Sentiment based matrix factorization with reliability for recommendation

Matrix Factorization with Topic and Sentiment Analysis for Rating Prediction

“…Hyun et al [8] proposed a CNN-based recommendation method that is guided to incorporate the sentiments when modeling the users and items. Shen et al [17] presented SBMF+R model based on the probability matrix factorization, incorporated the ratings, sentiment intensities, and helpful votes from other users for prediction simultaneously.…”

“…Let R 5 and R 1 be N × M user-item rating matrices such that r 5 ij ∈ R 5 and r 1 ij ∈ R 1 respectively. Also, we denote the textual review of user u i on item v j as d ij , and the sentiment intensity of d ij extracted by VADER [7] method in the third-party toolkit NLTK or any method based on lexicon [17,1,9] as s 5 ij and s 1 ij , where s 1 ij is original sentiment intensity in the scale of [-1, 1] and s 5 ij is in the scale of [1,5] converted from s 1 ij according to the following formula:…”

“…SBMF+R [17] is an improved model based on PMF, which adds the sentiment intensity extracted from user reviews, and takes the reliability measure into account simultaneously as shown in Figure 1. We can clearly see in the middle of the figure that the reliability measure is extracted from the user's feedback and used in the SGD training.…”

“…Further, we introduce the reliability measure for rating r 5 ij , based on the helpful votes of its corresponding review d ij and item consistency t j . Inspired by previous study [17], we define it as wv ij as follows:…”

“…In order to solve this problem, efforts are made to recognize and characterize such opinions into sentiments and topics, to enhance the performance of MF. Existing studies include the applications of the topic model [2,3,19], sentiment analysis [21,17], and their combination [14,22,20].…”