Recommendation systems in large-scale online marketplaces are essential to aiding users in discovering new content. However, state-of-the-art systems for item-to-item recommendation tasks are often based on a shallow level of contextual relevance, which can make the system insufficient for tasks where item relationships are more nuanced. Contextually relevant item pairs can sometimes have problematic relationships that are confusing or even controversial to end users, and they could degrade user experiences and brand perception when recommended to users. For example, the recommendation of a book about one sports team to someone reading a book about that team’s biggest rival could be a bad experience, despite the presumed similarities of the books. In this paper, we propose a classifier to identify and prevent such problematic item-to-item recommendations and to enhance overall user experiences. The proposed approach utilizes active learning to sample hard examples effectively across sensitive item categories and employs human raters for data labeling. We also perform offline experiments to demonstrate the efficacy of this system for identifying and filtering problematic recommendations while maintaining recommendation quality.

Massive amount of spatio-temporal data is generated as a result of user movement and data access activities, from both smart mobile devices and network infrastructure. We aim to be able to derive mobility intelligence from spatiotemporal data at scale with efficiency and yet meet the needs of diverse applications. In this paper, we propose Mobility GenomeTM, a computational framework that enables efficient and extensible discovery of mobility intelligence from large-scale spatio-temporal data. The framework is organised as multiple layers composed of fundamental and extensible computational units. We describe several algorithms and models, such as Human Daily Activity to demonstrate the derivation of mobility intelligence from these units, together with validation results. The framework has been integrated into our Mobility Intelligence platform, processing hundreds of millions of records per day in our production environment. We also show the possibility of building more advanced applications on top of the framework such as detection of home and work location, origin-destination trips and footfall analysis. With its deployment, the framework helps us avoid duplicated efforts on repeating common data processing and algorithmic tasks while serving a diverse set of mobility intelligence applications.

Accurate estimations of footfall count have several important uses, examples include monitoring crowd movement, allocation of emergency services, retail planning, transport planning, and so on. To estimate footfall count, data logs from telco's Location Based Services (LBS) system can be used. LBS records indicate the cellular towers that mobile phone users are connected to at a given date and time. However, such data only indicates the geospatial coordinates of the cellular tower that a user is connected to and does not accurately reflect the actual geolocation of the user. In this work, we first build a dataset comprising of the observations (cellular tower locations) and the actual ground truth (corresponding GPS locations). Next, we propose several schemes to improve the accuracy of footfall count estimation given the cellular tower coordinates that mobile phone users are connected to. This includes estimating footfall count based on different grids, as well as redistributing observed locations to an area surrounding the cellular tower based on a random distribution. Through our experiments, we found that the most effective scheme is to firstly compute the posterior probability that a user is in an area, conditioned on her being connected to a particular cellular tower. After the posterior probability distributions are obtained, we redistribute these cellular tower points based on this distribution table. Our experimental results show that redistribution of the observed cellular tower points based on posterior probabilities is able to improve the footfall count estimation accuracy by up to 83.7%.

Extraction of mobility patterns of people in a city helps in urban planning, traffic control, transportation management, etc. We present a demo that shows how to extract interesting individual and collective mobility patterns from mobile phone cellular data. Individual patterns are captured using a next place prediction algorithm that builds different Dynamic Bayesian Network (DBN) models and chooses the best DBN model. Collective patterns are captured by aggregating individual patterns using efficient query processing methods. Each query requires three arguments, i.e., current location cl, current time ct and query time qt. These parameters are varied to capture information at different granularity level resulting in many interesting patterns. Cellular data, although less accurate, irregular and sparser than GPS and less detailed than CDR, is cheap and abundantly available and is used by all sections of people. Demo results show that these shortcomings can be overcome, at least to some extent, by exploiting the underlying repetitive patterns. As part of individual patterns, the system predicts the next place (along with a confidence measure) for a given combination of current location, Time-of-Day and Day-of-week of a user. Collective mobility patterns can start from a planning area or a latlon (with a radius). It visually shows the major destinations along with their intensity. An interesting finding was that planning areas with higher population has higher predictability. The UI system also shows heat maps that show continuous mobility of people. These types of patterns can have significant bearing on urban planning and related applications.

Understanding how people use public transport is important for the operation and future planning of the underlying transport networks. We have therefore developed and deployed a traffic measurement system for a key player in the transportation industry to gain insights into crowd behavior for planning purposes. The system has been in operation for several months and reports, at hourly intervals, (1) the crowdedness of subway stations, (2) the flows of people inside interchange stations, and (3) the expected travel time for each possible route in the subway network of Singapore. The core of our system is an efficient algorithm which detects individual subway trips from anonymized real-time data generated by the location based system of Singtel, the country's largest telecommunications company. To assess the accuracy of our system, we engaged an independent market research company to conduct a field study--a manual count of the number of passengers boarding and disembarking at a selected station on three separate days. A strong correlation between the calculations of our algorithm and the manual counts was found. One of our key findings is that travelers do not always choose the route with the shortest travel time in the subway network of Singapore. We have therefore also been developing a mobile app which allows users to plan their trips based on the average travel time between stations.

In wireless sensor networks powered by ambient energy harvesting (WSNs-HEAP), sensor nodes' energy harvesting rates are spatially heterogeneous and temporally variant, which impose difficulties for medium access control (MAC). In this paper, we first derive the necessary conditions under which channel utilization and fairness are optimal in a WSN-HEAP, respectively. Based on the analysis, we propose an earliest deadline first (EDF) polling MAC protocol, which regulates transmission sequence of the sensor nodes based on the spatially heterogeneous energy harvesting rates. It also mitigates temporal variations in energy harvesting rates by a prediction and update mechanism. Simulation results verify the performance tradeoff predicted by our analysis for the proposed HEAP-EDF protocol. In the presence of spatial heterogeneity and temporal variations in energy harvesting rates, our proposed protocol exhibits significant performance advantages compared to the existing MAC protocols for WSNs-HEAP in the literature.

Fingerprint-based methods are widely adopted for indoor localization purpose because of their cost-effectiveness compared to other infrastructure-based positioning systems. However, the popular location fingerprint, Received Signal Strength (RSS), is observed to differ significantly across different devices' hardware even under the same wireless conditions. We derive analytically a robust location fingerprint definition, the Signal Strength Difference (SSD), and verify its performance experimentally using a number of different mobile devices with heterogeneous hardware. Our experiments have also considered both Wi-Fi and Bluetooth devices, as well as both Access-Point(AP)-based localization and Mobile-Node (MN)-assisted localization. We present the results of two well-known localization algorithms (K Nearest Neighbor and Bayesian Inference) when our proposed fingerprint is used, and demonstrate its robustness when the testing device differs from the training device. We also compare these SSD-based localization algorithms' performance against that of two other approaches in the literature that are designed to mitigate the effects of mobile node hardware variations, and show that SSD-based algorithms have better accuracy.

Existing approaches to indoor tracking have various limitations. Location-fingerprinting approaches are labor intensive and vulnerable to environmental changes. Trilateration approaches require at least three line-of-sight beacons for coverage at any point in the service area, which results in heavy infrastructure cost. Dead reckoning (DR) approaches rely on knowledge of the initial location and suffer from tracking error accumulation. Despite this, we adopt DR for location tracking because of the recent emergence of affordable hand-held devices equipped with low-cost DR-enabling sensors. In this paper, we propose an indoor pedestrian tracking system that comprises of a DR subsystem implemented on a mobile phone and a ranging subsystem with a sparse infrastructure. A particle-filter-based fusion scheme is applied to bound the accumulated tracking error by fusing DR with sparse range measurements. Experimental results show that the proposed system is able to track users much better than DR alone. The system is robust even when: 1) the initial user location is not available; 2) range updates are noisy; and 3) range updates are intermittent, both temporally and spatially.

Due to the widespread deployment and low cost, WLAN has drawn much attention for indoor localization. In this poster, an efficient indoor localization algorithm, which utilizes the WLAN received signal strength from each Access Point (AP), has been proposed. The algorithm is based on the Extreme Learning Machine (ELM), a Single layer Feed-forward neural Network (SLFN). It is competitive fast in offline learning and online localization. Also, compared with existing fingerprinting approach, it does not need the fingerprinting database in the online phase, which can substantially reduce the required storage space of the terminal devices.

The emergence of personal mobile device with low cost sensors, such as accelerometer and digital compass, has made dead-reckoning (DR) an attractive choice for indoor pedestrian tracking. In this paper, we propose a robust DR pedestrian tracking system on top of such commercially accessible sensor sets capable of DR. The proposed method exploits the fact that, multiple DR systems, carried by the same pedestrian, have stable relative displacements with respect to the center of motion, and therefore to each other. We first formulate the robust tracking task as a generalized maximum a posteriori sensor fusion problem, and then we narrow it to a simple computation procedure with certain assumptions. A prototype is implemented and evaluated with a benchmark system that collects ground truth efficiently and accurately. In a practical indoor testbed, the proposed scheme has exhibited robust tracking performance, with reduction in average tracking error up to 73.7%, compared to traditional DR tracking methods.

Accurate indoor pedestrian tracking has wide applications in the healthcare, retail, and entertainment industries. However, existing approaches to indoor tracking have various limitations. For example, location-fingerprinting approaches are labor-intensive and vulnerable to environmental changes. Trilateration approaches require at least three Line-of-Sight (LoS) beacons to cover any point in the service area, which results in heavy infrastructure cost. Dead Reckoning (DR) approaches rely on knowledge of the initial user location and suffer from tracking error accumulation. Despite this, we adopt DR for location tracking because of the recent emergence of affordable hand-held devices equipped with low cost DR-enabling sensors. In this paper, we propose an indoor pedestrian tracking system which comprises a DR sub-system implemented on a mobile phone, and a ranging sub-system with a sparse infrastructure. A probabilistic fusion scheme is applied to bound the accumulated tracking error of DR when new range measurements are available from sparsely deployed beacons. Experimental results show that the proposed system is able to track users much better than DR alone, with reductions in average error by up to 71.9%. The system is robust and works well even when the initial user location is not available and range updates are intermittent. This highlights the potential of using sparse but reasonably accurate partial information to limit location tracking errors.

In this paper, we propose a fingerprint-based localization scheme that exploits the location dependency of the channel impulse response (CIR). We approximate the CIR by applying inverse Fourier transform to the receiver's channel estimation. The amplitudes of the approximated CIR (ACIR) vector are further transformed into the logarithmic scale to ensure that elements in the ACIR vector contribute fairly to the location estimation, which is accomplished through nonparametric kernel regression. As shown in our simulations, when both the number of access points and density of training locations are the same, our proposed scheme displays significant advantages in localization accuracy, compared to other fingerprint-based methods found in the literature. Moreover, absolute localization accuracy of the proposed scheme is shown to be resilient to the real time environmental changes caused by human bodies with random positions and orientations.

Indoor localization techniques using location fingerprints are gaining popularity because of their cost-effectiveness compared to other infrastructure-based location systems. However, their reported accuracy fall short of their counterparts. In this paper, we investigate many aspects of fingerprint-based location systems in order to enhance their accuracy. First, we derive analytically a robust location fingerprint definition, and then verify it experimentally as well. We also devise a way to facilitate under-trained location systems through simple linear regression technique. This technique reduces the training time and effort, and can be particularly useful when the surrounding or setup of the localization area changes. We further show experimentally that because of the positions of some access points or the environmental factors around them, their signal strength correlates nicely with distance. We argue that it would be more beneficial to give special consideration to these access points for location computation, owing to their ability to distinguish locations distinctly in signal space. The probability of encountering such access points will be even higher when we denote a location's signature using the signals of multiple wireless technologies collectively. We present the results of two wellknown localization algorithms (K-Nearest Neighbor and Bayesian Probabilistic Model) when the above factors are exploited, using Bluetooth and Wi-Fi signals. We have observed significant improvement in their accuracy when our ideas are implemented.