Behavioural Clustering and Profile Embeddings for Buyer-Seller Matchmaking in Digital Marketplaces. Focus: Using Unsupervised Learning for Personalized B2B Recommendations

In the online B2B marketplace, proper buyer-seller matching is crucial for improving communication, conversion rate, and corporate relationships. Conventional recommendation models are typically based on supervised learning, where labels must be provided and fixed features are required, which can be less effective in fast-changing, high-dimensional systems. To address these challenges, this research proposes a hybrid machine learning framework that includes preprocessing, dimensionality reduction, and ensemble classification. Data normalization is performed by Min-Max scaling with the aim of normalizing features. That is, Principal Component Analysis (PCA) is then used to obtain a reduced dimensionality, such that key behavioural insights are not lost. The PCA focused with finding the biggest data variation sources and eliminating noise, although nonlinear patterns may exist. The method can be used to match behavioral trends that are dominant and most important by reducing many correlated features of behavior into a few principal components. To decrease the dimensionality of the linear classification and make it much quicker and more reliable without missing the essential B2B interaction signals. The optimized data is then passed through a hybrid classification model that combines Support Vector Machine (SVM) and Random Forest (RF). SVM provides a high level of accuracy in complex spaces, while RF gives it robustness and prevents overfitting based on ensemble learning. Such a two-pronged strategy refines the quality of the matching accuracy. The actual B2B dataset has approximately 50,000 records with numerical, categorical and behavioral characteristics that provide the interaction between the buyer and seller, i.e., the history of transactions and types of products. The reduction of features PCA is based on normalizing features and reduces dimensions to maintain the important insights. The model evaluation is conducted based on five performance metrics: precision, accuracy, recall, F1-score, and AUC. Experimental results obtained on both synthetic and real-world B2B data scenarios show that the hybrid model can outperform traditional methods significantly. This framework provides a solution for intelligent B2B matching that is scalable and dynamic.