The data

To predict an individual carbon footprint from daily behavior we needed to estimate the effect different behaviors would have on the carbon footprint. Thus, we required data to determine this relationship. The data we used for this project came from a kaggle challenge on calculating the individual carbon footprint (https://www.kaggle.com/datasets/dumanmesut/individual-carbon-footprint-calculation).

Please note, however, that the data was simulated and did not stem from an investigation of an actual sample. Nevertheless, the dataset contained data on the demographic (body type, sex, diet), the daily life and activities (shower habits, social activities, spending behavior, produced waste and recycling, internet and tv usage), the energy efficiency of their household (heating source, household appliances and their efficiency), the means of transportation (mode of transportation, distance, air travel) as well as the dependent variable of carbon emission in kg CO2 emission per month of 10,000 individuals.

Exploratory Data Analysis

• We found a strong correlation (r = .78) between car usage and kilometers traveled by vehicle in the last month. Both variables also show strong correlations with the target variable carbon emission (r = .59 and r = .49, respectively). However, since all correlations were below the threshold signaling multicollinearity (r > 0.80; Field, 2024) we continued with the analysis.

Data processing

• Handling Missing Data: We created a new variable by combining two existing ones to handle the NaN values.

• Encoding: To handle the different types of variables, we utilized the
  ‘ColumnTransformer’ from scikit-learn. For the linear regression, we dummy encoded
  the categorical variables (OneHotEncoder) and scaled the discrete variables to a range
  between 0 and 1 (MinMaxScaler). For the Decision Tree model, we mapped ordinal
  variables to numerical values based on their logical order. Notably, LightGBM efficiently
  handles categorical encoding internally, eliminating the need for one-hot encoding. The
  encoders were exported for use in Streamlit.

• Data Validation: While carefully examining the dataset, we identified that "cooking with a
  grill" always coincided with "cooking with an airfryer", suggesting a possible unintended
  linkage in the simulation. Consequently, we excluded airfryer, as this redundancy did not
  contribute additional information.

Modelling

• Linear Regression: Testing the model’s assumptions we identified problems with heteroscedasticity caused probably by a bias for underestimating extreme values. All other assumptions were satisfied. We achieved an R² 93.3% using all 19 variables.

• Decision Tree Regression: We used GridSearchCV with RepeatedKFold cross-validation to fine-tune hyperparameters and select the best model configuration. We achieved an R² 85% using all
  19 variables.

• LightGBM Regression Model: Answering 20 questions is time-consuming for the user, so we aimed to reduce the variables and select a better-performing model to maintain prediction accuracy.

  Model Selection: We used the Lazy Predict library to quickly evaluate several machine-learning models. The LGBM regressor emerged as the top performer with R² 98% when
  using all variables.

  Feature Evaluation: We identified the 10 most impactful variables on Carbon Emissions using LGBM feature importance scores. We calculated the R² for all subsets of these 10 variables and selected the model that showed a good R² (~ 90%) with the fewest possible number of features. We achieved an R² 88.6% using only 6 variables (equivalent to 7 questions).
  As an alternative to the questionnaire we made a printable A4-sized flowchart of a decision tree,

• Decision Tree Classification: To create a binary classification model, we median split the carbon emission data. The model achieved an accuracy of 85% using 18 variables.

Image Classification

Our initial vision was to enable users to take pictures of their homes and use image classification to extract relevant variables for calculating their carbon emissions. However, many variables in our dataset could not be accurately captured through images. For example, classifying body type as 'overweight' from a photo could be uncomfortable, and variables like kilometers driven cannot be determined from images alone. Therefore, while image classification remains a key component, users must also complete a questionnaire to ensure accurate predictions. We used the `duckduckgo_search library` to search for and download images, which were then resized for efficiency. We subsequently applied Fast.ai for model training.

Questionnaire/Streamlit Application (For finished app click here)

We used the Streamlit tool in Python to create an application that collects all the necessary information from the user in a questionnaire and then applies the estimated models to the individual carbon footprint. For demonstration purposes only we included an option to choose between our different models at the start. Note, this option would not be part of a finished product. The questionnaire part uses Streamlit’s interactive input widgets (e.g., streamlit.radio) to gather the user’s carbon emission data and stores the answers within the application’s “session state” (streamlit.session_state) which works similar to a Python dictionary. The stored data is then put into a Pandas.DataFrame with the same structure as the cleaned data used for modeling. To encode and model the individual data, we import the estimated encoders and models from pkl(“pickle”)-files into the app. With their help the app can estimate the user’s carbon footprint, which is displayed in kg of CO2 emission per month.