User opinions driven social recommendation system

Abstract

Recommendation systems provide reliable and relevant recommendations to users and also enable users’ trust on the website. This is achieved by the opinions derived from reviews, feedbacks and preferences provided by the users when the product is purchased or viewed through social networks. This integrates interactions of social networks with recommendation systems which results in the behavior of users and user’s friends. The techniques used so far for recommendation systems are traditional, based on collaborative filtering and content based filtering. This paper provides a novel approach called User-Opinion-Rating (UOR) for building recommendation systems by taking user generated opinions over social networks as a dimension. Two tripartite graphs namely User-Item-Rating and User-Item-Opinion are constructed based on users’ opinion on items along with their ratings. Proposed approach quantifies the opinions of users and results obtained reveal the feasibility.

Keywords

Tripartite graph user opinions item ratings recommendation systems

1. Introduction

A recommendation system with many items produces a wide variety of personalizing behaviors for users. This effects applications along with users in wide variety of domains such as movies, shopping, music, research articles and videos [1]. At instance, firms like Amazon, Netflix etc. provides recommendations to a user based on the user activity over such networks for promoting affinity of users towards these networks and increase sales profit. Further they limit the managing costs of picking things and selecting things among a variety of online setting production. Mostly, these systems are ached with cold-start problem and data sparsity [9]. Also, user choice or decision gets affected majorly by their friends’ opinions. Social Recommendation Systems (SRSs) model social behavior of user and recommends products or items to users based on user behavior over social networks. Such systems act on the social relations among users over social networks and help in recommendations.

As a part of literature review social information or opinions generated by the users are captured for providing recommendations (e.g., [1, 5]). Limited methods used ratings given by user for deriving recommendations [6]. These ratings are normally visualized as number of stars against the product. This rating gives a quick review over a product to a user, which seldom guarantees the quality and usage of the product. Reviews or opinions are expressed in a lot more extensive scope of structures and have the choice of being considerably more informative. Analyzing a review usually takes more time. Also reviews have potential to capture accurate experiences and more social data [22]. Hence, a review can be combined with a rating, or can be discrete as an Internet based social network post, blog entry, a report, or even a video [4].

In this paper, we present a new approach User-Opinion-Rating (UOR) for SRS using collaborative filtering [10] and community detection techniques [23] that include user opinions along with the user ratings for arriving at a better recommendation system. User opinion on products is analyzed using natural language processing techniques, and recommendations are predicted by combining rating as well as user opinions. In the proposed approach, we constructed two tripartite graphs namely user-item-rating and user-item-opinion and applied cosine similarity [5] to find the closeness of items. Prediction matrices, which help in determining the recommendations made to the users, are constructed using the tripartite graphs. We also constructed User-User social graph and derived communities using Louvain’s community detection algorithm [23]. The main contributions in this work are:

•
Proposed an SRS system that considers opinions of the users on items, and constructed two tripartite graphs: (a) user-item-rating and (b) user-item-opinion.
•
Derived two types of prediction matrices (based on ratings and opinions) using User-User social graph, and two tripartite graphs.
•
Performed experiments on Movielens, Amazon and email-Eu-core datasets, and predicted top–k recommendations for the existing users and cold start user.

The rest of the paper is organized as follows. In Section 2, we proposed background techniques associated with an approach and the related work. In Section 3, we present methodology for SRS. Section 4 presents the outcomes and Section 5 concludes the paper.
2. Background and related work

The major challenges in traditional recommendation approaches are [5, 17] and [20].

•
Sparsity problem: This rises when most of the users do not rate the item, results in sparse information of item and also effects over quality of recommendations. We addressed the sparsity issue by dividing the whole data in the form of communities and producing top-k recommendations from the community that the user belongs to. We modelled the data as User-Opinion-Rating tripartite graph and matrices are derived from the tripartite graphs respectively.

Sparsity can be reduced by using the community information to which user belongs to. According to the community to which a user belongs to the unrated item can be rated approximately. This helps in improving the quality of recommendations by reducing the sparsity in the data.

•
Cold-start problem: Recommendations for a new user is difficult as his/her behaviour is new and yet to be learned.
•
Scalability: Capacity to deal with huge scale information.
•
Independent and identically distributed: The traditional recommendation systems assume that every one of the clients are detached and identically distributed; this idea dismisses the relations among users that is not as indicated by the significant world suggestions.

A fruitful recommendation system may be developed by considering the following:

•
Individual users have their own attributes on different items such as movies, books, music, articles and food with different choices.
•
Users might be just guided by the companions they trust.
•
A user’s final judgment is the balance between his/her own opinion and his/her trusty friend’s reviews.

2.1 Related work

Recommendation systems are classified widely into four groups [7, 12, 21, 22]: Collaborative Filtering, Content-based, Knowledge-based and Hybrid recommendation systems:

1.
Collaborative filtering (CF): At first recommendation systems are based on CF techniques [12, 18, 27], which were relied on ratings given by user over products. These systems are divided into two kinds: user based CF and item based CF. Conversely, these systems suffer from cold start problem.

Table 1
Social recommendation systems

Model Based on Weaknesses Benefits

Tag-based models [15] Preferences of users and on social tags.
–
Suffers from cold start problem.
–
Sparsity.

–
Better performance.
–
Diversity of the recommendation.

Context-based models Users preferences.
–
Suffers from cold-start user problem.
–
Sparseness.

–
Diversity of the recommendation.
–
More accurate recommendations are generated.

Social influence-based models User relationship in the social network.
–
Suffer from cold-start user and Sparseness.
–
Diversity of recommendation not achieved.

–
Results more accuracy and better performance.

Trust-based models Trust measure of the similar users.
–
Suffer from cold-start user.
–
Lower performance.
–
Diversity of recommendation is not achieved.

–
More accurate recommendations are generated.

a) a)
User based collaborative filtering: The working principle of these systems is ratings given by user and neighbourhood over products. Recommendations to user are recommended based on the rating given by other users within the community and the kind of behaviour of user. These recommendations are generated to active user when opinions of similar users were published. This technique uses User-Item-Matrix for predicting the rating of active user based on ratings of similar users over an item.
b)
Item based collaborative filtering: This working principle of these systems is based on the similarity of items. For illustration, if a user purchases a product X then future recommendations will be recommended to user based on products which are similar to the product X. Item Based Collaborative Filtering includes two steps:

ii) i)
Similarity computation: Computes similarity between/ among the items.
ii)
Prediction computation: Predicts an item to a user.

Item Based CF overcomes the cold start problem of User Based CF.
2.
Content-based-filtering: The major attribute of this method is description of content. This content is user and item profile, keywords used and the time attribute associated with the content.
3.
Knowledge-based: This method uses knowledge of user’s needs and priorities for producing quality recommendations. This method is not responsive for short term variations, as this is static.
4.
Hybrid recommender systems: It is an integration of several recommendation Systems. It has points of interest of all the methods in combination, and addresses the issues such as sparsity and loss of data. Wide variety of algorithms are used as a result of which the hybrid recommenders systems’ architecture will be more complex.

2.2 Social recommendation systems

Model	Based on	Weaknesses	Benefits
Tag-based models [15]	Preferences of users and on social tags.	– Suffers from cold start problem. – Sparsity.	– Better performance. – Diversity of the recommendation.
Context-based models	Users preferences.	– Suffers from cold-start user problem. – Sparseness.	– Diversity of the recommendation. – More accurate recommendations are generated.
Social influence-based models	User relationship in the social network.	– Suffer from cold-start user and Sparseness. – Diversity of recommendation not achieved.	– Results more accuracy and better performance.
Trust-based models	Trust measure of the similar users.	– Suffer from cold-start user. – Lower performance. – Diversity of recommendation is not achieved.	– More accurate recommendations are generated.

Relationship such as user-user was not used in traditional recommendation systems although it is very prominent in social networks. These relationships when used, produce quality recommendations. The reason behind this is, a user usually exhibits affinity towards a similar product which was liked or purchased by his neighbours or friends. To predict the missing values, the relationships of users, their opinions and ratings are studied. Models implemented in various SRSs are given in Table 1.

Zhou et al. [25] projected a method SVD which is Singular Value Decomposition method utilizes incremental methodology, which combined the Approximating SVD (ApproSVD) algorithm with Incremental SVD algorithm, called the Incremental ApproSVD.

Park et al. [26] used a k-nearest neighbour (k-NN) graph for implementing a Reversed CF (RCF), a rapid CF algorithm. This methodology switches the way toward discovering k neighbours, rather than discovering $k$ comparative neighbours of unrated things, RCF finds the k-NN of evaluated things. In any case, RCF was not suitable for the situation where the strategy needs to foresee the evaluations for the majority of the unrated items.

To counter the cold start user problem Deepika et al. [5] proposed Collaborative filtering and Community based Social Recommendation System (CCSRS). Two prediction matrices derived from User-User social graph and User-Item bipartite graph were used in this approach. In this work, CCSRS work is extended by considering the users’ opinion. To relate users and items of interest communities are derived using Louvain’s community detection algorithm [23] for SRS. The mechanism of Louvain’s community detection algorithm is built on the modularity optimization and can investigate suggestively over larger networks. This algorithm is optimized using greedy approach which targets in discovering the best solution at each phase of building the large system of networks. The Louvain algorithm begins with every hub in a system that has a place with its very own locale. At first every network is a singleton, comprising of just a single hub. The algorithm at that point utilizes a separate measure depending on nearby moving heuristic to get an improved network structure. Singular hubs are moved starting with one network then onto the next until no further increment in measured quality is watched. Now, a decreased system is built. Every hub in the decreased system relates to a network in the first organize. In light of the subsequent network structure, a second decreased system is developed. The Louvain algorithm proceeds along these lines until a network is obtained that cannot be reduced further.

2.3 Motivation

Every real world entity is associated with lots of information over Internet. Practically a web user before purchasing will look into the ratings, opinions and reviews of the product. The available information may not reflect the quality of product. Hence there is a need to provide recommendations that reflect the quality of the product or item. These systems are based on not only user ratings but also opinions. There are number of methods or techniques available to build recommendation systems. To determine the item ranking, recommendation systems based on CF techniques uses User-Item bipartite graph [5, 6, 14, 15]. The concept of bipartite graphs used for recommendation systems is explained in the next subsection.

2.3.1 Bipartite graph

A Graph $G=(V,E)$ where $V=\{V_{1},V_{2}\}$ and $E$ , an edge that joins a vertex in $V_{1}$ to a vertex in $V_{2}$ is said to be bipartite graph if no edge exists within each partition. Vertices of bipartite graph are divided into two disjoint sets and no two vertices within the same set are adjacent.

Figure 1.

Bi-Partite graph between users and items.

In the traditional recommendation systems, the two sets correspond to Users and Items and the bipartite graph reveals the items in the Items set appraised by the users in the Users set [5]. Figure 1 shows a sample bipartite graph and its adjacency matrix; where, $u_{1}$ , $u_{2}$ and $u_{3}$ are users and $i_{1}$ , $i_{2}$ , $i_{3}$ and $i_{4}$ are items. Items purchased by each user are denoted using edges. Although this structure processes the ratings, it is not adequate to represent the opinions or reviews.

All in all, when a customer needs to purchase a thing from the market, he will search for number of parameters, for example, quality, durability and performance. Likewise, purchasers rely upon thing opinions and reviews that were posted on sites or from their companions’ inputs. Therefore, recommendation of a thing does not exclusively rely upon either opinion or rating.

•

Rating of a product is determined by the user built on either ranking a particular product based on its quality or performance compared with other products of the same category.

•

A review (or opinion) is a detailed description shared by the user on his/her experience with the product and aid as recommendation for the other users.

Ratings and reviews both offer knowledge, and both are various approaches to recognize which item, administrations or organizations that shoppers desire to buy and acquire. Reviews can be written in an assortment of structures and have the alternative of being considerably more instructive. A review can be joined by a rating, or can be discrete as a web based social post, blog entry, a report, or even a video. A review may set aside more effort to examine. Therefore, surveys have the potential for genuine bits of knowledge and more data. Star ratings do not give profound experiences into the item related characteristics. Rating is a form of generalization that offers faster access to how a business or service will perform without investing excessive amount of time. These generalizations may have a few constraints. They give the result however not the reasoning behind the rating. Reviews are significant as they uncover all the real world insights. The most concerning issue with surveys is that this data is a point of view, and a sentiment. A mix of the two reviews and ratings is a decent measurement to recognize the genuine estimation of the item. By allowing shoppers the chance to rate and survey the items, customers get the vital data alongside the outlined rating. Hence, for creating suggestions, in this work, we combined the trends of both the reviews and ratings and constructed the two tripartite graphs.

3. Proposed approach for SRS

Figure 2 shows the high-level view of proposed approach. The proposed approach mainly consists of three steps:

1.
Detect user communities from User-User social graph.
2.
Prediction matrices (see Section 3.3) are computed using user ratings and user opinions.
3.
Generate recommendation of items to users based on the predication matrices values.

3.1 Tripartite graph representation for social recommendations

A graph is termed as tripartite graph if and only if a graph $G=(V,E)$ has vertices decomposed into three disjoint subsets and within the same set, no two vertices are adjacent [2, 24]. In the tripartite graph, an edge between two vertices is permitted strictly when two vertices of the same edge belong to different subsets.

Figure 2.

Block diagram for proposed SRS approach.

In this proposed work, four sets are taken into account namely Users, Items, Ratings and Opinions to build tripartite graphs.

Consider a set of three users $U=\{u_{1},u_{2},u_{3}\}$ and set of four items $I=\{i_{1},i_{2},i_{3},i_{4}\}$ . Each user $u_{i}\in U$ gives an opinion value set $O_{u,i}\in O$ for a corresponding subset of item $i\in I$ . $O$ is opinion value domain which is set of quantified (transformed into numerical value) descriptive views. Figure 3 represents an example user-item-opinion tripartite graph, which shows the relationship between users, items and their respective opinion.

3.2 User communities from social graphs

Community discovery targets at partitioning the network into a set of communities of densely connected nodes; with the nodes belonging to different communities being only sparsely connected. Louvain’s community detection algorithm is used to find these sparsely connected communities.

Example: Figure 4 shows sample user-user social graph consisting of two communities $C_{1}$ and $C_{2}$ . $C_{1}$ the larger of the two contains users $U=\{u_{1},u_{2},u_{3},u_{4}\}$ . Table 2 shows the sample data consisting of the ratings as well as reviews from users. Table 3 shows polarity score for each movie. In the computation of prediction matrices and to generate recommendations for a user, user rating along with the polarity score obtained from user reviews [3] are used.

Table 2
User-movie-review-rating data

User Id	Movie ID	Movie review	Movie rating
$u_{1}$	001	This movie is very good and awesome characters with strong story line	5.0
$u_{2}$	002	This movie is not that bad and average flick with moderate story	3.0
$u_{3}$	001	This movie is very good and eye feast for movie lovers	4.0
$u_{4}$	–	–	–
–	–	–	–
$u_{100}$	523	An average routine movie one time watchable	2.0

Table 3

User-movie-review polarity-rating data

User Id	Movie ID	User review polarity (1–5)	Movie rating
$u_{1}$	001	5.0	5.0
$u_{2}$	002	3.25	3.0
$u_{3}$	001	4.1	4.0
$u_{4}$	–	–	–
–	–	–	–
$u_{100}$	523	2.25	2.0

Figure 3.

Example tripartite graph and its representation.

Table 4

Steps for computing recommendations for a new user

Step	Description	Computations
1	Take the largest community C1 {1, 2, 3, 4} to generate recommendations	User-movie-review polarity-rating data
User Id	Movie ID	User review polarity (1–5)	Movie rating
		$u_{1}$	001	5.0	5.0
		$u_{2}$	002	3.25	3.0
		$u_{3}$	001	4.1	4.0
2	Divide the matrix into user review polarity matrix and user rating matrix	User review polarity matrix			User rating matrix
		User Id	Movie ID	Review polarity	User Id	Movie ID	Movie rating
		$u_{1}$	001	5.0	$u_{1}$	001	5.0
		$u_{2}$	002	3.25	$u_{2}$	002	3.0
		$u_{3}$	001	4.1	$u_{3}$	001	4.0
		$u_{1}$	002	3.5	$u_{1}$	002	3.2
		$u_{3}$	002	‘2.75	$u_{3}$	002	2.75
		$u_{2}$	001	4.0	$u_{2}$	001	4.0
		$u_{4}$	–	–	$u_{4}$	–	–
3	Compute prediction matrices	Prediction matrix			Prediction matrix
	Movie Ids with polarities
		User Id	$M_{1}$	$M_{2}$	User Id	Movie Id	$P_{\text{avg}}$
		$u_{1}$	5.0	3.5	$u_{1}$	5.0	3.2
		$u_{2}$	4.0	3.25	$u_{2}$	4.0	3.0
		$u_{3}$	4.1	2.75	$u_{3}$	4.0	2.5
		$u_{4}$	?	?	$u_{4}$	?	?
		User Id	Movie ID	$P_{\text{avg}}$	User Id	Movie ID	$P_{\text{avg}}$
		u4	M1	4.37	u4	M1	4.33
		u4	M2	3.16	u4	M2	2.90
4.	Adding the two prediction matrices	User ID	Movie ID	$P_{\text{avg}}$
		u4	M1	4.35
		u4	M2	3.03
5	Recommendations to new user $u_{4}$	Top recommendations for user $u_{4}$ (Cold-start)
		$M_{1}$
		$M_{2}$

Table 5

Steps for computing recommendations for an existing user

Step	Description	Computations
1	Take the largest community $C_{1}\{1,2,3,4\}$ to generate Recommendations	User-movie-review polarity-rating data
User Id	Movie ID	User review polarity (1–5)	Movie rating
		$u_{5}$	001	5.0	5.0
		$u_{6}$	002	3.25	3.0
		$u_{5}$	002	3.5	3.25
2	Divide the matrix into user review polarity matrix and user rating matrix	User review polarity matrix			User rating matrix
User Id	Movie ID	Review polarity (1–5)	User Id	Movie ID	Movie rating
		$u_{5}$	001	5.0	$u_{5}$	001	5.0
		$u_{6}$	002	3.25	$u_{6}$	002	3.0
		$u_{5}$	002	3.5	$u_{5}$	002	3.25
3	Compute prediction matrices	Prediction matrix			Prediction matrix
		Movie Ids with polarities		Movie Ids with ratings
User Id	M1	M2	User Id	M1	M2
		$u_{5}$	5.0	3.5	$u_{5}$	5.0	3.25
		$u_{6}$	?	3.25	$u_{6}$	?	3.0
User Id	Movie Id	$P_{avg}$	User Id	Movie Id $P_{avg}$
$u_{6}$	$M_{1}$	5.0	$u_{6}$	$M_{1}$	5.0
4	Adding the two prediction matrices	User Id	Movie Id	$P_{avg}$
u6	M1	5.0
5	Recommendations to existing user $u_{6}$	User Id	Movie Id	$P_{CF}$	User Id	Movie Id	$P_{CF}$
$u_{6}$	$M_{1}$	3.51	$u_{6}$	$M_{1}$	3.42
6	Adding the two prediction matrices	User Id	Movie Id	$P_{CF}$
$u_{6}$	$M_{1}$	3.465
7	Adding the $P(u,i,c)=\alpha P_{\text{avg}}(u,i,c)+\beta P_{\text{CF}}(u,i,c)$	User Id	Movie Id	$P$
$u_{6}$	$M_{1}$	4.23

Consider first 3 users $u_{1}$ , $u_{2}$ and $u_{3}$ from Table 3, and $u_{4}$ is a new user. The steps to compute recommendation for the new user $u_{4}$ are shown in Table 4, and for an existing user, the computations are shown in Table 5.

3.3 Predicting recommendations

Calculating prediction matrix values based on user ratings: Using rating model for each user community obtained from the community detection model a

User-Item-Rating tripartite graph is first constructed (Fig. 3).

Figure 4.

Communities $C_{1}$ and $C_{2}$ from user-user social graph.

Table 6 shows the notations used in proposed approach. The description of User-Item-Rating tripartite graph model is given by the following equations:

$\displaystyle{S}^{R}_{i,j,c}=$ $\displaystyle\frac{\sum\forall u\in U({R}_{u,i,c}\times{R}_{u,j,c})}{\sqrt{{% \sum\forall u\in U({R}_{u,i,c})}^{2}}\sqrt{{\sum\forall u\in U({R}_{u,j,c})}^{% 2}}}$ (1)

Using user rating Prediction matrix $P^{R}_{\text{avg}}(u,i,c)$ is generated with the Eq. (2):

$\displaystyle{P}^{R}_{\text{avg}}(u,i,c)=i_{c}=\overline{R_{i,c}}$ (2)

where $\overline{R_{i,c}}$ is the average rating of the item $i$ in community $C$ .

Using Eq. (3.3) Prediction matrix $P^{R}_{\textit{CF}}(u,i,c)$ is generated as,

$\displaystyle{P}^{R}_{\textit{CF}}(u,i,c)=$ $\displaystyle\frac{\sum\text{all similar items $j$ (${R}_{u,j,c}\times{{S}^{R}% }_{i,j,c}$)}}{\sum\text{all similar items $j{S}^{R}_{i,j,c}$ }}$ (3)

Table 6

Notations

Notation	Description
$G(V,E)$	Social network with node set $V$ , edge set $E$
$n$	Quantity of items
$P_{\text{avg}}$	Opinion and item average rating for prediction matrix
${P^{R}}_{\text{avg}}$	Average rating of item based on prediction matrix
${P^{O}}_{\text{avg}}$	Opinion for item average based on prediction matrix
$P_{\textit{CF}}$	Item based collaborative filtering for prediction matrix
${P^{R}}_{\textit{CF}}$	Prediction matrix based on item based collaborative filtering using rating
${P^{O}}_{\textit{CF}}$	Prediction matrix based on item based collaborative filtering using opinion
$R_{u,i,c}$	Rating of item $i$ by user $u$ in community $C$
$S_{i,j,c}$	Similarity of $i$ and $j$ in community $C$
${S^{R}}_{i,j,c}$	Similarity between $i$ and $j$ in community $C$ Based on rating
${S^{O}}_{i,j,c}$	Similarity between $i$ and $j$ in community $C$ Based on opinion
$P_{u,i,c}$	User $u$ predicted rating on item $i$ in community $C$
$i_{c}$	Item $i$ belongs to community $C$
$u_{c}$	User $u$ belongs to community $C$
$I_{c}$	Group of all items belong to community $C$
$U_{c}$	Group of all users belong to community $C$
$P(u,i,c)$	Prediction of opinion-rating for item $i$ to user $u$ in community $C$

3.3.1 Calculating prediction matrix values based on user opinions

Analogous to user ratings, a User-Item-Opinion tripartite graph can be attained using opinion model for each user community. The description of User-Item-Opinion tripartite graph model is given by the following equations.

Similarity computation [5, 13] is performed as,

$\displaystyle{S^{O}}_{i,j,c}=$ $\displaystyle\frac{\sum\forall u\in U({O}_{u,i,c}*{O}_{u,j,c})}{\sqrt{\sum% \forall u\in U({O}_{u,i,c})^{2}}\sqrt{\sum\forall u\in U({O}_{u,j,c})^{2}}}$ (4)

Prediction matrix $P^{O}_{\text{avg}}(u,i,c)$ is generated as,

$\displaystyle{{{P}^{O}}_{\text{avg}}(u,i,c)=i}_{c}=\overline{O_{i,c}}$ (5)

Second prediction matrix $P^{O}_{\textit{CF}}(u,i,c)$ is generated using Eq. (3.3.1) as,

$\displaystyle{P}^{O}_{\textit{CF}}(u,i,c)=$ $\displaystyle\frac{\sum\text{all similar items $j({O}_{u,j,c}\times{S}^{O}_{i,% j,c})$ }}{\sum\text{all similar items $j{S}^{O}_{i,j,c}$ }}$ (6)

3.3.2 Adding prediction matrices values obtaining from both user ratings and opinions

For attaining optimal prediction values for recommendations, the overall $P_{\text{avg}}$ and $P_{\textit{CF}}$ values are calculated by using weighted average of values obtained from steps 2 and 3 as given below:

$\displaystyle P_{\text{avg}}(u,i,c)=w1\times P^{R}_{\text{avg}}(u,i,c)+w2% \times{P}^{O}_{\text{avg}}(u,i,c)$ (7) $\displaystyle P_{\textit{CF}}(u,i,c)=w3\times P^{R}_{\textit{CF}}(u,i,c)+w4% \times{P}^{O}_{\textit{CF}}(u,i,c)$ (8)

where $w1$ , $w2$ , $w3$ and $w4$ are weightages determined based on the application. Here, $w1+w2=1$ and $w3+w4=1$ . For experiments in this work, we considered equal weightage that is $w1=w2=w3=w4=0.5$ .

3.3.3 Recommendation of items to users

Two types of users are considered for giving item recommendations: (a) cold start user and (b) existing user. A new user can be added at any time posing a challenge to find appropriate recommendations to the newly joined user. To obtain recommendations to cold start users, top n items from the maximum of average opinion-rating prediction values are considered.

a) a)
Recommendation to cold start user

To attain at recommendations to cold start users top $n$ items from the maximum of average opinion-rating prediction values are considered. Recommendation to cold start user $=$

$\displaystyle\max^{n}_{i_{c}\to I}P_{\text{avg}}(u,i,c)$ (9)

where $P_{\text{avg}}(u,i,c)$ is taken from Eq. (7) and $n$ represents number of recommended items.
b)
Recommendation to existing user

For recommendations to existing user, we combine the values of average opinion-rating prediction values and opinion-rating collaborative filtering values and then top $n$ items are recommended.

$\displaystyle P(u,i,c)=\alpha P_{\text{avg}}(u,i,c)+\beta P_{\textit{CF}}(u,i,c)$

where ${P}_{\text{avg}}(u,i,c)$ and $P_{\textit{CF}}(u,i,c)$ are taken from Eqs (7) and (8) respectively.

Here, $\alpha$ and $\beta$ are weights, which are application specific. In this proposed work, equal weights for $\alpha=0.5$ and $\beta=0.5$ are considered.

4. Experimental results and discussion

Experimental results show the evaluation of proposed UOR approach using real world datasets namely Movielens dataset (https://grouplens.org/datasets/movielens/), Amazon instant video review dataset (http://jmcauley. ucsd.edu/data/amazon/), and social network dataset (Networks with Groundtruthcommunities) (http://snap. stanford.edu/data/email-Eu-core.html). Three main points are considered for evaluation. First point discusses how well the more number of predictions are made by proposed approach. Prediction Coverage [5] measure has been used to discuss first point. Second point discusses how best the accuracy has been achieved by proposed method. Two measures Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) have been used to discuss second point. Third point discusses how better the effectiveness and performance have been maintained by proposed approach. Precision and Recall measures have been used for the third point. Detailed analysis of these measures is discussed in the following.

4.1 Prediction coverage

Coverage is defined as the percentage of recommendations produced by the system. Let’s assume $I$ denote the set of available items and $I p$ denotes the set of items for which prediction can be made, then the prediction coverage is computed as a ratio of $|Ip|$ over $|I|$ . Figure 5 depicts the comparison of proposed UOR approach with the existing approaches CF and CCSRS [5].

Figure 5.

Prediction coverage comparisons for CF, CCSRS and UOR.

Results for First-N items ( $N=1,3,5,10$ ) were shown. From each user, they are randomly collected with users’ highest rated items list. The diagram clearly states that CCSRS has given good predictions compared to CF approach. Finally, the proposed UOR approach has dominated the CCSRS approach in covering the predictions.

4.2 MAE

Mean absolute error is a metric used to compute the average of all the absolute value differences between the true and the predicted rating [5]. Lower the MAE, higher the prediction accuracy. MAE precisely calculates the variation between the expected prediction and the actual rating preferably assigned by the user of the expected item. MAE generally measures the average magnitude of the errors in a bunch of predictions [5, 11]. MAE measures the average magnitude of the errors in a set of predictions. The Eq. (11) describes the MAE as follows [19].

$\displaystyle\text{MAE}=\frac{1}{n}\sum^{n}_{i=1}{|p_{i}-a_{i}|}$ (11)

where $n$ represents number of users, $p_{i}$ represents prediction and $a_{i}$ represents real ratings taken from the users. Figure 6 shows the comparison of MAE values generated by proposed UOR approach with the existing approaches SVD and KNN for the datasets Movielens, Amazon review data and social network datasets.

Figure 6.

MAE comparisons for SVD, KNN and UOR.

Results show that KNN has better MAE value compared to SVD approach whereas proposed UOR approach has given better value compared to KNN. It specifies that proposed approach outperformed the existing methods.

4.3 RMSE

RMSE computes the mean value of all the differences squared between the true and the predicted ratings and then proceeds to calculate the square root out of the result [5, 8]. Consequently, large errors may affect the RMSE rating, making it very useful metric to bring out the large errors. RMSE is $n^{\text{th}}$ part of root of sum of squares of differences of between the actual and predicted results of all items. RMSE could be elucidated as:

$\displaystyle\text{RMSE}=\frac{1}{n}\sqrt{\sum^{n}_{i=1}{{(p_{i}-a_{i})}^{2}}}$ (12)

Results are compared with the Singular Value Decomposition (SVD) [16, 25] and K-Nearest Neighbor (KNN) [26] approaches. Review information for data sets used in existing methods is not present. By combining data sets of Movielens and Amazon reviews data, a new synthesized data set is created. The new data set consists of fields like User ID, Movie Name, Movie Rating and Movie Review. Using Sentimental Analysis, reviews of users present in ‘Amazon Review dataset’ are converted into rating of scale 5.0. The other fields such as user ID are transformed to match the format in Movielens dataset. Figure 7 shows the comparison of RMSE values for the proposed approach with the existing methods SVD and KNN.

Figure 7.

RMSE comparisons for SVD, KNN and UOR.

The diagram depicts that KNN has better RMSE value compared to SVD approach whereas proposed UOR approach has shown better results compared to both SVD and KNN approaches.

4.4 Precision and recall

Precision is absolutely the fraction of the various items that we actually recommended the users concerned [5]. Precision is computed as:

$\displaystyle\text{Precision}=\frac{\text{Correctly recommended items}}{\text{% Total recommended items}}$ (13)

Recall represents the fraction of relevant items that are also part of the set of recommended items [5]. Recall is computed as:

$\displaystyle\text{Recall}=\frac{\text{Number of hits}}{|\text{testset}|}$ (14)

Figure 8.

Precision and recall for the existing and cold start user.

If an item $m$ is recommended, then it is referred to as a hit.

Figure 8 shows the values of Precision and Recall for the existing user and cold start user respectively. In addition to precision and recall, the parameters in the Table 7 give more insight into the performance of the proposed approach. It shows the top-10 recommendations for the cold-start user with user ID: 944.

Table 7

Top 10 recommendations for a new user with ID 944

S. No	Movie name
1	Three Lives and Only One Death (1996)
2	Princess Caraboo (1994)
3	Pompatusof Love, The (1996)
4	Waiting to Exhale (1995)
5	Walk in the Sun, A (1945)
6	Babyfever (1994)
7	New York Cop (1996)
8	Til There Was You (1997)
9	PennySarenade (1941)
10	Jackie Chan’s First Strike (1996)

5. Conclusion

In this paper, we addressed two main challenges. The first challenge that was highlighted by many recommendation systems is the cold start problem. The second challenge is to improve the accuracy of recommendations made by the system. This paper proposed a new approach (UOR) to handle the issues with the existing recommendation systems. Three datasets are used for carrying out the experimental work, Movielens, Amazon and email-Eu-core. This work uses user rating and user opinion to predict recommendations. Experiments on synthesized dataset created using Movielens and Amazon instant video review datasets, demonstrates an improved movie recommendation system. The performance of the proposed approach is measured using parameters MAE, RMSE, Precession, Recall and Coverage. The outcomes of the experiment on the datasets indicate that User-Opinion-Rating (UOR) provides better performance with respect to coverage and accuracy when compared with the existing approaches.

References

Anandhan

Shuib

Ismail

M.A.

and Mujtaba

, Social media recommender systems: Review and open research issues, (2018), IEEE Access.

Chen

Wang

Huang

and Mei

, Personalized video recommendation through tripartite graph propagation, Proceedings of the ACM Multimedia Conference & Co-located Workshops (MM’12) (2012), Nara, Japan.

Bhadane

Dalal

and Doshi

, Sentiment analysis: Measuring opinions, Procedia Computer Science 45 (2015), 808–814, Elsevier.

Devika

Sunitha

and Ganesh

, Sentiment analysis: A comparative study on different approaches, Procedia Computer Science 87 (2016), 44–49, Elsevier.

Lalwani

Somayajulu

D.V.L.N.

and Krishna

P.R.

, A community driven social recommendation system, IEEE Big Data 2015 (2015), 1070–1075, Santa Clara, CA, USA.

Zaho

Zhang

and Zaho

, Genre-based link prediction in bipartite graph for music recommendation, Procedia Computer Science 91 (2016), 959–965, Elsevier.

Ricci

Rokach

Shapira

and Paul Kantor

, eds, Recommender Systems Handbook, ISBN 978-0-387-85819-7, Springer New York Dordrecht Heidelberg London.

Fang

Jiang

and Wu

, Group recommendation systems based on external social trust networks, Hindawi, Wireless Communications and Mobile Computing Volume 2018, doi: 10.1155/2018/6709607.

Suryana

H.N.

and Hasanbasari

A.B.

, An understanding and approach solution for cold start problem associated with recommender system: A literature review, Journal of Theoretical and Applied Information Technology 96(9) (2018).

10.

Lee

and Lee

, Scalable deep learning-based recommendation systems, ICT Express (2018), Elsevier, doi: 10.1016/j.icte.2018.05.003.

11.

Palomares

and Kovalchuk

S.V.

, Multi-view data approaches in recommender systems: An overview, Procedia Computer Science 119 (2017), 30–41, Elsevier.

12.

Bobadilla

Ortega

Hernando

and Gutierrez

, Recommender systems survey, Knowledge-Based Systems 46 (2013), 109–132. Elsevier.

13.

J.H.

Chang

W.Y.

and Tseng

V.S.

, Personalized music recommendation by mining social media tags, Procedia Computer Science 22 (2013), 303–312, Elsevier.

14.

and Zhang

, A new algorithm based on bipartite graph networks for improving aggregate recommendation diversity, IOP Conf Series: Journal of Physics: Conf Series 887 (2017), 012056.

15.

Campana

M.G.

and Delmastro

, Recommender systems for online and mobile social networks: A survey, Online Social Networks and Media 3-4 (2017), 75–97, Elsevier.

16.

RamakrishnaMurty

Murthy

J.V.R.

Prasad Reddy

P.V.G.D.

and Satapathy

S.C.

, A new approach for finding appropriate number of clusters using SVD along with determining best initial centroids for K-means algorithm, International Journal of Enhanced Research in Science and Technology & Engineering 3(3) (2014), 430–437.

17.

RamakrishnaMurty

Murthy

J.V.R.

Prasad Reddy

P.V.G.D.

and Satapathy

S.C.

, Dimensionality reduction text data clustering with prediction of optimal number of clusters, International Journal of Applied Research on Information Technology Computing (IJARITAC) (2011), 41–49.

18.

Zeng

Cao

Chen

and Li

, User behaviour modeling, recommendations, and purchase prediction during shopping festivals, Electronic Markets (2018), doi: 10.1007/s12525-018-0311-0318.Springer.

19.

Katarya

and Verma

O.P.

, An effective collaborative movie recommender system with cuckoo search, Egyptian Informatics Journal (2016), doi: 10.1016/j.eij.2016.10.002.

20.

Panigrahi

L.R.

and Stitipragyan

, A hybrid distributed collaborativefiltering recommender engine using apache spark, Procedia Computer Science 83 (2016), 1000–1006, Elsevier.

21.

Sehgal

Chaudhry

Biswas

and Jain

, A new genre of recommender systems based on modern paradigms of data filtering, Procedia Computer Science 92 (2016), 562–567, Elsevier.

22.

Sohaila

S.S.

Siddiquia

and Ali

, Feature extraction and analysis of online reviews for the recommendation of books using opinion mining technique, Published by Elsevier (2016), doi: 10.1016/j.pisc.2016.06.0792213-0209.

23.

Blondel

V.D.

Guillaume

Lambiotte

and Lefebvre

, Fast unfolding of communities in large networks – louvain method, Journal of Statistical Mechanics: Theory and Experiment, Online at stacks.iop.org/JSTAT/2008/P10008, (2008), IOP Pub. Ltd and SISSA.

24.

Shiwei

Yonghua

and LinJu

G.H.

, The research of recommendation algorithm based on complete tripartite graph model, International Journal of Hybrid Information Technology 8(12) (2015), 205–214.

25.

Zhou

Huang

and Zhang

, SVD-based incremental approaches for recommender systems, Journal of Computer and System Sciences (2015).

26.

Park

Jung