Machine learning and sentiment analysis for intelligent project selection: design and experimental validation

Abstract

Relevance. The rapid growth of crowdfunding campaigns on platforms such as Kickstarter, combined with information asymmetry between founders and investors, creates a need for intelligent decision-support tools that account for investor sentiment rather than relying solely on structured financial indicators. Purpose. To design and experimentally validate a decision support workflow for predicting crowdfunding campaign success that combines several classification algorithms with sentiment analysis, using backers' comments as an auxiliary consistency check. Tasks. To formalize a hybrid feature set uniting structured project metadata with textual sentiment indicators, train and compare classification algorithms across several feature configurations, assess the impact of sentiment on prediction quality, interpret sentiment differences between successful and failed campaigns, cross-check blurb sentiment against comment sentiment, and forecast outcomes for live campaigns. Methods. Kickstarter data was processed through deduplication, outlier removal, logarithmic transformation, and one-hot encoding. Five classification algorithms, Logistic Regression, k-Nearest Neighbors, Decision Tree, Random Forest, and Multilayer Perceptron, were trained in three configurations combining structured features with lexicographic and transformer-based sentiment scores; Pearson and Spearman correlations linked blurb and comment sentiment. Results. About 194,000 cleaned projects (2009-2026) were analyzed, together with 1.64 million backer comments on the 2,085 overlapping projects. Random Forest performed best; sentiment features gave no meaningful improvement. Failed campaigns showed systematically higher sentiment scores than successful ones, an "over-promising" effect (a weak association, not a causal relationship). A consistency check across projects found a weak but significant correlation between blurb and comment sentiment. Scientific novelty. Sentiment is formalized not as a direct predictor but as an inverse quality signal within a hybrid, explainable decision support workflow that jointly models structured metadata, textual sentiment, and community feedback, using backers' comments as an auxiliary consistency signal. Practical significance. The workflow can support crowdfunding platforms and investors by providing automated campaign screening, feature-based interpretation of recommendations, and a composite investment-attractiveness index that combines model predictions, sentiment, and expert assessments. Conclusions. The workflow provides accurate and interpretable screening of crowdfunding projects. At the same time, its sentiment component requires further validation using full campaign texts and comment time series, as well as testing on other platforms.