Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Intelligent Data Processing for the Internet of Things

2,741 views

Published on

Guest lecture, International “IoT 360″ Summer School
October 29th – November 1st, 2014 – Rome, Italy

Published in: Education
  • Be the first to comment

Intelligent Data Processing for the Internet of Things

  1. 1. 1 Intelligent Data Processing for the Internet of Things Payam Barnaghi Institute for Communication Systems (ICS) University of Surrey Guildford, United Kingdom International “IoT 360″ Summer School October 29th – November 1st, 2014 – Rome, Italy
  2. 2. 2 Key characteristics of IoT devices −Often inexpensive sensors (actuators) equipped with a radio transceiver for various applications, typically low data rate ~ 10-250 kbps. −Deployed in large numbers −The sensors should coordinate to perform the desired task. −The acquired information (periodic or event-based) is reported back to the information processing centre (or some cases in-network processing is required) −Solutions are often application-dependent. 2
  3. 3. 3 Beyond conventional sensors − Human as a sensor (citizen sensors) − e.g. tweeting real world data and/or events − Software sensors − e.g. Software agents/services generating/representing data Road block, A3 Road block, A3 Suggest a different route
  4. 4. Internet of Things: The story so far P. Barnaghi, A. Sheth, “The Internet of Things: The story so far”, IEEE IoT Newsletter, September 2014.
  5. 5. 5 The benefits of data processing in IoT − Turn 12 terabytes of Tweets created each day into sentiment analysis related to different events/occurrences or relate them to products and services. − Convert (billions of) smart meter readings to better predict and balance power consumption. − Analyze thousands of traffic, pollution, weather, congestion, public transport and event sensory data to provide better traffic and smart city management. − Monitor patients, elderly care and much more… − Requires: real-time, reliable, efficient (for low power and resource limited nodes), and scalable solutions. Partially adapted from: What is Bog Data?, IBM
  6. 6. Not just Volume… … but also Data Dynamicity: How can we efficiently deal with: - Large amounts of (heterogeneous/distributed) data? - Both static and dynamic data? - In a re-usable, modular, flexible way? - Integrate different types of data - Provide hypothesis and create more context-aware solutions Adapted from: M. Hauswirth. A. Mileo, Insight, National University of Ireland, Galway.
  7. 7. Data Volume AnyThing AnyPlace AnyTime Security, Reliability, Trust and Privacy Societal Impacts, Economic Values and Viability Services and Applications Networking and Communication
  8. 8. Data is not what we want or is it?
  9. 9. What we need are insights and actionable-information
  10. 10. Diffusion of innovation image source: Wikipedia IoT
  11. 11. Problem #1 Data: We seem to have lots of it… Real World Data: it is always difficult to get (silos, format, privacy, business interests or lack of interest!...)
  12. 12. Problem #2 Data: interoperability and metadata frameworks… Real World Data: there are solutions for service based (RESTful) access, meta-data/ semantic representation frameworks (e.g. W3C SSN, HyperCat,…) but none of them are still widely adapted.
  13. 13. Problem #3 Data: quality, reliability… Real World Data: data can be noisy, crowed source data can be inaccurate, contradictory, delay in accessing/processing the data…
  14. 14. Problem #4 Data: having too much data and using analytics tools alone won’t solve the problem… Real World Data: in addition to the HPC issues, we need new methods/solutions that can provide real-time analysis of dynamic, variable quality and multi-modal streams…
  15. 15. Problem #5 Data: abstraction, discovering the associations… Real World Data: co-occurrence vs. causation; we need hypothesis, background knowledge,… After all data is not what we are really after…
  16. 16. We need more linked open data
  17. 17. Sometimes it’s even better if we have: (near) real-time linked open data Streams
  18. 18. or even better than that if we have: (near) real-time linked open data Streams + meta-data (semantic annotations) + Adaptable and scalable analytics tools + Sufficient background knowledge
  19. 19. IoT Data 19 ?
  20. 20. Current focus on Big Data − Emphasis on power of data and data mining solutions − Technology solutions to handle large volumes of data; e.g. Hadoop, NoSQL, Graph Databases, … − Trying to find patterns and trends from large volumes of data…
  21. 21. Myths About Big Data − Big Data is only about massive data volume − Big Data means Hadoop − Big Data means unstructured data − If we have enough data we can draw conclusions (enough here often means massive amounts) − NoSQL means No SQL − It is about increasing computational power and taking more data and running data mining algorithms. 21 Some of the items are adapted from: Brain Gentile, http://mashable.com/2012/06/19/big-data-myths/
  22. 22. What happens if we only focus on data − Number of burgers consumed per day. − Number of cats outside. − Number of people checking their facebook account. − What insight would you draw? 22
  23. 23. It is also important to note what type of problems we expect to solve.
  24. 24. Back to the future 24
  25. 25. Future cities: a view from 1998 Source LAT Times, http://documents.latimes.com/la-2013/ 25
  26. 26. Source: http://robertluisrabello.com/denial/traffic-in-la/#gallery[default]/0/ 26 Source: wikipedia
  27. 27. 27
  28. 28. 101 Smart City Use-case Scenarios http://www.ict-citypulse.eu/page/content/smart-city-use-cases-and-requirements
  29. 29. 29 Data alone is not enough − Domain knowledge − Machine interpretable meta-data − Delivery, sharing and representation services − Query, discovery, aggregation services − Publish, subscribe, notification, and access interfaces/services − More open solutions for innovation and citizen participation − Efficient feedback and control mechanisms − Social network and social system analysis − In cities, interactions with people and social systems is the key.
  30. 30. 30 We need an Integrated Approach
  31. 31. 31 Processing steps
  32. 32. 32 Storing, handling and processing the data Image courtesy: IEEE Spectrum
  33. 33. 33 Technical Challenges − Discovery: finding appropriate device and data sources − Access: Availability and (open) access to data resources and data − Search: querying for data − Integration: dealing with heterogeneous devices, networks and data − Large-scale data mining, adaptable learning and efficient computing and processing − Interpretation: translating data to knowledge that can be used by people and applications − Scalability: dealing with large numbers of devices and a myriad of data and the computational complexity of interpreting the data.
  34. 34. 34 IoT Data Access − Publish/Subscribe (long-term/short-term) − Ad-hoc query − The typical types of data request for sensory data: − Query based on − ID (resource/service) – for known resources − Location − Type − Time – requests for freshness data or historical data; − One of the above + a range [+ Unit of Measurement] − Type/Location/Time + A combination of Quality of Information attributes − An entity of interest (a feature of an entity on interest) − Complex Data Types (e.g. pollution data could be a combination of different types)
  35. 35. 35 IoT Data in the Cloud Image courtesy: http://images.mathrubhumi.com http://www.anacostiaws.org/userfiles/image/Blog-Photos/river2.jpg
  36. 36. Comparing IoT data streams with conventional multimedia streams Source: P. Barnaghi, W. Wang, L. Dong, C. Wang, "A Linked-data Model for Semantic Sensor Streams", in the Proc. of IEEE International Conference on Internet of Things (iThings 2013), August 2013.
  37. 37. 37 Describing IoT Data: An example Time Location Type Value Link to QoI metadata UTC #GeoHash #Hash [DataType, Value] URI Ontology for common types
  38. 38. 38 Observation and Measurement Value GeoHash UTC time Standard XSD data type UTC time (in Java) : The time indicated is returned represented as the distance, measured in milliseconds, of that time from the epoch (00:00:00 GMT on January 1, 1970).
  39. 39. 39 GeoHashing − For example Guildford: lat: 51.235401 and long: -0.574600 can be hashed as: gcpe6zjeffgp − It can be used as: − A unique identifier − represent point data as hash string − It uses Base 32 encoding and bit interleaving − It’s used for geo-tagging (and is a symmetric technique) − Place close to each other will have similar prefix (string similarity) − Limitations: − We could have Geohash codes with no common prefix − Edge case (locations close to each other but on opposite sides of the Equator) − A meridian point (line of longitude)
  40. 40. 40 GeoHash Example Sample locations on a Google Map and their equivalent geohash strings; - close locations have similar prefixes
  41. 41. IoT Data Processing WSN WSN WSN WSN WSN Network-enabled Devices Network-enabled Devices Network services/storage and processing units Data/service access at application level Data collections and processing within the networks Data Discovery Service/ Resource Discovery
  42. 42. 42 In-network processing − Mobile Ad-hoc Networks can be seen as a set of nodes that deliver bits from one end to the other; − WSNs, on the other end, are expected to provide information, not necessarily original bits − Gives additional options − e.g., manipulate or process the data in the network − Main example: aggregation − Applying aggregation functions to a obtain an average value of measurement data − Typical functions: minimum, maximum, average, sum, … − Not amenable functions: median Source: Protocols and Architectures for Wireless Sensor Networks, Protocols and Architectures for Wireless Sensor Networks Holger Karl, Andreas Willig, chapter 3, Wiley, 2005 .
  43. 43. 43 In-network processing − Depending on application, more sophisticated processing of data can take place within the network − Example edge detection: locally exchange raw data with neighboring nodes, compute edges, only communicate edge description to far away data sinks − Example tracking/angle detection of signal source: Conceive of sensor nodes as a distributed microphone array, use it to compute the angle of a single source, only communicate this angle, not all the raw data − Exploit temporal and spatial correlation − Observed signals might vary only slowly in time; so no need to transmit all data at full rate all the time − Signals of neighboring nodes are often quite similar; only try to transmit differences (details a bit complicated, see later) Source: Protocols and Architectures for Wireless Sensor Networks, Protocols and Architectures for Wireless Sensor Networks Holger Karl, Andreas Willig, chapter 3, Wiley, 2005 .
  44. 44. Data Aggregation − Computing a smaller representation of a number of data items (or messages) that is extracted from all the individual data items. − For example computing min/max or mean of sensor data. − More advance aggregation solutions could use approximation techniques to transform high-dimensionality data to lower-dimensionality abstractions/representations. − The aggregated data can be smaller in size, represent patterns/abstractions; so in multi-hop networks, nodes can receive data form other node and aggregate them before forwarding them to a sink or gateway. − Or the aggregation can happen on a sink/gateway node.
  45. 45. Aggregation example − Reduce number of transmitted bits/packets by applying an aggregation function in the network 1 1 3 1 1 6 1 1 1 1 1 1 Source: Holger Karl, Andreas Willig, Protocols and Architectures for Wireless Sensor Networks, Protocols and Architectures for Wireless Sensor Networks, chapter 3, Wiley, 2005 .
  46. 46. Efficacy of an aggregation mechanism − Accuracy: difference between the resulting value or representation and the original data − Some solutions can be lossless or lossly depending on the applied techniques. − Completeness: the percentage of all the data items that are included in the computation of the aggregated data. − Latency: delay time to compute and report the aggregated data − Computation foot-print; complexity; − Overhead: the main advantage of the aggregation is reducing the size of the data representation; − Aggregation functions can trade-off between accuracy, latency and overhead; − Aggregation should happen close to the source.
  47. 47. Publish/Subscribe − Achieved by publish/subscribe paradigm − Idea: Entities can publish data under certain names − Entities can subscribe to updates of such named data − Conceptually: Implemented by a software bus − Software bus stores subscriptions, published data; names used as filters; subscribers notified when values of named data changes Publisher 1 Publisher 2 Software bus Subscriber 1 Subscriber 2 Subscriber 3 Source: Holger Karl, Andreas Willig, Protocols and Architectures for Wireless Sensor Networks, Protocols and Architectures for Wireless Sensor Networks, chapter 12, Wiley, 2005 .
  48. 48. MQTT Pub/Sub Protocol − MQ Telemetry Transport (MQTT) is a lightweight broker-based publish/subscribe messaging protocol. − MQTT is designed to be open, simple, lightweight and easy to implement. − These characteristics make MQTT ideal for use in constrained environments, for example in IoT. −Where the network is expensive, has low bandwidth or is unreliable −When run on an embedded device with limited processor or memory resources; − A small transport overhead (the fixed-length header is just 2 bytes), and protocol exchanges minimised to reduce network traffic − MQTT was developed by Andy Stanford-Clark of IBM, and Arlen Nipper of Cirrus Link Solutions. Source: MQTT V3.1 Protocol Specification, IBM, http://public.dhe.ibm.com/software/dw/webservices/ws-mqtt/mqtt-v3r1.html
  49. 49. MQTT − It supports publish/subscribe message pattern to provide one-to-many message distribution and decoupling of applications − A messaging transport that is agnostic to the content of the payload − The use of TCP/IP to provide basic network connectivity − Three qualities of service for message delivery: − "At most once", where messages are delivered according to the best efforts of the underlying TCP/IP network. Message loss or duplication can occur. − This level could be used, for example, with ambient sensor data where it does not matter if an individual reading is lost as the next one will be published soon after. − "At least once", where messages are assured to arrive but duplicates may occur. − "Exactly once", where message are assured to arrive exactly once. This level could be used, for example, with billing systems where duplicate or lost messages could lead to incorrect charges being applied. Source: MQTT V3.1 Protocol Specification, IBM, http://public.dhe.ibm.com/software/dw/webservices/ws-mqtt/mqtt-v3r1.html
  50. 50. MQTT Message Format − The message header for each MQTT command message contains a fixed header. − Some messages also require a variable header and a payload. − The format for each part of the message header: — DUP: Duplicate delivery — QoS: Quality of Service — RETAIN: RETAIN flag —This flag is only used on PUBLISH messages. When a client sends a PUBLISH to a server, if the Retain flag is set (1), the server should hold on to the message after it has been delivered to the current subscribers. —This allows new subscribers to instantly receive data with the retained, or Last Known Good, value. Source: MQTT V3.1 Protocol Specification, IBM, http://public.dhe.ibm.com/software/dw/webservices/ws-mqtt/mqtt-v3r1.html
  51. 51. Sensor Data as time-series data − The sensor data (or IoT data in general) can be seen as time-series data. − A sensor stream refers to a source that provide sensor data over time. − The data can be sampled/collected at a rate (can be also variable) and is sent as a series of values. − Over time, there will be a large number of data items collected. − Using time-series processing techniques can help to reduce the size of the data that is communicated; − Let’s remember, communication can consume more energy than communication;
  52. 52. Sensor Data as time-series data − Different representation method that introduced for time-series data can be applied. − The goal is to reduce the dimensionality (and size) of the data, to find patterns, detect anomalies, to query similar data; − Dimensionality reduction techniques transform a data series with n items to a representation with w items where w < n. − This functions are often lossy in comparison with solutions like normal compression that preserve all the data. − One of these techniques is called Symbolic Aggregation Approximation (SAX). − SAX was originally proposed for symbolic representation of time-series data; it can be also used for symbolic representation of time-series sensor measurements. − The computational foot-print of SAX is low; so it can be also used as a an in-network processing technique.
  53. 53. 53 In-network processing Using Symbolic Aggregate Approximation (SAX) SAX Pattern (blue) with word length of 20 and a vocabulary of 10 symbols over the original sensor time-series data (green) Source: P. Barnaghi, F. Ganz, C. Henson, A. Sheth, "Computing Perception from Sensor Data", in Proc. of the IEEE Sensors 2012, Oct. 2012. fggfffhfffffgjhghfff jfhiggfffhfffffgjhgi fggfffhfffffgjhghfff
  54. 54. Symbolic Aggregate Approximation (SAX) − SAX transforms time-series data into symbolic string representations. − Symbolic Aggregate approXimation was proposed by Jessica Lin et al at the University of California –Riverside; − http://www.cs.ucr.edu/~eamonn/SAX.htm . − It extends Piecewise Aggregate Approximation (PAA) symbolic representation approach. − SAX algorithm is interesting for in-network processing in WSN because of its simplicity and low computational complexity. − SAX provides reasonable sensitivity and selectivity in representing the data. − The use of a symbolic representation makes it possible to use several other algorithms and techniques to process/utilise SAX representations such as hashing, pattern matching, suffix trees etc.
  55. 55. Processing Steps in SAX − SAX transforms a time-series X of length n into the string of arbitrary length, where typically, using an alphabet A of size a > 2. − The SAX algorithm has two main steps: − Transforming the original time-series into a PAA representation − Converting the PAA intermediate representation into a string during. − The string representations can be used for pattern matching, distance measurements, outlier detection, etc.
  56. 56. Piecewise Aggregate Approximation − In PAA, to reduce the time series from n dimensions to w dimensions, the data is divided into w equal sized “frames.” − The mean value of the data falling within a frame is calculated and a vector of these values becomes the data-reduced representation. − Before applying PAA, each time series to have a needs to be normalised to achieve a mean of zero and a standard deviation of one. − The reason is to avoid comparing time series with different offsets and amplitudes; Source: Jessica Lin, Eamonn Keogh, Stefano Lonardi, and Bill Chiu. 2003. A symbolic representation of time series, with implications for streaming algorithms. In Proceedings of the 8th ACM SIGMOD workshop on Research issues in data mining and knowledge discovery (DMKD '03). ACM, New York, NY, USA, 2-11.
  57. 57. SAX- normalisation before PAA Timeseries (c): 2, 3, 4.5, 7.6, 4, 2, 2, 2, 3, 1 Mean (μ): μ= (2+3+4.5+7.6+4+2+2+2+3+1)/10= 3.11 Standard deviation (σ): (2-3.11)2 = 1.2321 (3-3.11)2 = 0.0121 (4.5-3.11)2 = 1.9321 (7.6-3.11)2 = 20.1601 (4-3.11)2 = 0.7921 (2-3.11)2 = 1.2321 (2-3.11)2 = 1.2321 (2-3.11)2 = 1.2321 (3-3.11)2 = 0.0121 (1-3.11)2 = 4.4521 1.2321+0.0121+ 1.9321+ 20.1601+ 0.7921+ 1.2321+ 1.2321+ 1.2321+ 1.2321+ 0.0121+4.4521 = 33.5211 σ = √ (33.5211/10) = 1.83087683911
  58. 58. Normalisation Timeseries (c): 2, 3, 4.5, 7.6, 4, 2, 2, 2, 3, 1 Normalised: zi = (ci – μ)/ σ σ = 1.83087683911 μ = 3.11 z1 = (2- 3.11)/1.83087683911 = -0.606 z2 = (3-3.11)/ 1.83087683911= -0.600 z3 = (4.5-3.11)/ 1.83087683911= 2.452 z4 = (7.6-3.11)/ 1.83087683911= -0.600 z5 = (4-3.11)/ 1.83087683911= 0.486 z6 = (2-3.11)/ 1.83087683911= -0.606 z7 = (2-3.11)/ 1.83087683911= -0.606 z8 = (2-3.11)/ 1.83087683911= -0.606 z9 = (3-3.11)/ 1.83087683911= -0.600 z10 = (1-3.11)/ 1.83087683911= -1.152 Normalised Timeseries (z): -0.606, -0.600, 2.452, -0.600, 0.486, -0.606, -0.606, -0.606 , -0.600, -1.152
  59. 59. PAA calculation Timeseries (c): 2, 3, 4.5, 7.6, 4, 2, 2, 2, 3, 1 Normalised Timeseries (z): -0.606, -0.600, 2.452, -0.600, 0.486, -0.606, -0.606, -0.606 , -0.600, -1.152 PAA (w=5): -0.603, 0.926, -0.06, -0.606, 0.273
  60. 60. PAA to SAX Conversion − Conversion of the PAA representation of a time-series into SAX is based on producing symbols that correspond to the time-series features with equal probability. − The SAX developers have shown that time-series which are normalised (zero mean and standard deviation of 1) follow a Normal distribution (Gaussian distribution). − The SAX method introduces breakpoints that divides the PAA representation to equal sections and assigns an alphabet for each section. − For defining breakpoints, Normal inverse cumulative distribution function
  61. 61. Breakpoints in SAX − “Breakpoints: breakpoints are a sorted list of numbers B = β 1,…, β a-1 such that the area under a N(0,1) Gaussian curve from βi to βi+1 = 1/a”. Source: Jessica Lin, Eamonn Keogh, Stefano Lonardi, and Bill Chiu. 2003. A symbolic representation of time series, with implications for streaming algorithms. In Proceedings of the 8th ACM SIGMOD workshop on Research issues in data mining and knowledge discovery (DMKD '03). ACM, New York, NY, USA, 2-11.
  62. 62. Alphabet representation in SAX − Let’s assume that we will have 4 symbols alphabet: a,b,c,d − As shown in the table in the previous slide, the cut lines for this alphabet (also shown as the thin red lines on the plot below) will be { -0.67, 0, 0.67 } Source: JMOTIF Time series mining, http://code.google.com/p/jmotif/wiki/SAX
  63. 63. SAX Represetantion Timeseries (c): 2, 3, 4.5, 7.6, 4, 2, 2, 2, 3, 1 Normalised Timeseries (z): -0.606, -0.600, 2.452, -0.600, 0.486, -0.606, -0.606, -0.606 , -0.600, -1.152 PAA (w=5): -0.603, 0.926, -0.06, -0.606, 0.273 Cut off ranges: {-0.67, 0, 0.67} Alphabet: a ,b ,c, d SAX representation: bdbbc
  64. 64. Features of the SAX technique − SAX divides a time series data into equal segments and then creates a string representation for each segment. − The SAX patterns create the lower-level abstractions that are used to create the higher-level interpretation of the underlying data. − The string representation of the SAX mechanism enables to compare the patterns using a specific type of string similarity function.
  65. 65. High-level information/ knowledge 65 A sample data processing framework Temporal data (extracted from descriptions) Day-time Night-time High-level abstractions Domain knowledge fggfffhfffffgjhghfff dddfffffffffffddd cccddddccccdddccc dddcdcdcdcddasddd aaaacccaaaaaaaacccc Raw sensor data stream Raw sensor data stream Raw sensor data stream PIR Sensor Light Sensor Temperature Sensor Attendance Phone Hot Temperature Cold Temperature Bright Office room BA0121 On going meeting Window has been left open …. Spatial data (extracted from descriptions) Thematic data (low level abstractions) Intelligent Processing Observations SAX Patterns Raw sensor data (or Annotated data) … …. Intelligent Processing/ Reasoning
  66. 66. Correlation analysis Slides: Maria Bermudez-Edo
  67. 67. Correlations: linear and nonlinear 67
  68. 68. Co-occurrence analysis 68
  69. 69. Correlation analysis 69
  70. 70. Challenges − Complexity − Co-occurrences vs. Causation 70
  71. 71. Data Discovery Slides: Syed Amir Hoseinitabatabaei
  72. 72. A discovery method in the IoT time location type Query formulating [#location || ##ttyyppee || ttiimmee]] Discovery ID Discovery/ DHT Server Data repository (archived data) #location #type #location #type Data hypercube #location #type Gateway Core network Logical Connection Network Connection Data
  73. 73. An example: a discovery method in the IoT S. A. Hoseinitabatabaei, P. Barnaghi, C. Wang, R. Tafazolli, L. Dong, "A Distributed Data Discovery Mechanism for the Internet of 73 Things", 2014.
  74. 74. An example: a discovery method in the IoT S. A. Hoseinitabatabaei, P. Barnaghi, C. Wang, R. Tafazolli, L. Dong, "A Distributed Data Discovery Mechanism for the Internet of 74 Things", 2014.
  75. 75. Adaptive learning Slides: Daniel Puschmann, Masound Hassanpour
  76. 76. Adaptable and dynamic learning methods http://kat.ee.surrey.ac.uk/
  77. 77. Equilibrium in transient and non-uniform world A D B C Image source for equilibrium diagram: John D. Hey, The University of York.
  78. 78. Social data analysis: A case study Slides: Pramod Anantharam, Kno.e.sis Centre, Wright State University
  79. 79. Social data analysis: A case study − Are people talking about city infrastructure on twitter? − Can we extract city infrastructure related events from twitter? − How can we leverage event and location knowledge bases for event extraction? − How well can we extract city events? Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University.
  80. 80. Do people talk about city infrastructure on Twitter? Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University. 80
  81. 81. Some challenges in extracting events from Tweets − No well accepted definition of “events related to a city”. − Tweets are short (140 characters) and its informal nature make it hard to analyze − Entity, location, time, and type of the event − Multiple reports of the same event and sparse report of some events (biased sample) − Numbers don’t necessarily indicate intensity − Validation of the solution is hard due to the open domain nature of the problem Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University. 81
  82. 82. Event extraction techniques − N-grams + Regression − Text analysis to extract uni- and bi-grams (event markers) − Feature selection to select best possible event markers − Apply regression to predict P(Y|X) where Y is the target (rainfall) and X is the input (event marker). − Clustering − Create event clusters incrementally over time − Identify clusters of interest based on its relevance (manual inspection) − Granularity remains at the tweet/cluster level (tweets are assigned to clusters of interest) − Sequence Labeling (CRFs) − Text analysis to extract features such as named entities, POS1 tagging − Each event indicator is modeled as a mixture of event types that are latent variables − Each type corresponds to a distribution over named entities n (labels assigned to event types by manual inspection) and other features Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University.
  83. 83. City event extraction − Event extraction should be open domain (no a priori event types) with event metadata. − Incorporate background knowledge related to city related events e.g., 511.org hierarchy, SCRIBE ontology, city location names. − Assess the intensity of an event using content and network cues. − Robust to noise, informal nature, and variability of data. Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University.
  84. 84. N-grams + Regression − Open domain: works best when there is a reference corpus to extract n-grams − Event metadata: cannot distinguish between entities and hence hard to extract event metadata − Background knowledge: incorporating domain vocabulary (e.g., subsumption) is not natural − Event intensity: regression maps the event indicators to some quantified values − Robustness: quite robust if there is a reference corpus Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University.
  85. 85. Clustering − Open domain: works well for domains with no a priori knowledge of events (may need human inspection) − Background knowledge: incorporating domain vocabulary is not natural − Event intensity: not captured − Robustness: quite robust for twitter data with enough data for each cluster Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University.
  86. 86. Sequence Labeling (CRFs) − Open domain: works well for domains with no a priori knowledge of events (may need human inspection) − Event metadata: event metadata extraction is captured naturally with the named entities − Background knowledge: incorporating domain vocabulary is quite natural − Event intensity: part-of-speech tag may indirectly capture intensity − Robustness: with a deeper model for capturing context, quite robust for twitter data Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University.
  87. 87. City event extraction solution architecture POS Tagging Tweets from a city POS City Infrastructure Tagging Hybrid NER+ Event term extraction Hybrid NER+ Event term extraction Impact Assessment Impact Assessment Temporal Estimation Temporal Estimation Event Event OOSSMM L Looccaatitoionnss SSCCRRIBIBEE o onntotolologgyy Aggregation Aggregation GGeeoohhaasshhiningg 551111.o.orrgg h hieierraarrcchhyy City Event Annotation City Event Extraction Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University.
  88. 88. City event extraction - Key insights a) Space: events reported within a grid (gi ∈G where G is a set of all grids in a city)at a certain time are most likely reporting the same event b) Time: events reported within a time Δt in a grid gi are most likely to be reporting the same event c) Theme: events with similar entities within a grid gi and time Δt are most likely reporting the same event Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University.
  89. 89. 0.6 miles Max-lat Min-lat Min-long 37.7545166015625, -122.40966796875 Max-long 0.38 miles 37.7490234375, -122.40966796875 37.7545166015625, -122.420654296875 37.7490234375, -122.420654296875 4 37.74933, -122.4106711 Hierarchical spatial structure of geohash for representing locations with variable precision. Here the location string is 5H34 0 1 2 3 4 5 6 7 8 9 B C D E F G H I J K L 0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 8 City event extraction – Geohashing
  90. 90. Implmentation − City event annotation − Automated creation of training data − Annotation task (our CRF model vs. baseline CRF model) − City event extraction − Use aggregation algorithm for event extraction − Extracted events AND ground truth − Dataset (Aug – Nov 2013) ~ 8 GB of data on disk − Over 8 million tweets − Over 162 million sensor data points − 311 active events and 170 scheduled events Source: Pramod Anantharam, Kno.e.sis Centre, Wright State University.
  91. 91. Evaluation – extracted events & ground truth
  92. 92. 92 Conclusions − A primary goal of interconnecting devices and collecting/processing data from them is to create situation awareness and enable applications, machines, and human users to better understand their surrounding environments. − The understanding of a situation, or context, potentially enables services and applications to make intelligent decisions and to respond to the dynamics of their environments. − Dynamicity, energy efficiency, multi-modality, heterogeneity and volume are among the key challenges. − We need to design adaptable and scalable solutions that combine knowledge and data engineering (semantics), background knowledge (ontologies) and machine learning techniques.
  93. 93. Acknowledgements − Some parts of the content are adapted from: − Holger Karl, Andreas Willig, Protocols and Architectures for Wireless Sensor Networks, Protocols and Architectures for Wireless Sensor Networks, chapters 3 and 12, Wiley, 2005 . − Jessica Lin, Eamonn Keogh, Stefano Lonardi, and Bill Chiu. 2003. A symbolic representation of time series, with implications for streaming algorithms. In Proceedings of the 8th ACM SIGMOD workshop on Research issues in data mining and knowledge discovery (DMKD '03). ACM, New York, NY, USA, 2-11. − JMOTIF Time series mining, http://code.google.com/p/jmotif/wiki/SAX
  94. 94. Q&A − Payam Barnaghi, University of Surrey/EU FP7 CityPulse Project http://www.ict-citypulse.eu/ @pbarnaghi p.barnaghi@surrey.ac.uk

×