From Interaction to Prediction: A Multi-Interactive Attention based Approach to Product Rating Prediction
This archive is distributed in association with the INFORMS Journal on Computing under the MIT License.
The software and data in this repository are a snapshot of the software and data that were used in the research reported on in the paper From Interaction to Prediction: A Multi-Interactive Attention based Approach to Product Rating Prediction by Yu L, Gong W, Zhang D, Ding Y and Fu Z.
To cite the contents of this repository, please cite both the paper and this repo, using their respective DOIs.
https://doi.org/10.1287/ijoc.2023.0131
https://doi.org/10.1287/ijoc.2023.0131.cd
Below is the BibTex for citing this snapshot of the repository.
@misc{Yu2024,
author = {Li Yu and Wei Gong and Dongsong Zhang and Yu Ding and Zhe Fu},
publisher = {INFORMS Journal on Computing},
title = {From Interaction to Prediction: A Multi-Interactive Attention based Approach to Product Rating Prediction},
year = {2024},
doi = {10.1287/ijoc.2023.0131.cd},
url = {https://github.com/INFORMSJoC/2023.0131},
note = {Available for download at https://github.com/INFORMSJoC/2023.0131},
}
Despite increasing research on product rating prediction, very few studies have considered user-item interaction relationships at multiple levels. To address this critical limitation, we propose a novel Rating Prediction method based on Multi-Interaction Attention (RPMIA) by learning user-item interaction relationships at three levels simultaneously from online consumer reviews for predicting product ratings with reasonable interpretability. Specifically, RPMIA first deploys a multi-head cross attention mechanism to capture the interaction between contexts of items and users. Then, it uses a bi-layer gate-based mechanism to extract aspects of users and items, and a self-attention mechanism is further used to learn their interaction at the aspect level. Finally, the aspects of users and items are coupled together to form meaningful user-item aspect pairs (UIAP) via a joint attention. A multi-task predictor that integrates a factorization machine and a feedforward neural network is designed to generate a rating prediction. We have empirically evaluated RPMIA with seven real-world datasets. The results demonstrate that RPMIA outperforms the state-of-the-art methods consistently and significantly. We also conduct a user study to assess the interpretability of the RPMIA method.
We used seven publicly available OCR datasets collected from Amazon.com on different products, including music instruments, office products, digital music, grocery & gourmet food,video games, tools & home improvement, and sports and outdoors products, which have been widely used for recommendation evaluation in previous studies.These datasets consist of consumers’ product ratings ranging from 1 to 5 and corresponding textual reviews. Each consumer or item has at least 5 reviews. All the datasets are in folder dataset/ (where dataset_D->Digital Music，dataset_G->Grocery and Gourmet Food，dataset_M->Musical Instruments，dataset_O->Office Products，dataset_S->Sports and Outdoors，dataset_T->Tools and Home Improvement，dataset_V->Video Games. In this repository, we use dataset_G as an example).
The code src/RPMIA.py implements and RPMIA model. In this implementation, we define five modules, including an embedding module, a context-aware module, an aspect-aware module, a UIAP-aware module, and a prediction module.
The package _auxiliaryTools/ implements data preprocessing of datasets, where code ExtractData.py implements extracting dataset from file, Evaluation.py defines a function named get_test_list_mask to batch process test data (including user ID, item ID, rating, etc.) and related user reviews, item reviews, and mask information according to a given batch size, and return multiple lists containing batches of different types of data.
Use the following command to run the code:
python src/RPMIA.py
- tensorflow 1.15
- python 3.6
- numpy 1.19.2
- pandas 1.1.5