Data science foundations 10

  • Data Science Lifecycle:

    The iterative loop of Ask, Collect, Clean, Explore, Model, Communicate and Decide; each stage feeds back into earlier ones. Lec 1

  • Data Types:

    Quantitative (continuous or discrete) versus categorical (nominal or ordinal) variables; the type decides which summaries and charts make sense. Lec 1

  • Unstructured Data:

    Data that does not sit in rows and columns, such as text, images and audio; roughly 80 percent of business data. Also in PS 06. Lec 1

  • Four Types of Analytics:

    Descriptive (what happened), diagnostic (why), predictive (what will happen) and prescriptive (what to do). Lec 1

  • Correlation vs Causation:

    Two variables moving together does not mean one causes the other; a hidden third variable can drive both. Seen in: Machine Learning. Lec 1

  • Confounding Variable:

    A hidden third variable that drives both observed variables and creates a misleading association between them. Lec 1

  • Survivorship Bias:

    Looking only at the successes and ignoring the failures, which distorts conclusions (Wald's WW2 aircraft armour example). Lec 1

  • Data Ethics:

    Consent, purpose limitation, data minimisation, deletion rights and explainability when collecting and using personal data. Lec 1

  • Google Colab:

    Free cloud-hosted Jupyter notebook that runs Python in the browser with no local setup. Lec 1

  • Garbage In, Garbage Out:

    Bad input data gives bad results whatever the model; cleaning can take up to 80 percent of project time. Lec 1

Data handling and exploratory analysis 13

Classification 11

Evaluation metrics and validation 9

Regression and regularisation 10

Clustering and dimensionality reduction 14

Recommendation systems 18

Natural language processing 9

Image analysis 6

  • Image Analysis:

    Extracting information from visual data (photos, video, scans), also called computer vision; images are already grids of pixel numbers. Also in PS 06. Lec 7

  • Pixel:

    Smallest image unit, holding a brightness or colour value. Lec 7

  • Filter:

    Small kernel grid of weights slid across an image to detect a pattern such as edges. Also in PS 06. Lec 7

  • Convolution:

    Slide a filter over an image, multiply with the pixels beneath, and sum to one output value. Also in PS 06. Lec 7

  • Feature Map:

    The output grid produced by applying one filter across an image. Lec 7

  • Convolutional Neural Network:

    Deep network whose filters are learned from data instead of hand-designed. Lec 7

Deep learning 17

  • Deep Learning:

    Machine learning with many layers that learn useful features automatically. Lec 8

  • Neural Network:

    Layers of connected units that transform inputs into outputs through learned weights. Also in PS 07. Lec 8

  • Perceptron:

    One unit: a weighted sum of its inputs plus a bias, passed through an activation. Lec 8

  • Weight:

    Number controlling how much one input matters to a neuron; learned in training. Lec 8

  • Bias:

    Constant added before the activation; shifts the decision boundary. Lec 8

  • Activation Function:

    Non-linear function applied after the weighted sum, which is what lets networks learn curves. Also in PS 07. Lec 8

  • ReLU:

    Activation max(0, z): passes positive values, zeroes negatives. Lec 8

  • Multi-Layer Perceptron:

    Network of fully connected layers: input, one or more hidden layers, output. Also in PS 07. Lec 8

  • Hidden Layer:

    Middle layer whose neurons learn increasingly abstract features. Lec 8

  • Forward Pass:

    Data flowing from input to output through the layers to produce a prediction. Lec 8

  • Width vs Depth:

    Neurons per layer (width) versus number of hidden layers (depth). Lec 8

  • Loss Function:

    Measures how wrong a prediction is; training tries to minimise it. Lec 8

  • Gradient Descent:

    Repeatedly step the weights downhill on the loss surface. Seen in: Machine Learning. Lec 8

  • Learning Rate:

    Size of each gradient descent step: too high overshoots, too low crawls. Lec 8

  • Epoch:

    One full pass through the training set. Lec 8

  • Backpropagation:

    Tracing the error backward through the network to work out how to adjust each weight. Lec 8

  • Early Stopping:

    Stop training when validation loss stops improving, to avoid overfitting. Also in PS 07. Lec 8