Keywords

1 Introduction

With the increasing growth of online information, recommender systems have become an essential tool to efficiently manage this information and ease users with their decision-making procedure [27, 28, 48, 53]. The main purpose of several e-commerce platforms, such as Amazon, Last.fm and Netflix is to monitor users’ behavior (e.g., likes, ratings, comments) to understand the user preference on a set of items and use this information to recommend related items that match with users’ interests.

Matrix Factorization (MF) is one of the most successful collaborative filtering approaches in single domain recommender systems which have been widely adopted in the literature [37]. MF tries to learn latent vector of user-item interactions and realize users’ interests on an unseen item. In this context, data sparsity can be a challenging problem. An example would be in a single domain scenario, when the limited number of users’ interactions are available and they are not able to capture users’ preferences comprehensively subsequently [30].

In real world, users may use different systems for different reasons. For example, users might prefer to use their Facebook account in order to make a new friend or choose LinkedIn for their business purposes, and choose Netflix to watch videos. Aggregation of these activities on different domains provides an opportunity to understand users’ behavior properly and generating cross-network recommendation. In particular, cross-network recommendations have emerged as a solution to cope with the long-standing data sparsity problem [47]. These systems are able to monitor users’ behaviors on multiple domains and discover users’ preferences completely; thus, improve the recommendation accuracy [31]. Although cross-network recommender systems have shown a great improvement to tackle the data sparsity problem, their performance is limited due to some difficulties. They assume that users’ preferences on items are likely to be constant over a period of time and provide users with similar items to those they preferred in the past, degrading recommendations diversity.

For instance, during the Olympic games, a user may be interested to watch wrestling matches on YouTube, expanding new interests. After Olympic, however, the user may have no further interest on wrestling videos and prefer to watch other types of videos. Accordingly, users’ preferences may change over time and therefore, there is a need for new approaches to analyze and understand the personality and behavior of users over time. This in turn will create an environment for users to get recommendations with a various set of interesting and unexpected items and can increase users’ satisfaction, business profit and loyalty.

To achieve this goal, in this paper, we propose a novel approach to detect users’ personality type implicitly, without any burden on users and incorporate it into matrix factorization in order to identify users’ interests completely and broaden users’ suggestions.

The rest of the paper is organized as follows: Sect. 2 presents the related work. We present an overview and the framework of the proposed approach in Sect. 3. In Sect. 4, we present the results of the evaluation of the proposed approach, before concluding the paper with remarks for future directions in Sect. 5.

2 Related Work

2.1 Recommender Systems (RSs)

Recommender Systems are known as techniques which help both users and companies. Their aim is to assist customers with decision-making procedure to find interesting items matching with their preferences. The growing number of available digital information and due to increasing popularity of visitors to the Internet create the information overload problem. Therefore, systems such as Google have appeared to deal with this problem and help users to discover their interested items. Here, there is an increasing need for system to solve this problem and assist users has emerged. Recommender systems have known as an information filtering systems which mitigate information overload problem by filtering crucial information from all collected information [2]. Recommender system trace user’s actions and history and collect information preferred items and rating pattern to predict which items are more likely to prefer in the future.

Recommender systems can be beneficial to both providers and customers. There are various reasons why recommender systems attract providers’ attention; firstly it can boost sales rate which can be an essential reason for service providers to recommend items with the highest possibility of acceptance, secondly suggest different items where might not be achieved without recommender system in which captures user’s interest and finally increase loyalty and user’s satisfaction. In user’s point of view, they are eased with their decision-making procedure as recommender systems filter their desired and interesting items to them. There are five different type of recommender systems in the literature which have been investigated widely [22].

figure a

Content-Based. Recommending items similar to those that a user likes before, say in movie recommender, if the user watched and liked a drama genre movie, another drama one will be recommended to this user. The main goal of content-based recommender system is to find items in which attributes are similar to users’ profile [41]. In order to discover similarity between items there are different either statistical analysis such as Naïve Bayes Classifier [26] or machine learning techniques like Decision Trees [50] or Neural Networks [17]. Item is a general concept which regarding to the recommender system suggestions can be CD, Movie, Book and etc.

Collaborative-Filtering. Methods belonging to this category can be divided in the two different classes known as memory-based and model-based techniques. Model-based approaches learn the user-item ratings to predict user’s interest on an unseen item, including Bayesian Clustering [19], Latent Dirichlet Allocation [18], Support Vector Machines [29] and Singular Value Decomposition [16, 36, 54, 55]. While methods in Memory-based class use similarity metric to measure similarity either between users or items [25, 40].

Knowledge-Based Recommender tries to acquire knowledge about domain from users to make recommendation more accurate [59]. These systems explicitly ask user’s preferences and then make an appropriate recommendation.

Demographic-Based Recommender aim to find demographic information about users such as age, nationality, gender to provide a better recommendation which suits user’s interests [39, 45].

Hybrid approaches merge mentioned techniques together to benefit from their advantages in one model [21].

Data sparsity is one of the shortcoming that Collaborative Filtering (CF) approaches are confronted with. Some works such as CTR integrates topic molding to use additional information like the contents of documents to make a recommendation [60] and TopicMF which not only uses ratings but also exploits review texts to discover more data from them [5]. Although resorting to extra information can create an environment for recommendation systems to better understand users’ preferences, but they might be infeasible in the real-world scenarios. In order to understand users’ preferences completely, some other studies provide a questionnaire for users to directly ask their interests on different items [49]. The major difficulty of these kinds of approach is that users may avoid to participate in filling a questionnaire as it is a time consuming task.

In contrast to recommender systems on single domain, cross-network approaches appear to mitigate data sparsity problem and improve recommendation accuracy. They enrich data and generate accurate user profile with the help of auxiliary domain [44]. Although, widely attempts have been done in the literature to alleviate the data sparsity problem, diversity is a key factor that has been neglected in the most of them (Table 1).

Table 1. Big five factor features
Full size table

2.2 What Is Personality?

Personality was explained as “consistent behavior pattern and interpersonal processes originating within the individual” [20]. From psychological point of view, people differ in their behaviours and attitudes, which can be explained by their personality type. Personality is a stable feature without no changes over time. In terms of psychological view, there are different personality traits which among all Five Factor Model (FFM) is “the dominant paradigm in personality research and one of the most influential models in all of the psychology” [42]. The Big Five structure does not imply that Personality differences can be reduced to only five traits. Yet, these five dimensions represent Personality at the broadest level of abstraction, and each dimension summarizes a large number of distinct, more specific Personality characteristics”  [34]. FFM has five principal dimensions Openness to experience, Conscientiousness, Extraversion, Agreeableness, Neuroticism (OCEAN). As it is clear from Table 3, FFM are accompanied with different features;

2.3 Personality and User’s Preferences

It is conducted that our personality type plays an important role in our preferences on music, movies, TV shows, books and magazines [52]. This correlation provides an opportunity for RSs to suggest a divers set of items to users. To extract personality, Linguistic Inquiry and Word Count (LIWC) tool is a successful platform to identify 88 categories of linguistic features relevant to each domain of the FFM [56]. Research findings confirm that there is a strong correlation between personality type and user’s preferences in various fields, like music [33] books and magazines [51].

Fig. 1.
figure 1

Personality detection

Full size image
Table 2. LIWC categorizes [46]
Full size table

2.4 Personality Recognition

Personality is a domain-independent and stable factor that can be extracted explicitly i.e., questionnaire or implicitly. In order to find personality type implicitly, we can analyze user’s behaviours, actions like posts, comments and etc. Moreover, by analyzing digital or language-based features of written texts we will unable to predict user’s personality type implicitly with no need to user effort [4]. While explicit personality detection is more easier, it is time-consuming task and participants might be unwilling to attend due to privacy concern. In this type of personality recognition, individuals are asked to answer questions regarding to specific psychological personality model. Below we list the popular questionnaire regarding to the Big Five Factors [57]:

  • 240-items NEO-PI-R [43];

  • 300-items NEO-IPIP [35];

  • 100-items FFPI [1];

  • 132-items BFQ [6];

  • 120-items SIFFM [58];

We detect the user’s personality type implicitly with the help of Linguistic Inquiry and Word Count (LIWC) tool to understand how many words of users’ reviews are related to each category of this tool. Below, we represent LIWC categorize based on the [46] (Fig. 1):

3 Methodology

3.1 Preliminary Knowledge

Assume we have K items \(V=\{v_1,v_2,\cdots ,v_K\}\), and L users \(U=\{u_1,u_2,\cdots ,\)\(u_L\}\) and is the rating matrix, and \(R_{ij}\) indicates ratings which have been given to item i by user j. Let represents the personality matrix, where \(W_{ij}=\{0,1\}\), and when \(W_{ij}=1\) it means that users \(u_i\) and \(u_j\) have a similar personality type.

3.2 Our Model

\(R_{ij}\) predicts the value for unrated items which user \(u_i\) will give to item \(v_j\);

$$\begin{aligned} \large R_{ij}= p_{i}^{T}q_{j} \end{aligned}$$
(1)

Where \(p_i\) and \(q_j\) are latent feature vector for user i and item j respectively,

$$\begin{aligned} \begin{aligned} \large&\ min {\dfrac{1}{2}}\sum _{i=1}^{O}\sum _{j=1}^{L}I_{ij}\Bigg ({R_{ij}}-\bigg (\gamma p_i^T q_j + \Big ((1-\gamma ) \sum _{t \in \theta ^{+}_i} W_{it}p_t^Tq_j\Big )\bigg )\Bigg )^2\\ +\,&\alpha _1\Vert P\Vert _F^2 + \alpha _2\Vert Q\Vert _F^2 \end{aligned} \end{aligned}$$
(2)

In the above equation, \(I_{ij}\) = 1, if user i has rated item j, otherwise \(I_{ij}\) = 0. Matrix W contains personality information, and \(\theta i ^{+}_i\) is the set of users who are in the same personality type network with user i, and \( \beta \) is a controlling parameter to control the weight of user’s preferences. To save space, we omit the detailed of the updating rules.

Table 3. Datasets statistics
Full size table

4 Experimental Settings and Analysis

4.1 Datasets and Evaluation

We have selected Amazon dataset, which consists of a large number of reviews. In this paper, we have used Amazon Instant-video, because of its high relation between user’s preferences on video and their personality types and leave other domains for cross-network recommender as our future works. The dataset includes 2000 users who wrote more than 3 reviews.

Evaluation. We select two popular evaluation metrics, Mean Square Error (MSE) and Root Mean Squared Error (RMSE);

$$\begin{aligned} \large MAE=\dfrac{\sum _{(u,i)\in R_{test}}\mid {\hat{R_{ui}}-R_{ui}}\mid }{\mid {R_{test}}\mid },\quad \quad \end{aligned}$$
(3)
$$\begin{aligned} \large RMSE=\sqrt{\dfrac{\sum _{(u,i)\in R_{test}}\mid {\hat{R_{ui}}-R_{ui}}\mid }{\mid {R_{test}}\mid }} \end{aligned}$$
(4)

where, \(R_{ui}\) and \(\hat{R_{ui}}\) are real and predicted ratings values respectively, and \(R_{test}\) represents the user-item in the test dataset.

In order to comparison, we select SVD++, a single model in which integrates both neighborhood and latent factor approaches [36], Hu which proposes a metric to use of user personality characteristics and rating information  [32] and Random model (Figs. 2 and 3).

Fig. 2.
figure 2

MAE comparison

Full size image
Fig. 3.
figure 3

RMSE comparison

Full size image

4.2 Performance Comparison and Analysis

As it can be seen from Table 2, we use different sets of the training data size \((60\%,70\%,\) \(80\%,90\%)\), when we increased the size of training dataset the performance of all methods was improved. Therefore, to have a fair comparison, we consider the results related to \(90\%\) training size. Our proposed model, CNR, shows the best performance in terms of both RMSE and MAE among all approaches. The performance of CNR is improved compared to the SVD++ by 64%, 50% in terms of MAE and RMSE respectively and SVD++ performs around 2 and 6 times better compared to the Hu and Random methods in both evaluation metrics.

Analysis and Summary. As it is clear from the results our proposed model performs better in both MAE and RMSE metrics. CNR shows the best performance over compared methods, which do not pay attention to users’ personality type which have a strong relation with their preferences. We further investigate the time-based relationships [7] and effects of time factor in our future works. Finally, we will also focus on Information extraction [12,13,14] and Natural language processing techniques to have a more accurate users’ reviews analysis (Table 4).

Table 4. Performance analysis on the Amazon dataset
Full size table

5 Conclusion and Future Work

5.1 Conclusion

In this paper we have proposed a novel recommender system, in which exploiting user’s written reviews to discover their personality type which plays an important role in users’ decision-making process. Extensive validation on Amazon dataset demonstrates the advantages of our approach compared to the other methods in terms of both RMSE and MAE. In our future work, we will discover users’ personality characteristics and make a recommendation in separate domains to have a cross-domain recommendation. Furthermore, according to the Sect. 4.2, CNR is able to recommend divers set of items to users.

5.2 Future Work : Behavioural Analytics and Recommender Systems

Behavioural Analytics, a recent advancement in business analytics that focuses on providing insight into the actions of people, has the potential to enable Recommender Systems understanding the personality, behaviour and attitude of its users and come up with more accurate and timely recommendations. As an ongoing and future work, we plan to:

  • link Behavioural Analytics and Recommender Systems by collecting the activities of Recommender Systems users. We will introduce the new notion of Behavioural Provenance [11, 15], to be able to trace the user activities back to their origin and provide evidences to understand the personality, behaviour and attitude of Recommender Systems users.

  • transform the collected (raw) user activity data into contextualized Behavioural data and knowledge. We will use our previous work, Knowledge Lake [8, 9] to automatically curate the Behaviour data and provenance, and to prepare it for analytics and reasoning.

  • introduce a new generation of smart Recommender Systems, by leveraging the advances in natural language processing [14] machine learning [3] and crowdsourcing [10], to leverage the contextualized data and knowledge (generated in previous steps), and to provide cognitive assistant to the Recommender System users.