[MLS-C01] [Algorithms] Classification Algorithms

Defined

Input a set of high-dimensional vectors including a numeric target, or label
Target is 0 or 1 for binary classification
Learns a linear threshold function and maps a vector to an approximation of the target
Good classification model based on Linear Learner optimizes
- Discrete objectives suited for classification, such as F1 measure, precision, recall, or accuracy.
Requires a data matrix of observations across dimension of features
Also requires a target column across the observations
Example use case: predict event outcome - win or lose
Important Hyperparameters
1. feature_dim: number of feature in the input
2. predictor_type: type of the target variable (binary_classifier or multiclass_classifier)
3. loss: specifies the loss function (auto, logistic, hinge_loss, softmax_loss)

Implements the Word2vec and text classification algorithms
Useful for many downstream natural language processing (NLP) tasks, such as sentiment analysis, named entity recognition, machine translation
Also useful for applications that perform web searches, information retrieval, ranking, and document classification
Maps words to vectors
Resulting vector representation of a word is called a word embedding
Words that are semantically similar correspond to vectors that are close together, resulting that word embeddings capture the semantic relationships between words
Example use case: spam detection using detection of common spam phrases like “you’re a winner !”
Important Hyperparameters
1. mode: Word2vec architecture for training (batch_skipgram, skipgram, cbow)

Implementation of gradient boosted trees algorithm
Supervised learning algorithm for predicting a target by combining the estimates from a set of simpler models
Requires a data matrix of observations across dimension of features
Also requires a target column across the observations
Can differentiate the importance of features through weights
Example use case: predict default on credit card payments
Important Hyperparameters
1. num_round: number of rounds the training runs
2. num_class: the number of classes
3. objective: learning task and learning objective (i.e. reg:logistic, multi:softmax)

Finds the k closest points to the sample point and gives a prediction of the average of their features
Index based
Objective: build k-NN index to allow for efficient determination of the distance between points
Train to construct the index
Use dimensionality reduction to avoid the “curse of dimensionality”
Example use case: predict wilderness tree types from geological and forest service data
Important Hyperparameters
1. feature_dim: number of feature in the input
2. k: number of nearest neighbors
3. predictor_type: classifier for classification problems
4. sample_size: number of data points to be samples from the training dataset
5. dimensionality_reductoin_target: target dimension to reduce to

Extension of linear model used on high dimensional sparse datasets
Typically used for sparse datasets such as click prediction and item recommendation
Scored using Binary Cross Entropy (Log Loss), Accuracy (at threshold=0.5) and F1 Score (at threshold=0.5)
Example use case: item recommendation
Important Hyperparameters
1. feature_dim: number of feature in the input
2. num_factors: dimensionality of factorization
3. predictor_type: binary_classifier for classification problems

Supervised learning algorithm that supports multi-label classification
Takes an image as input and outputs one or more labels assigned to that image
Uses a convolutional neural network (ResNet) that can be trained from scratch or trained using transfer learning when a large number of training images are not available
Recommended input format is RecordIO, but can also use raw images in .jpg or .png format.
Example use case: classify image content as offensive or not for Twitter feed posts
Important Hyperparameters
1. num_classes: the number of output classes
2. num_training_samples: Number of training examples in the input dataset

Unsupervised algorithm for detecting anomalous data points within a data set
Uses an anomaly score
- Low score indicates data point is considered normal, high score indicates the presence of an anomaly in the data
- Definition of “low” and “high” depend on the application, common practice: scores beyond three standard deviations from the mean score are considered anomalous
Requires a target column across the observations
Example use case: find exceptions in streaming trade data
Important Hyperparameters
1. feature_dim: number of feature in the input
2. eval_metrics: List of metrics used to score a labeled test data set (accuracy, precision_recall_fscore)
3. num_trees: Number of trees in the forest

本文由 Oscaner 创作, 采用知识共享署名4.0 国际许可协议进行许可
本站文章除注明转载/出处外, 均为本站原创或翻译, 转载前请务必署名