Data Science Interview Preparation Guide

Explore the bias-variance tradeoff by defining bias and variance, illustrating underfitting with high bias and overfitting with high variance, and identifying the optimal model complexity that minimizes total error.

Use the elbow method to determine the optimal number of clusters in k-means by analyzing the distortion—the sum of squared distances to cluster centers.

Data science lecture on k-means highlights its simple, fast, and explainable clustering, while noting the need to choose k, sensitivity to initialization and outliers, struggles with non-spherical shapes, and standardization.

Explain how the ROC curve plots the true positive rate against the false positive rate, linking sensitivity and specificity, and summarize classifier performance with the area under the curve (AOK).

Explain the difference between type one error and type two error, including false positives and false negatives, through the null hypothesis framework and the shepherd and wolf example.

Learn how k-fold cross validation estimates model skill by shuffling data, splitting it into k groups, training on k-1 groups, and testing on the remaining group to produce evaluation scores.

Discover when to use random forests instead of SVMs, emphasizing feature importance, faster and simpler training, and scalable multiclass handling for data science interview preparation.

Dimensionality reduction lowers model variance and overfitting by reducing features, speeds training, and saves storage. It also mitigates multicollinearity, aids visualization in 2D–3D, and combats the curse of dimensionality.

Learn the key assumptions of linear regression, including representative samples, linear X–Y relationship, homoscedastic residuals, independent observations, and normally distributed errors.

Perform data profiling, visualize distributions, and standardize or normalize features to prepare data for machine learning, while handling null values, outliers, and removing duplicates and irrelevant features.

Understand multicollinearity in multiple regression, where highly correlated independent variables count the same feature twice and undermine statistical significance. Use VIF with a threshold of five to remove one variable.

Evaluate algorithms on unbalanced data using metrics like precision, recall, F1, MCC, and ROC AUC to avoid misleading accuracy. Use undersampling, oversampling, SMOTE, or cost-sensitive models to balance classes.

Explain why aggregating percentiles across time windows is unreliable. Averaging percentiles yields meaningless results, and the 100th percentile equals the max; accurate percentiles require all raw data.

Use forward chaining for time series cross-validation, training on past data and testing on future data to respect chronological order and avoid random k-fold splits.

Unlock the non-seasonal arima components: P (autoregressive lags), d (differencing), and Q (moving average order), and see how 000 yields white noise while 011 and 022 relate to smoothing models.

Keep models simple to reduce variance and noise, using a basic baseline to gauge expected error. Apply k-fold cross validation and L1/L2 regularization to prevent overfitting and shrink parameters.

Learn to handle missing data with deletion or imputation, using mean, median, mode, interpolation, and multiple imputation, plus time series considerations and predictive models like logistic regression and ANOVA.

Explore the kernel trick: kernel functions compute inner products in a feature space, enabling higher dimensionality without explicit coordinates and cheaper decision surfaces.

Compare gradient descent and stochastic gradient descent: gradient descent suits small datasets, while SGD converges faster on large datasets; in practice, use SGD for speed and memory efficiency.

Learn to determine whether offering coupons affects purchase decisions by using one-way ANOVA to compare means across three groups (two coupon types and no coupon) using the F distribution.

Explore the central limit theorem: as sample size grows, the sampling distribution of the means approaches a normal curve, regardless of population shape, guiding hypothesis testing and confidence intervals.

Sampling is a technique that selects a representative subset of data to identify patterns in a population, using probability-based methods with random numbers or non-probability approaches such as consecutive sampling.

Explore four probability-based sampling methods: simple random sampling, systematic sampling, stratified sampling, and cluster sampling, and learn how each method selects subjects from a population.

Explore four nonprobability sampling methods: convenience, consecutive, purposive or judgmental, and quota sampling, and explain how they work, including bias concerns and the absence of random sampling.

Explore the advantages and disadvantages of neural networks and deep learning, including state-of-the-art performance across image, language, translation, speech, and time series tasks, and challenges like interpretability and data needs.

The popularity of deep learning rises from vast data and faster GPUs, enabling the training of deep neural networks for state-of-the-art results in image recognition.

Explore why convolutional layers excel with images by exploiting local structure and translation equivariance, unlike fully connected layers that ignore spatial shifts and lack weight sharing, requiring many parameters.

Pooling, including max pooling, drives translation invariance in CNNs by summarizing nearby outputs, so a shifted input yields similar pooled values, while convolution and activation create features used by pooling.

Explore the four main CNN layers: convolutional, ReLU, pooling, and fully connected, and how they build feature learning toward a softmax-classified image output.

Explore the three types of spatial pooling, max pooling, average pooling, and sum pooling, and how they downsample feature maps to reduce dimensionality while retaining important information.

Explore how the stride in convolutional layers determines how far the filter moves over the input matrix, with stride one and two examples, and pooling options.

Explore why segmentation cnns use an encoder-decoder structure, where the encoder extracts features and the decoder upscales to classify each pixel into road, footpath, sky, cars, and pedestrians.

Batch normalization stabilizes training by normalizing each layer's mini-batch inputs to zero mean and a standard deviation of one, accelerating learning by feeding next layer outputs normalized before activation.

Use many small 3x3 kernels to achieve the same receptive field with fewer parameters and computations, while enabling more filters and activation functions for a more discriminative CNN.

Explain how generative adversarial networks, consisting of a generator and a discriminator, train to produce images indistinguishable from real ones by fooling the discriminator through their competing objectives.

Apply the softmax non-linearity as the last network operation to convert a real-valued vector into a probability distribution by exponentiating each element and normalizing by the sum of exponentials.

Explore the relu activation function, noting its computational efficiency and non-linear backpropagation-friendly nature, while recognizing the dying relu drawback where gradients vanish for zero or negative inputs, hindering learning.

Learn about the leaky relu activation function, which adds a slope for negative inputs to enable back propagation and avoid dying relu, with inconsistency in applications like breast cancer screening.

Explore how backpropagation trains neural networks by performing a forward pass, computing error, propagating gradients to adjust weights, and updating them across batches and epochs.

Explore common hyperparameters shaping neural network structure, including hidden layer count, dropout probability and layer-wise dropout, activation functions like sigmoid and ReLU, and weight initialization methods.

Explore four hyperparameter tuning methods for machine learning, including manual tuning, grid search, random search, and Bayesian optimization, and learn how each affects model accuracy.

Explore neural network architectures by analyzing input, hidden, and output cells alongside memory and convolutional layers. Understand how a classic feedforward network uses backpropagation to model logic gates.

Explore how recurrent neural networks introduce time-dependent memory cells that feed back through time, enabling sequence processing and auto completion. Understand gradient challenges and upcoming solutions like LSTMs and GRUs.

Compare gated recurrent units with lstms by examining the middle cell, its update and reset gates, and the speed-expressiveness trade-offs in time series forecasting and auto completion.

Explain variational autoencoders, highlighting probabilistic hidden cells, Bayesian inference and reparameterization, and discuss their generative applications in images, faces, music, molecules, and drug design.

Explore Hopfield networks, a fully connected recurrent architecture where every neuron connects to all others, with fixed weights that converge to learned patterns as associative memory for pattern recognition tasks.

Explore Boltzmann machines, their stochastic training with contrastive divergence, and how constrained connectivity enables practical use in combinatorial optimization, image recognition, and neural network design.

Explore echo state networks as a variant of recurrent networks with random, nonlayered connections. Train only observer-to-hidden links, with input priming, for efficient time series forecasting and sequence prediction.