ICACT20260275 Question.1
Questioner: hripsime.hovhannisyan@rau.am 2026-02-11 오전 8:08:35 |
ICACT20260275 Answer.1
Answer by Auhor yennt@hvnh.edu.vn 2026-02-11 오전 8:08:35
Chrome  Click!! |
| Why SBERT instead of fine-tuning a domain model? |
We adopt SBERT instead of fine-tuning a domain-specific language model because SBERT provides semantically rich and stable sentence-level embeddings with significantly lower computational cost.
Reviews contain fine-grained semantic information that cannot be captured by item IDs alone, and SBERT is specifically optimized to map semantically similar reviews into nearby representations. |
ICACT20260275 Question.2
Questioner: hripsime.hovhannisyan@rau.am 2026-02-11 오전 8:13:14 |
ICACT20260275 Answer.2
Answer by Auhor yennt@hvnh.edu.vn 2026-02-11 오전 8:13:14
Chrome Click!! |
| How many reviews per item, and how were they selected? |
The number of reviews per item is determined dynamically in each batch based on the item with the largest number of reviews; other items are padded accordingly. |
ICACT20260275 Question.3
Questioner: maho.hoshiai@gmail.com 2026-02-13 오전 9:14:07 |
ICACT20260275 Answer.3
Answer by Auhor yennt@hvnh.edu.vn 2026-02-13 오전 9:14:07
Chrome Click!! |
| How much improvement comes specifically from reviews compared to using item IDs only? |
The improvement specifically attributed to review information can be observed by comparing HiSeR with ID-only sequential models such as FPMC, GRU4Rec, and SASRec. While ID-only models rely purely on behavioral transitions, review-enhanced representations capture richer product characteristics and user preferences (e.g., quality, functionality, sentiment).
|
ICACT20260275 Question.4
Questioner: maho.hoshiai@gmail.com 2026-02-13 오전 9:17:39 |
ICACT20260275 Answer.4
Answer by Auhor yennt@hvnh.edu.vn 2026-02-13 오전 9:17:39
Chrome Click!! |
| What is the key reason HiSeR improves performance — richer representation, better sequential modeling, or the joint training objective? |
The key factor is that HiSeR performs sequential modeling on semantically enriched item embeddings—combining ID-based information with review-derived features—rather than on pure ID embeddings alone. |
ICACT20260275 Question.5
Questioner: maho.hoshiai@gmail.com 2026-02-13 오전 9:22:01 |
ICACT20260275 Answer.5
Answer by Auhor yennt@hvnh.edu.vn 2026-02-13 오전 9:22:01
Chrome Click!! |
| How does HiSeR perform when reviews are noisy or very short? |
HiSeR remains robust because review information is processed through both sentence-level embeddings and an attention mechanism. We employ Sentence-BERT to generate sentence-level embeddings and the attention layer assigns lower weights to less informative or noisy reviews. Moreover, since review features are combined with ID-based representations, the model does not rely solely on textual content, ensuring stable performance even when reviews are short or noisy. |
