Authors

Maksym Vasyliuk, Dominika Bykonia

Abstract

FlowKeep addresses customer churn and lifetime-value optimization in subscription-basedB2B software-as-a-service businesses. The project applies machine learning to productusage,financial, subscription, support, and account data to identify clients who may leavebefore cancellation occurs. It is intended to help customer-success managers prioritizeattention across a large account base, replacing subjective manual screening with datainformedrisk rankings and explanations. The project also examines how churn predictionscould inform retention actions, support escalation, product planning, and pricing decisions.

FlowKeep was presented as an analytical machine-learning prototype for predicting churnamong B2B SaaS customers. The team used a Kaggle dataset containing 3,001 client recordsand a reported churn rate of 22 percent. Its variables represented account profiles, financialmeasures, subscription events, product-usage logs, support-ticket histories, and the churntarget. Exploratory analysis considered relationships among recurring revenue, login andusage behavior, feature adoption, ticket volume, response time, and satisfaction.

The preprocessing pipeline converted 37 raw columns into 49 model-ready features. Missingnumerical values were filled with medians, while missing categorical values received anunknown category. Numerical variables were standardized and categorical variables wereone-hot encoded. Two engineered variables were described: text length and a binary indicatorfor the presence of tickets. The team used a stratified 80:20 train-test split so that both subsetsretained the original 78:22 class distribution. ROC-AUC and F1 were selected as primarymetrics because a classifier predicting retention for every account could otherwise obtain 78percent accuracy without identifying any churners.

Logistic regression served as a baseline, while Random Forest, XGBoost, and CatBoost werealso trained and evaluated. CatBoost was selected as the principal candidate because it achieved the strongest reported metrics and supported mixed data types. SHAP explanationswere used to examine influential variables. The intended operating pattern is companyspecific:a business supplies its own CSV data and trains the model for its accounts. Proposedoutputs include ranked at-risk accounts and explanations that direct customer-success teamstoward cases requiring attention.

license

Creative Commons BY-SA 4.0