Prediction of User Opinion for Products - A Bag-of-Words and Collaborative Filtering based Approach

García-Cuesta, Esteban; Gómez-Vergel, Daniel; Expósito, Luis Gracias; Vela-Pérez, María

doi:10.5220/0006209602330238

Cited by 3 publications

(1 citation statement)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The dictionary can be of constant length-common to all products-containing the most commonly used opinion-words in the whole review dataset [42], or as proposed of variable length solving the problems of: (i) using a large set of words that are not shared by different products and thereof introducing unnecessary sparsity in the input matrix, (ii) this same increase on sparsity worsens to some extent the prediction capabilities of the algorithm, and (iii) the only use of opinion-words as vocabulary allows the overall opinion prediction for a product but not for the sentiment associated with its characteristics. Since different products have truly different features, we must consider a distinctive set of opinion-words for each one of them, reducing the number of features and the sparsity of the input matrix.…”

mentioning

confidence: 99%

Prediction of Opinion Keywords and Their Sentiment Strength Score Using Latent Space Learning Methods

et al. 2020

Self Cite

View full text Add to dashboard Cite

Most item-shopping websites give people the opportunity to express their thoughts and opinions on items available for purchasing. This information often includes both ratings and text reviews expressing somehow their tastes and can be used to predict their future opinions on items not yet reviewed. Whereas most recommendation systems have focused exclusively on ranking the items based on rating predictions or user-modeling approaches, we propose an adapted recommendation system based on the prediction of opinion keywords assigned to different item characteristics and their sentiment strength scores. This proposal makes use of natural language processing (NLP) tools for analyzing the text reviews and is based on the assumption that there exist common user tastes which can be represented by latent review topics models. This approach has two main advantages: is able to predict interpretable textual keywords and its associated sentiment (positive/negative) which will help to elaborate a more precise recommendation and justify it, and allows the use of different dictionary sizes to balance performance and user opinion interpretability. To prove the feasibility of the adapted recommendation system, we have tested the capabilities of our method to predict the sentiment strength score of item characteristics not previously reviewed. The experimental results have been performed with real datasets and the obtained F1 score ranges from 66% to 77% depending on the dataset used. Moreover, the results show that the method can generalize well and can be applied to combined domain independent datasets.

show abstract

mentioning

confidence: 99%