In this GitHub repository, you'll find code and documentation for our Data Mining course project. Our project focuses on exploring and analyzing data (EDA) and creating regression models. We aim to provide a deep understanding of the dataset we've been given and build and test regression models using a step-by-step approach. This project allows us to gain practical experience in working with data, cleaning it, and making it ready for analysis. We also get to create and improve regression models as part of our learning process.
This dataset includes data on adult's diet, physical activity, and weight status from Behavioral Risk Factor Surveillance System. This data is used for DNPAO's Data, Trends, and Maps database, which provides national and state specific data on obesity, nutrition, physical activity, and breastfeeding. In this analysis, we have performed exploratory data analysis on the dataset and thereby partitioned the dataset in various categories and analyzed and predicted obesity rates based on those factors. We have employed various machine learning techniques along with hyperparameter tuning to improve our test results. We aim to achieve a good prediction through our model.
Our dataset contains the following atrributes :-
YearStart - Year start
YearEnd - Year End (for single-year indicator, year end=year start)
LocationAbbr - Location abbreviation
LocationDesc - Location description
Datasource - Name or abbreviation of Data Source
Class - Class Description
Topic - Topic Description
Question - Question Description
Data_Value_Unit - Description of unit e.g. %, etc
Data_Value_Type - Description of type of data e.g. Value, Percentage, Number
Data_Value - Data value (percentage, text)
Data_Value_Alt - Numeric representation of data value
Data_Value_Footnote_Symbol - Symbol that would be used to flag footnotes
Data_Value_Footnote - Footnote text
Low_Confidence_Limit - Low 95% Confidence Interval value
High_Confidence_Limit - High 95% Confidence Interval value
Sample_Size - Sample Size
Total - Total/Overall breakout category
Age(years) - Age (years) breakout category
Education - Education breakout category
Gender - Gender breakout category
Income - Income breakout category
Race/Ethnicity - Race/Ethnicity breakout category
GeoLocation - Latitude & Longitude to be provided for formatting GeoLocation or Geocode in the format (latitude, longitude)
ClassID - Lookup identifier value for Class
TopicID - Lookup identifier value for Topic
QuestionID - Lookup identifier value for Question
DataValueTypeID - Lookup identifier value for Data_Value_type
LocationID - Lookup identifier value for Location
StratificationCategory1 - Lookup Identification value, such as Age Group, Gender
Stratification1 - Data stratified by this value, such as Male, Female, Total
StratificationCategoryId1 - Lookup identifier value for Stratification1
StratificationID1 - Lookup identifier value for StratificationCategory1
Note: This dataset description was sourced from- https://chronicdata.cdc.gov/Nutrition-Physical-Activity-and-Obesity/Nutrition-Physical-Activity-and-Obesity-Behavioral/hn4x-zwk7
- The EDA provided valuable insights into our dataset, enabling us to proceed with confidence to the prediction phase.
- We observed patterns in the distribution of Data_Value, Data_Value_Footnote, LocationID, and StratificationCategory1, which will inform our modeling efforts.
- Obesity rates were analyzed across states, revealing varying rates and highlighting states with the highest obesity prevalence.
- Different classes (Obesity/Weight Status, Fruits and Vegetables, Physical Activity) showed variations in average obesity rates, suggesting the influence of topic categories.
- We noted consistent obesity rates across various subgroups.
- High correlations among certain variables guided our decision to drop redundant features for model simplification.
These findings collectively equip us for the next phase of our project, where we will build predictive models based on these insights.
The dataset was divided based on stratification categories, and missing values were handled by removing irrelevant features.
- Histograms and boxplots helped analyze data distribution and range.
- Bar plots revealed insights such as an inverse relationship between income and obesity rates, variation in rates among racial/ethnic groups, lower obesity rates with higher education levels, and lower rates for females compared to males.
- Scatter plots showed that the non-Hispanic white demographic group had the most samples with large sample sizes, females had lower obesity rates than males, and lower data values correlated with higher income and education levels.
Additionally, question and class counts were examined, and correlation analysis with label encoding was performed. Notably, 'low_confidence_limit' and 'high_confidence_limit' were highly correlated and should be excluded from modeling to prevent overfitting. These findings will guide feature selection in subsequent machine learning tasks.
Note: Several plots in our EDA_category.ipynb file utilize the Plotly library, which creates interactive visualizations not viewable directly in the GitHub file. To access and interact with these visualizations, you can simply click the "Open in Colab" link provided at the beginning of the document.
In this module, we engage in data processing tasks, which encompass the selection of essential columns and the handling of missing values. Drawing from the insights gained through both of the conducted Exploratory Data Analysis (EDAs), we proceed to partition our datasets into six subdatasets, each categorized based on stratification criteria. Subsequently, we carry out additional preprocessing on each of these subdatasets, ultimately converting them into CSV files. These CSV files serve as the foundational data for our subsequent model prediction tasks.
The machine learning pipeline comprises key stages. It begins with data splitting and standardization, where we divide the dataset, standardize it, and move to model selection. In this stage, we consider models like linear regression, polynomial regression, Ridge regression, k-Nearest Neighbors (kNN), and Random Forest.
Models used for Evaluation
Model 1: Linear Regression Hyperparameters: None (No specific hyperparameters tuned)
Model 2: Polynomial Regression Hyperparameters: degree(2, 3)
Model 3: Ridge Regression Hyperparameters: Alpha (Regularization strength): [0.01, 0.1, 1.0]
Model 4: K-Nearest Neighbors Regression Hyperparameters: Number of Neighbors (n_neighbors): [3, 5, 7]
Model 5: Random Forest Regression Hyperparameters: Number of Estimators (n_estimators): [100, 200, 300] Maximum Depth of Trees (max_depth): [5, 10, 20, 30] Minimum Samples for Split (min_samples_split): [2, 5, 10]
Subsequently, hyperparameter tuning and model evaluation optimizes model hyperparameters using GridSearchCV, with evaluation on the validation set. Selecting the best model and evaluating on the test set follows, where the top-performing model on the validation set is assessed on an independent test dataset. Finally, model comparison showcases results, highlighting the best hyperparameters and critical metrics like RMSE, MAE, and R2 Score. These stages encompass data preparation, model selection, and thorough evaluation.
In summary, our project was a deep exploration of a dataset. We began by looking closely at the data and finding important patterns and relationships through graphs and analysis. This helped us get ready for the next step, where we used what we learned to make predictions.
After that, we carefully got the data ready, making sure it was in good shape for our analysis. We also split the data into groups based on certain characteristics. This part was important for our machine learning.
The machine learning part involved trying different models, like Linear Regression and others, to see which one worked best. We made these models even better by adjusting their settings. We tested them on some data we kept separate to see how well they could predict obesity rates.
Our results showed that our approach works well. It helps us understand and predict obesity rates for different groups of people. This isn't just interesting; it can also be helpful for future research and efforts to improve public health.