مها الهدو

This thesis proposes A Hybrid Approach Integrating Similarity Metrics to Document Clustering to meet the shortcomings of single-metric document clustering methods that can often fail to provide the complete range of lexical, statistical, an‎d semantic relationships between documents. This model starts with pre-processing such as text normalization, stopwords removal, punctuation removal, an‎d lemmatization, an‎d then documents are represented using Count Vectorizer an‎d TF-IDF. Then Agglomerative Hierarchical Clustering is applied on the distance matrices based on several different similarity measures. The proposed framework combines Cosine, BM25, Dice, an‎d Minkowski similarities into a single hybrid similarity, unlike traditional methods that can only use a single similarity measure, an‎d thus utilize their complementary features. BBC News dataset, which included 2,225 documents in five categories was used to carry out the experimental eva‎luation. Findings, showed that the hybrid similarity model was better than the single measures in that it yielded congruent an‎d distinct clustering. In particular, the optimal model had an average pairwise similarity of 0.1400, stan‎dard deviation of 0.1113 an‎d better internal eva‎luation scores with higher Silhouette Score an‎d a lower Davies-Bouldin Index than the single-metric baselines. Visual analysis through similarity heatmaps confirmed that the hybrid approach effectively balances intra-cluster cohesion with inter-cluster separation. These findings validate the robustness an‎d interpretability of the hybrid similarity framework, offering a reliable solution for document clustering in heterogeneous text corpora.

رويكرد تركيبي , خوشه بندي اسناد , شباهت تركيبي , متن كاوي , پردازش زبان طبيعي (NLP) , ارزيابي خوشه بندي

hybrid approach , Document Clustering , Hybrid Similarity , Text Mining , Natural language processing (NLP) , Clustering eva‎luation

Maha Alhado

Dr. Hassan Naderi