This project aims to predict students' midterm scores based on their homework scores using two different machine learning models: Linear Regression and Long Short-Term Memory (LSTM).

Linear Regression is a statistical method to model the relationship between a dependent variable and one or more independent variables. In this project, we used students' scores from Homework 1 to 10 (HW1 to HW10) as independent variables to predict the mid1 score.

• Data Cleaning: Handled NaN values by filling them with 0.
• Model: Used LinearRegression from sklearn.
• Evaluation Metrics: Mean Squared Error (MSE) and $R^{2}$ Score. Additionally, we calculated the accuracy within a range of ±12.5 points.

• Accuracy: 0.5
• Mean Squared Error: 454.59267933162926
• $R^{2}$ Score: 0.18515071572160635

LSTM is a type of recurrent neural network (RNN) which is capable of learning long-term dependencies and is used for sequence prediction problems. Here, the sequence of homework scores is used to predict the midterm scores.

• Data Cleaning: Filled NaN values with 0.
• Model: Defined an LSTM model using PyTorch.
• Evaluation Metrics: Loss (MSE) and accuracy within a range of ±2.5 points were monitored during training.
• Special Handling: The model is saved every 1000 epochs.

• on Training Set:
  • Best_Accuracy: 0.7289
  • Epoch: 80000
• on Validation Set:
  • Best_Accuracy: 0.9412
  • Epoch: 1000

The datasets consist of scores from 10 homework assignments (HW1 to HW10) and a midterm (mid1) for a set of students, each identified by a unique randNumCol.

• Entries with a Total Score less than or equal to 1 were removed.
• Cleaned data was saved as HW{i}_clean.csv.
• The data was merged on randNumCol to create a consolidated dataset data_total.csv.

To improve the baseline models:

1. Data Imputation: NaN values were replaced with 0.
2. Data Preprocessing: Data was cleaned and only relevant features were kept.
3. Model Saving: For the LSTM model, the state was saved every 1000 epochs to monitor progress and potentially use transfer learning later.

This project showcases two different approaches to predict midterm scores based on homework scores. By comparing the performances of Linear Regression and LSTM, we can gain insights into the predictability of the midterm scores and potentially improve the teaching strategies to assist students in better preparation.