Data collected from mobile phones has the potential to provide insight both into the relational dynamics of individuals (Eagle et al., Proceedings of the National Academy of Sciences, 106:15274-15278, 2009) and into human mobility patterns (Gonzalez et al., Nature, 453: 779-78, 2008). The high degree of temporal and spatial regularity showed in previous work prompts us to explore the potential of social networks as an infrastructure for transportation networks of small-portable physical goods. We propose an exchange platform where people lend each other objects of small value - such as ski boots or a star-shaped screwdrivers, which are transported by the owners or by carriers (friends or acquaintances of the owner in the social network). Participants in this exchange platform are not required to change their daily routines in order to deliver the items. By running simulations on mobile Call Detail Records - which include location information - from a large metropolitan area, we evaluate the performance of several transportation strategies such as direct transfer, nearest neighbor/only friends, nearest neighbor/everyone, and shortening the distance to the destination, in terms of the percentage of successfully transported objects, time to dispatch, number of hops and traveled distance. For the sake of the simulation we fixed that the delivery of an object was failed if it was still circulating in the network after a month since the injection in the simulation. Results show that completely unoptimized routing heuristics could successfully deliver an average of 3,908 objects --over 10,000 injected objects-- with an average delivery time of 0.59 days. These preliminary results suggest that, under considerably general assumptions, social networks may indeed be an effective and inexpensive infrastructure for transportation networks. This may have two important implications for sustainability. First, a service designed according to this principle might support people in reusing objects and tools that lie unused in one's premises (it was calculated that the average power tool is used 15 minutes in its entire life (Takara, 'In The Bubble', MIT Press, 2005)), therefore reducing the environmental pollution due to the production and transportation of new items. Second, leveraging people's routines to transport items might avoid additional environmental impact due to the transport of the shared items as if they were going to be delivered using a standard service.
More than Just Lines on a Map: Best Practices for U.S Bike Routes
Exploring Social Networks as an Infrastructure for Transportation Networks
1. Exploring Social Networks as
an Infrastructure for
Transportation Networks
Mauro Cherubini, Mengxiao Zhu (Northwestern),
Manuel Cebrian (MIT), and Nuria Oliver
5. use, not own
The average consumer power tool is used for
10 minutes in its entire life –but it takes hundreds of
times its own weight to manufacture such an object.
[Takara, 2005]
intro → SwapRelay → scientific relevance → methodology → results → conclusion
6. SwapRelay
is a platform where people lend each other objects or
services and earn “swap points” that they can reuse to
borrow other items.
intro → SwapRelay → scientific relevance → methodology → results → conclusion
7. some details
• items are insured by credit-card
• objects are transported by owners or by
carriers
• carriers earn swap points
• swap points might be converted in real
money for specific services
intro → SwapRelay → scientific relevance → methodology → results → conclusion
8. challenges
• we do not want to change people’s daily
routines
• we would like to reuse existing
infrastructure
direct transfer
indirect
intro → SwapRelay → scientific relevance → methodology → results → conclusion
9. scientific contribution
• Exploring laws governing human
motion particularly focusing on
transfer of physical items
• Defining a stochastic optimization strategy
to exploit regularities in human mobility
Home Work Home Work Home Work
10pm 8am 10p 8am 10pm 8am
10pm
intro → SwapRelay → scientific relevance → methodology → results → conclusion
10. system questions
• would the system be sustainable (can it be
compared with snail mail)?
• what is the best incentive mechanism that
can persuade people to inject a minimum
number of items?
intro → SwapRelay → scientific relevance → methodology → results → conclusion
11. research questions
• How many days for an item to travel from
point A to point B?
• What’s the average distance an item has to
travel?
• What is the average number of hops that
the item has to go through?
• Trusted nodes vs. untrusted nodes
intro → SwapRelay → scientific relevance → methodology → results → conclusion
12. analysis of CDR data
We analyzed anonymous CDR data of a large
metropolitan area in a country with an
emerging economy.
Using this information we reconstructed the
social network and the mobility
patterns of customers.
Then, we simulated the peer-to-peer
exchange of items as if the service
existed and measured transfer performance.
intro → SwapRelay → scientific relevance → methodology → results → conclusion
13. data cleansing
dataset: #users 300K, #links 350K,
observation period 6 months
• took only most active users in the dataset:
at least 5 phone calls in a week to ensure
enough datapoints
• considered only mobile customers for
which we could detect a social network
• removed business and service numbers
• removed short calls, dropped calls, spam
intro → SwapRelay → scientific relevance → methodology → results → conclusion
14. constructing Social
Networks
• Exchange of phone calls (Martinez et al.,
2009)
• Number of calls between two users
exceed a certain number, for example, 5
interactions (JP Onnela et al., 2007)
• Direct calling neighbor and Indirect
calling neighbor (Kianmehr & Alhajj, 2009)
intro → SwapRelay → scientific relevance → methodology → results → conclusion
15. constructing Social
Networks
• Exchange of phone calls (Martinez et al.,
2009)
• Number of calls between two users
exceed a certain number, for example, 5
interactions (JP Onnela et al., 2007)
• Direct calling neighbor and Indirect
calling neighbor (Kianmehr & Alhajj, 2009)
Call duration and/or call frequency
intro → SwapRelay → scientific relevance → methodology → results → conclusion
16. criterium for SN
dataset: #components 350
intro → SwapRelay → scientific relevance → methodology → results → conclusion
17.
18. inferring mobile patterns
Calculated
central point
Displacement
Radius of gyration Diameter of users
(Gonzalez et al., 2008) (Martinez et al., 2009)
intro → SwapRelay → scientific relevance → methodology → results → conclusion
19. criteria for location
• we divided the day into time slots of 1 hour and
we modeled each slot individually
• for the time slots for which we do not have data
we assume the person has not moved.
• probability of being in a location decreased by half
with time (to fill gaps)
• peers were in the same location if calling using the
same BTS
Antenna
B
A
C
intro → SwapRelay → scientific relevance → methodology → results → conclusion
20.
21. details of the simulation
• 10K transactions over the course of 1 year
• for each injected item, we picked at
random the “lender” and the “borrower”
from the same social network
• we used 3 different heuristics for the
delivery with 2 variations: with friends and
with everybody
intro → SwapRelay → scientific relevance → methodology → results → conclusion
22. constraints
• we considered an item not delivered if it
was circulating in the simulation after 1
month from the injection in the simulation
• peers could do the exchange if they were
within 1 Km away from each other
(parameter)
• maximum 10 hops (parameter)
intro → SwapRelay → scientific relevance → methodology → results → conclusion
23. h1: direct transfer (dt)
The item can only be transferred between
the ‘lender’ and the ‘borrower’
direct transfer
intro → SwapRelay → scientific relevance → methodology → results → conclusion
24. h2: nearest neighbor (nn)
The current carrier of the item first checks if the target is
in the range. If yes, the item will be handed over to the
target, otherwise, the item will be given to the nearest
neighbor of user depending on whether or not non-friend
transfer is allowed.
intro → SwapRelay → scientific relevance → methodology → results → conclusion
25. h3: hot potato (hp)
The transfer rules are similar to nn. The only difference is
that here a carrier list is maintained to make sure that the
item is not given back to any one of the previous carriers.
intro → SwapRelay → scientific relevance → methodology → results → conclusion
33. conclusion
social networks may indeed be an effective
and inexpensive infrastructure
for transportation networks
intro → SwapRelay → scientific relevance → methodology → results → conclusion
34. implications
1) a service designed according to this principle might
support people in reusing objects and tools that lie unused
in one's premises (it was calculated that the average power
tool is used 10 minutes in its entire life [Takara, 2005])
2) leveraging people's routines to transport items might
avoid additional environmental impact due to the transport
of the shared items as if they were going to be delivered
using a standard service
intro → SwapRelay → scientific relevance → methodology → results → conclusion
35. future work
• extend simulation to different datasets
coming from different regions of the world
• test heuristics with more strict
assumptions on how SN are formed and on
the spatial proximity
• develop an optimization algorithm using the
history of phone calls
intro → SwapRelay → scientific relevance → methodology → results → conclusion
36. take away message:
social networks may indeed be an effective
and inexpensive infrastructure
for transporting physical goods
Q&A
Mauro Cherubini
mauro@tid.es
http://research.tid.es
Editor's Notes
Let me start by acknowledging my co-authors, Mengxiao Zhu is a PhD candidate at Northwestern university, Manuel Cebrian was working at Telefonica when we started this project and now is is completing a Post-Doc at MIT. Finally, Nuria Oliver is my colleague at Telefonica Research.
So, the codename of this project is SwapRelay. Basically, we were interested in understanding whether networks of mobile customers could be used to transport physical items.
We were interested in this topic because the recent economical crisis has brought the attention again on our lifestyle. To introduce the project I would like to illustrate three interesting aspect that influenced us while developing this work.
The first one is a trend in scientific research. We see more and more scientist proposing technologies that could provoke positive changes in human behavior. We believe it is important to raise awareness on the fact that we do not live in isolated bubbles but that we inhabit a single ecosystem and that resources are limited.
The second idea that influenced this work is that of virtual economy. This project was carried out by Nokia research in rural Africa. Ethnographers have discovered that because people live very far away, they developed an alternative form of payment that consists in paying with airtime. So, they scratch the number of a top-up card for prepay phones and they send the number via SMS.
Here it is interesting to think that because the “real” economy is failing people can somewhat self-organize to create an alternative economy that perhaps is grounded on more tangible assets.
The final idea that influenced this project and that I would like to share here today is a recent statistics that was published about the way we produce and consume goods.
Our consumeristic approach to life encourage us to possess many objects that we need very little. So, if we could somewhat create a shift towards a culture of sharing and reusing these items we could reduce pollution.
So, the idea behind SwapRelay is very simple: encourage people to share little-used objects and to borrow items from each others instead of buying new items. The currency in this system are “swap points” a form of virtual currency that should keep the system working and prevent “free riders” to take only advantage of the system.
We have developed also a full concept of how the system should work. However I am not going to focus on SwapRelay in the rest of the presentation. Instead, I will focus on the scientific challenges that underlie the construction of a service such as SwapRelay.
Many people in the audience might be thinking that a service like this already exists on the Internet and so forth. However, the main issue with existing solutions is that of the COST users incur into when they have to transfer one item from the lender to the borrower. Often the cost of shipping the item through snail mail exceed the cost of the item itself.
So we take a radically different approach. [read the slide]
For the chosen Senders and Receivers Direct transfer or indirect transfers through friends or even strangers: eventually we would like to minimize the number of “Hops” (intermediate persons) and minimize the distance the item has to travel.
The high degree of temporal and spatial regularity showed in previous work (e.g., Gonzalez el al., 2008) encouraged us to explore the potential of social networks as an infrastructure for transportation networks of small-portable physical goods.
The future goal of this work will be that of defining a stochastic optimization strategy to optimize the routing of the obects into the system.
There might be many questions that we can ask if we were to build the commercial product. However, in this talk we will concentrate more on the scientific challenges of SwapRelay.
So, in the results that I will present I will touch upon these questions. For instance, ...
[read slide]
To answer our research questions we analyzed the Call Detail Records of a large metropolitan area of a developing country.
With this data we tried to do two things: first, to reconstruct the social network of the customers from the call graph, and second we tried to infer their urban mobility patterns.
Finally, on top of this data, we simulated the exchange of items AS IF SwapRelay was an existing service. Finally, we measured the performance of transferring items using the natural movements of customers.
The dataset contained 300K users ... [read slide]
As you can imagine, some noise had to be removed from the dataset. [read]
In the last years different approaches on how to reconstruct the social network from the call graph have been proposed. Here, I list only a few. But basically [CLICK] all of them refer to the duration of the phone calls and the call frequency.
As for us, we used a similar approach: we said that two customers are in the social network if they called each other and if the duration of their phone call was above 1 minute.
Criteria for friendship links between users.
Indirect calling neighbor: A, B call the same number C.
Given this criterium, we could reconstruct 350 components of the SN from the dataset that we have. The degree distribution seemed to follow a power law and the density of the networks tended towards 1, meaning that most of our networks seemed to be fully connected. This might be an artifact due to the fact that we had access to phone call data of only one part of the social network (we could not include customers of other operators).
The other challenge that we had to face was that of inferring the movements of people in the city space given their calling behavior and the location data we had access to. As you can imagine there were lots of “holes” in the dataset that we had to account for.
Other scholars have proposed in the past to use different metrics to infer mobility patterns. I am including these two examples here just for your reference.
Human movement show patterns of returning to the same location after an interval of 24h, 48h, 72h, etc. (Gonzalez et al, 2008)
Divide each 24 hours to several time slots and model each slot individually, for example, Fine-grain (30 mins timeslots) vs. coarse-grain (Farrahi & Gatica-Perez, 2008)
In our case we used a more simplistic approach to infer the mobility patterns. We basically divided the day into time slots of 1 hour (we did not considered the night hours because there was little activity in the dataset) and we modeled each time slot individually. We assumed that for a time slot for which we did not have data, the person did not move from the previous known location. However, we associated a probability function to this last position that decreased by half with time. We basically considered peers to be in the same location if they wer calling using the same base station.
Have you ever played SimCity? SimCity is a city-building simulation game, first released in 1989 and designed by Will Wright. Well, we did not do anything that sophisticated but the underlying idea was that of simulating the exchange of items using the data that we had prepared.
To give you some details of the simulation, we basically simulated 10K transaction over the course of 1 year. [read slide]
We also set a number of constraints in the simulations. For instance ... [read]
So, this is the first heuristic that we used to rely the items in the simulation. Basically, the lender was giving the item to the borrower only then they were “bumping” into each other.
The second heuristic was a bit more complex and it involved also indirect transfers. It worked as follows ... [read]
Finally, the third heuristic, the hot potato. The transfer rules are similar to nn. The only difference is that here a carrier list is maintained to make sure that the item is not given back to any one of the previous carriers. We were expecting that this method could help to avoid inefficient transfers among several users.
This strategy makes sure that the agents are always trying to minimize the distance between the item and the target at the next time step. This is also a strategy based on local information. We further suppose that the users only know their friends’ future locations at the next time step.
Now, I would like to give you a demo of the simulation. The simulation was implemented in NetLogo, a programming language developed by Uri Wilensky here at the Media Lab and then continued at Northwestern.
Ok. Let’s look at the results now. The first results I would like to show you has to do with the performance of the system. Here I show the number of item that were successfully delivered and those who failed delivery. The sum of the two columns is 10K items. As you can see [CLICK] the winner here is the hot potato heuristic involving only the peers in the SN. As you can see half of the items got lost in the process. However, it is important to notice that this is a completely self-organizing system that was controlled with a limited amount of information. Introducing an optimization algorithm and a dedicated software could change dramatically this figures.
The second result that I would like to present is the distance the items had to travel to reach their final destination. As you can see here the winner is Nearest Neighbor heuristic involving everybody in the simulation. The average distance travelled in this case was 15 Km (about 9 miles).
Finally, the last results that I will present involve the number of hops that the items had to go through and the number of days that they had to travel. [CLICK] The winner is here again the Nearest Neighbor heuristic involving everybody. As you can see in this case the items had to go through an average of 3.1 hops and it took an average of 0.48 days for them to be delivered. These results clearly demonstrate that if we were to build such system this could clearly outperform any standard delivery system such as snail mail, fedex, etc.
So, to conclude, we have presented encouraging results that demonstrate how social networks might be used effectively to transfer physical good. Of course, our results have to be taken carefully because they have been obtained under general assumptions that will have to be verified. The assumptions that we imposed deal essentially with the way we calculate social networks and spatial proximity between users. Also, I would like to stress again that these results have been obtained from heuristic which are completely unoptimized. Finally, the topology of the GSM network of the city from where we got the data is very sparse so the results I presented might very well be a lower-bound in relation to what it could be really obtained from a fully optimized system running in a geographical region with more antennas.
This initial results have two important implications for sustainability. First, a service designed according to this principle might support people in reusing objects and tools that lie unused in one's premises. This therefore can reduce environmental pollution due to the reduced production of items.
Second, leveraging people's routines to transport items might avoid additional environmental impact due to the transport of the shared items as if they were going to be delivered using a standard service.
We are certainly continuing this work with the objective of building a working prototype. But before we get there we would like to ... [read]
Thanks for your attention. And if you have questions that we cannot answer straight away, please do send me a message.