Keywords

1 Introduction

Since the emergence of social networks and microblogging sites, such as Twitter and Sina Weibo, hundreds of millions of users have become to use the microblogging service as a tool to propagate and share information on the Internet [1]. Personality affects everyone’s thinking, behavior and decision-making and maintain its long-term stability and affect its social relations [2]. The behavior of users in the online world and in the real world is the same [3]. However, there are few researches on personality theory in the previous studies.

The researcher combines personality psychology and social network analysis to analyze and predict the user’s personality through the user’s behavior and other data. The potential value of personality prediction will be very helpful in solving these problems.

The rest of this paper is organised as follows. In Sect. 2, we discuss related work. Section 3 presents the followee recommendation method and the UPS model we propose. In Sect. 4, we evaluate the performance of our design experiment, present the proposed weights for quantitatively evaluating users’ personality and how to combine it with other recommendation and compare to existing schemes. Section 5 summarises the conclusions obtained from the performed experiment evaluation.

2 Related Work

Personalized recommendation technology is the core and critical technology of E-commerce recommendation system. However, these traditional methods only focus on how to improve the accuracy of recommendation, while ignoring the inherent characteristics of the user behavior is determined by their personality characteristics [4].

Some more related worked on followee recommendation has also been proposed. Most of the existing followee recommendation systems on micro-blogging platforms rely on either topological or content-based factors. Hannon et al. [5] use bag-of-word model to exploit the content information created by a user and recommend him/her followees based on the content similarity between the candidate followees and the target user. Sun et al. [6] construct a diffusion graph to select a small subset of tweets as recommended emergency new feeds for regular users. There are a few people use personality trait in followee recommendation systems. Wu et al. [7] generate personalized tags for Twitter users to label their interest by extracting keywords from tweets they post. Armentano et al. [8] propose a followee recommender system using social relation features, including user popularity and number of common friends, to measure the relevance among users in Twitter. However, the recommendation effect of adding personality traits has improved compared with the traditional recommendation method, but there is much room for improvement.

3 Our Method

3.1 User-Based Factor

The Analysis of User Attributes of Microblogging. In this section, we discuss the representation of the micro-blogging user model. In this work, we consider 4 properties of user information, which are location information, micro-blogging text, social information and interactive information. Thus, the user model of the user u is represented as follows:

$$\begin{aligned} Profile_{user}(u)=\{Place(u), Posts(u), Relation(u), Interaction(u)\} \end{aligned}$$
(1)

where Place(u) is the location information of the user u, both of which are short text and can be represented as a string. Post(u) represents the long text that the user u has released into the micro-blogging, which is represented as a text vector. Relation(u) represents the social information of u, including two kinds of attribute information, which is concerned about attention information, fans information. Interaction(u) represents the interactive information of user u, including two kinds of attribute information, that is, forwarding information and comments.

Firstly, we define the user(u, v) in the various attribute similarity and then weighted and combined. The similarity is given by the following equation:

$$\begin{aligned} \begin{aligned} sim(u, v)&=\omega _{1}sim(Place(u), Place(v))+\omega _{2}sim(Post(u), Post(v)) \\&\quad +\omega _{3}sim(Relation(u), Relation(v)) \\&\quad + \omega _{4}sim(Interaction(u), Interaction(v)) \end{aligned} \end{aligned}$$
(2)

In essence, \(\omega _{i}\) is the weight of each attribute similarity, and \(\omega _{1}+\omega _{2}+\omega _{3}+\omega _{4}=1\). The sim(Place(u), Place(v)) = 1 if the information of the provincial and city of the Place(u) and Place(v) are the same. The sim(Place(u), Place(v)) = 2/3 if the information of provincial of the Place(u) and Place(v) is the same, but the city information is different. For two users(u) and user(v), their micro-blogging text can be represented as two text feature vectors: \( Post(u)=(w_{u1},w_{u2},...,w_{un}), Post(v)=(w_{v1}, w_{v2},..., w_{vn})\). The text similarity sim(Post(u), Post(v)) is calculated by cosine similarity, which used TF-IDF. The similarity of the relationship information between the two users are computed weighted averages of Followee(u) and Followee(v). Also, we use forwarding and comment numbers to compute the similarity of sim(Interaction(u), Interaction(v).

3.2 Personality-Based Factors

The Big-Five Personality Model. The “Big Five” model of personality dimensions has emerged as one of the most well-researched and well-regarded measures of personality structure in recent years. The model’s five domains of personality, Openness, Conscientiousness, Extroversion, Agreeableness, and Neuroticism, were conceived by Tupes and Christal as the fundamental traits that emerged from analyses of previous personality across age, gender, and cultral lines.

The Matching Calculation of the Personality Traits. TextMind is a Chinese language psychological analysis system developed by Computational Cyber-Psychology Lab, Institute of Psychology, Chinese Academy of Sciences. Through the relationship between the function of these words and the text, in this paper, we can obtain the relationship between each word in TextMind Chinese psychological analysis system dictionary and each specific factor in the big five personality. The character factor values for the ith dimension in word w are defined as follows:

$$\begin{aligned} BFM(\omega _{i})=\frac{\sum _{j=1}^{n}P_{ij}}{n} \end{aligned}$$
(3)

In Eq. (3), \(\omega _{i}\) denotes the ith personality factor in word w. n indicates that word w has n functional nouns. \(P_{ij}\) represents the relationship between the jth word function and the ith personality factor. The functional word of TextMind and BFM corresponding a correspondence table can be divided into five groups.

In this paper, we calculate the personality score vector of each dimension of the big five personality according to the BFM correspondence table [9]. The user u’s personality score in the ith dimension is calculated as follows:

$$\begin{aligned} Score(u)_{i}=\frac{\sum _{j=1}^{N_{i}}k_{\omega _{j}}\cdot BFM(\omega _{ji})}{N_{i}} \end{aligned}$$
(4)

Where \( BFM(\omega _{ji}) \) denotes the personality factor value of the jth-class word in the ith dimension in the micro-blogging text published by the user u. \(k_{\omega _{j}}\) represents the word frequency of the jth-class word \(\omega \) of the microblogging text published by the user u. \(N_{i}\) is the total number of functional words that are statistically relevant under the ith personality dimension. We use the Eq. (4) to calculate the average score between five dimensions in the table, which were \(\mu _{E}\), \(\mu _{A}\), \(\mu _{C}\), \(\mu _{N}\), \(\mu _{O}\), and then we put the user scores compared with the average correlation score.

The User Personality-Similarity Model. In this Section, we proposed the user personality-similarity model to make the final recommendation. The personality matching score between a user u and the potential blogger pf is expressed as follows:

$$\begin{aligned} TPM(u, pf)=\mu (\sum MS(u, pf,dim)) \end{aligned}$$
(5)

In Eq. (5), TPM(u, pf) is the total personality matching (TPM) score between user u and potential followee pf. \(\mu \) is the average value of each dimension. MS(u, pf, dim) indicates the personality matching score of the user u and the potential recommendation followee pf in a certain dimension.

Next, the matching score calculation formula for each dimension is defined as follows [10]:

$$\begin{aligned} scoreAgreement(u, pf, dim) = \left\{ \begin{array}{ll} 0.5 &{} \text {both u and pf are dimension}\\ 0.25 &{} \text {either u or pf are dimension}\\ 0 &{} \text {None is dimension} \end{array} \right. \end{aligned}$$
(6)

Finally, we formally define the evaluation of the formula in the UPS model as follows,

$$\begin{aligned} FScore=\gamma _{1}sim(u, v)+\gamma _{2}TPM(u, pf) \end{aligned}$$
(7)

Where \(\gamma _{i}\) is the weight value of the formula, and \(\gamma _{1}+\gamma _{2}=1\). Finally, we use this formula as the final evaluation formula in followee recommendation.

4 Experiment

In this experiment, We use the microblogging users between the various attributes of information similarity calculation method for user recommendation experiment. In this paper, two evaluation indicators are used to evaluate the user similarity algorithm which are p@N and average precision.

We use the large-scale dataset crawled from Sina Weibo, the most popular microblogging system in China. The dataset contains 256.7 million users’ social link information and 550 million tweets. The dataset include tweets, user relations and user background information.

Fig. 1.
figure 1

Recommendation results based on user similarity

Full size image
Fig. 2.
figure 2

Personality-aware of compared methods

Full size image

4.1 Results

In the experiment, the weight of the attribute similarity calculation formula between the user(u) and user(v) is calculated by Analytic Hierarchy Process (AHP).

Figure 1 shows the experimental results of the above similarity on P@N, which reflects the motivation of the experiment is to examine the user information and its four kinds of attribute information (place, tweets, relation, interaction) in the calculation of user similarity performance. The experiment results show that compared with the four attribute information of user information, the P@N results of the similarity of relation are best in the previous P@25 and the overall similarity of the users is the second.

Unlike when computing the personality profiles in which ever term appearing in tweets was considered, terms in tweets were filtered according to a processing approach. The approach considered the full-text of tweets (named Original). The effect of adding personality as a factor in user-based followee recommendation is shown in Fig. 2. The figure summarises the average precision for each of the predefined N, including results for six linear combinations of factor’s weights. As a result, it can be stated that considering a quantitative analysis of personality in combination with user-based factors could help to correctly place the most important or interesting users in the first positions of the ranking of suggested users.

5 Conclusions

In this paper, we propose a User Personality-Similarity (UPS) model for followee recommendation, a novel personality followee recommendation scheme over microblogging systems based on user attribute and the big-five personality model. This paper analysed how user personality conditions the followee selection process by combining a quantitative analysis of personality traits with the most commonly used predictive factors for followee recommendation. The combined attributes were insert into a recommendation algorithm that computed the similarity among target users and potential followees. We conduct experiments using large-scale traces form Sina Weibo to evaluate our design. Results show that UPS model greatly outperforms existing recommendation schemes.