Electronic Theses and Dissertation

Minyak kenanga (cananga odorata) merupakan minyak atsiri bernilai ekonomi tinggi yang komposisi kimianya dipengaruhi oleh varietas dan asal geografis. penelitian ini bertujuan mengevaluasi kinerja model k-nearest neighbors (k-nn) dalam mengklasifikasikan minyak cananga odorata forma macrophylla dan var. fruticosa berdasarkan data absorbansi ftir, membandingkannya dengan partial least squares-discriminant analysis (pls-da), mengevaluasi kemampuan generalisasi model, serta mengidentifikasi pita spektrum yang berkontribusi terhadap klasifikasi. spektrum ftir dipreproses menggunakan standard normal variate (snv), turunan pertama savitzky–golay, dan standardisasi, kemudian dianalisis menggunakan principal component analysis (pca), hierarchical cluster analysis (hca), dan algoritma k-nn. evaluasi model dilakukan menggunakan rezim a dan rezim b (cross bebas kebocoran data), kemudian divalidasi secara eksternal menggunakan sampel asal singkil. analisis gc-ms digunakan sebagai data pendukung untuk menginterpretasikan perbedaan komposisi kimia. hasil penelitian menunjukkan bahwa model k-nn memperoleh balanced accuracy 1,000 pada rezim a untuk klasifikasi varietas dan lokasi. pada evaluasi cross (rezim b), nilainya menurun menjadi 0,325 untuk klasifikasi varietas dan 0,500 untuk klasifikasi lokasi, menunjukkan kemampuan generalisasi model yang masih terbatas. dibandingkan pls-da, k-nn memiliki performa lebih rendah pada klasifikasi varietas (Δ balanced accuracy = −0,125), sedangkan pada klasifikasi lokasi kedua metode menunjukkan performa yang sama (balanced accuracy = 0,500). validasi sampel eksternal berhasil mengklasifikasikan seluruh sampel singkil–fruticosa dengan recall 1,000. analisis gc-ms menunjukkan bahwa trans-α-bergamotene dominan pada varietas fruticosa, sedangkan linalool dan benzil benzoat dominan pada forma macrophylla. perbedaan komposisi kimia tersebut mendukung perbedaan spektrum ftir sebagai dasar klasifikasi minyak kenanga berdasarkan varietas dan asal geografis, meskipun kemampuan generalisasi model k-nn terhadap data baru masih perlu ditingkatkan. kata kunci: minyak kenanga, ftir, k-nearest neighbors, pls-da, klasifikasi geografis

ABSTRACT Ylang-ylang oil (Cananga odorata) is a high-value essential oil whose chemical composition is influenced by plant variety and geographical origin. This study aimed to evaluate the performance of the k-Nearest Neighbors (k-NN) model in classifying Cananga odorata forma macrophylla and var. fruticosa based on FTIR absorbance data, compare its performance with Partial Least Squares–Discriminant Analysis (PLS-DA), evaluate the model's generalization ability, and identify the FTIR spectral bands contributing to the classification. FTIR spectra were preprocessed using Standard Normal Variate (SNV), the first derivative of the Savitzky–Golay method, and standardization, followed by analysis using Principal Component Analysis (PCA), Hierarchical Cluster Analysis (HCA), and the k-NN algorithm. Model performance was evaluated using Regime A and Regime B (CROSS without data leakage) and externally validated using ylang-ylang oil samples from Singkil. GC-MS analysis was employed as supporting data to interpret differences in chemical composition. The results showed that the k-NN model achieved a balanced accuracy of 1.000 in Regime A for both variety and geographical origin classification. However, under CROSS evaluation (Regime B), the balanced accuracy decreased to 0.325 for variety classification and 0.500 for geographical origin classification, indicating limited model generalization. Compared with PLS-DA, k-NN exhibited lower performance in variety classification (Δ balanced accuracy = −0.125), whereas both methods showed similar performance for geographical origin classification (balanced accuracy = 0.500). External validation successfully classified all Singkil–fruticosa samples with a recall of 1.000. GC-MS analysis revealed that trans-α-bergamotene was the dominant compound in var. fruticosa, whereas linalool and benzyl benzoate predominated in forma macrophylla. These differences in chemical composition support the observed variations in FTIR spectra as the basis for classifying ylang-ylang oil according to variety and geographical origin, although the generalization capability of the k-NN model for new data still requires improvement. Keywords: ylang-ylang oil, FTIR, k-Nearest Neighbors, PLS-DA, geographical classification.