Recommendation System Based on Semantic Analysis and Network Models

In this work, we implement a hybrid recommendation system for news articles that combines two primary approaches: semantic analysis via TF–IDF vectorization of headlines and Nearest Neighbors search, and network analysis using an article-similarity graph constructed from shared tags. To improve recommendation quality, rare tags were filtered out, and the number of articles per tag was capped to balance the dataset. A weighted combination of semantic and graph-based scores was also employed with parameter tuning. Precision was adopted as the evaluation metric, measuring the proportion of correctly predicted tags in the recommended articles against the ground-truth tags in the test set. Experimental results show that the hybrid model effectively leverages both semantic headline features and network connections between articles. When increasing the per-tag article cap from 1,000 to 3,000, precision rose from 0.168 to 0.187—an 11% improvement—while training time increased from 31.8 s to 506 s. This trade-off confirms the value of expanding the data scope and demonstrates the strength of the hybrid approach. The achieved precision on the test set indicates that integrating semantic and network analyses yields more accurate and well-grounded recommendations tailored to user interests.