Recommendation system, currently a highly sought-after technology, is extensively utilized to assist users in discovering preferred content from vast amounts of information by learning individual interactions. However, existing RS methods rely on text-based interactions (e.g., user-item), overlooking the potential preference signals carried by users’ voice, which could enhance accessibility, provide more diverse data insights, and increase user engagement. In this paper, we propose a novel voice-empowered recommendation framework named VGRec, which captures the intrinsic semantic correlation between voice and historical interactions, representing voice preference within a collaborative filtering paradigm. To better encode the voice characteristics of users, VGRec establishes a semantic conduit between voice and text feature information through feature retrieval. Moreover, all the voice and text representational information is smooth upon high-order graph convolutional networks, which is to learn user-voice-item interaction patterns and thereby facilitate the prediction of user preferences. Extensive experiments validate the feasibility and consistent superiority of our method over existing text-based recommendation models. Compared with the strongest baseline, VGRec yields relative improvements of 12.57%, 5.18%, and 2.78% on Coat, ML1M-mini, and ML1M, respectively.
conda create -n VGRec python=3.9
conda activate VGRec
pip install -r requirements.txt
You can get the audio datasets from GoogleDrive. Please extract the data under the ./data/{dataset_name}/mp3/
- Coat
python run_MFE.py --dataset=coat --epochs=35 --lr=0.0005 --test=1 --neg_num=3
- Movielens-1m
python run_MFE.py --dataset=movielens1m --epochs=35 --lr=0.0005 --test=1 --neg_num=5
- Coat
python main.py --weight_decay=1e-4 --lr=0.001 --n_layers=3 --dataset=coat --recdim=64 --neg_num 3 --ens_ratio 0.8
- Movielens-1m
python main.py --weight_decay=1e-4 --lr=0.001 --n_layers=3 --dataset=movielens1m --recdim=64 --neg_num 5 --ens_ratio 0.6
Coat:
| Metrics | VERS |
|---|---|
| F1@10 | 12.68 |
| Precision@10 | 8.24 |
| Recall@10 | 27.47 |
| NDCG@10 | 18.68 |