Julio De Melo Borges

EstaVis: A Real-World Interactive Platform for Crowdsourced Visual Urban Analytics

Fueled by the increasing proliferation of citizen generated spatio-temporal data -- especially in participatory urban infrastructure monitoring -- municipal authorities are in need for ways to process and understand increasingly overwhelming amounts of data. However, duplicate issue reporting by citizens such as broken traffic lights, potholes or garbage can lead to bottlenecks in manual processing of such data. As contribution this paper examines which city issue report presentation methods are useful to support a human in analyzing and processing them. We compare presentation methods such as automatically clustered information, manual clustered information and mixes of both. Automatically clustering of information is performed by a data analytics algorithm which is also presented in this paper together with EstaVis, a prototype of an interactive visual urban analytics platform. Evaluation studies with 282 crowd-workers show how the platform can potentially help to speed-up report processing by detecting and aggregating duplicate reports by up to 3 orders of magnitude and discuss which lessons can be learned in terms of features and user experience pitfalls for this kind of system.

Interestingness Classification of Association Rules for Master Data

High quality of master data is crucial for almost every company and it has become increasingly difficult for domain experts to validate the quality and extract useful information out of master data sets. However, experts are rare and expensive for companies and cannot be aware of all dependencies in the master data sets. In this paper, we introduce a complete process which applies association rule mining in the area of master data to extract such association dependencies for quality assessment. It includes the application of the association rule mining algorithm to master data and the classification of interesting rules (from the perspective of domain experts) in order to reduce the result association rules set to be analyzed by domain experts. The model can learn the knowledge of the domain expert and reuse it to classify the rules. As a result, only a few interesting rules are identified from the perspective of domain experts which are then used for database quality assessment and anomaly detection.

Urban Anomaly Detection: a Use-Case for Participatory Infra-Structure Monitoring

Internet-enabled, location aware smart phones with sensor inputs have led to novel urban infra-structure monitoring applications exploiting unprecedented high levels of citizen participation in dense metropolitan areas. For policy makers, it is a key task to keep track of trends and developments of reported infra-structure issues for understanding and effectively reacting to problems around a city, specially in their early stages. In contrast to previous strategies which consider only limited information such as text and geographic locations, we analyze the urban dynamics of crowdsourced collected data using an existing approach that considers a novel modeling of heterogeneous attributes and relationships in the data. First, the underlying data is modeled into a heterogeneous network, in which its measured for each node its current level of anomalousness for a desired time interval (e.g. a week) and then the most anomalous networks subgraph is extracted and described by means of problem category, geographical area, time and participants. First experiments illustrate the effectiveness and efficiency of leveraging this anomaly detection approach in our use-case (participatory infra-structure monitoring).

Feature engineering for crime hotspot detection [in press]

San-Francisco in the US and Natal in Brazil are two coastal cities which are known rather for its tech scene and natural beauty than for its criminal activities. We analyze characteristics of the urban environment in these two cities, deploying a machine learning model to detect categories and hotspots of criminal activities. We propose an extensive set of spatio-temporal & urban features which can significantly improve the accuracy of machine learning models for these tasks, one of which achieved Top 1% performance on a Crime Classification Competition by kaggle.com. Extensive evaluation on several years of crime records from both cities show how some features — such as the street network — carry important information about criminal activities.

Event detection for smarter cities [in press]

Detecting what is going on around in near real time through analysis of social network data has become part of assessing the pulse of a city. The advances in event detection techniques enable cities to give a real-time overview of the events — ranging from music concerts and exhibitions to emergencies like fires and car accidents — and activities happening in a smart city. While previous work mostly focused on large-scale, e.g. global or national level, event detection recent development focuses on hyper-local events, that are occurring in a small region, e.g. a street corner or a certain venue rather than city-or country-level area. This paper offers a broad survey and classification of event detection techniques, identifying the key features of recent techniques and their usage in the context of smart cities. This is done by introducing them as well as comparing and categorizing different up to date techniques regarding their event definition, their mode of operation, and their qualitative and quantitative evaluation approaches. Research gaps are highlighted and future work in the area is identified.

Predicting Target Events in Industrial Domains

In industrial environments, machine faults have a high impact on productivity due to the high costs it can cause. Machine generated event logs are a abundant source of information for understanding the causes and events that led to a critical event in the machine. In this work, we present a Sequence-Mining based technique to automatically extract sequential patterns of events from machine log data for understanding and predicting machine critical events. By experiments using real data with millions of log entries from over 150 industrial computer numerical control (CNC) cutting machines, we show how our technique can be successfully used for understanding the root causes of certain critical events, as well as for building monitors for predicting them long before they happen, outperforming existing techniques.

A predictive policing application to support patrol planning in smart cities

Recently, patrol planning and other predictive policing strategies were improved for Smart Cities applications. In such public safety context, some police departments have been recording crime events in their databases to compose better strategies to understand and predict crime incidence. This work presents the ROTA-Analytics, a web-based application which aims to provide crime incidence forecasting as outputs. This crime incident forecasting helps patrol supervisors to elaborate the list of predefined locations (points) and staying time at which each police vehicle must patrol. ROTA-Analytics supports multiple machine and statistical learning methods selection to create an environment of crime prediction in different areas of the city. All the phases from time series creation to the automatic machine selection are discussed and exemplified with real data from Natal City in Brazil. Finally, we evaluated our architecture by using two regression strategies for different spatial granularity levels.

Sequential pattern mining — A study to understand daily activity patterns for load forecasting enhancement

Load forecasting at appliance-level or house-level is a key to develop efficient Demand Side Management programs. Lots of recent research work have pointed out that load curves at households level depend highly on human behaviors and activities. However, the state-of-the-art load modeling approach takes only individual human activities with appliance-level time-of-use data into account. There are few studies about influence of sequences of activities performed throughout a day on power consumption at households level. In this work, we conduct a broad study of activity sequences in daily life that influence power consumption of individual households. A context-rich data set including daily activity information and power consumption measurements from 23 households is collected across Japan. The contribution of this paper is twofold: 1) a set of insights into household-specific activity sequences influencing power consumption derived from a sequence mining algorithm, in order to identify significant associations between power consumption and household-specific activity sequences; 2) a load forecasting study using identified frequent activity sequences as an enhancement. Our analysis on sequence-based rules shows potential for inferring future activities and the power consumption of the corresponding activity. Finally, we demonstrate how very short-term load forecasting, like 15 minutes ahead, can benefit from activity sequences for individual households.