Personalized Transformer for Explainable Recommendation

Li, Lei; Zhang, Yongfeng; Chen, Li

doi:10.48550/arxiv.2105.11601

Cited by 2 publications

(7 citation statements)

References 30 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…• PETER+ [2]: This is an explanation-generating model that uses Transformer. We used it to generate an explanation from the user ID, item ID, features, and reviews for comparison with the effectiveness of the proposed method.…”

Section: Comparison Of Methodsmentioning

confidence: 99%

“…The complexity of this model makes it more accurate and superior to previously popular RNN-based networks. However, the existing explanation generation method only generates an explanation based on a Transformer [2] or RNN [7,47], ignoring the order of user behaviors that play important roles in increasing recommendation accuracy. This study was aimed at addressing this issue.…”

Section: Explanation-generating Approachesmentioning

confidence: 99%

“…Researchers have investigated different approaches for providing explanations, ranging from using predefined templates that can be customized to specific context [31,17], generating helpful tips to guide users [25], and generating sentences tailored to individual recommendations automatically [2,7,32]. We focused on the last task, which has recently gained significant research attention because of the advancements in natural language generation techniques, including recurrent neural networks (RNN), Transformer [14], and pre-trained language models [26,29,50].…”

Section: Introductionmentioning

confidence: 99%

“…Li et al [7] introduced a NEural TEmplate (NETE) explanation generation framework that learns sentence templates from data and generates template-controlled sentences, providing comments on specific features. Based on this work, Li et al [2] proposed a model called PETER, which incorporates the use of IDs to predict words in the target explanation, enabling personalization within the Transformer architecture. However, explanations generated by these methods are related only to the features of the same item, and they ignore the time-series information of the purchase history of the target user.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

Generating Explanations for Explainable Recommendations Using Time-Series Information

Nobuhara

2023

Preprint

View full text Add to dashboard Cite

Generating explanations in natural language processing methods such as review generation and explainable recommendations plays an essential role in personalization. These methods can provide explanations that can help target users better understand the recommended items and enhance their awareness of the system’s knowledge about their preferences. However, existing explanation-generating methods for recommendations ignore time-series information, which is important to improve the effectiveness of the recommendation. To address this issue, we propose a Transformer-based explanation-generating method designed to generate recommendation reasons based on time-series information from sequential recommendations. This method bridges the time-series information from the target user’s purchase history and the item they may want to purchase to assign it a linguistic meaning and generate explanations for the recommended item. We divided our proposed method into two parts, sequential recommendation and reason generation, to maximize the benefits. Extensive experiments on three datasets show that the proposed approach can generate explanations for the results of sequential recommendations both reasonably and effectively compared to state-of-the-art explanation generation methods in most cases. Further experiments on time-series information and case studies verify the effectiveness of our approach. Our proposed model demonstrates its ability to generate explanations that incorporate time series information, as supported by our proposed evaluation metric. This also affirms the significant role of time series information in improving the performance and quality of the generated explanations.

show abstract

Section: Comparison Of Methodsmentioning

confidence: 99%

Section: Explanation-generating Approachesmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Generating Explanations for Explainable Recommendations Using Time-Series Information

Nobuhara

2023

Preprint

View full text Add to dashboard Cite

show abstract

“…For sequential recommendation, we adopt Caser [50], HGN [34], GRU4Rec [21], BERT4Rec [48], FDSA [63], SASRec [24] and S 3 -Rec [67] as baselines for comparison. For explanation generation, we utilize Attn2Seq [15], NRT [30] and PETER [29] as baselines. For review summarization, we adopt pretrained T0 [45] and GPT-2 [41] as baselines.…”

Section: Baselines For Multiple Tasksmentioning

confidence: 99%