"attrs":{}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"游戏平台上的游戏推荐","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/26/26b940d43271615781f27467e968c2ab.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这张图看到的是Steam平台上的一个游戏推荐的应用。Steam平台主要是使用标签的推荐方法，它的标签主要是基于用户选择去收集的信息。","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/53/53e24436b47a12a3719a9d10d78871e5.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Steam平台的特点在于游戏的内容比较多，深度也比较深一些。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果item比较多，而且使用用户比较多，那就可以通过用户来选择一些标签，能够抽象出其实本来是需要协同过滤ALS的算法计算的因变量的特征向量。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"也就是说，Steam是用人工选择的方式，用集体智慧的方式来抽象出特征向量。","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/b3/b33da44a5bd61b5b1fda47436dec59a7.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这是Wegame平台上的一个活动。它的推荐算法并没有使用人工抽取标签的方式，而使用了CF算法，以及在用户行为数据的基础上使用Deep FM算法。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"Tesla平台上的推荐算法","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/54/543d2c8c63865c0be8350d1b41660a73.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这个是特斯拉平台上的推荐算法。如果你要在外网试用，可以到下面的网址进行试用：","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"https://cloud.tencent.com","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"在试用的过程中，根据wiki文档生成相关的模型，定义相关的参数，就可以使用了。","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/d9/d9f92e430aaad420d234ee49cd468f13.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特斯拉平台上也可以使用一些传统的算法，就像CF-ALS算法，可以把Rank、Lambda、Alpha这几个重要的参数，通过一个迭代的方式去选择合适的超参数，就可以使用了。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"经典算法的线性特征","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/78/780f9cbd530396e91d916bf5e515ea99.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"如果是基于内容的标签推荐，不用集体智慧的抽象，在很多情况下就是一个主观的先行判断。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"另一种是协同过滤——或基于物品，或基于用户，或者是联合的协同过滤，会存在稀疏矩阵和长尾推荐的问题。","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/ac/ac7111110e5f4c1d5eb3bcb11c8471c4.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"此时会出现一个问题：这是基于物品点击率比较高的协同过滤，但游戏大部分推荐的并不是都是热门物品。如果我们想在经典算法的基础上对一些点击比较少的物品进行推荐，也就是对这些长尾的物品进行推荐，如何去实现呢？","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"此时可以先做一个聚类，再应用一个协同过滤的算法。如果是经典的算法，有一个很大的核心的问题：怎么去结合用户画像，还有物品画像，参与到DeepFM的建立？","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"DeepFM的非线性特征","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/f3/f3fdcaf8685c556f8b69c14ce1a022e5.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"其实CF算法中输入的是Item Id和 User ID，但是Deep FM在输入UserID和Item Id之外，还可以输入用户或者物品特征。另外也可以对特征进行一个二次交叉：二次标签的特征组合，在本质上还是比较依赖于原来所采集的特征，而且特征组合也仅仅是二次的一个特征表达。","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/c6/c6404900f4e8c94d72660f1c5b82a5f7.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"DeepFM的各种非线性特征，其实比较类似于CNN算法里面，在识别分类过程中所作的权重分解。上图是在CNN的过程中，做的一个热力图。其中CNN是对图像做了一个分类。这个分类依据，是Class之前所做的权重分布，从Deep FM的一部分的模型研发而来的一部分。","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/2a/2a3cbf4caabd533443fba10f3564cc49.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Deep FM模型的random部分有很多类别特征的。类别特征是用人工选择出来的categorical Features。而Deep FM在自动归类分群的过程中，有一个自动embedding 离散维度的推荐过程。一般我们可以根据误差的自动调整，来调整这个特征的维度组合。类似于推荐过程中自动聚类分群的过程，这就解决了传统算法的问题。如果只输入一个User ID或者是Item ID，就无法使用用户画像和物品画像，进行信息的分群，这样Deep FM就成为了一个可以根据误差进行自动调整，自动聚类分群的推荐工具。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"DeepFM应用过程","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/10/105ac1f58bb5a13ba09f3fc01a54c475.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"另一种情形则是在Deep FM的过程中进行编码，如果有单列的类别值，如何进行、转换成一个稀疏矩阵？一般来说可以使用手工编码，但这个方式比较低效。这时可以使用Feature Hasher，可以自动把几列类别值转换成一个技术矩阵，把稀疏矩阵和由连续值组成的vector和Hash feature组成vector，合成一个Vector，再把这个Vector作为一个Deep FM的输入，就可以比较方便进行数据计算了。","attrs":{}}]},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/ae/aefec1bd3642564ff230489462b8cdf9.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"这里面有几个经验：","attrs":{}}]},{"type":"bulletedlist","content":[{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Validation AUC与点击率的对应关系：就是说在算法的训练的过程中，会输出一个Validation AUC，这个AUC很大情况下会和点击率有对应关系。在算法的模型训练过程中AUC比较低的话，点击率可能比较高，但是如果AUC比较高，其点击率一般情况下也比较高。","attrs":{}}]}],"attrs":{}},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"特定场景下的target 数据产生的模型只能应用到特定场景下的点击率评估：你可能会需要使用到3个采集位置。这三个位置前面的Target数据就是用户在item下面点击数据，如果你运用3个模型，每一个模型应用在3个特定位置下，应该会得到一个更好的效果，而尽量不要把3个位置下的Target的数据合并为一个Target数据。因为这和Deep FM的特征的抽象能力相关，也就是说Deep FM有比较强的特征抽取能力，Target数据会反向传递到特征的采集过程中，会比较依赖于这个特征的场景。","attrs":{}}]}],"attrs":{}},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"Deep FM 的参数调整比较简单，因为一旦收敛，AUC 的提升跟训练数据集的大小最密且：因为增加更多的训练数据，AUC提升会比较明显一点。","attrs":{}}]}],"attrs":{}},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"数据精排需要过滤已拥有，已玩过,并且根据业务需要重排数据，比如重点展示新游，热销等。","attrs":{}}]}],"attrs":{}},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"FM 需要输入 user_id 作为训练数据：Deep FM是从FM演化过来的，FM的训练和使用的过程中，一般会是要用User Id做一个输入，真正FM的User ID会在大数据情况下形成一个大的标签。","attrs":{}}]}],"attrs":{}},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"FM 的user_id 在大数据环境下会形成很大的标签数据，例如：如果有1000万用户，一台一个输入进去，在水平层这边特征就会比较大。","attrs":{}}]}],"attrs":{}},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"FM 的user_id 会形成冷启动的瓶颈，并且导致更新频率受到限制：传统的FM算法和CM算法，是一定要输入User ID的，那样就会产生一个冷启动的问题。也就是说，如果用户没有参与到每天的活动中，其实是没法在预测阶段做输入的。但是Deep FM则不同，如果对自己的特征比较自信，也就是说你的输入除了User ID之外，还有很多特征。因为Deep FM还会有比较强的特征抽取的能力，也就是自动特征组合能力。此时可以不输入User ID，只输入用户的特征或者是物品的特征，就会避免冷启动问题，而且更新频率会更快。","attrs":{}}]}],"attrs":{}},{"type":"listitem","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"DeepFM 的特征输入可以替代user_id 的输入：比如在一个活动刚开始没多久，收集了一定数量的数据的情况下，特征是比较丰满的，就可以对用户和User ID，Item ID等没有接触到的数据进行一个预测。","attrs":{}}]}],"attrs":{}}],"attrs":{}},{"type":"image","attrs":{"src":"https://static001.geekbang.org/infoq/4b/4baa92592d220e5cf3de32ae55947894.png","alt":"图片","title":null,"style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":null,"fromPaste":true,"pastePass":true}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong","attrs":{}}],"text":"以上就是今天的分享，谢谢大家。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"王培军，腾讯高级工程师，主要负责wegame平台的广告系统和深度学习与系统的结合探索。","attrs":{}}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"原文链接：","attrs":{}},{"type":"link","attrs":{"href":"https://mp.weixin.qq.com/s/jtUh-f9zrTtLdW4zc6w8FQ","title":""},"content":[{"type":"text","text":"Angel推荐算法在游戏推荐中的应用","attrs":{}}]}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}}]}