1. NLP
1.1. Text preparation
1.1.1. Remove punctuation
1.1.2. Lower case
1.1.3. Tokenize words
1.1.4. Remove stop words
1.1.5. Remove blanks
1.1.6. Remove single letter words
1.1.7. Remove/translate non-english words
1.1.8. Stemming/Lemitization
1.1.8.1. Snowball ()
1.2. Tool-kits
1.2.1. NLTK (python)
1.2.2. Spacy (python)
1.3. Text classifiction
1.3.1. 1. Prepare text data (see text preparation)
1.3.2. 2. CountVectorize each feature (word) into a matrix
1.3.3. 3. Apply TD-IDF (Term Frequency, Inverse-Term-Frequency) to account for different length of documents
1.3.4. 4. Split data set into variables (countvector of text) and target (category label of the text)
1.3.5. 5. Deploy standard ML classification process (model, evaluate, iterate/tune)
1.4. Topic modelling
1.5. Entity recognition
2. Network Analysis
2.1. Metrics
2.1.1. Centricity
2.1.2. Betweenness
2.2. Data format
2.2.1. Node_df = NAME, NODE_ATTRIBUTE_1 ,NODE_ATTRIBUTE_N Relation_df = FROM, TO, EDGE_ATTRIBUTES_1, EDGE_ATTRIBUTE_N
3. Front-End (Web application tools)
3.1. Flask (python)
3.2. Shiny (R)
3.3. Dash (Python)
3.4. Tableau
3.5. Carto
3.6. Angular/React (JS)
3.7. Django (python)
4. Linear Algebra
4.1. Vector/Matrix operations
4.1.1. Matrix/Matrix Addition
4.1.2. Matrix/Matrix Multiplication
4.1.3. Matrix/Vector Multiplication
4.2. Matrix properties
4.2.1. Matrices are not commutative (A*B != B*A)
4.2.2. Matrices are associative (A*B)*C = A*(B*C)
4.2.3. Matrices with the identity matrix are commutative (AI = IA)
4.2.4. SHAPE(M) = ALWAYS Row,Columns (R,C) (e.g. 2,3)
4.3. Inverse & Transposed Matrices
4.3.1. Inverse: A*A^-1 = A^-1*A = I
4.3.1.1. (A^-1 is the inverse matrix of A, though not all matrices have an inverse)
4.3.2. Transpose: A -> AT (where A is a m*n matrix and AT is an n*m, where Aij = ATji) First column becomes first row basically.
4.3.2.1. X
5. Data Vizualisation libraries
5.1. GGPLOT2(R)
5.2. MATPLOTLIB(PYTHON)
5.3. SEABORN(PYTHON)
5.4. PLOTLY(PYTHON)
6. Anomaly detection
6.1. https://raw.githubusercontent.com/ritchieng/machine-learning-stanford/master/w9_anomaly_recommender/anomaly_detection8.png
6.2. Can be an unsupervised problem (looking for points with high p(x) standard deviation away from the mean of many of the features), from but mostly setup as a supervised problem with a training set with labels of anomalies
6.2.1. Premise
6.2.1.1. Premise: assume features follow normal distribution. Find the u, sd & p(x) for each feature and use this to create new derived p(x) features. Then use these to predict anomalies
6.2.2. Process
6.2.2.1. Create a 'good training' set with 60% of all non-anomaly (y=0) examples and use this to create p(x) derived features from each of the original features (see formula p(x) below).
6.2.2.1.1. If you complete this process and still find anomaly y=1 samples which are not detected then it is a good idea to look into these specific example to see if there are new derived features that can be create to help detect it
6.2.2.2. Put the remaining 20% of non-anomalously records with 50% of the anomalously records (y=1) into a training set, and the last 20% of non-anomalously and last 50% of anomalously records into a test set
6.2.2.3. Use 'good training' set to create the p(x) derived features, use the training set to predict y=0 good, y=1 anomaly, and optimize the model, then finally use test set to do cross-validation performance
6.2.2.4. We can then use standard supervised performance metrics to evaluate the model - though due to imbalanced classes must use a more robust metric (like F1) rather than accuracy!
6.2.3. Pros (supervised / anomaly detection)
6.2.3.1. AD preferable when we have a very small set of positive (y=1) examples (as we want to save this just for training and test set and can 'expend' many y=0 examples to fit the p(x) model)
6.2.3.2. When anomalies may follow many different 'patterns' so fitting a standard supervised model may not be able to find a good separation boundary, but the pattern of their probability distribution (i.e. the fact they are very different from normal) will be a constant pattern
6.2.4. Examples
6.2.4.1. Spam detection
6.2.4.2. Manufacturing checks
6.2.4.3. Machine/data monitoring
6.2.5. Formula for p(x)
6.2.5.1. Using set of y=0 data points create new derived features which model the original features as a normal distribution and calculate the sample mean, sd, and p(x) as new derived features
6.2.5.1.1. https://yyqing.me/2017/2017-08-09/anomaly-detection.png
6.2.5.2. Assumes features are Normally distributed (x~(u,s2)
6.2.5.2.1. To check this assumption more-or-less holds true it is highly recommended to graph the features first
6.2.5.2.2. Even if this does not hold true AD algorithms generally work OK
6.2.6. Multivariate Gaussian Distribution (AD)
6.2.6.1. Premise
6.2.6.1.1. Standard AD uses single-variance Gaussian distribution - essentially creating a circle radius of p(x) around the mean. However often it may be better to have a more complex shape around the mean - to do this we simply use a multi-var gaussian formula to calculate p(x)
6.2.6.2. Formula
6.2.6.2.1. https://notesonml.files.wordpress.com/2015/06/ml51.png
6.2.6.3. Advantages
6.2.6.4. Disadvantages
7. Time-series analysis
8. Recommendation engines
8.1. Content based
8.2. Collaborative filter
9. Labeling Data
9.1. Manual Labeling
9.1.1. Calculate approximate time it would take (e.g. 10s to label one, ergo...)
9.2. Crowd Source
9.2.1. E.g. Amazon Mechanical Turk / Chiron
9.3. Synthetic Labeling
9.3.1. Introducing distortions to smaller training set to amplify it (but only if distortions are what we would expect to find in real training set not just random noise)
10. Data Preparation
10.1. Unbalanced Classes
10.1.1. Collect more data
10.1.2. Change performance metrics
10.1.2.1. Confusion matrix
10.1.2.2. Precision
10.1.2.3. Recall
10.1.2.4. F1
10.1.2.5. Kappa
10.1.2.6. ROC Curves
10.1.3. Resampling data
10.1.3.1. Up sampling
10.1.3.1.1. 'Oversampling'
10.1.3.2. Down sampling
10.1.4. Generate Synthetic samples
10.1.4.1. SMOTE
10.1.5. Try different algorithms
10.1.6. Try Penalized Models
10.1.6.1. e.g. penalized-LDA
10.1.6.2. Weka CostSensitive wappers
10.1.7. Try different approaches
10.1.7.1. Anomaly detection
10.1.7.2. Change detection
10.1.8. Get creative
10.1.8.1. Split into smaller problems
10.2. Scaling/normalizing (feature scaling)
10.2.1. best for numeric variables which are on different scales (e.g. height = 178m, score = 10,000, shoesize = 5).
10.2.2. This will make gradient descent work much better! as less back and forth as it tries to find local minimum between the parameters.
10.2.3. Many variations but generally we want to get all features into approximately a -1 < x < 1 range
10.2.4. MEAN NORMALIZATION: X - Xu / Xmax - Xmin
10.2.4.1. Will have a Xu ~= 0
10.2.4.2. Can also use standard deviation as denominator (X / s)
10.3. Feature construction
11. Data Project Management
11.1. CRISP-DM
11.1.1. 1. Business understanding
11.1.2. 2. Data understanding
11.1.3. 3. Data preperation
11.1.4. 4. Modelling
11.1.5. 5. Evaluation
11.1.6. 6. Deployment
11.1.7. https://pbs.twimg.com/media/DNF5vACVQAAxOWD.jpg
11.2. Ceiling Analysis
11.2.1. Assess which part of the pipeline is most valuable to spend your time?
11.2.2. To do this, override each module/step with the perfect output (e.g. replace predictions with correct labels) for each module and assess where getting closer
12. Machine Learning
12.1. Generic ML approaches
12.1.1. ML Diagnostics (assess algorithms)
12.1.1.1. Over-fitting (high variance)
12.1.1.1.1. The hypothesis equation is 'over fit' to the training data (e.g. complex polynomial equation that passes through each data point) meaning it performs very well in training but fails generalize well in testing
12.1.1.2. Under-fitting (high bias)
12.1.1.2.1. The hypothesis equation is 'under fit' meaning it over generalized the problem (e.g. using a basic linear separation line for a polynomial problem), meaning if cannot identify more complex cases well
12.1.1.3. Approaches
12.1.1.3.1. Cross-validation
12.1.1.3.2. Learning curves
12.1.1.3.3. General diagnostic options
12.1.2. Generic ML algorithm Methodology
12.1.2.1. Input: x, the input variable that predicts y
12.1.2.2. target: y, a labelled outcome
12.1.2.3. hypothesis: h(x), the function line that is a function of x
12.1.2.4. Parameter: θ, the parameter(s) we choose with the objective of minimising the cost function
12.1.2.5. Cost function: J(θ) a function of the parameters that we try to reduce to get a good prediction (e.g. MSE). We can plot this to see the minimum point.
12.1.2.5.1. https://raw.githubusercontent.com/ritchieng/machine-learning-stanford/master/w1_linear_regression_one_variable/2_params.png
12.1.2.5.2. e.g. RMSE
12.1.2.6. Goal: minimize J(θ), the goal of the algorithm to minimize the error of the cost function through changing the parameters
12.1.2.7. Gradient decent (cost reduction mechanism): Repeat θj := θj - α dθj/d J(θ)
12.1.2.7.1. := assignment operator, take a and make it b
12.1.2.7.2. α = learning rate = how big steps to take, if it is too small then baby-steps will take a lot of time, if too big can fail to converge, or even diverge. The learning rate impact varys depending of slope of the derivative - This means that closer to convergence the steps will be smaller anyway.
12.1.2.7.3. Simultaneously updates all parameters!
12.1.2.7.4. dθj/d J(θ) = derivative function, the slope of the straight line at the tangent of the curve at each point (derivative). If slope is positive then it is θ - positive number makes θ less, if slope is negative then makes θ more until we get to a point where derivative is 0 (local minimum).
12.1.2.7.5. Sometimes called "Batch" gradient decent as it looks at all the available examples in the training set (compared to cross-validation where we look at a sub-set of samples)
12.1.2.7.6. Pros: works well even when you have a large number of features - so scales well.
12.1.2.7.7. Cons: you need to choose a learning rate (α) and you need to do lots of iterations
12.1.2.7.8. There are however other ways of solving this problem
12.1.2.8. Prediction: a predict value of y using a new x sample and a θ trained by reducing the cost function for the training set
12.1.3. The phenomenon of increasing training data
12.1.3.1. X 2001
12.1.3.2. This only holds if the features X hold enough information to predict y (i.e. predicting missing word from a specific sentence compared to trying to predict house prices from only having the square feet ... not possible even for human experts)
12.2. Supervised (predictive models)
12.2.1. Classification models
12.2.1.1. Performance Metrics
12.2.1.1.1. Confusion matrix http://www.dataschool.io/content/images/2015/01/confusion_matrix2.png
12.2.1.1.2. Simple Metrics
12.2.1.1.3. Advanced Metrics
12.2.1.1.4. Other considerations
12.2.1.2. Classification Model Types
12.2.1.2.1. Logistic Regression
12.2.1.2.2. SVMs
12.2.1.2.3. KNN
12.2.1.2.4. Decision Trees
12.2.1.2.5. Random Forest
12.2.1.2.6. XGBoost
12.2.1.3. Classification types
12.2.1.3.1. Binary class
12.2.1.3.2. Multi class
12.2.2. Regression models
12.2.2.1. Performance Metrics / Cost function
12.2.2.1.1. We can measure the accuracy of our hypothesis function by using a cost function. This takes an average difference (actually a fancier version of an average) of all the results of the hypothesis with inputs from x's and the actual output y's.
12.2.2.1.2. We can measure the accuracy of our hypothesis function by using a cost function. This takes an average difference (actually a fancier version of an average) of all the results of the hypothesis with inputs from x's and the actual output y's.
12.2.2.1.3. Cost functions
12.2.2.2. Regression Model Types
12.2.2.2.1. Linear Regression
12.2.2.2.2. Decision Trees for Regression
12.2.2.2.3. Random Forest for Regression
12.2.3. Reinforcement models
12.2.3.1. Performance Metrics
12.2.3.2. Neural Networks
12.2.3.2.1. Architectures
12.2.4. Ensemble modeling
12.2.4.1. Definition
12.2.4.1.1. Ensembling is a technique of combining two or more algorithms of similar or dissimilar types called base learners
12.2.4.2. Types
12.2.4.2.1. Averaging:
12.2.4.2.2. Majority vote:
12.2.4.2.3. Weighted average:
12.2.4.3. Methods
12.2.4.3.1. Bagging
12.2.4.3.2. Boosting
12.2.4.3.3. Stacking
12.2.4.4. Advantages/Disadvantages of ensembling
12.2.4.4.1. Advantages
12.2.4.4.2. Disadvantages
12.3. Unsupervised (descriptive models)
12.3.1. Clustering
12.3.1.1. KNN
12.3.1.1.1. Process
12.3.1.2. DBscan
12.3.1.3. Auto-encoders (Neural Nets)
12.3.2. Dimensionality reduction
12.3.2.1. PCA
12.3.2.1.1. Reduce the dimensions of a dataset by finding a plane between similar variables than can be used to express the original variables in a lower-dimensional space
13. Statistics
13.1. Distributions
13.1.1. Gaussian (normal) distribution
13.1.1.1. Described by the mean (u) and variance (σ2) - middle is mean, width is 95% in 2σ
13.1.1.1.1. https://upload.wikimedia.org/wikipedia/commons/7/74/Normal_Distribution_PDF.svg
13.1.1.2. 'Bell shaped curve'
13.1.1.3. probability distribution = 1
13.2. Statistical tests
13.2.1. t-test
13.2.2. ANOVA
14. HL Programming Languages
14.1. R
14.2. Python
14.2.1. Vectorization
14.2.1.1. Matrix / for loops
14.2.1.1.1. Matrix multiplications applied across an entire dataset is much more efficient that a for loop as do not have to reset and find memory space for each variables each time and has pre-indexed order for column vector
14.3. Octave
15. Data gathering
15.1. APIs
15.2. Web Scrapers
15.2.1. Selenium/PhantonJS
15.2.1.1. Good when info is behind JS or when you need to interact with the browser (e.g. login as a human)
15.2.2. BeautifulSoup
15.2.2.1. Simple scraper than you can use directly in a python script
15.2.3. Scrapy
15.2.3.1. Most developed and efficient scraper for large trawling. Also offers lots of functionality to customize (e.g. IP masking). Though needs to be setup with correct directory and class structures.
15.3. Manual Labeling
15.3.1. Manual
15.3.2. Services
15.3.2.1. Mechanical turk (etc.)
15.3.3. Exotic sampling
15.4. Major file types
15.4.1. CSV
15.4.2. JSON
16. Optimization algorithms
17. Big Data
17.1. Big Data technologies
17.1.1. Hadoop
17.1.2. Spark
17.2. ML on large datasets
17.2.1. Gradient descent
17.2.1.1. Stochastic gradient descent