myriam phd
TRANSCRIPT
Contextualise Sensors with Linked Data
To Improve Relevancy, Data Quality and Network Adaptability
Myriam LeggieriPhD Thesis
Sensor out of Context
2
Contextualise Sensors with Linked Data
3
dB, Km, µPa? dBs in water have a different relative value than in air
Q1. How to model it?
—> Is it worth it?
4
4
dB, dB, Km, µPa?
Yes, because
Q1. Contextualised Model for
Q3. Relevancy Prediction
Q4. Enriched Web Content
Q5. Network Adaptability
Q2. Cross-Network Communication
Research Questions
5
Q1. How to model Linked Sensor Data
for
Q2. Cross-Network Communication
Q3. Relevancy Prediction
Q4. Web Data Quality
Q5. Network Adaptability
Outline
1. Linked Sensor Data Model [Q1]
2. LD4Sensors Web Service [Q2]
3. Sensor Relevancy Prediction [Q3]
4. Enriched Web Content [Q4]
5. Network Adaptability [Q5]
6. Research Answers
7. Lessons Learned and Future Work
6
Core Research and
Results
Conclusion
Q1. How can contextual information be used to enrich sensor data?
Linked Sensor Data Model
7
Ontology Modularisation
ContextNetwork
Components
Energy Conservation
Linked Sensor Data Model
8
Application Ontology
Domain Ontology Task Ontology
Upper Ontology
Ontology Aligning Inheritance & Reuse
Dolce+DnS Ultralite(DUL)
W3C Semantic Sensor Network (SSN)
Our Ontology
Quantities, Units, Dimensions and Data Types (QUDT)
Linked Sensor Data Model
9
Dolce+DnS Ultralite(DUL)
W3C Semantic Sensor Network (SSN)
Provenance (PROV)
Event Model-F (EVENT)
Unified Code for Units of Measure (UCUM)
Friend Of A Friend (FOAF)
Measurement Unit (MUO)
Online Presence (OPO)
Review Vocabulary (REV)
Quantities, Units, Dimensions and Data Types (QUDT)
Ontology Aligning Inheritance & Reuse
10
Linked Sensor Data Model
spt:Agent
spt:Activity
ssn:Device
ssn:Sensor
EventParticipation
ssn:Stimulus
spt:Place
11
Linked Sensor Data Model
OWL FullSymmetric
Transitive Inverse
Equivalent
room A floor 2 house H
spt:containedIn
Asserted, Inferred, Direct Relations
12
Comment &
Rate
from
to
title
date-time
link motivation
same thing
same domain
same date-time
same location
Linked Sensor Data Model
Social Feedback and
Sharing
Outline
1. Linked Sensor Data Model [Q1]
2. LD4Sensors Web Service [Q2]
3. Sensor Relevancy Prediction [Q3]
4. Enriched Web Content [Q4]
5. Network Adaptability [Q5]
6. Research Answers
7. Lessons Learned and Future Work
13
Core Research and
Results
Conclusion
Q1. How can contextual information be used to enrich sensor data?
Q2. How can sensors communicate across different platforms without ad-
hoc solutions?
14
Dolce+DnS Ultralite(DUL)
W3C Semantic Sensor Network (SSN)
Provenance (PROV)
Event Model-F (EVENT)
Unified Code for Units of Measure (UCUM)
Friend Of A Friend (FOAF)
Measurement Unit (MUO)
Online Presence (OPO)
Review Vocabulary (REV)
Quantities, Units, Dimensions and Data Types (QUDT)
Which ontologies?
Which links? How to enable inference?
How to enable cross-communication?
Non-experts Average users
Non-experts Average users
Automate Inference
Automate Reasoning
Automate Link Search &
Creation
Easy Browsing
REST API
GUI
LD4Sensors (LD4S) Web Service
Easy Storing
16
LD4Sensors (LD4S) Web Service
17
LD4S: Evaluation
Goals
1. Actual Gain in Building Automation
2. Uptake, Usability,Utility
Performance
3. Implementation Quality
Users feedbackDeployment
18
LD4S: Evaluation
1. Actual Gain in Building Automation
Deployment
80%Accuracy Matching the real
consumption
19
LD4S: Usability Evaluation
2. Uptake, Usability,Utility
Users feedback
Tot. Participants: 38
1% had previously interacted with sensors
SurveyGUI Usable and clear API to be improved Applicability to be made explicit
20
LD4S: Utility Evaluation
Time of usage
# Data accessed
# Data transmitted
Amount
Type
Uniqueness
Location
Quality
Time Sensitivity
Relevance
Web Service Resources Linked Data output
Purpose to be made more explicit Links relevancy to purpose to be improved Highlight importance of network/context metadata
21
LD4S: Uptake Evaluation
Unique accessesTot # accesses
Per day (over the 30 days period)
2. Uptake, Usability,Utility
Users feedback
Satisfying for pilot evaluation Project-driven modelling: positive feedback from partners To be repeated over a longer time period
22
LD4S: Performance EvaluationThreshold = # requests / response time (sec)
compared to
Payload size sent + received by LD4S
Performance
3. Implementation Quality Decrease of throughput as payloads increases
But not exponential Improvable by implementing a cache
Outline
1. Linked Sensor Data Model [Q1]
2. LD4Sensors Web Service [Q2]
3. Sensor Relevancy Prediction [Q3]
4. Enriched Web Content [Q4]
5. Network Adaptability [Q5]
6. Research Answers
7. Lessons Learned and Future Work
23
Core Research and
Results
Conclusion
Q2. How can sensors communicate across different platforms without ad-
hoc solutions?
Q3. How to identify which sensors are more relevant sources of information to define a
specific small scope of interest?
24
Relevancy Predictionin Activity Logging
25
LexicalRealisation
ConceptualObjects Cooking
Concept
Objects
FridgeMicrowave
OvenSink
...
Locations
Kitchen...
Fridge
13
12
11
10
9
8
7
6
5
4
3
2
L
5V
A0
ANALO
G IN
AREF
1
GND
TXRX
RESET
3V3
A1
A2
A3
A4
A5
VIN
GND
GND
DIG
ITAL (PWM
=)
ArduinoTM
IOREF
ICSP
ICSP2
ON
POW
ER
0
1TX0
RX0
RESET
Sensor<switch,fridge>
Relevancy Predictionin Activity Logging
Distributional Semantics
Hierarchical Clustering
Feature of interest
(FoI)
26
Relevancy Predictionin Activity Logging
1. from DataHub:Algorithm
Sensors sharing Location & Time
Activated Sensors
2. EasyESA Similarity (X,Y)
X
Y
3. Add to Distance Matrix
4. Clustering
Sensors in the same cluster are relevant for the same activity. Activity = Cluster
Predicting Sensor Relevancy for ADLs Logging 127
of the rows corresponds to a word that occurs inS
i=1...n di. An entry T [i, j] in the table
corresponds to the TF-IDF value of term ti in document dj.
T [i, j] = tf(ti, dj) ⇤ logn
dfi(5.1)
where tf(ti, dj) is the term frequency of the term ti in the document dj defined as
tf(ti, dj) =
8<
:1 + log(count(ti, dj)) if count(ti, dj) > 0
0 otherwise(5.2)
while dfi = |dk : ti 2 dk| is the document frequency, i.e., the total amount of documents
that contain ti.
EasyESA . The size of the textual corpus on which semantic models rely upon is critical
to the quality of the results. This leads to high hardware and software requirements on the
implementation side (e.g., the English version of Wikipedia 2013 contains 43 GB of article
data). For simplicity, we use EasyESA [Carvalho et al., 2014], a JSON webservice which
implements ESA based on Wikiprep-ESA10. It can be queried for either the semantic
relatedness measure, concept vectors or the context windows. In particular, we query
the online available instance11 which run ESA on the English version of Wikipedia 2013.
The query asks for semantic relatedness of pairs of sensors represented as tuples of terms
like ¡ switch, fridge ¿.
5.6.2. Unsupervised Hierarchical Clustering
Unsupervised methods . We chose unsupervised methods because we believe that
given the amount of di↵erent activities and sensors involved, supervised methods are not
likely to scale with the expansion of the Internet of Things phenomenon. In particular,
we chose hierarchical clustering because it is the approach that has so far achieved the
better precision [Kwon et al., 2014].
10https://github.com/faraday/wikiprep-esa11http://vmdeb20.deri.ie:8890/esaservice
Relevancy Prediction:Distributional Semantics
Term frequency Inverse document frequency
model
term frequency of term t in document d
tot documentstot documents containing the term t
Relevancy Prediction:Hierarchical Clustering
Unweighted Pair Group Method with Arithmetic mean (UPGMA)
Weighted Pair Group Method with Arithmetic mean (WPGMA)
Farthest Point or VoorHees (VH)
Reflection of Semantic Distribution
Reflection of Structural Subdivision
Reflection of Centrality
29
Predicting Sensor Relevancy for ADLs Logging 131
with the sensors manually annotated as part of such activity logging. These annotations
and readings are taken from the public14 dataset MITes [Tapia et al., 2004] and were
collected during live experiment settings. We pre-processed such dataset (i.e., CSV files
of sensor readings and metadata about both sensors and activities) to form HTTP PUT
requests to the LD4S API for annotating and storing the data, as in Listing 5.7. Based
on such comparison, the overall accuracy and precision of our system are calculated when
applying either of the clustering algorithms UPGMA, WPGMA or VH.⇤ �1 PUT ld4s:device/2_99
2
3 payload: {’observed_property ’: ’switch’,
4 ’location -name’: [’Kitchen ’],
5 ’foi’: [’Fridge ’]}
6
7 headers: {’Content-type’: ’application/json’,
8 ’Accept ’: ’application/x-turtle ’}⇥ �Listing 5.1: HTTP PUT request forwarded to the LD4S RESTful API.
DataHub (see Section 5.5) was then queried for all the sensor datasets available15
thus returning a JSON list of details of these datasets such as their ID, title, tags, license
and endpoint URIs. The system filters only those datasets that either have no license or
grant an open-access 1. expose a SPARQL endpoint and forward the query in Listing 6.1
towards each of them.Since LD4S triple store is published on DataHub, its endpoint
is also mentioned in such JSON list. Consequently, our query will be forwarded to the
LD4S endpoint as well, so that we will actually get all the data that we had annotated
and stored in the pre-processing step but while also assuring that any other potential
dataset is considered.
The results obtained from each endpoint are XML files - as by W3C standard
recommendation - that the system merged and parsed to distinguish between sensors
that sensed a change in status and the others who just happened to share the same
location. In this experiment we evaluated the worse case: only one sensor has recently
sensed a change in status. The semantic relatedness must be calculated between the
higher amount of possible pairs that share the same location at the same time. This is
used to fill a distance matrix on which the hierarchical clustering algorithms were applied.
In addition to precision and overall accuracy, we also evaluated the performances in
14http://courses.media.mit.edu/2004fall/mas622j/04.projects/home/thesis_data_txt.zip15http://ckan.net/api/3/action/package_search?q=sensor
Predicting Sensor Relevancy for ADLs Logging 133
Table 5.1.: Activities labelled in the MITes dataset.
Number of Examples per Class
Activity Subject 1 Subject 2
Preparing dinner 8 14
Preparing lunch 17 20
Listening to music - 18
Taking medication - 14
Toileting 85 40
Preparing breakfast 14 18
Washing dishes 7 21
Preparing a snack 14 16
Watching TV - 15
Bathing 18 -
Going out to work 12 -
Dressing 24 -
Grooming 37 -
Preparing a beverage 15 -
Doing laundry 19 -
Cleaning 8 -
Internet of Things expansion. The growth of time cost is analysed more thoroughfully in
Section 5.7.4.
The lowest semantic similarity value calculated was �1.0 for the pair ¡switch, tv¿ and
¡ switch, hamper ¿, followed by 0.00036 for the pair ¡ switch, jewelry box¿ and ¡ switch,
microwave ¿. While the highest similarity value was 0.75839 for the pair ¡ switch, cabinet
¿ and ¡ switch, medicine ¿, followed by 0.11285 for the pair ¡ switch, refrigerator ¿ and ¡s
witch, freezer ¿.
5.7.3. Algorithms Comparison
The hypothesis we wanted to verify by applying the chosen algorithms were 1. UPGMA:
is the distance of the semantic distribution of similarities relevant in predicting the
sensor-activity association? 2. WPGMA: does considering the structural subdivision
of the sensor objects positively influence such prediction? 3. VH: can we rely on the
Relevancy Prediction:Evaluation Data
27 FoIs —> 351 Similarity Pairs
132 Predicting Sensor Relevancy for ADLs Logging
terms of execution time for the di↵erent HTTP requests, the SPARQL queries, the whole
pre-processing step and the overall system.
5.7.1. MITes Dataset
Tapia et al. [Tapia et al., 2004] published the MITes dataset from an experiment where
human activity was collected for two weeks. They installed 200 switch sensors deployed
on 27 di↵erent features of interest (FoIs) in two single-person apartments. The sensors
were installed in everyday objects such as drawers, refrigerators, containers, etc. to
record opening-closing events (activation deactivation events) as 2 subjects carried out
everyday activities. The subjects used a software application while they were performing
an activity, to manually annotate it. This resulted in the annotated activities associated
with readings as in Table 5.1. In our experiment we used the data from both subjects
combined together, since evaluating the system di↵erently according to the subject at
end was out of the scope of this paper.
5.7.2. Similarity Results
We considered the worse case in which only one of the sensors sharing the same location
at the same time range has recently sensed a change in status for the current ongoing
activity, while all the other nearby ones which will likely do so in the near future have to
be predicted. In this case, given n sensors, the amount of pairs to check for semantic
relatedness is the binomial coe�cient as in Equation 5.10. In our case since there are 27
di↵erent features of interest, there are 27 di↵erent types of sensors and 351 distinct pairs.
✓n
2
◆=
n!
2!(n� 2!)(5.10)
Even though the binomial coe�cient grows quickly, it only depends on the amount
of features of interest rather than on the amount of actually deployed sensors. At the
same time, the amount of ICOs is expected to grow but the amount of ”types” of
sensors is not, since there is only so much in the real world that can be monitored by
sensors. Our method then is not expected to hinder the system from scaling during the
Worst case scenario: only one of the sensors sharing the same location at the same time range has recently sensed a change in status for the current ongoing activity
Predicting Sensor Relevancy for ADLs Logging 135
Figure 5.4.: Clustering performed by the Voor Hees algorithm.
When comparing our results with the annotated dataset, since we do not perform
cluster labelling, it was not possible to directly map our clusters to the labels in Table 5.1.
However, we considered the match verified whenever the sensors belonging to the same
cluster according to our system (i.e., predicted class) were the ones that sensed the same
activity in the MITes annotations (i.e., actual class). Consequently, we considered a
2-class classification problem, i.e., whether the sensors actually part of the same activity
had been clustered in the same cluster. As a result a separate confusion matrix (Table 5.2)
was created for each of the annotated activity.
Table 5.2.: Confusion matrix displaying number of true positives, true negatives, false positivesand false negatives for a 2-class classification problem.
Predicted vs Actual Actual class
1 2
Predicted class1 TP11 FP12
2 FN21 TN22
With such settings, we calculated precision and overall accuracy.
Precision =TP11
TP11 + FP12(5.11)
Accuracy =TP11 + TN21
TP11 + TN22 + FP12 + FN12(5.12)
Predicting Sensor Relevancy for ADLs Logging 135
Figure 5.4.: Clustering performed by the Voor Hees algorithm.
When comparing our results with the annotated dataset, since we do not perform
cluster labelling, it was not possible to directly map our clusters to the labels in Table 5.1.
However, we considered the match verified whenever the sensors belonging to the same
cluster according to our system (i.e., predicted class) were the ones that sensed the same
activity in the MITes annotations (i.e., actual class). Consequently, we considered a
2-class classification problem, i.e., whether the sensors actually part of the same activity
had been clustered in the same cluster. As a result a separate confusion matrix (Table 5.2)
was created for each of the annotated activity.
Table 5.2.: Confusion matrix displaying number of true positives, true negatives, false positivesand false negatives for a 2-class classification problem.
Predicted vs Actual Actual class
1 2
Predicted class1 TP11 FP12
2 FN21 TN22
With such settings, we calculated precision and overall accuracy.
Precision =TP11
TP11 + FP12(5.11)
Accuracy =TP11 + TN21
TP11 + TN22 + FP12 + FN12(5.12)
Relevancy Prediction:Evaluation: Precision
Dressing Cleaning Toileting Laundry Dinner WashingUp Snack Lunch
Precision of the Activity Clustering
Performance%
0
20
40
60
80
100WPGMAUPGMAVH
Predicting Sensor Relevancy for ADLs Logging 135
Figure 5.4.: Clustering performed by the Voor Hees algorithm.
When comparing our results with the annotated dataset, since we do not perform
cluster labelling, it was not possible to directly map our clusters to the labels in Table 5.1.
However, we considered the match verified whenever the sensors belonging to the same
cluster according to our system (i.e., predicted class) were the ones that sensed the same
activity in the MITes annotations (i.e., actual class). Consequently, we considered a
2-class classification problem, i.e., whether the sensors actually part of the same activity
had been clustered in the same cluster. As a result a separate confusion matrix (Table 5.2)
was created for each of the annotated activity.
Table 5.2.: Confusion matrix displaying number of true positives, true negatives, false positivesand false negatives for a 2-class classification problem.
Predicted vs Actual Actual class
1 2
Predicted class1 TP11 FP12
2 FN21 TN22
With such settings, we calculated precision and overall accuracy.
Precision =TP11
TP11 + FP12(5.11)
Accuracy =TP11 + TN21
TP11 + TN22 + FP12 + FN12(5.12)
Predicting Sensor Relevancy for ADLs Logging 135
Figure 5.4.: Clustering performed by the Voor Hees algorithm.
When comparing our results with the annotated dataset, since we do not perform
cluster labelling, it was not possible to directly map our clusters to the labels in Table 5.1.
However, we considered the match verified whenever the sensors belonging to the same
cluster according to our system (i.e., predicted class) were the ones that sensed the same
activity in the MITes annotations (i.e., actual class). Consequently, we considered a
2-class classification problem, i.e., whether the sensors actually part of the same activity
had been clustered in the same cluster. As a result a separate confusion matrix (Table 5.2)
was created for each of the annotated activity.
Table 5.2.: Confusion matrix displaying number of true positives, true negatives, false positivesand false negatives for a 2-class classification problem.
Predicted vs Actual Actual class
1 2
Predicted class1 TP11 FP12
2 FN21 TN22
With such settings, we calculated precision and overall accuracy.
Precision =TP11
TP11 + FP12(5.11)
Accuracy =TP11 + TN21
TP11 + TN22 + FP12 + FN12(5.12)
Dressing Cleaning Toileting Laundry Dinner WashingUp Snack Lunch
Accuracy of the Activity Clustering
Accuracy%
0
20
40
60
80
WPGMAUPGMAVH
Relevancy Prediction:Evaluation: Accuracy
Relevancy Prediction:Hierarchical Clustering
Unweighted Pair Group Method with Arithmetic mean (UPGMA)
Weighted Pair Group Method with Arithmetic mean (WPGMA)
Farthest Point or VoorHees (VH)
Reflection of Semantic Distribution
Reflection of Structural Subdivision
Reflection of Centrality
Relevancy Prediction:Evaluation: Comparison with SoTA
Predicting Sensor Relevancy for ADLs Logging 137
Dressing Cleaning Toileting Laundry Dinner WashingUp Snack Lunch
Accuracy of the Activity Clustering
Accuracy%
0
20
40
60
80
WPGMAUPGMAVH
Figure 5.6.: Comparison between accuracy percentages achieved by the clustering algorithmsfor some of the activities.
Table 5.3.: Comparison between the experiment setup and results for our own approach andthe previous closest research e↵orts.
Kwon et al. Wyatt et al. Ours
# Sensors 3 100 200
# Activities 5 26 16
Collection Time 50 mins 360 mins 2 weeks
Goal AR AI RSP
Algorithms HIER HMM UH
Precision 79% 70% 89%
Accuracy - 52% 69%
Our results are relevant as we can notice that our system improved the accuracy by 32%
and the precision by 5% with respect to such previous e↵orts from the state of the art.
5.7.4. Performance
The evaluated system run on a laptop equipped with Intel CoreTM2 Duo and 305GB
of disk space. We used the LD4S and EasyEsa service instances running on external
servers in order to support and test a modular and distributed architecture. These were
Increase of 32% accuracy and 5% precision
Relevancy Prediction:Evaluation: Performance
50 100 150 200
2040
6080
Features of Interest (FoIs)
Tim
e (m
sec)
●
●
●
Time Complexity Growth
Time Growth per Amount of FoIs
●
●
●
#FoIs275481112135162189216
HTTP PUT requests: 3ms
Overall Execution: 18ms
Dataset Discovery on DataHub: 3ms (20 datasets)
LD4S SPARQL response: 246ms
ESA: 14ms (351 similarity pairs)
Easy-ESA response: 9ms
Highest time cost = 1 min 26 sec for comparing 216 FoIs
Possibility of updating sensors similarities at run-time CoRE devices (RAM 4 kB and ROM 128 kB): pre-compute offline clustering
Outline
1. Linked Sensor Data Model [Q1]
2. LD4Sensors Web Service [Q2]
3. Sensor Relevancy Prediction [Q3]
4. Enriched Web Content [Q4]
5. Network Adaptability [Q5]
6. Research Answers
7. Lessons Learned and Future Work
35
Core Research and
Results
Conclusion
Q3. How to identify which sensors are more relevant sources of
information to define a specific small scope of interest?
Q4. How can contextualised sensors improve the quality of traditional
Web content?
36
Enriched Web ContentBetween Web and Real Place
37
Hospital Dublin
?
38
Short-lived Data
Long-lived Data
Cost
Enriched Web ContentBetween Web and Real Place
39
1. from DataHub:Algorithm
Sensors sharing
Location & Time
2. Extract Google Search results representing Real Places
3. Live Data Fetching
4. Result Dictionary Update
Bridging the gap between Web and Real Places
Enriched Web ContentG-Sensing
40
Enriched Web ContentG-Sensing Frontend
41
Enriched Web ContentEvaluation Deployment
DataHub
Clinic
Clinic
Clinic
30 sensors
1Km
LD4S
PUT <JSON sensor metadata>
G-Sensing
Google Places
3.692 locations 1.455 (39.4%) have a website
Query: Acupuncture Galway Salthill
42
Enriched Web ContentEvaluation Coverage
How much of the area defined by the virtual locations overlaps with the city of
Galway within radius r=150m
Coverage percentage as we vary the vicinity radius
Added value of our approach for integrating live data into physical locations' websites:
! We divided the areas of Galway divided into squares with different side lengths l
! We counted the number of virtual locations within each square.
Enriched Web ContentEvaluation Distribution
! # virtual locations per square and their respective frequency shows a power-law relationship
! while most squares only contain a small set of locations, a few squares contain a very large number of locations
! (e.g., city centres, business parks).
44
Enriched Web ContentEvaluation Performance
! Google search result page: ~145 KB ! After enabling G-Sensing: ~175 KB (~20% increase)
! At browser start-up: query to DataHub for data source discovery: 3 ms ○ 20 sensor datasets discovered ○ 3 sensor datasets have an open license + expose a SPARQL endpoint ○ 1 sensor dataset’s SPARQL endpoint was accessible (LD4S): 246 ms
G-Sensing does not impede on a user's browsing experience
Bandwidth Overhead
Response Time
Outline
1. Linked Sensor Data Model [Q1]
2. LD4Sensors Web Service [Q2]
3. Sensor Relevancy Prediction [Q3]
4. Enriched Web Content [Q4]
5. Network Adaptability [Q5]
6. Research Answers
7. Lessons Learned and Future Work
45
Core Research and
Results
Conclusion
Q4. How can contextualised sensors improve the quality of traditional
Web content?
Q5. How can contextualised sensors improve the adaptability of
mobile constrained and heterogeneous sensor networks?
Network Adaptability: Demo
46
LD4S
Fuzzy Logic
6LoWPAN + CoAP
Automated LD4S Annotation of new Sensors entering a network
Outline
1. Linked Sensor Data Model [Q1]
2. LD4Sensors Web Service [Q2]
3. Sensor Relevancy Prediction [Q3]
4. Enriched Web Content [Q4]
5. Network Adaptability [Q5]
6. Research Answers
7. Lessons Learned and Future Work
47
Core Research and
Results
Conclusion
Q5. How can contextualised sensors improve the adaptability of
mobile constrained and heterogeneous sensor networks?
Research Answers
48
Future Work• Filtering
• of links according to the resource rating/review LD4S system• of sensor data injected into Google Search results according to user’s prefs &
context• Extending
• derive labels of activities beyond the per-activity sensor clustering• sensor data injected into any Web page and content• sensor data sources extended to include, e.g., TripAdvisor and other user-
generated content• collect user’s feedback on auto-derived annotations for incremental learning
• Evaluation• Long-term large-scale user study to gather insights into how users really use
the current functionalities offered by LD4S• Other areas of research
• Sensor-triggered data can feed back to Linguistic Linked Data knowledge
49