Please enable JavaScript.

Coggle requires JavaScript to display documents.

1 - Data Preparation (Categorical Encoding - converting categorical values…

- - - - Multicategorical
      - Binary Categorical
- - - - N-Gram with size of n=1 produces the same output as Bag-of-Words
      - N-gram uses sliding window
      - E.g. n=2 applied to "He is a jedi now" would produce {"he is", "is a", "a jedi", "jedi now","he","is","a","jedi","now"}
      - It also includes n-grams of all sizes less than n, too
      - N-Grams can be used in spam filter to try to identify common 2-3 word combinations of words that appear in spam emails (e.g. "Click here" or "You're a winner")
      - extension of Bag-of-Words, which produces groups of words of n size. unigram (n=1), bigram (n=2), trigram (n=3), etc.
    - - E.g. OSB n=4 applied to "he is a jedi and he will save us" produces: {"he|is", "he||a", "he|||jedi"},{"is|a", "is||jedi", "is|||and"}, {"a|jedi", "a||and", "a|||he"}, {"jedi|and", "jedi||he", "jedi|||will"}, {"and|he", "and||will", "and|||save"}, {"he|will", "he||save", "he|||us"}, {"will|save", "will||us"}, {"save|us"}
      - Creates groups of words of size n that always include the first word and some character is used as a delimiter (e.g. pipe)
    - - shows the popularity of a word or words in text data by assigns weights to them and making common words like "the" or "and" less important
      - Vectorized tf-idf is in format:
        (number of documents, number of unique n-grams)
        E.g. (2,7) represents 2 documents and 7 unique n-grams
      - Use Case: Apply tf-idf to filter less important words in PDFs, then find common word combinations repeated throughout PDFs. This can help us determine what topic the document is about
    - - only punctuation between words gets removed. Punctuation within words does not (e.g. "you're" stays as is
    - - creates a new feature from the combination of two or more text or categorical values. i.e. combines sets of words together and uses underscore "_" as a combining character
      - E.g. Cartesian product of "Python Data Science Handbook" and "Softcover" would produce:
        {"Python_Softcover", "Data_Softcover", "Science_Softcover", "Handbook_Softcover"}
  - - - Normalization
        
        scaling = feature scaling = normalization
        
        Assigns smallest values to 0 and largest value to 1, then scales all the values in between
        
        Note that outliers can throw off normalization. Apply Random Cut Forest to eliminate outliers or use Standardization instead
      - Standardization
        
        sets average value to 0 and then applies z-score to all the other values. Unlike Normalization, Standardization can have values above 1 and under -1. z-score takes into account average value and standard deviation to smooth out the values so that outliers don't throw off the entire scale
        
        z score is calculated as (x - average value) divided by standard deviation
      - Scaling is often required by many algos like linear/non-linear regression, clustering, neural networks, etc.
    - - Changes numeric values into groups or buckets of similar values
      - Quantile Binning aims to assign the same number of features to each bin
      - Can be used when we know that numeric values in our dataset do not have a linear role in prediction or our target attribute
      - E.g. use person's age to perform binning
      - Often used in addition to other feature engineering methods. E.g. quantile binning can be applied based on person's age to assign each person to one of the 3 age groups and then One-hot encoding can be applied to represent 3 age groups as 3 flag columns (one for each group)
  - - - transforms images to find out useful information
      - E.g. use a picture of a handwritten number and break it into individual pixels, then translate each pixel to 0 or 1 (depending on whether it's blank or filled). The produced grid of values can then be compared to other grids of values in some repository (e.g. MNIST) to determine whether the character is a number or not and what number it is
    - - transform audio data into useful information
      - E.g. take an Analog signal and map its amplitude over time (sampling). Then convert each sample point to a numeric value (for instance, you can represent amplitude of the audio signal with values between -10 and 10)
    - - File
        
        Loads the entire dataset from S3 directly to your training instance volumes
        
        Supports CSV, json, Parquet, Image files (png and jpg)
      - Pipe
        
        datasets are streamed from S3
        
        Provides faster start times for the model jobs and better throughput
        
        data type is recordIO-protobuf :star:, which will create a tensor (i.e. a multidimensional array)
        
        Amazon's built-in python and sdk libraries can be used to convert a more traditional file format (e.g. csv) to recordIO-protobuf
- - - - Supervised Learning - Predicts missing values based on the values of other features. It's most difficult to use, but can yield best results. Use if you have a large number of values missing. E.g. MICEImputer (Multiple Imputation by Chained Equations)
      - Mean - the average value. Quick and easy, but results can vary. Use if you have a small number of values missing
      - Median - Orders values, then chooses value in the middle. Quick and easy, but results can vary. Use if you have a small number of values missing
      - Mode - Most common value. Quick and easy, but the results can vary. Use if you have a small number of values missing
      - Synthetic Minority Over-sampling Technique (SMOTE) adds synthetic data points to the minority classes. Great option when datasets are not fully populated with data from minority groups
- - - - use Gluon to distribute training onto multiple GPUs.
      - To ensure that you fully use each GPU, you must increase the mini-batch size linearly with each additional GPU. Mini-batch size has an impact not only on training speed, but also on the quality of the trained model.
      - To increase the rate of convergence with larger mini-batch size, you must increase the learning rate. Increasing the learning rate with an increase in mini-batch size has been shown to reduce the impact of having a large mini-batch on model quality.