Portfolio Assignment 1

Note: M3 - Group Assignment 1 Deadline: Monday, 10th Feb at 12:00

Introduction

In this assignment, you will deepen your understanding of feature engineering, data preprocessing, and neural network modeling in PyTorch. While we suggest using a house pricing dataset for illustrative purposes, you are free to choose any dataset of interest. Your goal is to build a neural network that can handle either:

  1. Cross-sectional data (i.e., data without a time component; a single “snapshot” of many observations). In this case, you will use an MLP (Multi-Layer Perceptron).
  2. Time-series data (i.e., data with a temporal or sequential component). In this case, you will use an RNN (Recurrent Neural Network) variant (e.g., simple RNN, LSTM) to capture the temporal relationships.

Your primary objectives are:

  1. Preprocess and engineer features from your chosen dataset.
  2. Build, train, and evaluate at least one neural network architecture in PyTorch:
    • MLP if you have cross-sectional data.
    • RNN if you have time-series/sequential data.
  3. Experiment with hyperparameters (e.g., number of layers/neurons, activation functions, learning rates, epochs, etc.).
  4. Evaluate and discuss your results, using appropriate performance metrics.

Task

  1. Feature Selection

    • Identify the most relevant features (columns) for predicting your target variable.
    • Provide a rationale for why these features are important or hypothesized to be predictive.

  2. Feature Engineering

    • Create or transform features that might enhance predictive power.
    • If you have time-series data, consider creating lag features or rolling statistics for the RNN.

  3. Data Preprocessing

    • Handle missing values (e.g., imputation, removal).
    • Encode categorical features (e.g., one-hot encoding).
    • Normalize/standardize numerical features (e.g., Min-Max or StandardScaler).

  4. Train-Test (or Train-Validation-Test) Split

    • Split your data into appropriate sets.
    • Consider using a validation set or cross-validation to fine-tune hyperparameters.

  5. Define Your Neural Network Architecture in PyTorch

    • MLP (if your data is purely cross-sectional):

      • At least 2 hidden layers.
      • Common activation functions (e.g., ReLU, Sigmoid, Tanh).
      • Consider using dropout or batch normalization if helpful.

    • RNN (if your data has a temporal/sequential dimension):

      • At least 2 hidden layers.
      • Use RNN, LSTM cells.
      • Experiment with different numbers of hidden units and sequence lengths.

  6. Training Loop

    • Implement a standard PyTorch training loop.
    • Specify an appropriate loss function (e.g., MSE for regression, CrossEntropy for classification).
    • Choose an optimizer (e.g., Adam, SGD).
    • Track metrics (training/validation loss, accuracy, or other relevant metrics).

  7. Hyperparameter Experiments

    • Test different hyperparameters (e.g., hidden layers, learning rate, batch size) to observe changes in performance.
    • Record and discuss these experiments.

  8. Model Evaluation

    • Evaluate your final model(s) on the test set.
    • Use appropriate metrics (e.g., RMSE, MAE, R² for regression tasks; accuracy, F1-score, etc. for classification tasks).
    • Provide plots of the training process (loss curves, etc.) and any model diagnostics.

  9. Documentation & Discussion

    • Summarize which model and hyperparameter settings worked best.
    • Discuss possible reasons for better or worse performance.
    • Mention limitations or potential improvements.

Data

  • Dataset:

    • You may use a house pricing dataset or any other dataset of your choosing. Examples:

      • Kaggle’s House Prices for cross-sectional data.
      • A stock prices or energy consumption dataset for time-series.

    • Make sure you have enough data to train and test your model adequately.

  • Possible Features:

    • Cross-sectional: purely numeric or categorical features.
    • Time-series: sequential numeric data with timestamps.

Delivery

  1. GitHub Repository

    • Create a repository containing your work.
    • Include a README.md with a brief description of your assignment and how to run the code/notebook.

  2. Colab Notebook

    • Save your notebook in the repository.

  3. Group Work

    • You may work in groups of up to 3 members.
    • Each group member’s contribution should be briefly outlined in the README or the notebook.

  4. Technical Explainer Video

    • Record a short (~5 minutes) technical explainer video presenting your main ideas, methodology, and results.
    • You may use Panopto, OBS Studio, Loom, or any other screen-recording tool.
    • Include the video link in your submission.

  5. Submission

    • Send an email to Hamid (hamidb@business.aau.dk) with the link to your GitHub repository (and video) by the deadline.

Good luck and have fun exploring neural networks in PyTorch—whether you choose to predict house prices or tackle another dataset!