如何使用ML.NET构建推荐系统？

{"type":"doc","content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推荐系统无处不在，从 Netflix、谷歌、亚马逊到小型网店，都能看到它的身影。实际上，推荐系统可能是机器学习最成功的商业应用之一。它具备预测用户喜欢阅读、观看和购买什么的能力。推荐系统不仅对企业有益，对用户也是有益的。对用户而言，推荐系统提供了一种探索产品空间的方式；对企业而言，推荐系统提供了一种增加用户参与的方式，同时也能让企业对客户有更多的了解。本文中，我们将了解如何使用 ML.NET 来创建推荐系统。"}]},{"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":"本文主要包括以下五个部分："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"数据集和前提"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"推荐系统的类型"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"协同过滤直觉"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"矩阵分解直觉"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"用 ML.NET 实现"}]}]}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"1. 数据集和前提"}]},{"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":"大家都喜欢 Netflix，其中一个原因就是他们的推荐做得很好，这家公司已经在推荐系统方面投入了大量资金。Netflix 因 “Netflix Prize”竞赛而闻名，工程师们在没有关于用户或电影的其他信息的情况下，根据先前的评级，试图预测用户对电影的评级，他们甚至提供了一个数据集，这个数据集包括 480189 个用户给 17770 部电影的 100480507 个评级。每个样本在数据集中被格式化为每组四个特征：用户 ID、电影 ID、评级、评级日期。用户 ID 和电影 ID 的特征是整数 ID，而评级则是从 1 到 5。这些数据看起来像这样："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/03\/95\/0327721667b8dff17b38e8ea7995a695.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"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":"本文的实现用 C# 语言完成，我们使用最新的 "},{"type":"text","marks":[{"type":"strong"}],"text":".NET 5"},{"type":"text","text":"，因此要确保你已安装此 SDK。若你正在使用 Visual Studio，则随附 16.8.3 版本。此外，确保你已安装下列软件包："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"python"},"content":[{"type":"text","text":"$ dotnet add package Microsoft.ML\n$ dotnet add package Microsoft.ML.Recommender\n"}]},{"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":"你可以在 Package Manager Console 中执行相同操作："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"Install-Package Microsoft.ML\nInstall-Package Microsoft.ML.Recommendation\n"}]},{"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":"注意，这也会安装默认的 Microsoft .ML 包。你可以使用 Visual Studio 的 Manage NuGetPackage 选项来执行类似操作："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/c0\/84\/c074f3a72f21056b98ce649911684084.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"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":"假如你想了解使用 ML.NET 进行机器学习的基本知识，请看这篇文章：《"},{"type":"link","attrs":{"href":"https:\/\/rubikscode.net\/2021\/01\/04\/machine-learning-with-ml-net-introduction\/","title":"xxx","type":null},"content":[{"type":"text","text":"使用 ML.NET 进行机器学习：简介"}]},{"type":"text","text":"》（Machine Learning with ML.NET – Introduction）"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"2. 推荐系统的类型"}]},{"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":"如前所述，Netflix 的数据集包含用户如何为电影评级的信息。在此基础上，我们如何为该用户创建推荐呢？在推荐同类项目前，我们需要考虑用户看过的电影的一些特征并进行排名。此外，我们可以基于这些排名寻找相似用户，并推荐这些用户购买的项目。但这两个项目的相似意味着什么呢？用户之间的相似意味着什么呢？怎样用数学术语计算和表达这种相似呢？"}]},{"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":"针对这些问题，不同类型的推荐系统采用不同的方法。通常有四种推荐系统："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"bulletedlist","content":[{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"基于内容的推荐系统"},{"type":"text","text":"：这种类型的推荐系统侧重于内容。也就是说，它们只使用项目的特征和信息，并在此基础上为用户创建推荐。它们忽略了其他用户的信息。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"协同过滤推荐系统"},{"type":"text","text":"：推荐系统最强大的地方在于，它们可以根据用户在特定平台上的行为，或者根据同一平台上其他用户的行为，来为用户推荐项目。举例来说，Netflix 会根据你以前看过的电视剧，在用户和你观看并喜欢相同内容的电视剧的基础上，为你推荐下一部精彩的电视剧。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"基于知识的推荐系统"},{"type":"text","text":"：这种类型的推荐系统使用明确的用户偏好、项目以及推荐标注知识。本例中，推荐系统询问用户的偏好，并基于这些反馈建立推荐。"}]}]},{"type":"listitem","attrs":{"listStyle":null},"content":[{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","marks":[{"type":"strong"}],"text":"混合解决方案推荐系统"},{"type":"text","text":"：通常情况下，我们在一些定制的解决方案中使用所有类型的组合。"}]}]}]},{"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":"如果你想进一步了解这些系统的工作原理，请参阅这篇文章《"},{"type":"link","attrs":{"href":"https:\/\/rubikscode.net\/2020\/04\/13\/introduction-to-recommendation-systems\/","title":"xxx","type":null},"content":[{"type":"text","text":"推荐系统简介"}]},{"type":"text","text":"》（Introduction to Recommendation Systems）。在这三种类型中，前两种使用最频繁，也最受欢迎。在实践中，可能出现的情况是，我们创建了混合解决方案来实现更好的结果。"}]},{"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":"ML.NET 仅支持协同过滤，或者更确切地说：矩阵分解。因此，在本文中，我们将重点放在这种类型的推荐系统上。让我们进一步了解这些系统是如何在幕后运作的。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"3. 协同过滤直觉"}]},{"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":"创建推荐系统的最流行的技术之一是协同过滤。不像基于内容的过滤，这种方法把用户和项目置于一个共同的嵌入空间中，沿着它们共有的维度（观看特征）。举例来说，让我们考虑两个 Netflix 用户及其对节目的评级。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/b7\/13\/b7826a4a207576543a6dc73405c6bc13.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"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 TensorFlow 中，我们可以这样表示（别担心，我们不会研究 TensorFlow 的细节，只是举例说明）："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"users_tv_shows = tf.constant([\n [10, 2, 0, 0, 0, 6],\n [0, 1, 0, 2, 10, 0]],dtype=tf.float32)\n"}]},{"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":"我们现在可以获取每个节目的特征，也就是这一类型的 k-hot 编码："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/be\/79\/be3ce372fa17c68dc517a55f14f53479.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"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":"或者在 TensorFlow 中："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"tv_shows_features = tf.constant([\n [0, 0, 1, 0, 1],\n [1, 0, 0, 0, 0],\n [0, 1, 0, 1, 0],\n [0, 0, 1, 0, 0],\n [0, 1, 0, 0, 0]],dtype=tf.float32)\n"}]},{"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":"这样就可以对这些矩阵进行简单的点乘，得到每个用户的相似度："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"users_features = tf.matmul(users_tv_shows, tv_shows_features)\nusers_features = users_features\/tf.reduce_sum(users_features, axis=1, keepdims=True)\n\ntop_users_features = tf.nn.top_k(users_features, num_feats)[1]\nfor i in range(num_users):\n feature_names = [features[int(index)] for index in top_users_features[i]]\n print('{}: {}'.format(users[i],feature_names))\n"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"User1: ['Comedy', 'Drama', 'Sci-Fi', 'Action', 'Cartoon']\nUser2: ['Comedy', 'Sci-Fi', 'Cartoon', 'Action', 'Drama']\n"}]},{"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":"我们可以看到，这两个用户的首要特征是喜剧，这意味着他们喜欢相似的东西。我们在这里做了什么？我们不只是用所提到的流派来描述项目，而是用相同的术语来描述每个用户。比如，用户 1 的意思是她喜欢喜剧片 0.5，而他喜欢动作片 0.1。注意，如果我们将用户的嵌入矩阵与转置项目嵌入矩阵相乘，我们会重新创建用户 - 项目交互矩阵。现在，这对于用户和项目很少的简单例子来说效果很好。但是，当向系统添加更多项目和用户时，它就变得不可扩展了。另外，我们如何确定我们所选择的特征是否相关？假如存在一些我们无法识别的潜在特征呢？我们怎样才能找到正确的特征？要想解决这些问题，就需要先了解矩阵分解。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"4. 矩阵分解直觉"}]},{"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":"我们提到，人为定义的项目和用户特征在整体上可能并不是最佳选择。幸好，这些嵌入可以从数据中学习。也就是说，我们不会手工地给项目和用户分配特征，而是用用户 - 项目交互矩阵来学习最佳因子分解的潜在因子。与之前的思维练习一样，这个过程会产生一个用户因子嵌入和项目因子嵌入矩阵。从技术上讲，我们对稀疏的用户 - 项目交互矩阵进行了压缩，提取了潜在因子（类似主成分分析）。这就是矩阵因子分解的意义所在，它能将一个矩阵因子分解成两个较小的矩阵，利用这些矩阵，我们可以重建原始矩阵："}]},{"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":"你的内容放在这里，在模块内容设置中编辑或删除此文本。你还可以在模块的设计设置中对这个内容的各个方面进行样式化，甚至在模块的高级设置中将自定义 CSS 应用到这个文本。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/16\/80\/160a7593ff48a73a20cf43d3a7062080.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"katexinline","attrs":{"mathString":" A \\approx U \\times V^{T} "}}]},{"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":"像其他降维技术一样，潜在特征的数量也是一个超参数，我们可以对其进行调整，使其在信息压缩和重建误差之间取得平衡。我们有两种方法进行预测，可采用用户与项目因子的点积，或项目与用户因子的点积。矩阵分解法也可以帮助我们求解另一个问题。假设你的系统有数千个用户，你希望计算它们之间的相似性矩阵，这是一个相当大的矩阵，矩阵分解能为我们压缩这些信息。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"4.1 矩阵分解算法"}]},{"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":"数年前，Netflix 举办了一场价值 100 万美元的推荐系统竞赛。目标是根据用户的评级来改进系统的准确性。获胜者使用了奇异值分解算法来获得最佳结果。这种算法现在仍然很流行。形式上，可以这样定义："}]},{"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":"设 A 是一个 m×n 矩阵，则 A 的奇异值分解（Singular Value Decomposition，SVD）为："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"katexinline","attrs":{"mathString":" A=U \\Sigma V^{T} "}}]},{"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":"其中 U 是 m×m 且正交，V 是 n×n 且正交，Σ 是一个 m×n 的对角矩阵，具有非负对角线项 σ1≥σ2≥--≥σp，p=min{m，n}，称为 A 的奇异值。"}]},{"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":"另一种非常流行的算法是交替最小二乘法或 ALS 及其变体。正如它的名字一样，该算法在保持 V 不变的情况下交替求解 U，然后在保持 U 不变的情况下求解 V，仅适用于最小二乘问题。但是，因为它是专用的，ALS 可以并行化，所以算法非常快速。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/8b\/36\/8b9837e05b89895f903c55fc6b38a836.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"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":"它的一个变体是加权交替最小二乘法或 WALS。其区别在于处理缺失数据的方式。正如我们在以前的文章中多次提到的那样，推荐系统最大的敌人之一是稀疏数据。WALS 为特定项目增加权重，并使用这些权重向量，它们可以通过线性或指数缩放来规范行和 \/ 或列的频率。"}]},{"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":"在矩阵分解中，NMF 是另一种常用算法。它表示非负矩阵分解。该算法基于获取低秩表示的矩阵，它包含有非负或正元素。NMF 使用迭代过程来修改 U 和 V 的初始值，使其积接近 V。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/70\/b1\/70375f536054e646e673f5304f1ef1b1.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"5 用 ML.NET 实现"}]},{"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":"ML.NET 目前仅支持带有随机梯度下降的标准矩阵分解。正如我们稍后将看到的，MatrixFactorization Trainer 支持这一点。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"5.1 高层架构"}]},{"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":"在深入研究这个实现之前，让我们先来考虑该实现的高层架构。正如之前的 ML.NET 指南一样，我们想要建立一个易于扩展的解决方案，并且我们能够轻易地扩展出 ML.NET 未来可能包含的新矩阵分解算法。本文所提供的解决方案是 Auto ML 的一种简单形式。下面显示了我们的解决方案的文件夹结构法："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/45\/87\/45fa57360b7a85c5ae8dea105d706c87.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"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":"Data 文件夹包含输入数据的 .csv，MachineLearning 文件夹包含我们使用算法所需的所有东西。可以这样表示架构概述："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"image","attrs":{"src":"https:\/\/static001.infoq.cn\/resource\/image\/4c\/cb\/4c6cf8cd955c3f259aa1yy8b0d6b5bcb.jpeg","alt":null,"title":"","style":[{"key":"width","value":"75%"},{"key":"bordertype","value":"none"}],"href":"","fromPaste":false,"pastePass":false}},{"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":"该解决方案的核心是一个抽象的 TrainerBase 类。这个类位于 Common 文件夹中，其主要目标是规范整个过程的完成方式。这是一个处理数据和进行特征工程的类，此类还负责训练机器学习算法。实现这个抽象类的类位于 Trainers 文件夹中，在这里我们可以找到多个利用 ML.NET 算法的类，这些类定义了应该使用哪种算法。在这种特殊情况下，我们只有一个 Predictor，位于 Predictor 文件夹中。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"5.2 数据模型"}]},{"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":"要从数据集中加载数据，并使用 ML.NET 算法，需要实现类来建模这些数据。在 Data 文件夹中可以找到两个文件：MovieRating 和 MovieRatingPredictions。MovieRating 类为输入数据建模，看起来如下所示："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"using Microsoft.ML.Data;\n\nnamespace RecommendationSystem.MachineLearning.DataModels\n{\n public class MovieRating\n {\n [LoadColumn(0)] \n public int UserId;\n\n [LoadColumn(1)] \n public int MovieId;\n\n [LoadColumn(2)] \n public float Label;\n }\n}\n"}]},{"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":"正如你所看到的，我们不使用来自数据集的数据。"}]},{"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":"MovieRatingPredictions 类为输出数据建模："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"namespace RecommendationSystem.MachineLearning.DataModels\n{\n public class MovieRatingPrediction\n {\n public float Label;\n public float Score;\n }\n}\n"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"5.3 TrainerBase 和 ITrainerBase"}]},{"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":"这个类是该实现的核心。从本质上讲，它有两个部分：一个是用于描述该类的接口；另一个是需要用具体实现重写的抽象类，不过，它实现了接口方法。下面是 ITrainerBase 的接口："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"using Microsoft.ML.Data;\n\nnamespace RandomForestClassification.MachineLearning.Common\n{\n public interface ITrainerBase\n {\n string Name { get; }\n void Fit(string trainingFileName);\n BinaryClassificationMetrics Evaluate();\n void Save();\n }\n}\n"}]},{"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":"该接口由 TrainerBase 类实现。但是，它是抽象的，因为我们希望注入特定的算法："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"using Microsoft.ML;\nusing Microsoft.ML.Data;\nusing Microsoft.ML.Trainers;\nusing Microsoft.ML.Trainers.Recommender;\nusing Microsoft.ML.Transforms;\nusing RecommendationSystem.MachineLearning.DataModels;\nusing System;\nusing System.IO;\n\nnamespace RecommendationSystem.MachineLearning.Common\n{\n \/\/\/

\n \/\/\/ Base class for Trainers.\n \/\/\/ This class exposes methods for training, evaluating and saving ML Models.\n \/\/\/ \n public abstract class TrainerBase : ITrainerBase\n {\n public string Name { get; protected set; }\n \n protected static string ModelPath => Path.Combine(AppContext.BaseDirectory, \n \"recommender.mdl\");\n\n protected readonly MLContext MlContext;\n\n protected DataOperationsCatalog.TrainTestData _dataSplit;\n protected ITrainerEstimator _model;\n protected ITransformer _trainedModel;\n\n protected TrainerBase()\n {\n MlContext = new MLContext(111);\n }\n\n \/\/\/

\n \/\/\/ Train model on defined data.\n \/\/\/ \n \/\/\/ \n public void Fit(string trainingFileName)\n {\n if (!File.Exists(trainingFileName))\n {\n throw new FileNotFoundException($\"File {trainingFileName} doesn't exist.\");\n }\n\n _dataSplit = LoadAndPrepareData(trainingFileName);\n var dataProcessPipeline = BuildDataProcessingPipeline();\n var trainingPipeline = dataProcessPipeline.Append(_model);\n\n _trainedModel = trainingPipeline.Fit(_dataSplit.TrainSet);\n }\n\n \/\/\/

\n \/\/\/ Evaluate trained model.\n \/\/\/ \n \/\/\/ RegressionMetrics object.\n public RegressionMetrics Evaluate()\n {\n var testSetTransform = _trainedModel.Transform(_dataSplit.TestSet);\n\n return MlContext.Regression.Evaluate(testSetTransform);\n }\n\n \/\/\/

\n \/\/\/ Save Model in the file.\n \/\/\/ \n public void Save()\n {\n MlContext.Model.Save(_trainedModel, _dataSplit.TrainSet.Schema, ModelPath);\n }\n\n \/\/\/

\n \/\/\/ Feature engeneering and data pre-processing.\n \/\/\/ \n \/\/\/ Data Processing Pipeline.\n private EstimatorChain BuildDataProcessingPipeline()\n {\n var dataProcessPipeline = MlContext.Transforms.Conversion.MapValueToKey(\n inputColumnName: \"UserId\",\n outputColumnName: \"UserIdEncoded\")\n .Append(MlContext.Transforms.Conversion.MapValueToKey(\n inputColumnName: \"MovieId\",\n outputColumnName: \"MovieIdEncoded\"))\n .AppendCacheCheckpoint(MlContext);\n\n return dataProcessPipeline;\n }\n\n private DataOperationsCatalog.TrainTestData LoadAndPrepareData(string trainingFileName)\n {\n IDataView trainingDataView = MlContext.Data.LoadFromTextFile\n (trainingFileName, hasHeader: true, separatorChar: ',');\n return MlContext.Data.TrainTestSplit(trainingDataView, testFraction: 0.1);\n }\n }\n}\n"}]},{"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":"这是一个大类，它控制着整个过程。我们把它拆开，看看它到底是怎么回事。首先，我们观察这个类的字段和属性："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"public string Name { get; protected set; }\n \n protected static string ModelPath => Path.Combine(AppContext.BaseDirectory, \n \"recommender.mdl\");\n\n protected readonly MLContext MlContext;\n\n protected DataOperationsCatalog.TrainTestData _dataSplit;\n protected ITrainerEstimator _model;\n protected ITransformer _trainedModel;\n"}]},{"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":"继承该属性的类使用 Name 属性为算法添加名称。ModelPath 字段用于定义模型训练完成后将其存储在何处。注意，文件名的扩展名是 .mdl。接下来是 MlContext，以便我们能够使用 ML.NET 的功能。不要忘记，这个类是一个单例，因此在我们的解决方案中只有一个。_dataSplit 字段包含加载的数据。该结构将数据分割成训练数据集和测试数据集。"}]},{"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":"子类将使用字段 _model。这些类定义了该字段中使用哪种机器学习算法。_trainedModel 字段是结果模型，应该对其进行评估和保存。从本质上讲，继承和实现此类的唯一工作是定义应该使用的算法，通过实例化作为 _model 的所需算法的对象。"}]},{"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":"现在让我们来探索 Fit() 方法："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"public void Fit(string trainingFileName)\n{\n if (!File.Exists(trainingFileName))\n {\n throw new FileNotFoundException($\"File {trainingFileName} doesn't exist.\");\n }\n\n _dataSplit = LoadAndPrepareData(trainingFileName);\n var dataProcessPipeline = BuildDataProcessingPipeline();\n var trainingPipeline = dataProcessPipeline.Append(_model);\n\n _trainedModel = trainingPipeline.Fit(_dataSplit.TrainSet);\n}\n"}]},{"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":"这个方法是训练算法的蓝图。它接收 .csv 文件的路径作为输入参数。确定文件存在之后，我们使用私有方法 loadAndPrepareData。该方法将数据加载到内存中，并将其分割成两个数据集，即训练数据集和测试数据集。在 _dataSplit 中保存返回值，因为我们需要一个用于评估阶段的测试数据集。接着我们调用 BuildDataProcessingPipeline()。"}]},{"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":"这就是进行数据预处理和特征工程的方法。对于这些数据，无需做大量工作，我们只需通过以下方式对其进行编码："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":" private EstimatorChain BuildDataProcessingPipeline()\n {\n var dataProcessPipeline = MlContext.Transforms.Conversion.MapValueToKey(\n inputColumnName: \"UserId\",\n outputColumnName: \"UserIdEncoded\")\n .Append(MlContext.Transforms.Conversion.MapValueToKey(\n inputColumnName: \"MovieId\",\n outputColumnName: \"MovieIdEncoded\"))\n .AppendCacheCheckpoint(MlContext);\n\n return dataProcessPipeline;\n }\n"}]},{"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":"接下来是 Evaluate() 方法："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"public RegressionMetrics Evaluate()\n{\n var testSetTransform = _trainedModel.Transform(_dataSplit.TestSet);\n\n return MlContext.Regression.Evaluate(testSetTransform);\n}\n"}]},{"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":"通过使用 _trainedModel 和测试数据集创建 Transformer 对象是一种非常简单的方法。接着，利用 MlContext 来检索回归指标。最后，让我们看看 Save() 方法。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"public void Save()\n{\n MlContext.Model.Save(_trainedModel, _dataSplit.TrainSet.Schema, ModelPath);\n}\n"}]},{"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":"这是另一个简单的方法，只是使用 MLContext 将模型保存到定义的路径中。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"5.4 训练器"}]},{"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":"因为我们在 TrainerBase 类中完成了所有繁重的工作，所以唯一的 Trainer 类非常简单，只专注于 ML.NET 算法的实例化。下面看看 RandomForestTrainer 类："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"using Microsoft.ML;\nusing Microsoft.ML.Trainers.Recommender;\nusing RecommendationSystem.MachineLearning.Common;\n\nnamespace RecommendationSystem.MachineLearning.Trainers\n{\n \/\/\/

\n \/\/\/ Class that uses Decision Tree algorithm.\n \/\/\/ \n public sealed class MatrixFactorizationTrainer : TrainerBase\n {\n public MatrixFactorizationTrainer(int numberOfIterations, \n int approximationRank, \n double learningRate) : base()\n {\n Name = $\"Matrix Factorization {numberOfIterations}-{approximationRank}\";\n\n _model = MlContext.Recommendation().Trainers.MatrixFactorization(\n labelColumnName: \"Label\",\n matrixColumnIndexColumnName: \"UserIdEncoded\",\n matrixRowIndexColumnName: \"MovieIdEncoded\",\n approximationRank: approximationRank,\n learningRate: learningRate,\n numberOfIterations: numberOfIterations);\n }\n }\n}\n"}]},{"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":"正如你所看到的那样，这个类非常简单。我们覆写了 Name 和 _model。在 Recommendation 扩展中使用 MatrixFactorization 类。注意，我们是如何使用这个算法所提供的一些超参数的。有了这个，我们可以做更多的实验。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"5.5 预测器"}]},{"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":"Predictor 类用于加载保存的模型并运行一些预测。通常，这个类与训练器不是同一个微服务的一部分。我们通常有一个微服务来执行模型的训练。该模型被保存到文件中，其他模型从该文件加载该模型，并基于用户输入运行预测。该类看上去如下："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"using RecommendationSystem.MachineLearning.DataModels;\nusing Microsoft.ML;\nusing System;\nusing System.IO;\n\nnamespace RecommendationSystem.MachineLearning.Predictors\n{\n \/\/\/

\n \/\/\/ Loads Model from the file and makes predictions.\n \/\/\/ \n public class Predictor\n {\n protected static string ModelPath => Path.Combine(AppContext.BaseDirectory, \n \"recommender.mdl\");\n private readonly MLContext _mlContext;\n\n private ITransformer _model;\n\n public Predictor()\n {\n _mlContext = new MLContext(111);\n }\n\n \/\/\/

\n \/\/\/ Runs prediction on new data.\n \/\/\/ \n \/\/\/ New data sample.\n \/\/\/ Prediction object\n public MovieRatingPrediction Predict(MovieRating newSample)\n {\n LoadModel();\n\n var predictionEngine = _mlContext.Model.CreatePredictionEngine(_model);\n\n return predictionEngine.Predict(newSample);\n }\n\n private void LoadModel()\n {\n if (!File.Exists(ModelPath))\n {\n throw new FileNotFoundException($\"File {ModelPath} doesn't exist.\");\n }\n\n using (var stream = new FileStream(ModelPath, FileMode.Open, FileAccess.Read, \n FileShare.Read))\n {\n _model = _mlContext.Model.Load(stream, out _);\n }\n\n if (_model == null)\n {\n throw new Exception($\"Failed to load Model\");\n }\n }\n }\n}\n"}]},{"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":"简单地说，模型是从已定义的文件加载，并预测新的样本。要做到这一点，我们需要创建 PredictionEngine 来执行此操作。"}]},{"type":"heading","attrs":{"align":null,"level":3},"content":[{"type":"text","text":"5.6 用法和结果"}]},{"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":"让我们把所有这些放在一起。"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"using RecommendationSystem.MachineLearning.Common;\nusing RecommendationSystem.MachineLearning.DataModels;\nusing RecommendationSystem.MachineLearning.Predictors;\nusing RecommendationSystem.MachineLearning.Trainers;\nusing System;\nusing System.Collections.Generic;\n\nnamespace RecommendationSystem\n{\n class Program\n {\n static void Main(string[] args)\n {\n var newSample = new MovieRating\n {\n UserId = 6,\n MovieId = 11\n };\n\n var trainers = new List\n {\n new MatrixFactorizationTrainer(10, 50, 0.1),\n new MatrixFactorizationTrainer(10, 50, 0.01),\n new MatrixFactorizationTrainer(20, 100, 0.1),\n new MatrixFactorizationTrainer(20, 100, 0.01),\n new MatrixFactorizationTrainer(30, 100, 0.1),\n new MatrixFactorizationTrainer(30, 100, 0.01)\n\n };\n\n trainers.ForEach(t => TrainEvaluatePredict(t, newSample));\n }\n\n static void TrainEvaluatePredict(ITrainerBase trainer, MovieRating newSample) \n {\n Console.WriteLine(\"*******************************\");\n Console.WriteLine($\"{ trainer.Name }\");\n Console.WriteLine(\"*******************************\");\n\n trainer.Fit(\".\\\\Data\\\\recommendation-ratings.csv\");\n\n var modelMetrics = trainer.Evaluate();\n\n Console.WriteLine($\"Loss Function: {modelMetrics.LossFunction:0.##}{Environment.NewLine}\" +\n $\"Mean Absolute Error: {modelMetrics.MeanAbsoluteError:#.##}{Environment.NewLine}\" +\n $\"Mean Squared Error: {modelMetrics.MeanSquaredError:#.##}{Environment.NewLine}\" +\n $\"RSquared: {modelMetrics.RSquared:0.##}{Environment.NewLine}\" +\n $\"Root Mean Squared Error: {modelMetrics.RootMeanSquaredError:#.##}\");\n\n trainer.Save();\n\n var predictor = new Predictor();\n var prediction = predictor.Predict(newSample);\n Console.WriteLine(\"------------------------------\");\n Console.WriteLine($\"Prediction: {prediction.Score:#.##}\");\n Console.WriteLine(\"------------------------------\");\n }\n }\n}\n"}]},{"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":"非 TrainEvaluatePredict() 方法在这里做的是重头戏。使用这个方法，我们可以注入继承 TrainerBase 类的一个实例，以及一个新的样本，以便进行预测。接着调用 Fit() 方法对算法进行训练，并调用 Evaluate() 方法、打印出指标。最后，我们保存该模型。这样做之后，我们创建一个 Predictor 的实例，用一个新的样本调用 Predict() 方法，并打印出预测结果。在 Main 中，我们创建一个训练器对象列表，然后在这些对象上调用 TrainEvaluatePredict。"}]},{"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":"我们根据这些超参数创建了算法列表中随机森林的一些变体。结果如下："}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null}},{"type":"codeblock","attrs":{"lang":"text"},"content":[{"type":"text","text":"*******************************\nMatrix Factorization 10-50\n*******************************\niter tr_rmse obj\n 0 1.4757 2.4739e+05\n 1 0.9161 1.2617e+05\n 2 0.8666 1.1798e+05\n 3 0.8409 1.1348e+05\n 4 0.8240 1.1079e+05\n 5 0.8100 1.0897e+05\n 6 0.7980 1.0736e+05\n 7 0.7847 1.0575e+05\n 8 0.7691 1.0405e+05\n 9 0.7549 1.0284e+05\nLoss Function: 0.77\nMean Absolute Error: .68\nMean Squared Error: .77\nRSquared: 0.29\nRoot Mean Squared Error: .88\n------------------------------\nPrediction: 3.94\n------------------------------\n*******************************\nMatrix Factorization 10-50\n*******************************\niter tr_rmse obj\n 0 3.1309 9.0205e+05\n 1 2.3707 5.4640e+05\n 2 1.7857 3.3435e+05\n 3 1.5459 2.6501e+05\n 4 1.4055 2.2888e+05\n 5 1.3103 2.0634e+05\n 6 1.2430 1.9129e+05\n 7 1.1902 1.8002e+05\n 8 1.1493 1.7159e+05\n 9 1.1185 1.6546e+05\nLoss Function: 1.27\nMean Absolute Error: .89\nMean Squared Error: 1.27\nRSquared: -0.17\nRoot Mean Squared Error: 1.13\n------------------------------\nPrediction: 4.01\n------------------------------\n*******************************\nMatrix Factorization 20-100\n*******************************\niter tr_rmse obj\n 0 1.5068 2.5551e+05\n 1 0.9232 1.2707e+05\n 2 0.8675 1.1773e+05\n 3 0.8426 1.1358e+05\n 4 0.8260 1.1082e+05\n 5 0.8116 1.0874e+05\n 6 0.7984 1.0705e+05\n 7 0.7849 1.0547e+05\n 8 0.7699 1.0374e+05\n 9 0.7556 1.0222e+05\n 10 0.7407 1.0084e+05\n 11 0.7252 9.9587e+04\n 12 0.7108 9.8130e+04\n 13 0.6962 9.6890e+04\n 14 0.6845 9.6048e+04\n 15 0.6718 9.4877e+04\n 16 0.6615 9.4167e+04\n 17 0.6510 9.3413e+04\n 18 0.6419 9.2767e+04\n 19 0.6322 9.1971e+04\nLoss Function: 0.75\nMean Absolute Error: .67\nMean Squared Error: .75\nRSquared: 0.31\nRoot Mean Squared Error: .86\n------------------------------\nPrediction: 4.06\n------------------------------\n*******************************\nMatrix Factorization 20-100\n*******************************\niter tr_rmse obj\n 0 3.1188 8.9340e+05\n 1 2.4196 5.6643e+05\n 2 1.8203 3.4467e+05\n 3 1.5710 2.7129e+05\n 4 1.4210 2.3212e+05\n 5 1.3245 2.0894e+05\n 6 1.2559 1.9343e+05\n 7 1.2024 1.8189e+05\n 8 1.1592 1.7289e+05\n 9 1.1247 1.6594e+05\n 10 1.0956 1.6027e+05\n 11 1.0717 1.5566e+05\n 12 1.0506 1.5171e+05\n 13 1.0326 1.4838e+05\n 14 1.0169 1.4550e+05\n 15 1.0032 1.4306e+05\n 16 0.9907 1.4085e+05\n 17 0.9798 1.3893e+05\n 18 0.9698 1.3718e+05\n 19 0.9610 1.3563e+05\nLoss Function: 0.99\nMean Absolute Error: .78\nMean Squared Error: .99\nRSquared: 0.09\nRoot Mean Squared Error: .99\n------------------------------\nPrediction: 3.92\n------------------------------\n*******************************\nMatrix Factorization 30-100\n*******************************\niter tr_rmse obj\n 0 1.4902 2.5094e+05\n 1 0.9364 1.2934e+05\n 2 0.8672 1.1737e+05\n 3 0.8428 1.1349e+05\n 4 0.8264 1.1104e+05\n 5 0.8114 1.0883e+05\n 6 0.7966 1.0681e+05\n 7 0.7836 1.0532e+05\n 8 0.7698 1.0378e+05\n 9 0.7540 1.0209e+05\n 10 0.7402 1.0089e+05\n 11 0.7248 9.9437e+04\n 12 0.7098 9.7999e+04\n 13 0.6966 9.6791e+04\n 14 0.6826 9.5745e+04\n 15 0.6687 9.4572e+04\n 16 0.6593 9.3841e+04\n 17 0.6480 9.3017e+04\n 18 0.6404 9.2448e+04\n 19 0.6321 9.1986e+04\n 20 0.6238 9.1298e+04\n 21 0.6160 9.0879e+04\n 22 0.6090 9.0430e+04\n 23 0.6025 9.0006e+04\n 24 0.5962 8.9550e+04\n 25 0.5909 8.9269e+04\n 26 0.5859 8.9011e+04\n 27 0.5809 8.8598e+04\n 28 0.5764 8.8393e+04\n 29 0.5714 8.8086e+04\nLoss Function: 0.74\nMean Absolute Error: .67\nMean Squared Error: .74\nRSquared: 0.32\nRoot Mean Squared Error: .86\n------------------------------\nPrediction: 3.98\n------------------------------\n*******************************\nMatrix Factorization 30-100\n*******************************\niter tr_rmse obj\n 0 3.1699 9.2239e+05\n 1 2.4110 5.6279e+05\n 2 1.8361 3.4988e+05\n 3 1.5652 2.6961e+05\n 4 1.4201 2.3188e+05\n 5 1.3248 2.0902e+05\n 6 1.2537 1.9291e+05\n 7 1.2017 1.8175e+05\n 8 1.1583 1.7271e+05\n 9 1.1237 1.6575e+05\n 10 1.0953 1.6017e+05\n 11 1.0711 1.5555e+05\n 12 1.0502 1.5162e+05\n 13 1.0324 1.4834e+05\n 14 1.0168 1.4549e+05\n 15 1.0036 1.4316e+05\n 16 0.9905 1.4080e+05\n 17 0.9795 1.3886e+05\n 18 0.9697 1.3715e+05\n 19 0.9607 1.3558e+05\n 20 0.9526 1.3418e+05\n 21 0.9452 1.3293e+05\n 22 0.9384 1.3175e+05\n 23 0.9322 1.3070e+05\n 24 0.9265 1.2976e+05\n 25 0.9211 1.2883e+05\n 26 0.9163 1.2802e+05\n 27 0.9118 1.2727e+05\n 28 0.9075 1.2653e+05\n 29 0.9036 1.2589e+05\nLoss Function: 0.9\nMean Absolute Error: .74\nMean Squared Error: .9\nRSquared: 0.17\nRoot Mean Squared Error: .95\n------------------------------\nPrediction: 3.86\n------------------------------\n"}]},{"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":"我们使用用户 ID-6 和电影 ID-11 进行测试。如果你看一下数据集，你会发现这一对和评级是 4。 正如你所看到的那样，大多数矩阵分解的变体都很好用。迭代 10 次，近似秩 50，学习率 0.01 的变化似乎最接近。而且，它的指标似乎也非常好。但是，还需要进一步的测试才能确定那种变体表现最佳。"}]},{"type":"heading","attrs":{"align":null,"level":2},"content":[{"type":"text","text":"结语"}]},{"type":"paragraph","attrs":{"indent":0,"number":0,"align":null,"origin":null},"content":[{"type":"text","text":"本文涉及了很多方面。我们了解了不同类型的推荐系，接着研究了协同过滤和矩阵分解。另外，我们也有机会了解如何将其应用于电影推荐。最终，我们使用 ML.NET 实现了这一切。"}]},{"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"}],"text":"作者介绍："}]},{"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":"Nikola M. Zivkovic，是 Rubik's Code 的首席人工智能官，也是《Deep Learning for Programmers》（尚无中译本）一书的作者。热爱知识分享，是一位经验丰富的演讲者，也是塞尔维亚诺维萨德大学的客座讲师。"}]},{"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"}],"text":"原文链接："}]},{"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":"https:\/\/rubikscode.net\/2021\/03\/15\/machine-learning-with-ml-net-recommendation-systems\/"}]}]}