kamran0153 / food-pantry-household-visit-forecasting

The goal of the project is to forecast the number of clients in a food pantry several weeks ahead.

Food Pantry Household Visit Forecasting

Our dataset contains 336 samples which contains number of client visits in Baptist Food Pantry, the number of COVID affected in Lee county, CPI, unemployment rate, average personal income, the number of school open days and the date of the client visit in every week.

Based on the observation of 4 most recent weeks of the data, our goal is to predict the number of client visits next week. In this prediction problem, the dataset is split into 80% of the training data and 20% of the inference data.

Procedure 1

The client visit forecasting algorithms used in this experiment are listed as follows, which include 7 machine learning algorithms and two baseline algorithms:

1. Linear Regression
2. LASSO (Regularization Parameter = 0.01)
3. Neural Network Based Regression (One Hidden Layer with 100 ReLU nodes)
4. Gradient Boosting (Max Depth 11)
5. Decision Tree Regression (Max Depth 19)
6. Bayesian Ridge Regression
7. Ridge Regression (Regularization Parameter 0.5)
8. Baseline Algorithm 1: The number of visits 2 months ago
9. Baseline Algorithm 2: The average number of visits during the last two months

Procedure 2

Data Generation Using Variational Auto Encoder (VAE):

By using 336 samples and a variational auto encoder, we have generated 835 more data samples. The total dataset is split into 80% for training and 20% for inference. The encoder network has one intermediate layer with 10 ReLU nodes and 2 output nodes. The output of the encoder is the input of the decoder. The decoder has one intermediate layer with 10 linear nodes. The loss function is the addition of reconstruction loss which is the cross-entropy loss and KL divergence loss. After that, we have used above-mentioned 9 client visit forecasting.

Performance

The client visit forecasting performance for the Lakeview Baptist Food Pantry is shown in the following table, where the forecasting accuracy is measured by using Mean Absolution Percentage Error (MAPE).

Algorithm	Training Error	Inference Error	Training Error with VAE	Inference Error with VAE
Linear Regression	38.06 %	45.08 %	25.45 %	26.4 %
LASSO	40.16%	43.35%	24.7%	26.6%
Neural Network Based Regression	33.91%	44.13%	24.5%	25.67%
Gradient Boosting	0%	38.44%	19.11%	26.6 %
Decision Tree Regression	0%	41.98 %	25.42%	26.51%
Bayesian Ridge	43.07%	42.36%	25.43%	26.4%
Ridge Regression	41.8%	41.6%	25.4%	26.62%
Baseline Algorithm 1		71.4%
Baseline Algorithm 2		57%