Data and Code to reproduce results in paper "Credit Risk Prediction via Graph Neural Networks with Homophily-Guided Graph Construction" This is a working paper. The restriction on codes will be removed after publication. This repository contains the necessary codes to reproduce results in the paper. Raw data: The paper applis its methodology on the 4 datasets from UCI Machine Learning Repository (Markelle Kelly et al.). 1. Hofmann, H. (1994). Statlog (German Credit Data) [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5NC77. 2. Quinlan, R. (1987). Statlog (Australian Credit Approval) [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C59012. 3. Yeh, I. (2009). Default of Credit Card Clients [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C55S3H. 4. South German Credit [Dataset]. (2020). UCI Machine Learning Repository. https://doi.org/10.24432/C5QG88. Data cleaning: 6th paper (1) data cleaning AUS.ipynb 6th paper (1) data cleaning DEF.ipynb 6th paper (1) data cleaning GER.ipynb 6th paper (1) data cleaning SGER.ipynb These four files generates the clened dataset ready form modleing, as described in section 4.1 in the paper. For consistency and comparability, we use the same four benchmark datasets as in \citet{shi_improved_2024} from UCI Machine Learning Repository \citep{markelle_kelly_uci_nodate}. These datasets—Australian Credit Approval (AUS) \citep{quinlan_statlog_1987} , German Credit Data (GER) \citep{hofmann_statlog_1994}, Default of Credit Card Clients (DEF) \citep{yeh_default_2009}, and South German Credit (SGER) \citep{noauthor_south_2019}—are all publicly available from open-access sources and have been widely used in prior research on credit risk prediction. They span a range of scales in terms of sample size, attribute count, and class imbalance, and collectively represent diverse real-world scenarios of credit application approval, default behavior, and credit scoring. To shape these datasets in the so-referred tabular structure for input into our graph-based modeling pipeline, standard preprocessing steps are applied. Missing values, if present, are treated through imputation or removal. Continuous and ordinal variables are normalized to the $[0,1]$ range using min–max scaling to ensure numerical comparability across features. While ordinal variables are discrete, they are commonly treated as continuous during scaling to retain their rank information without inflating dimensionality. Categorical (nominal) variables are transformed into binary dummy variables to preserve their semantic meaning and ensure compatibility with distance-based similarity metrics. The original study states that: “The four experimental datasets in our study are numeric...” \citep{shi_improved_2024}, but provides no further details on how nominal variables were handled. This ambiguity in feature type treatment may partially explain the differences observed in baseline model performance during our replication. In contrast, we explicitly identify and process nominal variables by generating dummy variables, while treating ordinal variables as continuous for the purpose of scaling. Table~\ref{tab:data_summary} summarizes the dimensional changes resulting from our preprocessing, along with the class distributions in each dataset. Modeling: The remaining notebooks. These notebooks generate the results presentated in the paper.