A Hybrid Approach of Prediction Using Rating and Review Data

Introduction

The need for the Recommender System (RS) has been increasing with the number of data companies generating these days. It's worth mentioning how these recommendation systems use user-specific and item-specific details to understand recommendations better. The key target remains traditional, that is to recommend personalized recommendations of a product to similar users or by finding similar products for the same user. To complete this objective, we use the filtering process. Filtering algorithms in a recommendation system are studied in two forms, one being content-based-filtering (CBF) and the second being, collaborative filtering (CF).

CBF focused on features such as profiles and items. With such specifications and the rating feedback of users for each set of things in the filtering process. However, this filtering process offers many advantages to the users but also has some limitations. These limitations are limited content analysis and over-specialization etc. On the other hand, the general philosophy is that similar users have similar tastes or related items that follow the same rating pattern.

To target the specified problem statement, there are a plethora of workflows available. Still, we introduce you to a novel rating prediction based recommendation system that surpasses either of the two CF-based algorithms that could be bifurcated as model-based CF or memory-based CF. In general, it is observed that model-based CF may provide better accuracy than memory-based, CF, but it has some limitations such as inflexibility and quality of predictions. Memory-based CF adopts a similarity computation method, which can be further bifurcated into user-based and item-based CF. Apart from this, we used user-similarity in predicting the rating in our proposed user-based CF approach, and at the same time, item similarity is used in item-based CF.

The intensive research on recommendation system components becomes very tedious when users generally do not provide the items' rating. Since we are dealing with the Amazon products daily, the number of users buying the same thing is comparatively very high than the number of ratings received in the user's feedback. Also, it remains null in many cases, which means the product has been rated even once. The above phenomenon increases the sparsity in the similarity matrix. Out of all the outstanding work yet proposed by many researchers to mitigate high sparsity weakness in CF-based RS. Most of all, utilize only the users' textual review Chen et al. (2015) Wu (2020) and ratings to mitigate the existing dataset's high sparsity level. As we propagated our research in the proposed model's direction, we observed that our proposed CF-based RS had significantly improved the performance. Our contribution in this paper has been summarized below:

• •

  Apply traditional CF-based methodology, i.e. matrix factorization using SVD to predict the ratings.

• •

  Apply proposed recommendation methodology, i.e. recommendation based on the users' reviews and rating as provided in the feedback process.

• •

  Comparison of traditional CF algorithm and proposed novel recommendation algorithm.

Background

Findings users' interest is always a challenging task due to the dynamic nature of a user. CF has become a solution to deal with this challenge described above since the early '90s. User-based CF is the most widely used technique in memory-based CF in the past few years. To achieve CF's general philosophy, a user-based CF utilizes a framework, as shown in figure 1. The detailed description of the framework of the user-based CF is as follows:

Data Collection

Let U= {α₁, α₂, . . . . α_m } and I={β₁, β₂. . . .. β_n} are the sets of m users and n items. A user-item rating matrix can be formed when a user gave a rating to a particular item. Implicit and explicit are the two methods utilized in data collection of CF Núñez-Valdez et al. (2018). In implicit data collection; systems gather data from users' activities such as searching behaviour, browsing history, time spent on a particular item, cursor movement, etc. On the other hand, in an explicit method, the system collects data directly from the user in reviews or ratings Berbatova (2019).

Figure 1.

Framework of the user-based Collaborative Filtering