Paavo Nevalainen

Business-to-business (B2B) sales forecasting is critical for organizational growth, but available data are often small, imbalanced, and noisy, complicating the identification of high-potential customers. This study proposes a hybrid, explainable framework for opportunity forecasting that integrates customer segmentation using Self-Organizing Maps (SOM), feature attribution via SHapley Additive exPlanations (SHAP), and probabilistic prediction with Random Forests (RF). Purchase probabilities are further refined through cost-sensitive ROC-based threshold optimization to account for class imbalance and business costs. The framework explicitly addresses noisy and heterogeneous data and unstable purchase signals via explainability-guided segmentation that isolates behaviorally consistent customer groups. Empirical evaluation on two telecommunications products demonstrates that high-probability segments, characterized by top-ranked features, achieve high predictive performance (F1-score = 0.964 for Product 1, 0.922 for Product 2), whereas heterogeneous segments show weaker signals. These results highlight that explainability-guided segmentation isolates actionable customer groups and improves decision-making reliability. Overall, the framework provides a modular, interpretable approach for identifying B2B sales opportunities, prioritizing resources, and supporting data-driven business decisions.

This study presents a comprehensive methodology for processing multilingual customer support data to prepare it for training AI‐based conversational systems. Using a dataset of 36,599 unique customer interactions from a Finnish energy company, we employed language‐specific BERT models to identify and analyse thematic patterns within customer inquiries. Our dual‐model approach separately processed Finnish and English content, utilizing specialized embeddings for each language before comparing topics across languages using cosine similarity measures. Our preprocessing pipeline employs confidence‐weighted language detection optimized for technical domain terminology, adaptive frequency filtering with domain preservation and morphology‐aware language processing to address the unique challenges of multilingual customer support data. Results revealed distinct topic distributions between languages with strategically important overlapping themes, particularly in technical domains where similarity coefficients reached 0.700. UMAP visualizations confirmed the computational findings, showing well‐defined clusters in Finnish content and more diffuse boundaries in English topics. The identified cross‐language similarities in contract and energy‐related topics provide promising opportunities for targeted knowledge transfer, while disparities in other domains highlight the need for language‐specific optimization. The methodology presented demonstrates a robust framework for analysing bilingual customer support corpora that can be extended to training specialized AI assistants while addressing language‐specific nuances.

Mineral prospectivity mapping is essential for identifying areas with significant potential for mineral deposits. Major challenges arise due to imbalanced data environments, which include sparse mineral occurrences and vast unexplored regions lacking confirmed "true negatives." This research presents a Bayesian Kernel Density Estimation (KDE) framework that incorporates spatial priors and geophysical feature similarity to tackle these challenges. The approach integrates: (1) spatial kernel density estimation (KDE) utilizing geographic proximity to established deposits, (2) feature-space KDE reflecting geophysical similarity, and (3) Bayesian fusion of these KDE models. The Bayesian KDE demonstrates a 72% reduction in false positives and a 240% enhancement in known-mineral detection relative to feature-only KDE. The evaluation of model performance was conducted using leave-one-out cross-validation (LOOCV) and intersection-over-union (IOU). LOOCV yielded an average percentile rank of 0.791, demonstrating robust predictive accuracy in the presence of data imbalance. The IOU analysis, performed at spatial resolutions ranging from 50 to 600 meters, demonstrated model stability and indicated strong concordance between full model predictions and individual leave-one-out cross-validation iterations. The approach necessitates prior spatial knowledge of mineral occurrences, which may introduce bias; however, extensive validation supports its practical applicability. The Bayesian KDE reduced potential zones to about 1% of the unexplored areas, thereby enhancing the efficiency of future exploration initiatives. The proposed method significantly surpassed standard KDE, enhancing the mean probability of known minerals by 240% and diminishing background noise, thereby offering an effective and validated strategy for mineral exploration in highly imbalanced datasets.

Real-time multi-object tracking plays a pivotal role in many computer vision applications such as spanning surveillance systems, crowd behavior analysis, and urban monitoring. Unlike existing datasets that predominantly focus on controlled environments or traffic scenarios, this paper introduces a novel, high-resolution dataset collected from Turku Market Square, Finland. The dataset captures diverse real-world conditions, including varying crowd densities, frequent occlusions, and dynamic pedestrian movements, providing a valuable benchmark for pedestrian tracking in unstructured urban environments. In addition to the dataset, we propose a real-time multi-object tracking system designed for object detection and tracking. Our system compares the performance of two well-known object detectors (YOLOv11 and Faster R-CNN) and evaluates their outputs using eight different tracking algorithms. The experimental results demonstrate that the proposed system achieves high precision scores of 0.88 and 0.87 for YOLO-based BoTSORT and ByteTrack, respectively. These findings underscore the effectiveness of deep learning-based tracking methods for realworld applications and establish a strong benchmark for future research in multi-person tracking within dynamic public spaces.