En jämförelse av dimensionsreducerande metoders påverkan på modellträffsäkerhet

This Master’s thesis investigates dimensionality reduction and feature selection techniques’ impact on prediction of Share of Wallet (SoW) in the grocery retail domain. SoW is a common metric to measure customer satisfaction and loyalty and can be predicted using supervised machine learning (ML) methods. However, the complex nature of predicting SoW requires a large amount of features. A large amount of features coupled with few training instances tend to lead to issues with the curse of dimensionality. Therefore, this main objective of this thesis and purpose is to investigate the effectiveness of dimensionality reductions and feature selection techniques to mitigate the curse of dimensionality when predicting SoW. The research question posed in this study is: How does the application of feature reduction or feature selection affect the performance of pre-defined models when it comes to estimating the Share of Wallet (SoW) for individual food retail customers? When the models are trained using features derived from a single retailer and payment card transaction as labels. The features are created using data collected the retailer and its consumers. The labels are constructed from categorised card transactions made by the customer. Experiments were performed using four conventional machine learning, and two different feature reductions techniques; Principal component analysis and Filter method. The null hypothesis of the experiment, was to not use any feature reductions techniques and the original feature set of 150 features. Principal component analysis was use as preprocess to transform the feature into 30 principal component. Filter method was used to select the 20 feature with the highest amount of absolute correlation with the label of the dataset. The experimental results obtained from the application of machine learning algorithms were similar across all three types of feature sets, but the datasets where dimensionality reduction methods had been applied indicated slighter better results. PCA and correlation filtering had comparable results in terms of MSE and R-squared. Although these results did not achieve the desired level of accuracy for predicting SoW necessary to provide tangible value to the business, the study still offers valuable insights regarding dimensionality reduction impacts on model accuracy.