https://github.com/z-fran/nzmsa-phase2-2024

https://github.com/z-fran/nzmsa-phase2-2024

Last synced: 3 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/z-fran/nzmsa-phase2-2024
• Owner: Z-Fran
• Created: 2024-06-17T06:13:11.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2025-03-26T13:03:17.000Z (3 months ago)
• Last Synced: 2025-03-26T14:24:26.637Z (3 months ago)
• Language: Jupyter Notebook
• Size: 105 MB
• Stars: 0
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# NZMSA-Phase2-2024 

Implementation of [Microsoft Student Accelerator New Zealand Programme 2024](https://github.com/NZMSA/2024-Phase-2)

## Part1 Sales Forecasting: 1_Analysis_and_Preprocessing

### 1. Fina all variables

- Merge three dataframes into one by matching store id.

### 2. Clean data

- Deal with missing values

  - **Filling missing values** of markdowns with 0 is better than deleting them.

  - `0` is meaningful because it can presents there is no markdown on that date.

- Deal with outliers

  - Find outliers by **Boxplot** and **IQR** method.

  - Deal with outliers of unemployment by replacing with nearest values.

### 3. Feature Engineering

- Map `Type` and `IsHoiday` to numeric features.

- Split Date into Year, Month and Day so that it will be numeric.

- Use Dept and Size **create new meaningful features** according to correlation anlysis:

  - `Dept_num`: departments quantity of a store may affects sales.

  - `Size_per_dept`: average size of departments of a store may affects sales.

### 4. EDA with Visualization with Power BI

- Use pie charts to visualize the proportion of each Type, Year, Month, and IsHoliday.

    - the distribution of Year and Month are balanced

    - the distribution of Type and IsHoliday is not balanced, which may affect the model performance.

- Visualize sales varying with dates of a specific Dept of a specific Store by filtter of Power BI.

    - sales data has a contrary relationship with temperature.

    - sales data has not obvious relationship with fuel prices and unemployment.

### 5. Correlation analysis

-  Size(0.24), Dept(0.15) and Type(-0.18) have stonger correation with Weekly_Sales.

-  New feature Dept_num(0.16) and Size_per_dept (0.24) have stonger correation with Weekly_Sales.

-  The data of `Type=0` and `Type=1` are similar, but they are different from `Type=2`. We can try reclassifying them into two types. 

## Part1 Sales Forecasting: 2_Train_and_Evalution

### 1. Split dataset

- Use high correlation columns in part 1 correlation analysis

- Fix zero values of Weekly_Sales.

- Split dataset acoording to date. train:test = 4:1.

  

### 2. Train and Test models

- Train LinearRegression, KNN, RandomForest, and get scores on test dataset.

### 3. Evalution

- Compare MAE, MAPE, RMSE, R2 metrics.

  - Performance of KNN and RandomForest are similar and they are vey higher than LinearRegresiion

### 4. Ensemble Learning

- Use KNN and RandomForest to build a VotingRegressor which has a better performance.

### 5. Visualize predicted data

- Using all columns to train have a better curve fitting

## Part2 Image Classification: 3_Deep_Learning

### 1. Data preprocessing & augmentation

- RandomResizedCrop: and RandomHorizontalFlip are useful.

- RandomRotation, RandomVerticalFlip and Normalizea are unuseful.

### 2.Define the model

- ResNet

  - Classic CNN network.

  - Very easy to achieve 90%+ score on this dataset.

- ViT

  - Use transformer framework on vision tasks.

  - Training more slowly than ResNet.

### 3. Pipelines

- Design unified pipelines to train model and test on test dataset.

### 4. Train model and hyperparameters tuning

- ResNet

  - larger framework is not useful.

  - Adam and StepLR is better.

  - larger learning rate is batter (1e-3 > 1e-4).

- ViT

  - Because of slow convergence speed, MultiStepLr can set a large learning rate in the beginning to accelarate training.

  

### 5. Evaluation

  - Compare ResNet and ViT with following metrics:

    - Precision

    - Recall

    - F1 score

    - Overall Accuracy

    - ROC curve

    - AUC

    - Confusion matrix

- Accuracy and AUC of Resnet are higher than ViT.

- Both of models are more likely to confuse airplane and dog.

### Ensemble Learning

- Using Ensemble Learning method, use Resnet and ViT vote for the pridiction result.

- Weights of them is Resnet:ViT = 3:1.