Rahman, Wahyu Arief (2026) OPTIMASI KINERJA ARSITEKTUR TWO-TOWER NEURAL COLLABORATIVE FILTERING UNTUK PENANGANAN COLD START PADA DATA IMPLICIT FEEDBACK E-COMMERCE. S1 thesis, Fakultas Teknik Universitas Sultan Ageng Tirtayasa.
![[img]](https://eprints.untirta.ac.id/style/images/fileicons/text.png) |
Text (Fulltext)
Wahyu Arief Rahman_3337210024_Fulltext.pdf Restricted to Registered users only Download (1MB) | Request a copy |
|
Text (Bab 1)
Wahyu Arief Rahman_3337210024_01.pdf Restricted to Registered users only Download (1MB) | Request a copy |
|
Text (Bab 2)
Wahyu Arief Rahman_3337210024_02.pdf Restricted to Registered users only Download (466kB) | Request a copy |
|
Text (Bab 3)
Wahyu Arief Rahman_3337210024_03.pdf Restricted to Registered users only Download (375kB) | Request a copy |
|
Text (Bab 4)
Wahyu Arief Rahman_3337210024_04.pdf Restricted to Registered users only Download (1MB) | Request a copy |
|
Text (Bab 5)
Wahyu Arief Rahman_3337210024_05.pdf Restricted to Registered users only Download (265kB) | Request a copy |
|
Text (Daftar Referensi)
Wahyu Arief Rahman_3337210024_Ref.pdf Restricted to Registered users only Download (332kB) | Request a copy |
|
Text (Lampiran)
Wahyu Arief Rahman_3337210024_Lamp.pdf Restricted to Registered users only Download (324kB) | Request a copy |
|
Text (Cek Plagiasi)
Wahyu Arief Rahman_3337210024_CP.pdf Restricted to Registered users only Download (15MB) | Request a copy |
Abstract
This study aims to develop and evaluate an efficient hybrid Neural Collaborative Filtering (NCF) architecture to address three major challenges in e-commerce recommendation systems: cold start, data sparsity, and computational efficiency. The research method used is quantitative experimental by developing a Two-Tower architecture model that integrates user-item interaction data from public datasets with additional features in the form of user demographic data and item category simulations. To address the computational challenges of large-scale data, a data pipeline optimization strategy is applied through the pre-generation negative sampling technique. Model performance was comprehensively evaluated in warm start, user cold start, and item cold start scenarios using the Hit Ratio@10 (HR@10) and Normalized Discounted Cumulative Gain (NDCG@10) metrics, and further validated through qualitative analysis using hold-out data. The results of the study show that the proposed architecture successfully provides an effective solution. First, the pipeline optimization strategy has been proven to drastically reduce training time from more than 2 hours to around 5-7 minutes per epoch on consumer-grade GPUs. Second, the hybrid model has successfully overcome the cold start problem by increasing HR@10 from 0.0053 (baseline) to 0.0096, representing an 81% improvement in performance. Third, in the user cold start scenario, despite a trade-off in the precision ranking metric, qualitative validation proves that the model has good generalization capabilities by achieving a category relevance rate of 65% for new users who have never interacted before.
| Item Type: | Thesis (S1) | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Contributors: |
|
|||||||||
| Additional Information: | Penelitian ini bertujuan untuk mengembangkan dan mengevaluasi arsitektur Neural Collaborative Filtering (NCF) hibrida yang efisien, guna mengatasi tiga tantangan utama sistem rekomendasi e-commerce: cold start, data sparsity, dan efisiensi komputasi. Metode penelitian yang digunakan adalah kuantitatif eksperimental dengan mengembangkan model berarsitektur Two-Tower yang mengintegrasikan data interaksi user-item dari dataset publik dengan fitur tambahan berupa data demografis pengguna dan simulasi kategori item. Untuk menjawab tantangan komputasi pada data berskala besar, diterapkan strategi optimasi pipeline data melalui teknik pre-generation negative sampling. Kinerja model dievaluasi secara komprehensif pada skenario warm start, user cold start, dan item cold start menggunakan metrik Hit Ratio@10 (HR@10) dan Normalized Discounted Cumulative Gain (NDCG@10), serta divalidasi lebih lanjut melalui analisis kualitatif menggunakan data hold-out. Hasil penelitian menunjukkan bahwa arsitektur yang diusulkan berhasil memberikan solusi efektif. Pertama, strategi optimasi pipeline terbukti mereduksi waktu pelatihan secara drastis dari lebih 2 jam menjadi sekitar 5-7 menit per epoch pada GPU kelas konsumen. Kedua, model hibrida berhasil mengatasi masalah item cold start dengan meningkatkan HR@10 dari 0,0053 (baseline) menjadi 0,0096, yang merepresentasikan peningkatan kinerja sebesar 81%. Ketiga, pada skenario user cold start, meskipun terdapat trade-off pada metrik peringkat presisi, validasi kualitatif membuktikan bahwa model memiliki kemampuan generalisasi yang baik dengan mencapai tingkat relevansi kategori sebesar 65% pada pengguna baru yang belum pernah berinteraksi sebelumnya. | |||||||||
| Uncontrolled Keywords: | NCF, Cold Start, Data Sparsity, Demografi, Two-Tower | |||||||||
| Subjects: | Q Science > QA Mathematics > QA75 Electronic computers. Computer science Q Science > QA Mathematics > QA76 Computer software T Technology > T Technology (General) |
|||||||||
| Divisions: | 03-Fakultas Teknik > 55201-Jurusan Teknik Informatika | |||||||||
| Depositing User: | Mr Wahyu Arief Rahman | |||||||||
| Date Deposited: | 09 Feb 2026 08:03 | |||||||||
| Last Modified: | 09 Feb 2026 08:03 | |||||||||
| URI: | http://eprints.untirta.ac.id/id/eprint/58151 |
Actions (login required)
 |
View Item |