<![CDATA[Recommendation systems help users identify products that match their needs amid the growing number of choices on digital platforms. The increasing variety of fitness and sports products can cause information overload and make relevant products difficult to find. This study develops and evaluates a content-based filtering recommendation system using Doc2Vec and cosine similarity. The study applies a quantitative experimental method to the Amazon Reviews'23 Dataset. Product_name, category, and clean_text are used as content attributes, while Product_id is used as the product identity. The data undergo selection, text cleaning, case folding, tokenization, and stopword removal, followed by an 80% training and 20% testing split. The Doc2Vec model uses the Paragraph Vector Distributed Bag-of-Words architecture with 100-dimensional vectors. Query vectors are produced through infer_vector and compared with product vectors using cosine similarity to create Top-N recommendations. Evaluation on four queries shows average Precision of 0.9500, 0.9000, and 0.8750; average Recall of 0.1100, 0.2078, and 0.6082; and average NDCG of 0.9732, 0.9584, and 0.9361 for Top-5, Top-10, and Top-30, respectively. The results show that smaller Top-N values provide higher precision and ranking quality, while larger Top-N values retrieve a broader set of relevant products. The system is implemented as a Flask-based website named Bolang.]]>
Copyrights © 2026
