
1. Data Preparation
1.1. Unbalanced Classes
1.1.1. Collect more data
1.1.2. Change performance metrics
1.1.2.1. Confusion matrix
1.1.2.2. Precision
1.1.2.3. Recall
1.1.2.4. F1
1.1.2.5. Kappa
1.1.2.6. ROC Curves
1.1.3. Resampling data
1.1.3.1. Up sampling
1.1.3.1.1. 'Oversampling'
1.1.3.2. Down sampling
1.1.4. Generate Synthetic samples
1.1.4.1. SMOTE
1.1.5. Try different algorithms
1.1.6. Try Penalized Models
1.1.6.1. e.g. penalized-LDA
1.1.6.2. Weka CostSensitive wappers
1.1.7. Try different approaches
1.1.7.1. Anomaly detection
1.1.7.2. Change detection
1.1.8. Get creative
1.1.8.1. Split into smaller problems
1.2. Scaling/normalizing (feature scaling)
1.2.1. best for numeric variables which are on different scales (e.g. height = 178m, score = 10,000, shoesize = 5).
1.2.2. This will make gradient descent work much better! as less back and forth as it tries to find local minimum between the parameters.
1.2.3. Many variations but generally we want to get all features into approximately a -1 < x < 1 range
1.2.4. MEAN NORMALIZATION: X - Xu / Xmax - Xmin
1.2.4.1. Will have a Xu ~= 0
1.2.4.2. Can also use standard deviation as denominator (X / s)
1.3. Feature construction
2. NLP
2.1. Text preparation
2.1.1. Remove punctuation
2.1.2. Lower case
2.1.3. Tokenize words
2.1.4. Remove stop words
2.1.5. Remove blanks
2.1.6. Remove single letter words
2.1.7. Remove/translate non-english words
2.1.8. Stemming/Lemitization
2.1.8.1. Snowball ()
2.2. Tool-kits
2.2.1. NLTK (python)
2.2.2. Spacy (python)
2.3. Text classifiction
2.3.1. 1. Prepare text data (see text preparation)
2.3.2. 2. CountVectorize each feature (word) into a matrix
2.3.3. 3. Apply TD-IDF (Term Frequency, Inverse-Term-Frequency) to account for different length of documents
2.3.4. 4. Split data set into variables (countvector of text) and target (category label of the text)
2.3.5. 5. Deploy standard ML classification process (model, evaluate, iterate/tune)
2.4. Topic modelling
2.5. Entity recognition
3. Network Analysis
3.1. Metrics
3.1.1. Centricity
3.1.2. Betweenness
3.2. Data format
3.2.1. Node_df = NAME, NODE_ATTRIBUTE_1 ,NODE_ATTRIBUTE_N Relation_df = FROM, TO, EDGE_ATTRIBUTES_1, EDGE_ATTRIBUTE_N
4. Optimization algorithms
5. Front-End (Web application tools)
5.1. Flask (python)
5.2. Shiny (R)
5.3. Dash (Python)
5.4. Tableau
5.5. Carto
5.6. Angular/React (JS)
5.7. Django (python)
6. Data gathering
6.1. APIs
6.2. Web Scrapers
6.2.1. Selenium/PhantonJS
6.2.1.1. Good when info is behind JS or when you need to interact with the browser (e.g. login as a human)
6.2.2. BeautifulSoup
6.2.2.1. Simple scraper than you can use directly in a python script
6.2.3. Scrapy
6.2.3.1. 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.
6.3. Manual Labeling
6.3.1. Manual
6.3.2. Services
6.3.2.1. Mechanical turk (etc.)
6.3.3. Exotic sampling
6.4. Major file types
6.4.1. CSV
6.4.2. JSON
7. HL Programming Languages
7.1. R
7.2. Python
7.2.1. Vectorization
7.2.1.1. Matrix / for loops
7.2.1.1.1. 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
7.3. Octave
8. Linear Algebra
8.1. Vector/Matrix operations
8.1.1. Matrix/Matrix Addition
8.1.2. Matrix/Matrix Multiplication
8.1.3. Matrix/Vector Multiplication
8.2. Matrix properties
8.2.1. Matrices are not commutative (A*B != B*A)
8.2.2. Matrices are associative (A*B)*C = A*(B*C)
8.2.3. Matrices with the identity matrix are commutative (AI = IA)
8.2.4. SHAPE(M) = ALWAYS Row,Columns (R,C) (e.g. 2,3)
8.3. Inverse & Transposed Matrices
8.3.1. Inverse: A*A^-1 = A^-1*A = I
8.3.1.1. (A^-1 is the inverse matrix of A, though not all matrices have an inverse)
8.3.2. 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.
8.3.2.1. X
9. Statistics
9.1. Distributions
9.1.1. Gaussian (normal) distribution
9.1.1.1. Described by the mean (u) and variance (σ2) - middle is mean, width is 95% in 2σ
9.1.1.1.1. https://upload.wikimedia.org/wikipedia/commons/7/74/Normal_Distribution_PDF.svg
9.1.1.2. 'Bell shaped curve'
9.1.1.3. probability distribution = 1
9.2. Statistical tests
9.2.1. t-test
9.2.2. ANOVA
10. Time-series analysis
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. Data Vizualisation libraries
12.1. GGPLOT2(R)
12.2. MATPLOTLIB(PYTHON)
12.3. SEABORN(PYTHON)
12.4. PLOTLY(PYTHON)
13. Machine Learning
13.1. Generic ML approaches
13.1.1. ML Diagnostics (assess algorithms)
13.1.1.1. Over-fitting (high variance)
13.1.1.1.1. 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
13.1.1.2. Under-fitting (high bias)
13.1.1.2.1. 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
13.1.1.3. Approaches
13.1.1.3.1. Cross-validation
13.1.1.3.2. Learning curves
13.1.1.3.3. General diagnostic options
13.1.2. Generic ML algorithm Methodology
13.1.2.1. Input: x, the input variable that predicts y
13.1.2.2. target: y, a labelled outcome
13.1.2.3. hypothesis: h(x), the function line that is a function of x
13.1.2.4. Parameter: θ, the parameter(s) we choose with the objective of minimising the cost function
13.1.2.5. 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.
13.1.2.5.1. https://raw.githubusercontent.com/ritchieng/machine-learning-stanford/master/w1_linear_regression_one_variable/2_params.png
13.1.2.5.2. e.g. RMSE
13.1.2.6. Goal: minimize J(θ), the goal of the algorithm to minimize the error of the cost function through changing the parameters
13.1.2.7. Gradient decent (cost reduction mechanism): Repeat θj := θj - α dθj/d J(θ)
13.1.2.7.1. := assignment operator, take a and make it b
13.1.2.7.2. 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.
13.1.2.7.3. Simultaneously updates all parameters!
13.1.2.7.4. 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).
13.1.2.7.5. 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)
13.1.2.7.6. Pros: works well even when you have a large number of features - so scales well.
13.1.2.7.7. Cons: you need to choose a learning rate (α) and you need to do lots of iterations
13.1.2.7.8. There are however other ways of solving this problem
13.1.2.8. Prediction: a predict value of y using a new x sample and a θ trained by reducing the cost function for the training set
13.1.3. The phenomenon of increasing training data
13.1.3.1. X 2001
13.1.3.2. 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)
13.2. Supervised (predictive models)
13.2.1. Classification models
13.2.1.1. Performance Metrics
13.2.1.1.1. Confusion matrix http://www.dataschool.io/content/images/2015/01/confusion_matrix2.png
13.2.1.1.2. Simple Metrics
13.2.1.1.3. Advanced Metrics
13.2.1.1.4. Other considerations
13.2.1.2. Classification Model Types
13.2.1.2.1. Logistic Regression
13.2.1.2.2. SVMs
13.2.1.2.3. KNN
13.2.1.2.4. Decision Trees
13.2.1.2.5. Random Forest
13.2.1.2.6. XGBoost
13.2.1.3. Classification types
13.2.1.3.1. Binary class
13.2.1.3.2. Multi class
13.2.2. Regression models
13.2.2.1. Performance Metrics / Cost function
13.2.2.1.1. 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.
13.2.2.1.2. 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.
13.2.2.1.3. Cost functions
13.2.2.2. Regression Model Types
13.2.2.2.1. Linear Regression
13.2.2.2.2. Decision Trees for Regression
13.2.2.2.3. Random Forest for Regression
13.2.3. Reinforcement models
13.2.3.1. Performance Metrics
13.2.3.2. Neural Networks
13.2.3.2.1. Architectures
13.2.4. Ensemble modeling
13.2.4.1. Definition
13.2.4.1.1. Ensembling is a technique of combining two or more algorithms of similar or dissimilar types called base learners
13.2.4.2. Types
13.2.4.2.1. Averaging:
13.2.4.2.2. Majority vote:
13.2.4.2.3. Weighted average:
13.2.4.3. Methods
13.2.4.3.1. Bagging
13.2.4.3.2. Boosting
13.2.4.3.3. Stacking
13.2.4.4. Advantages/Disadvantages of ensembling
13.2.4.4.1. Advantages
13.2.4.4.2. Disadvantages
13.3. Unsupervised (descriptive models)
13.3.1. Clustering
13.3.1.1. KNN
13.3.1.1.1. Process
13.3.1.2. DBscan
13.3.1.3. Auto-encoders (Neural Nets)
13.3.2. Dimensionality reduction
13.3.2.1. PCA
13.3.2.1.1. 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
14. Anomaly detection
14.1. https://raw.githubusercontent.com/ritchieng/machine-learning-stanford/master/w9_anomaly_recommender/anomaly_detection8.png
14.2. 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
14.2.1. Premise
14.2.1.1. 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
14.2.2. Process
14.2.2.1. 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).
14.2.2.1.1. 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
14.2.2.2. 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
14.2.2.3. 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
14.2.2.4. 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!
14.2.3. Pros (supervised / anomaly detection)
14.2.3.1. 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)
14.2.3.2. 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
14.2.4. Examples
14.2.4.1. Spam detection
14.2.4.2. Manufacturing checks
14.2.4.3. Machine/data monitoring
14.2.5. Formula for p(x)
14.2.5.1. 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
14.2.5.1.1. https://yyqing.me/2017/2017-08-09/anomaly-detection.png
14.2.5.2. Assumes features are Normally distributed (x~(u,s2)
14.2.5.2.1. To check this assumption more-or-less holds true it is highly recommended to graph the features first
14.2.5.2.2. Even if this does not hold true AD algorithms generally work OK
14.2.6. Multivariate Gaussian Distribution (AD)
14.2.6.1. Premise
14.2.6.1.1. 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)
14.2.6.2. Formula
14.2.6.2.1. https://notesonml.files.wordpress.com/2015/06/ml51.png
14.2.6.3. Advantages
14.2.6.4. Disadvantages
15. Recommendation engines
15.1. Content based
15.2. Collaborative filter
16. Big Data
16.1. Big Data technologies
16.1.1. Hadoop
16.1.2. Spark
16.2. ML on large datasets
16.2.1. Gradient descent
16.2.1.1. Stochastic gradient descent
17. Labeling Data
17.1. Manual Labeling
17.1.1. Calculate approximate time it would take (e.g. 10s to label one, ergo...)
17.2. Crowd Source
17.2.1. E.g. Amazon Mechanical Turk / Chiron
17.3. Synthetic Labeling
17.3.1. 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)