Integrasi Pemodelan Frekuensi Item Temporal Dan Large Language Model Untuk Penalaran Konteks Musiman Pada Rekomendasi Belanja Berikutnya

Sujono, Yasmin Putri (2026) Integrasi Pemodelan Frekuensi Item Temporal Dan Large Language Model Untuk Penalaran Konteks Musiman Pada Rekomendasi Belanja Berikutnya. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221273_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Sistem Rekomendasi Keranjang Berikutnya atau Next Basket Recommendation (NBR) bertujuan untuk memprediksi sekumpulan barang yang akan dibeli pengguna pada transaksi berikutnya berdasarkan riwayat belanja historis. Dalam penelitian ini saya mengusulkan untuk mengintegrasikan pendekatan Temporal Item Frequency-based User-kNN (TIFU-KNN) dengan menghitung PIF dan pembobotan peluruhan waktu yang menyaring daftar barang rekomendasi berbasis riwayat transaksi dan Large Language Model (LLM) sebagai modul post-processing penalar konteks musiman (seasonal). Evaluasi dilakukan pada 3.540 pengguna uji yang keranjang terakhirnya mengandung setidaknya satu produk musiman dengan menggunakan tiga metrik evaluasi, yaitu Recall, F1-Score, dan nDCG. Berdasarkan hasil proses grid search pada konfigurasi terbaik TIFU-KNN dengan parameter decay rate r = 0.9, α = 0.95, k_nearest = 40, dan metrik Jaccard Similarity sebagai metrik kemiripan terbaik. LLM diintegrasikan melalui teknik prompt engineering, kemudian skor kNN diperkuat menggunakan mekanisme Soft Scaling (perkalian faktor λ) pada item musiman pilihan LLM. Hasil percobaan menunjukkan bahwa sistem hybrid bergantung pada ukuran daftar rekomendasi (Top-K) dan metrik kemiripan yang digunakan. Pada metrik Cosine, baseline memiliki performa terbaik dengan nilai Recall sebesar 0.62990, F1-Score sebesar 0.52215, dan nDCG sebesar 0.63586, karena ini penambahan LLM hanya membutuhkan λ yang kecil (λ=1.05 hingga λ=1.15) untuk mencapai titik optimal. Di sisi lain, metrik Jaccard menunjukan kontribusi peningkatan penalaran musiman LLM dengan peningkatan F1-Score mencapai 0.50164 pada Top-5 dengan λ=1,25. Dengan demikian, penelitian ini menunjukkan bahwa integrasi LLM sebagai pasca-pemrosesan (post-processing) mampu memberikan peningkatan kualitas rekomendasi.
=====================================================================================================================================
Next Basket Recommendation (NBR) aims to predict the set of items that a user is likely to purchase in their next transaction based on historical shopping records. In this study, we propose an integration of the Temporal Item Frequency-based User-kNN (TIFU-kNN) approach, which computes Personalized Item Frequency (PIF) and applies time-decay weighting to generate recommendation lists from transaction histories, with a Large Language Model (LLM) serving as a post-processing module for seasonal contextual reasoning. The evaluation was conducted on 3,540 test users whose most recent baskets contained at least one seasonal product. Performance was assessed using three evaluation metrics: Recall, F1-Score, and nDCG. Based on the grid search results, the best-performing TIFU-kNN configuration employed a decay rate of r = 0.9, α = 0.95, k_nearest = 40, and Jaccard Similarity as the similarity metric. The LLM was integrated through prompt engineering techniques, and the kNN recommendation scores were enhanced using a Soft Scaling mechanism by multiplying the scores of LLM-selected seasonal items with a scaling factor (λ). Experimental results indicate that the effectiveness of the hybrid system depends on both the recommendation list size (Top-K) and the similarity metric employed. In the Cosine metric, the baseline has the best performance with a Recall value of 0.62990, F1-Score value of 0.52215, and nDCG value of 0.63586, because of this the addition of LLM only requires a small λ (λ = 1.05 to λ = 1.15) to reach the optimal point. On the other hand, the Jaccard metric shows a more measurable contribution to the increase in seasonal reasoning of LLM with an increase in F1-Score reaching 0.50164 in the Top-5 with λ = 1.25. Thus, this study shows that the integration of LLM as post-processing is able to provide an increase in the quality of recommendations.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Next-Basket Recommendation, TIFU-KNN, Large Language Model, Sistem Rekomendasi, Soft Scaling, Recommendation System.
Subjects:	Q Science > QA Mathematics > QA75 Electronic computers. Computer science. EDP Q Science > QA Mathematics > QA76.9.I58 Recommender systems (Information filtering)
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Yasmin Putri Sujono
Date Deposited:	24 Jul 2026 07:39
Last Modified:	24 Jul 2026 07:39
URI:	http://repository.its.ac.id/id/eprint/137244

Actions (login required)

View Item