Predicting Activity Priorities with Machine Learning

Author

Rica Rebusit

Random Forest + CHC logo

My capstone project focused on using machine learning to predict activity priorities for Chico State’s Center for Healthy Communities (CHC) based on text descriptions. The CHC is a non-profit organization that promotes food security, nutrition education, and various community health initiatives.

The challenge was to develop a system that could automatically classify activities by priority level, helping the organization optimize resource allocation and better serve campus partners.

Project Goal

The Center for Healthy Communities conducts numerous activities and programs throughout the year. Understanding which activities should be prioritized is crucial for effective resource management and maximizing community impact.

The goal: Build a machine learning model that predicts activity priority based on text descriptions, using:

  • Latent Dirichlet Allocation (LDA) for topic modeling and feature extraction
  • Random Forest for classification and priority prediction
  • Natural language processing techniques for text preprocessing

The Technical Approach

The first step involved cleaning and preparing the activity description data. This included removing special characters, standardizing text format, and handling missing values.

# Load required libraries
library(tidyverse)
library(tm)
library(tokenizers)

# Data cleaning function
clean_text <- function(text) {
  text %>%
    tolower() %>%                           # Convert to lowercase
    str_replace_all("[[:punct:]]", " ") %>% # Remove punctuation
    str_replace_all("\\s+", " ") %>%        # Normalize whitespace
    str_trim()                               # Trim leading/trailing spaces
}

# Apply cleaning to activity descriptions
df_clean <- df %>%
  mutate(descrip_clean = clean_text(descrip))

Breaking down text into meaningful tokens is essential for text analysis. This process converts sentences into individual words while removing common stopwords.

library(tidytext)

# Tokenize and remove stopwords
tokens <- df_clean %>%
  unnest_tokens(word, descrip_clean) %>%
  anti_join(stop_words) %>%
  filter(!str_detect(word, "^\\d+$"))  # Remove numbers

# Create document-term matrix
dtm <- tokens %>%
  count(activity_id, word) %>%
  cast_dtm(activity_id, word, n)

LDA discovers hidden topics within the activity descriptions, creating new features for the classification model.

library(topicmodels)

# Fit LDA model with 5 topics
lda_model <- LDA(
  dtm,
  k = 5,                    # Number of topics
  method = "Gibbs",
  control = list(
    seed = 123,
    burnin = 1000,
    iter = 2000,
    thin = 100
  )
)

# Extract topic probabilities for each activity
topic_probs <- posterior(lda_model)$topics %>%
  as.data.frame() %>%
  rename_with(~paste0("topic_", 1:5))

# View top words for each topic
terms(lda_model, 10)

The Random Forest classifier uses the extracted topics and other features to predict activity priorities.

library(randomForest)
library(caret)

# Combine features: topics + other variables
model_data <- df %>%
  bind_cols(topic_probs) %>%
  select(priority, shortcode, quarter, topic_1:topic_5,
         train_meet, partner, level)

# Split data into training and testing sets
set.seed(123)
train_index <- createDataPartition(model_data$priority,
                                   p = 0.8,
                                   list = FALSE)
train_data <- model_data[train_index, ]
test_data <- model_data[-train_index, ]

# Train Random Forest model
rf_model <- randomForest(
  priority ~ .,
  data = train_data,
  ntree = 500,
  mtry = 3,
  importance = TRUE
)

# Make predictions
predictions <- predict(rf_model, test_data)

# Evaluate model performance
confusionMatrix(predictions, test_data$priority)

Key Variables

The project utilized several important variables to predict activity priorities:

  • shortcode: Brief identifier for the activity type
  • quarter: Academic term (Fall, Spring, Summer)
  • contract_year: Fiscal year of the activity
  • train_meet: Type of training or meeting conducted
  • partner: Campus or community partner organization
  • descrip: Full text description of the activity
  • level: Activity complexity or engagement level
  • topic_1 to topic_5: Probability distributions from LDA topic modeling

Impact & Applications

Question

How does this project make a difference for the Center for Healthy Communities?

Answer

Resource Optimization: Enables data-driven decisions about where to allocate staff time and budget.

Strategic Planning: Provides insights into which activity types have historically been high-priority, informing future programming.

Scalability: Automates the classification process as CHC expands its programs, saving time on manual categorization.

Evidence-Based Decision Making: Helps stakeholders understand activity impacts through quantitative analysis rather than intuition alone.

Platforms & Tools Used

  • R and RStudio: Primary environment for data analysis and model development
  • tidyverse: Data manipulation and visualization
  • tm & tidytext: Text mining and natural language processing
  • topicmodels: Latent Dirichlet Allocation implementation
  • randomForest: Machine learning classification
  • caret: Model training and evaluation
  • Quarto: Reproducible research and documentation
  • GitHub: Version control and collaboration

Lessons Learned

Working on this project taught me several valuable lessons about applied machine learning:

  • Data Quality Matters: Text data requires extensive cleaning and preprocessing. The quality of your model is only as good as the quality of your input data.

  • Feature Engineering is Crucial: LDA topic modeling transformed unstructured text into meaningful features that significantly improved model performance.

  • Iterate and Refine: The initial cleaning and tokenization process needed several iterations to handle edge cases and improve accuracy.

  • Balance Complexity and Interpretability: While more complex models might offer marginal accuracy gains, Random Forest provided a good balance of performance and interpretability for stakeholders.

Future Enhancements

Several opportunities exist to expand and improve this project:

  1. More Labeled Data: Collecting additional labeled examples would improve model accuracy and generalizability.

  2. Model Comparison: Evaluate alternative algorithms like XGBoost, Support Vector Machines, or neural networks.

  3. Real-Time Deployment: Create a Shiny application or API for real-time priority predictions.

  4. Hyperparameter Tuning: Systematic optimization of model parameters using grid search or Bayesian optimization.

  5. Cross-Validation: Implement k-fold cross-validation for more robust performance estimates.

About the Repository

The complete project code is available on GitHub and includes:

  • scripts/LDA.qmd: Data cleaning, preprocessing, and topic modeling
  • scripts/Random_Forest.qmd: Model training and evaluation
  • documentations/: Meeting notes and project documentation
  • presentation/: Poster and stakeholder reports
Warning

Note: Access to the project data requires permission from the Center for Healthy Communities to protect sensitive program information.

Thank you to Robin Donatello, the Chico State Mathematics and Statistics Department, and the Center for Healthy Communities team for their guidance and support throughout this project!