Social friendship and crowd mobility: User mobility in a location-based social network (i)

Original title: Friendship and Mobility:user Movement in location-based social Networks

Author unit: Stanford University Published: 2011

Conference: 17th Annual ACM SIGKDD International Conference--Knowledge discovery and data mining

Reference:Cho E, Myers S A, Leskovec J. Friendship and Mobility:user movement in location-based social networks[c]//Proceed Ings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and data mining. ACM, 2011:1082-1090.

Absrtact: Despite the high degree of freedom and variability of human mobility and migration, there are still certain structures due to geographical and social constraints. This paper uses telephone location data and two location-based social networks to understand the basic human activity patterns. Short-haul movement does not affect social structure, long-distance travel is restricted by social relations.

Social relations explain 10%-30% human migration activities, while periodic behavior reveals 50%-70%. In this paper, a human mobile model is established to predict the position and dynamics of human activities in the future.

1. Introduction

related research. "7,23" has a strong cyclical habitual movement of humans as between home and work units. The "12" human movement is limited by the geographic distance that can be moved in a day. "11,26" mobile can be further restricted by social relations, such as visiting a friend's good residence.

However, due to the difficult collection of reliable human mobile data, these studies and assumptions have limitations. With the advent of emerging social networks such as Foursquare, Facebook, Gowalla, and the collection of landing sites (checking-in), human mobile data can be obtained. The traditional way is to use mobile phone, LAN positioning to determine the mobile location. But when you move from the second-floor studio to the café on the first floor, it's clear that location positioning is more accurate than mobile positioning. Location positioning is more sporadic, mobile positioning is concentrated. In either case, network information can be collected. The former applies to the Friendship network, the latter applies to the communication network. The data mentioned later will be used to investigate the three main human activities: Where to move (geographic movement), the frequency of movement (temporal dynamics), the impact of social relations on the Movement (Social network). We will start with these three aspects of the study.

Broadly speaking, there are many applications for understanding human mobility patterns. For example, help improve large-scale computing, content-based publishing sites, urban planning, understanding human migration patterns, and disease transmission.

work today : Investigate the links between mobile location, frequency and social relations, analyze the role of geographic location and activity routes in human activities, and the impact of social relationships, such as meeting a friend; Identify the fundamental factors that govern human activity: how likely are humans to move to meet old friends? The likelihood that humans will move in order to meet new friends; the farther away from home, the more likely it is to increase or decrease.

Results: Empirical conclusions. Get data from two popular location-based social networks: Gowalla and Brightkite, and tracking a 2 million-person phone call from European countries. It is observed that people are basically active in a geographical area, and occasionally travel long distances. Long-distance travel is more likely when there are friends in a place, and short trips are rarely affected by this social relationship. In short, for the friendship of the mobile frequency is to meet new friends of the mobile frequency of twice times. At the same time, the login data and call data show strong consistency and robustness.

In general, the use of friendship to predict the individual mobile location has advantages and disadvantages. For example, a person may land at the place where his friend landed, and the likelihood decreases as time varies. 84% of people had fewer than 20% landings before their friends visited. All in all, collecting data can explain 10% of human movement, and landing data can explain 30%.

Results Overview: Model creation. based on empirical findings, we build a predictive and social mobility model to predict individual movement. First set the first destination and second destination, such as home and company. The model will contain three components: (1) a spatial location model where users often log on. (2) The mobile model under the influence of the temporary movement Model (3) between these positions. The daily move mode is a home and workplace conversion, based on which a weekly social mobility model is added.

The model can predict the probability of User location movement is 40%, the average distance of mobile phone data error rate of 0.23%. The error rate for logging data records is 2.7%. And the consistency and stability of the two kinds of data are observed.

Further related work: The human movement as a diffusion process "2", or a random process around a central point "12", our model of human activity as a random process around a few fixed points. This flexibility leads to more flexibility. There are also some studies focused on mobile detection "18,27" in wireless networks. In the same vein, there are GPS-based human position monitoring, which is limited by the specific road location "16". GPS and wireless positioning can track user locations for a long time, and this research has been limited to a small number of users and regions.

2. Role of landing position

We use different data sets to capture the human movement: the oct.2010 for Gowalla, Apr. to Oct. Gowalla the total number of landings is 6.4 million, brightkite 4.5 million. The former friendship relations constitute the non-direction diagram, the latter is the graph. For the sake of simplicity, we consider brightkite as a graph of the two-way side only. Gowalla A total of 196,591 nodes, 950,327 edges. The Brightkite has 58,228 nodes and 214,078 edges.

In order to ensure accurate data, the introduction of mobile phone tracking data. Provided by Europe, including nearly 200 million users, 450 million call logs. Track for 455 days on average. The nearest phone base station will record the location of each phone. This means we have nearly 900 million accuracy of 3 km of landing detection data. Only the city-wide landings are considered, and when each pair of individuals calls more than 5 times (total 10 times), a contact edge is established for them, including 200 million nodes, 450 million sides.

Brightkite data is shown in blue, Gowalla is red, and phone data is green.

User login behavior. Users may tend to be far away from home, and how much of a possibility they have to meet friends of social networks at the point of travel. This is our point of interest. The user home location is not explicitly given, and we assume the average "29" for 25 landing sites. This method has 85% accuracy for manual examination.

First, we measured how far the user tended to go from home, figure 1 marked the Brightkite, Gowalla and the cell phones data showing the distance from home, with exponential characteristics. When leaving home over 100km, the distribution is fast attenuated.

Figure 1 100 km range, B and G land landing possibilities and mobile phone contact possibilities vary with distance

Figure 2 (a) the distribution of residential distances between friends, (b) The distance distribution of all user dwellings, (c) distances between 200 large cities, (d) when one accounts for the non-uniform population density, as the probability of friendship as a distance function

3. Friendship and Mobility

Assuming B is a friend, B's position affects the movement of a, and we measure the likelihood of this movement to investigate the sociality of human activity.

: A in the residence of B for the center, RADIUS r within the range of the landing, the possibility of P. The distance from home is d. Change r value to Long group experiment.

Figure 3 (a) shows that the closer you are from home, the more likely you are to visit a friend and move. If we are 100km away from home, there is a 30% chance of visiting friends. Outside of 100km, the possibility of visiting friends remains the same. As distances become larger, a person may arrive at an increased position and the number of friends decreases. What we have observed is to remain the same, and we suspect that the greater the influence of a friend, the greater the distance. We compared this model to a null model that ignores social structures, which is drawn with dashed lines. 3 (b) describes the ratios.

Obviously, if the social structure is not affected, people will not be able to meet their friends. 10 times times the impact of a friend when traveling 1000km

The influence of friends in the movement of individuals

Because friends can be produced before and after a trip, there are two kinds of assumptions. The first one is that friends affect travel, and the latter is the movement that affects social relationships.

In order to differentiate between the two, we took the Gowalla social network distance of three months T1 and T2. The CA indicates that a landing place within a day of T1 was compared to a friend in his network to prove whether he was travelling because of a friend. CB said T1 before the landing, and three months after the relationship network, to prove whether the trip has created new friends. The range is limited to a radius of 25km. It turns out that there are 61% possible visits to existing friends, and 24% of potential new friends. The former is about 2.5 times times the latter. Cell phone data show that the former is higher than the latter 70%.

Move to a friend's landing place before

Travel more than 100km,10% may land in a similar location to friends. The farther away the distance is, the higher the probability.

Using friendships to predict movement limits

Data show,Gowalla have 9.6%, Brightkite have 4.1% probability of a friend log in first after the user login. This indicates that only a small number of users overlap with friends.

We create a trajectory vector for the I position:

Figure 4 (a) shows the consistency of a and friend trajectory vectors, when a pair of users simultaneously log in more than 40%, they friendship possibility 0.3. Figure 4 (b) Describes a large number of users who do not log in before landing sites. gowalla,80% is less than 20% and 52% is 0 times. This indicates that at least 50% of users have no information to prove that his social connections affect the movement.

Temporal and geographical projections of human activities

4. Human activity Model

The paper mainly presents two models: Periodic Mobility model (PMM) and Periodic & Social Mobility model (PSMM) Social motion prediction model.

will be in social friendships and human activities: social friendships and crowd mobility: User mobility in a location-based social network (ii) continue to translate.

Social friendship and crowd mobility: User mobility in a location-based social network (i)