Inferring User Demographics and Social Strategies in Mobile Social Networks

In this paper, we employ a real-world large mobile network comprised of more than 7,000,000 users and over 1,000,000,000 communication records (CALL and SMS) as the basis of our study, which we use to systematically investigate the interplay of user communication behaviors and demographic information. Through the study, we first reveal several intriguing social strategies that users of different ages and genders use to maintain their social connections. Based on the discoveries, we then develop the WhoAmI method to predict user demographics based on their communication behaviors. To the best of our knowledge, we are the first to study the problem of inferring user demographics and social strategies in such a real-large mobile network.

The mobile dataset used in this paper is extracted from a collection of more than 1 billion (1,000,229,603) call and text-message records of an anonymous country, which spans from Jul. 28, 2008 to Sep. 30, 2008. We construct two undirected and weighted mobile communication networks: call network (referred to as CALL) and messaging network (referred to as SMS). The CALL network consists of 7,440,123 nodes and 32,445,941 edges and the SMS network is composed of 4,505,958 nodes and 10,913,601 edges.

Human communications reveal interesting demographic-based social strategies. Specifically, we highlight several of our key findings as follows.

1. Young people are very active in broadening their social circles, while seniors have the tendency to maintain small but close connections. (Cf. Fig. 1).
2. People tend to communicate with others of both similar age gender, i.e., demographic homophily. (Cf. Fig. 2).
3. Frequent cross-generation interactions are maintained to bridge age gaps in family, workplace, education, and human society as a whole. (Cf. Fig. 3).
4. Cross-generation interactions between females are more frequent than those between males. (Cf. Fig. 3.b and Fig. 3.c).
5. Young male users maintain broader social connections than young female users. (Cf. Fig. 3.b and Fig. 3.c).
6. Opposite-gender communications are more frequent than those between same-gender users. (Cf. Fig. 3.d).
7. People expand both same-gender social groups ('FFF': Female-Female-Female and 'MMM': Male-Male-Male triads) and opposite-gender social groups (`FMM' and `FFM' triads) during the dating and reproductively active period (18-34 years old). However, upon entering into middle-age, people's attention to opposite-gender triads quickly disappears, and the insistence and social investment on same-gender social groups lasts for a lifetime. (Cf. Fig. 4)

All the observations above clearly demonstrate the strong demographic interrelations between users’ gender and age. For example, a 20-year-old female’s behaviors are distinct from not only a 20-year-old male’s, but also from a 50-year-old female’s. Based on this, we formalize the demographic prediction as a double dependent-variable classification problem, i.e., we infer gender and age simultaneously. 

Problem Definition: Demographic Prediction

Given a partially labeled network G = (VL, VU, E, YL, ZL) and the attribute matrix X, where VL is a set of users with labeled demographic information YL and ZL, and VU is a set of unlabeled users, the objective is to learn a function

f : G = (VL, VU, E, YL, ZL), X --> (YU, ZU)

to predict users’ gender and age simultaneously, where YU, ZU are the demographic information for the unlabeled users VU.

Prediction Framework: WhoAmI

We propose the WhoAmI method, a Double Dependent-Variable Factor Graph (DFG) Model. to solve this problem by leveraging not only the correlations between features and gender/age, but also the interrelations between gender and age.

Figure 5 shows an illustration of our proposed model, which consists of two layers of nodes. The bottom layer contains random variables and the upper layer contains the three kinds of factors introduced above. The joint distribution over the whole set of random variables can be factorized as the product of all factors. Specifically, the joint distribution can be factorized as follows:

Prediction Results: 

Table 1 shows the prediction results of different algorithms on the four prediction cases, i.e., gender and age predictions in the CALL and SMS networks, respectively. 

Clearly our DFG model yields better performance than the other alternative methods in the four cases. The Bag algorithm achieves the best prediction results among all non-graphical methods. In terms of weighted Precision, Recall, and F1-Measure, the DFG model achieves up to 10% improvements compared with the baselines for the prediction of users' gender and age. As for Accuracy, the DFG model can infer 80% of the users' gender in the CALL network and 73% of the users' age in the SMS network correctly.

The Double Dependent-Variable Factor Graph Model is implemented in C++. This CODE can also be used as a general Conditional Random Field or Factor Graph Model, please check the readme. Download the code here. 

Yuxiao Dong, Yang Yang, Jie Tang, Yang Yang, Nitesh V. Chawla. Inferring User Demographics and Social Strategies in Mobile Social Networks, KDD'14, 2014.

The full paper can be downloaded here [PDF download].