|
所在平台: Udemy |
课程主页: https://www.udemy.com/course/supervised-learning-traditional-methods/
课程评论：没有评论
课程名称：监督学习 - 传统方法 课程概述： 本课程聚焦于数据挖掘中的监督学习，探讨传统的机器学习模型。监督学习是CRISP-ML(Q)方法论中模型构建步骤的一个子领域，它主要包括分类模型、浅层机器学习模型、集成模型、回归模型以及黑箱技术。我们将深入讨论概率、联合概率、贝叶斯定理及简单贝叶斯分类器，并通过案例进行说明。由于简单贝叶斯对于大规模数值特征并不理想，因此需要将数值特征转换为类别特征。该算法假设类条件独立，对训练数据中未见的新词，其概率计算为零，而拉普拉斯估计器则可以用来解决这一问题。 课程还包括K近邻分类器的介绍，它被称为懒惰学习者或基于实例的学习。我们将了解k均值算法与kNN之间的区别，并研究不同的邻居数（1、2、3和7）。k的最小值为1，最大值为观察数量，同时也是一个超参数。我们还会了解基准模型及其准确率，并讨论偏差-方差权衡的概念。课程最后将探讨k-NN的应用及其重要性。 接着，我们介绍决策树算法，这是一种基于规则的算法。课程涵盖决策树的构建方法，包括贪婪算法、如何选择最佳决策树及属性选择。决策树的结构类似于树，内部节点表示属性，分支代表最佳结果，叶节点表示类标签。我们将讨论如何构建决策树，包括使用训练数据建立模型、确定切分变量及其值，以及终止节点的分配。课程将结合贪婪方法、熵和信息增益进一步分析决策树的构建过程。 通过本课程，学习者将获得传统监督学习方法的全面理解与实践经验。
Data Mining Supervised Learning - Traditional ML Models.Supervised Learning is a sub-division of the Model Building step of CRISP-ML(Q)Methodology. Supervised learning is a type of Predictive Modeling that involvesClassification Models, Shallow Machine Learning Models, Ensemble Models, RegressionModels, and the Black Box technique. We have numerous divisions of each modelingtechnique.We thoroughly discuss Probability, Joint Probability, Bayes Rule, and Naive Bayes using ause case. Naive Bayes is not ideal for larger numeric features because numeric featuresmust be converted into categorical ones through discretization or bining. It allows thedeletion of missing value entries. This algorithm assumes class-conditional independence.The probability is zero for new words not seen in training data, making the entire calculationzero. To encounter this problem, we use Laplace Estimator. French MathematicianPierre-Simon Laplace created this algorithm. The default value of the Laplace estimator is 1.Any value can be used for the Laplace estimator.K-Nearest Neighbor Classifier is also called Lazy Learner, Memory-Based Reasoning,Example-Based Reasoning, Instance-Based Learning, Case-Based Reasoning, RoteLearning, etc. We understand the differences between the k-means algorithm and kNN. Wethen understand 1, 2, 3, and 7 Nearest Neighbors. The minimum k value equals 1, and themaximum equals the number of observations. k is a hyperparameter. We then understandwhat a baseline model is, where accuracy is equal to the majority class, and for predictionmodels, accuracy is greater than 80%. We further understand the Bias-variance trade-off.We jump into the applications and importance of k-NN at the end.The Decision Tree algorithm is a Rules-based algorithm. We understand what a decisiontree is, followed by learning how to build decision trees, then we dive into the greedyalgorithm, building the best decision tree and attribute selection. A decision tree is a tree-likestructure in which an internal node represents an attribute, each branch represents theoutcome of the best and each leaf node represents a class label. There are 3 types of nodesa root node, a branch node, and a leaf node. So how do we build a decision tree? First, weuse training data to build a model, and then the tree generator determines the following- Which variable has to be split at a node, and the value of the split- The decision to stop or split again has to be made- Assigning terminal nodes to a label- A basic or Greedy algorithm is a tree constructed in a top-down recursivedivide-and-conquer manner.- Further, we analyze Greedy Approach, Entropy, and Information gain.
