Machine learning process and method

Problem modeling
Feature Engineering
Model selection
Model fusion
Model online application

Problem modeling
1、 Must be clear , What are my evaluation indicators ？
2. Select the sample set
3. Cross validation

• What are the evaluation indicators ？
  • P ,R ,PRC,ROC&AUC ,LOSS,mAP
• How to select a sample subset ？
  • Stratified sampling
• How to cross verify ？
  • k-fold

Feature Engineering
1.EDA
2. Data cleaning （ Missing value , outliers ） And normalization , Continuous data discretization （ Points barrels ）
3. feature selection

• Feature Engineering First On data EDA( Exploratory data analysis )
  • Learn about datasets General information , Ratio of missing values , The type of feature
  • Box figure , Histogram , Stem and leaf , Correlation matrix heat map ,PCA Dimension reduction and so on
• EDA after , Understand the data , Also right data Conduct Cleaning and screening .
  • data defect value
  • data abnormal value
  • Distinguishing features species ：
    • The number features
      • Continuous feature
        • normalization
        • discretization
          • Points barrels
      • Discrete features
    • Category features
      • code
        • Hot coding alone
        • Count rank code
          • （ Effective for both linear and nonlinear ）
          • （ The outliers are not sensitive ）
          • （ Feature values do not conflict , Ranking does not conflict ）
        • Natural coding
        • Layered coding
          • Postal Code , ID number
    • Time features
      • Specific date
      • Minutes and seconds
      • Sunday system
      • Is it a weekend 、 Month end 、 Whether quarter end 、 Whether it is business hours 、 Holidays, etc
      • The last time distance … The time interval
    • Space features
      • GPS coordinate
      • Country ID, City ID, Administrative region ID、 The street ID etc.
      • Space distance
    • Text features
      • Regularization
        • Alphanumerics are unified into alphanumerics of one language
      • Corpus construction
        • file ： Description of the item
      • Text cleaning
        • Remove space , Punctuation, etc
      • participle
        • Part of speech tagging
          • Verb
          • Noun
          • Adjective
        • Semantic restoration
          • Can express semantics
        • 3-Gram Model
          • Convert text to a continuous sequence , Three consecutive words make up a sentence ：
            • ABCDE => (ABC, BCD CDE)
        • Skip-Gram Model
          • The word bag model
            • First form a vector
            • Each component of a vector represents , The frequency of words appearing in the document TF
          • TF-IDF
            • First put all the documents , Form a vocabulary （ The dictionary ）
            • For a document , In the dictionary key Form a vector , The value of the vector is tf*idf
            • tf It's a local parameter , Express Words in text d Word frequency in
            • idf It's a global parameter , by log( Total number of documents / The number of documents with this word )
• feature selection ： Select the matching feature combination
  • How to measure the quality of features ？
    • Target label scalar and The characteristics of the distance / Similarity degree
  • Distance and similarity indicators
    • L-p norm distance （ length ）
    • Cosine similarity （cos）（ angle ）
    • Pearson correlation coefficient ( comprehensive )
    • Jaccard Similarity degree （ Number of set intersections / Number of union sets ）
      • Jaccard distance = 1- Jaccard Similarity degree
    • Fisher score
    • Mutual information KL(p(x,y) ||p(x)p(y)）
    • Hypothesis testing
    • CFS Correlation feature selection
  • Three kinds of methods for feature selection ：（ according to Whether the feature selection interacts with the machine learning algorithm ）
    • Filter Method
      • Complete set of features => feature selection => Machine learning algorithm => Model effect
        • Feature selection and machine learning algorithms do not interact , It's independent . So it's simple and effective .
      • Univariate filtering
        • Just think about relevance , according to Relevance ranking , To filter out The least relevant feature
      • Multivariable filtering
        • Consider not only relevance , Also consider consistency ？
          • CFS Relevant feature selection , Including the correlation of cross features
          • MBF
          • FCBF
    • encapsulation Method
      • Complete set of features => | feature selection <=> Machine learning algorithm | => Model effect
      • For possible feature subsets , Consider the suitability of machine learning algorithms , Use algorithms , Verification set To select the best feature subset .
    • The embedded Method
      • Complete set of features =>| feature selection <=> Machine learning algorithm + Model effect |
      • direct The fusion Feature selection and machine learning algorithms , At the same time, evaluate the effect
      • Cross validation is required
      • Decision tree , Random forests , Gradient lifting tree ,SVM, Lasso