Mood classification from Spotify audio descriptors is frequently reported with accuracies above 90%, yet several studies derive the mood label from valence thresholds and then supply valence to the classifier as an input feature, a form of target leakage that overstates generalization. This study quantifies that inflation and establishes a leakage-free baseline for Gaussian Naïve Bayes on numerical audio features. Following CRISP-DM, 6,000 popular tracks from six genres in the Spotify Tracks Dataset were labeled into four valence bands (Angry, Sad, Cheerful, Happy); the class distribution (17.3%, 28.6%, 32.0%, and 22.1%) and the 32% majority baseline are reported in full. Two configurations were evaluated on a stratified 80:20 split with five-fold cross-validation: Model A with all ten descriptors including valence, and Model B with the nine remaining descriptors. Model A reached 89.17% test accuracy (macro-F1 0.89), whereas Model B reached 42.33% accuracy, 44.47% balanced accuracy, and macro-F1 0.43, an inflation of 46.8 percentage points attributable to leakage alone. Model B still exceeds the majority baseline by about 10 points, with the weakest recall for Sad (0.26). Permutation importance ranks danceability first, followed by instrumentalness, duration, and energy, indicating that the remaining descriptors carry rhythm- and arousal-related information rather than pleasantness. Errors concentrate in adjacent valence bands and in the Sad/Angry pair, which share low valence and differ only in arousal, a dimension the labeling rule does not observe. The findings argue for reporting leakage-free results and class distributions in Spotify mood studies, and for two-dimensional (valence by energy) labeling..
The demand for Liquefied Petroleum Gas (LPG) in Indonesia continues to increase along with the growing population, economic activities, and energy conversion policies. On the other hand, the demand pattern for LPG is fluctuating and influenced by seasonal factors, making the distribution planning process difficult. Inaccuracy in forecasting LPG demand can lead to an imbalance between supply and demand, resulting in either supply shortages or stockpiling. This study aims to compare the performance of the Autoregressive Integrated Moving Average (ARIMA) and Seasonal Autoregressive Integrated Moving Average (SARIMA) methods in forecasting LPG demand using daily historical data from PT Pertamina Patra Niaga for the period from January 2023 to December 2025. The research stages include data preprocessing, stationarity testing using the Augmented Dickey-Fuller (ADF) test, model formation using Auto ARIMA, and evaluation using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and Akaike Information Criterion (AIC). The research results show that the ARIMA (2,1,4) model produces an MAE value of 114,427.20, an RMSE value of 219,563.30, and an MAPE value of 55.48%, while the SARIMA (3,1,3) (1,0,0,7) model produces an MAE value of 114,663.22, an RMSE value of 219,532.32, and an MAPE value of 55.46%. Although the ARIMA model yields a slightly lower MAE value, the SARIMA model provides lower RMSE, MAPE, and AIC values, making it the best model in this study because it better represents seasonal patterns in the historical data. The research results indicate that the SARIMA (3,1,3) (1,0,0,7) model is more suitable for supporting short-term LPG distribution planning based on historical data.
Mobile banking applications generate large volumes of user reviews that can be used to evaluate digital banking service quality. This study analyzes sentiment in Indonesian mobile banking reviews on the Google Play Store in 2025 and compares the performance of Naive Bayes (NB), Support Vector Machine (SVM), and Random Forest (RF). A total of 270,977 reviews were collected from five applications, namely BRImo, Livin’ by Mandiri, myBCA, BYOND by BSI, and wondr by BNI. The data were processed through text cleaning, case folding, tokenization, word normalization, stopword removal, and negation retention. Sentiment labels were generated automatically using the InSet Lexicon and validated on 450 stratified review samples. The validation results showed moderate agreement between automatic and manual labels, with an accuracy of 0.6111, macro-F1 score of 0.5384, and Cohen’s Kappa of 0.4167. Review texts were represented using TF-IDF with unigram and bigram features, and model performance was evaluated across six stratified data-splitting scenarios using accuracy, macro precision, macro recall, macro-F1, confusion matrix, and macro ROC-AUC. The results show that RF in Scenario 6 achieved the best performance, with an accuracy of 0.9524, macro-F1 score of 0.8498, and macro ROC-AUC of 0.9798. Dominant-term analysis indicates recurring complaints related to login failures, transaction errors, system disruptions, and application updates. This study contributes a cross-application comparison of machine learning algorithms for Indonesian mobile banking sentiment analysis and provides practical insights for improving service reliability. However, the findings should be interpreted within the limitation of automatic lexicon-based labeling, particularly for neutral and context-dependent reviews.
Complementary product recommendation plays an important role in E-commerce platforms by assisting users in discovering products that are likely to be purchased together. Many existing approaches treat this task primarily as a predictive modeling problem and directly apply machine learning models to transactional data. However, such approaches often overlook the underlying information structure of product relationships. This study formulates complementary recommendation as a relational information modeling problem in which transactional purchase data are transformed into structured product interactions. Products are first represented symbolically and mapped into numeric indices through vocabulary-based encoding functions. The encoded identifiers are then transformed into vector-space representations that capture semantic proximity between products and categories. Complementary prediction is performed by applying an interaction inference function to structured product-pair representations derived from these vectors. Experimental results using real-world E-commerce transaction data demonstrate stable learning convergence with a Mean Absolute Error (MAE) of 0.0106 and a Mean Squared Error (MSE) of 0.0035, indicating that the model effectively captures complementary relationships. From an informatics perspective, this work highlights how recommendation tasks can be formulated as structured information transformation processes rather than solely predictive modeling tasks to support cross-selling and product bundling strategies.
