Simulation For The Sensing Scheduling Algorithm

We evaluated performance of the proposed sensing scheduling algorithm in large cases via simulation. In the simulation, the duration of sensing scheduling period was set to 3 hours, which is divided by 1080 time instants. The arrival (leaving) times of mobile users were randomly generated, following a uniform distribution between 0 (the corresponding arrival time) and 10800s. We used a bell-shaped Gaussian distribution (with µ = 0 and σ = 10s) to model coverage, as discussed in section 3.3. A simple scheduling algorithm served as the baseline: a smartphone starts to sense every 10s since its arrival for NB

k times, where NkB is the corresponding budget. The average

coverage probability was used as performance metric, which is the sum of coverage probabilities (objective function) divided by the total number of time instants in the scheduling period (i.e., 1080). In the first simulation scenario, we changed the number of mobile users from 10 to 50 with a step size of 5 and the budgets of all mobile users were fixed to 17. In the second scenario, we changed the budget from 15 to 25 with a step size of 1 and the numbers of mobile users were fixed to 40. We presented the results in Fig. 3-12. Note that every number in the figure is an average over 10 runs. From the figure, we can see that on average, our scheduling algorithm outperforms the baseline algorithm by 65% in terms of average coverage probability. From Fig. 3- 12(a), we can see that when 55 users participate in sensing, our algorithm leads to almost 100% coverage. In order to achieve an average coverage probability of 80%, our scheduling algorithm need no more than 40 users (with a budget of 17) while the baseline algorithm can only reach an average coverage probability of 50% with 40 users. Similar observations can be made from Fig. 3-12(b). No matter which method is used, the average coverage probability always increases with the number of mobile users and budget as expected. In addition, we observe that the variance of the coverage probability given by our scheduling algorithm is always less than that given by the baseline algorithm, which means our algorithm is more stable and is suitable

10 15 20 25 30 35 40 45 50 55 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 # of mobile users

Average coverage probability

Ours Baseline

(a) Varying # of mobile users

15 16 17 18 19 20 21 22 23 24 25 0.4 0.5 0.6 0.7 0.8 0.9 1 Budget

Average coverage probability

Ours Baseline

(b) Varying budget

Figure 3-12: Performance of the sensing scheduling algorithm for various situations.

Chapter 4

Application II: LI festyle Learning

Via P hone S ensing (LIPS )

4.1

Overview

As analyzed in the previous chapters, the sensors of a smartphone can easily detect the context (such as location, local weather, activities, etc.) of its mobile user. How- ever, how to leverage this unique capability for learning lifestyles of mobile users and making their life better has not yet been fully exploited. One of the early efforts in this regard is that some urban guide mobile applications (such as Yelp [77] and Urbanspoon [70]) have used locations (collected by smartphones) to recommend local businesses (such as restaurants, bars, local applications, etc.) to mobile users. We believe that with smartphones and their sensors, we can do much better and do much more than this.

Some related works are focusing on analyzing location data collected by GPS. In the project TraClass [37], Lee etc. studied how to predict moving objects’ types by analyzing their trajectories and other features. They proposed to generate a hierarchy of features by partitioning trajectories and exploring different types of clustering.

In [22], the authors proposed a sequential pattern mining paradigm that can be used to analyze the trajectories of moving objects. In [76], the authors proposed a data mining framework to retrieve association patterns from raw individual GPS data. Note that all these works leveraged only location and time information for analysis without using other features (such as moving states, weather, etc. ) that can be collected via smartphone sensing.

In this chapter, we present the design, implementation and evaluation of a sensing application, LIPS, which can learn LIfestyles of mobile users via smartPhone Sensing (LIPS). According to businessdictionary.com, “Lifestyle is expressed in both work

and leisure behavior patterns and (on an individual basis) in activities, attitudes, interests, opinions, values, and allocation of income.” By leverage multiple sensors

on a smartphone, we can obtain a comprehensive view of a mobile user’s context (such as location, local weather, activities, etc.) over a long period. Based on those context information, we aim to find out what a mobile user likes to do (characterization) and what he/she will do next (prediction) based on the collected sensor data. Such a lifestyle learning application can be used to support a large variety of applications for improving life quality. For example, a major application is to recommend local businesses to mobile users based on not only his/her location but also his/her lifestyle. This work represents one of the first efforts along this line, which is focused on lifestyle learning, while leaving lifestyle-aware recommendation or lifestyle-based applications for future research.

LIPS is built on top of the platform. Both mobile frontend and cloud backend need to be extended with new application modules. The LIPS frontend application module on the mobile frontend reports the context information collected by sensors of the smartphone to the learning application module on the backend server periodically. Based on this information, the learning application module builds models for lifestyles of mobile users.

Combining both unsupervised and supervised learning, we propose a hybrid scheme for lifestyle learning, which consists of two parts: characterization and prediction. Specifically, we present a two-stage algorithm to characterize the lifestyle of a mo- bile user using Places of Interest (PoIs), which leverages two different algorithms for coarse-grained and fine-grained clustering in two stages respectively. Based on dis- covered PoIs, we present a supervised learning based algorithm to build a model for predicting the future activities of a mobile user.

In addition, operating smartphone sensors (such as GPS) could be energy consum- ing. Note that even though some sensors (such as accelerometer) are always active, a thread needs to be spawned to collect its readings, which will prevent the smartphone enter sleep mode and affect the energy consumption too. A phone’s main job is not sensing after all. Extensive smartphone sensing may drain its battery quickly, leav- ing it dead when it is needed to perform its regular duties such as phone calls, web surfing, etc. To enable green lifestyle learning, we present an adaptive sampling algo- rithm, which adaptively control the sampling rate according to discovered PoIs and the lifestyle model. To the best of our knowledge, we are the first to build a smart- phone sensing based system to learn and analyze lifestyles of mobile users based on various context information (collected from smartphones).

We build a novel smartphone sensing based application for lifestyle learning, and propose practical and effective solutions to fundamental problems (learning and energy-efficient sampling). Specifically, we summarize our contributions in the fol- lowing:

• We present design and implementation of LIPS application, which learns lifestyles

of mobile users via smartphone sensing.

• We present an effective hybrid scheme for lifestyle learning, which combines

• We present an energy-efficient sampling algorithm, which leverages the discov-

ered PoIs and the lifestyle model for adaptively controlling the sample rate.

• We performed extensive field tests to validate and evaluate LIPS. The exper-

imental results well justify the effectiveness and efficiency of LIPS on lifestyle learning.

The rest of the chapter is organized as follows: We present the software archi- tecture and implementation details of the proposed LIPS application in Section 4.2. The proposed learning scheme and adaptive sampling algorithm are presented in Sec- tion 4.3 and Section 4.4 respectively. Experimental results are presented and analyzed in Section 4.5.