Nokia Mobile Data Challenge Research Articles

Personalized location recommendation using mobile phone usage information

Location recommendation has become a hot research area in recent years. The cold-start problem is still a great challenge in personalized location recommendation, which makes it difficult to infer a new user’s preferences, because a new user generally has never visited any location at the start. To address this problem, the existing studies usually exploit other information, e.g., demographic features, to characterize users. However, such little information is not sufficient to profile users accurately. In addition, abundant mobile phone usage information can be recorded when users are using their phones, e.g., the use frequency of Apps, which can fully reveal the diverse characteristics of different users. In this paper, we propose a personalized location recommendation method using mobile phone usage information, which transforms the location recommendation problem into a regression task, and extracts six types of mobile phone usage features to profile users. Demographic features and location features are also extracted. To efficiently reduce model parameters, factorization machines are employed to construct the recommendation model, which models feature interactions as the inner products of latent vectors with matrix factorization. We evaluate the proposed method using the open dataset of Nokia Mobile Data Challenge, and experimental results show the effectiveness of the proposed method in personalized location recommendation.

Mobile crowd location prediction with hybrid features using ensemble learning

With the explosive growth of location-based service on mobile devices, predicting users’ future locations and trajectories is of increasing importance to support proactive information services. In this paper, we model this problem as a supervised learning task and propose to use ensemble learning methods with hybrid features to solve it. We characterize the properties of users’ visited locations and movement patterns and then extract feature types (temporal, spatial, and system) to quantify the correlation between locations and features. Finally, we apply ensemble methods to predict users’ future locations with extracted features. Moreover, we design an adaptive Markov Chain model to predict users’ trajectories between two locations. To evaluate the system performance, we use a real-life dataset from the Nokia Mobile Data Challenge. Experiment results unveil interesting findings: (1) For individual predictors, Bayes Networks outperform all others when data quality is good, while J48 delivers the best results when data quality is bad; (2) Ensemble predictors outperform individual predictors in general under all conditions; and (3) Ensemble predictor performance depends on the user movement patterns.

Generic context adaptation for mobile cloud computing environments

Markets for mobile applications offer myriads of apps ranging from simple to quite demanding ones. The latter are on the rise since every new generation of smartphones is equipped with more resources (CPU, memory, bandwidth, energy) to even allow re-source-demanding services like speech- or face recognition to be executed locally on a device. But compared to their stationary counterparts, mobile devices remain comparatively limited in terms of resources. Because of this, current approaches aim at extending mobile device capabilities with computation and storage resources offered by cloud services or other nearby devices. This paradigm, known as mobile cloud computing (MCC), is challenged by the dynamically changing context of mobile devices, which developers are required to take into account to decide, e.g., which application parts are when to offload. To rise to such and similar challenges we introduce the concept of Generic Context Adaptation (GCA), a data mining process that facilitates the adaptation of (mobile) applications to their current and future context. Moreover, we evaluate our approach with real usage data provided by the Nokia Mobile Data Challenge (MDC) as well as with CloudAware, a context-adaptive mobile middleware for MCC that supports automated and context-aware self-adaptation techniques.

Predicting Users’ Mobility Using Monte Carlo Simulations

Made possible by the availability of spatio-temporal data collected by smart phones and other smart devices, understanding people’s mobility patterns has become one of the most promising location-based services in the past few years, providing various businesswise application possibilities. The simplest version of possibilities is to predict where a user will go next. In this paper, we present a novel approach that goes beyond predicting users’ next location and is able to predict their entire mobility patterns. Building on previous work, our models are based on statistical Markov state-space models. In our approach, however, we add temporal information ( “arrival profiles” and “probability of stay” profiles) explicitly to the models. Using Monte Carlo simulations on these models enables us to predict multiple future locations, including residence times. We evaluate the models on real-world data sets (1.5 year personally collected mobile phone raw GPS logs and a publicly available Nokia Mobile Data Challenge data set) using different evaluation metrics. An extensive evaluation shows that our proposed methods have better predictive power (higher recall) than standard state-space models.

Generic Context Adaptation for Mobile Cloud Computing Environments

Markets for mobile applications offer myriads of apps ranging from simple to quite demanding ones. The latter are on the rise since every new generation of smartphones is equipped with more resources (CPU, memory, bandwidth, energy) to even allow resource-demanding services like speech- or face recognition to be executed locally on a device. But compared to their stationary counterparts, mobile devices remain comparatively limited in terms of resources. Because of this, current approaches aim at extending mobile device capabilities with computation and storage resources offered by cloud services or other nearby devices. This paradigm, known as Mobile Cloud Computing (MCC), is challenged by the dynamically changing context of mobile devices, which developers are required to take into account to decide, e.g., which application parts are when to offload. To rise to such and similar challenges we introduce the concept of Generic Context Adaptation (GCA), a data mining process that facilitates the adaptation of (mobile) applications to their current and future context. Moreover, we evaluate our approach with real usage data provided by the Nokia Mobile Data Challenge (MDC) as well as with CloudAware, a context-adaptive mobile middleware for MCC that supports automated context-aware self-adaptation techniques.

Sequential behavior prediction based on hybrid similarity and cross-user activity transfer

The proliferation of smart phones has opened up new kinds of data to model human behavior and predict future activity but this prediction can be tempered by the relative sparsity of data. In this paper, we integrate a time-dependent instance transfer mechanism, driven by a hybrid similarity measure, into learning and predicting human behavior. In particular, transfer component analysis (TCA) is utilized for domain adaptation from different data types to overcome data sparsity. The hybrid user similarity measure is developed based on three different characteristics: eigen-behavior, longest common behavior (LCB), and daily common behavior (DCB). Extensive comparisons are made against state-of-the-art time series prediction algorithms using the Nokia Mobile Data Challenge (MDC) dataset and the MIT Reality Mining dataset. We compare the prediction performance given (i) no additional data, (ii) only data from identical behavior from other users, and (iii) data from any type of behavior from other users. Experimental results show that our proposed algorithm significantly improves the performance of behavior prediction.

When Enough is Enough: Location Tracking, Mosaic Theory, and Machine Learning

When Enough is Enough: Location Tracking, Mosaic Theory, and Machine Learning

Generating storylines from sensor data

We present an approach for producing narratives, or storylines, from sensor data collected from a mobile phone. Given a training set of English-language descriptions of events and a set of corresponding sensor data, we learn a probabilistic translation model. Then, given new sensor data, our model can produce English-language descriptions of the corresponding events. Our approach is evaluated on the data provided as part of the Nokia Mobile Data Challenge (MDC), focusing, in particular, on location labelling. We also present a set of tools for visualizing the MDC data, which were used to generate training data for storyline creation. Finally, we present a quantitative analysis of wifi data in a busy office setting, and present a novel model for wifi signals that more accurately matches the properties of signals in a natural environment.

Feature engineering for semantic place prediction

We present in this paper our winning solution to Dedicated Task 1 in Nokia Mobile Data Challenge (MDC). MDC Task 1 is to infer the semantic category of a place based on the smartphone sensing data obtained at that place. We approach this task in a standard supervised learning setting: we extract discriminative features from the sensor data and use state-of-the-art classifiers (SVM, Logistic Regression and Decision Tree Family) to build classification models. We have found that feature engineering, or in other words, constructing features using human heuristics, is very effective for this task. In particular, we have proposed a novel feature engineering technique, Conditional Feature (CF), a general framework for domain-specific feature construction. In total, we have generated 2,796,200 features and in our final five submissions we use feature selection to select 100 to 2000 features. One of our key findings is that features conditioned on fine-granularity time intervals, e.g. every 30 min, are most effective. Our best 10-fold CV accuracy on training set is 75.1% by Gradient Boosted Trees, and the second best accuracy is 74.6% by L1-regularized Logistic Regression. Besides the good performance, we also report briefly our experience of using F# language for large-scale (∼70 GB raw text data) conditional feature construction.

Interdependence and predictability of human mobility and social interactions

Previous studies have shown that human movement is predictable to a certain extent at different geographic scales. The existing prediction techniques exploit only the past history of the person taken into consideration as input of the predictors.In this paper, we show that by means of multivariate nonlinear time series prediction techniques it is possible to increase the forecasting accuracy by considering movements of friends, people, or more in general entities, with correlated mobility patterns (i.e., characterised by high mutual information) as inputs. Finally, we evaluate the proposed techniques on the Nokia Mobile Data Challenge and Cabspotting datasets.

Where to go from here? Mobility prediction from instantaneous information

We present the work that allowed us to win the Next-Place Prediction task of the Nokia Mobile Data Challenge. Using data collected from the smartphones of 80 users, we explore the characteristics of their mobility traces. We then develop three families of predictors, including tailored models and generic algorithms, to predict, based on instantaneous information only, the next place a user will visit. These predictors are enhanced with aging techniques that allow them to adapt quickly to the users’ changes of habit. Finally, we devise various strategies to blend predictors together and take advantage of their diversity, leading to relative improvements of up to 4%.

A visual exploration of mobile phone users, land cover, time, and space

A series of maps were produced that together form a type of atlas of the Nokia Mobile Data Challenge (MDC). Like in a traditional geographic atlas, a limited number of base map configurations is generated, onto which various thematic elements are then overlaid. Two of those base maps are themselves derived from MDC data; the third is referenced in geographic space. Thematic overlays serve several purposes, including elaborating different elements from which the base map geometry had been derived, as well as linking other data to it. The core of the study presented here is an intersection of high-dimensional concepts, dimensionality reduction, geographic analysis, and visualization, intended as a point of departure towards an integrated, attribute-centered understanding of people’s movement patterns. Among the advances put forth is a new time-weighted kernel density model approach derived from journey vertices captured via GPS and WLAN.

Analysis of Proximity Networks from Multiple Mobile Sensor Data

In this paper, we report our data-driven investigation on measuring proximity networks derived from multiple mobile sensor data: Bluetooth, GPS, and WLAN coordinates. By exploiting the dataset from NOKIA Mobile Data Challenge 2012, we propose our methods to generate sensor data based proximity networks between mobile users and evaluate these networks with a call log network as ground truth. Our results can be useful to understand that Bluetooth, GPS, and WLAN AP are of significant sensor data to quantify the pattern and strength of proximate relation between mobile users.