In this project, we will collect, extract, transform, load, and analyse sensor data transmitted from large amount of sensors installed in school buildings across three European countries. The sensor data report multiple types of information including temperature, humidity, outdoor weather, electronic consumption, human activities, and etc.
Using python library such as Pandas, Matplotlib, Numpy and Comfort Tools from Berkeley , we aim to predict the overall comfort and other KPI for indoor environmental quality in the school buildings
Based on our findings, school management can optimise the environment quality.
From six until we are ready to step out to get to work, most of our time spent at school, sometimes even after school we still stay there for sports activities, or stay in the library for studying.
What should be an ideally school? Enough illumination, comfortable temperature not too hot not too cold, enough fresh air to cool you head down before you crashed down by study and/or stress?
The governments spend a lot of money on education, but is that enough money or is there some way to put the money on the better place instead of just paying the electricity bills?
If we really want to do something for making a better place to study, where should we start to do or look at? We need some first-hand data to tell us which part is the most expensive one so we can start to work on it.
In this situation, sensor data is good cutting point for having an observation and do the analysis.
- Locations in Google Map
| Country | Parameter | Number | Comment |
|---|---|---|---|
| Greece | Sensing endpoints | 872 | Each sensor equals 1 sensing endpoint |
| Sensing rate | 1 minute | Can be modified | |
| Educators | 294 | Greek public schools in GAIA | |
| Students | 2267 | Greek public schools in GAIA | |
| Italy (Roma) | Sensing endpoints | 118 | Will soon be augmented |
| Sensing rate | 1 minute | Can be modified | |
| Educators | 120 | University faculty and Post Doc | |
| Students | 1706 | University students | |
| Italy(Prato) | Sensing endpoints | 117 | |
| Sweden(Soderhamn) | Sensing endpoints | 3 | create on 2017-07-12T12 more to add, no data yet |
Total : 16 sites and 1922 sensors are on the record till 2017/09/16. Part of them are newly installed in this year and some have history data from 2015. The data is collected under different weather condition, from different cultures , with different user behaviour pattens It is a good start we could see the difference and find the similarity.
-
Power consumption
-
Calculated Power Consumption : mWh
-
Power Consumption : mWh
-
Electrical Current : mA/A
-
Active Power : mW
-
Apparent Energy : Vah
-
Apparent Power : VA
-
Voltage : V
-
Power Factor : Raw Value
-
Reactive Energy : VARh
-
Reactive Power : VAR
-
-
Environmental parameters
-
Noise : Raw Value
-
Motion : Raw Value
-
Movement : Raw Value
-
Luminosity : Raw Value
-
Light : lux
-
Atmospheric Pressure : kPa
-
External Relative Humidity : %
-
Relative Humidity : %
-
Rain Height : mm
-
Wind Direction : degrees
-
Wind Speed : m/sec
-
Temperature : Centigrade
-
External Temperature : C
-
Radiation ：uSv/h ( be careful, too high, you will die )
-
External Air Contaminants : Raw Value
-
External Ammonia Concentration : Raw Value
-
External Carbon Dioxide Concentration : Raw Value
-
External Carbon Monoxide Concentration : Raw Value
-
External Oxygen Concentration : Raw Value
-
Carbon Monoxide Concentration : Raw Value
-
Methane Concentration : Raw Value
-
- WiFi
- 2G/3G mobile network connection
- Ethernet
- Low-rate wireless personal area networks(IEEE 802.15.4)
Take a close look on the raw data and seek for the changing patterns :
- Temperature
- Light
- Motion
- Power consumption
- Power Consumption, 10 schools in Greece
There are 3 other schools without power consumption sensors or no data.
- Temperature, 12 schools in Greece
-
Temperature for 4 weeks in the main building with building floor plan
So here are the pattens
- the room(4ce) at ground floor , heading to the north has lowest temperature all the time.
- the two on the first floor, two classrooms to west (class 1 and 2) are next to each other and have similar pattern of temperature changing and the "warmest" rooms in the whole school building.
- the rooms at the north are cooler
-
Temperature at the main building with building floor plan
Patterns : Temperature in Computer Lab is more stable than the rest of others , but still fit our expectation.
-
Humidity at the main building with building floor plan
Patterns : we can see the basement has the highest humidity and the music class is stable and remain in a good dry condition for preserving the music instruments
-
Luminosity at the main building and the sub-site building
Pattern: Most of the rooms use natural light but there is always light only turn off during the weekend The rooms facing south exposed in the longer daylight have maximum luminosity compared with the lab in the basement.
Intuition:
- Some sensor data has reasonable zero value as true value,like motion while no one walking around
- Some sensor data should never be zero , like Temperature and humidity always above zero, or some others type might below zero.
In general , the summary for one point of sensors even there is(/are) some sensor data has "legal" zero the summary based on the same timestamps should always be above zero as long as it is active.
Process :
- Put all [value != 0] = 1
- Sum for all sensor data in one points of sensors
- Normalized all [value > 0] = 1
Output : 1 = active , 0 = inactive for each points of sensors
This table includes large range of data which are missing due to sensors no longer working or the whole sites are power off during [2015-Nov-1,2017-Oct-30]
| ID | Name | Inactive | start time | outlier | Total number of measurements |
|---|---|---|---|---|---|
| 144024 | Δημοτικό Σχολείο Λυγιάς | 30.48 % | before 2015-10-30 | 16.49% | 73,000 |
| 144242 | 1ο Γυμνάσιο Ν. Φιλαδέλφειας | 2.94 % | before 2015-10-30 | 10.90% | 94,900 |
| 144243 | Δημοτικό Σχολείο Μεγίστης | 24.49 % | before 2015-10-30 | 15.80% | 64,970 |
| 155076 | Gramsci-Keynes School | 4.56 % | 2016-08-04 | 39.77% | 39.77% |
| 155077 | Sapienza | 58.22 % | 2016-10-29 | 51.01% | 177,390 |
| 155849 | 6ο Δημοτικό Σχολείο Καισαριανής | 22.98 % | before 2015-10-30 | 16.76% | 52,560 |
| 155851 | 5ο Δημοτικό Σχολείο Νέας Σμύρνης | 29.23 % | 2016-08-02 | 39.25% | 75,190 |
| 155865 | 46ο Δημοτικό Σχολείο Πατρών | 38.72 % | 2016-09-22 | 38.28% | 53,290 |
| 155877 | 2ο Δημοτικό Σχολείο Παραλίας Πατρών | 35.18 % | 2017-02-01 | 45.80% | 46,720 |
| 157185 | Ελληνογερμανική Αγωγή | 3.31 % | 2017-02-01 | 40.50% | 123,370 |
| 159705 | Soderhamn | 0.00 % | 2017-09-21 | 48.24% | 70,810 |
| 19640 | Γυμνάσιο Πενταβρύσου Καστοριάς | 1.72 % | before 2015-10-30 | 20.34% | 112,420 |
| 27827 | 8ο Γυμνάσιο Πατρών | 3.71 % | before 2015-10-30 | 10.71% | 71,540 |
| 28843 | 2ο ΕΠΑΛ Λάρισας | 43.08 % | before 2015-10-30 | 22.17% | 117,530 |
| 28850 | 55o Δημοτικό Σχολείο Αθηνών | 21.23 % | before 2015-10-30 | 22.36% | 91,250 |
| Name | Inactive | outlier | Total number of measurements |
|---|---|---|---|
| Libelium for outdoor weather | 15.16% | 10.61% | 31,390 |
| Synfield for outdoor weather | 14.40 % | 9.28% | 10,950 |
| Electrical Power Consumption | 18.68 % | 33.40 % | 73010 |
-
Clean the times period which all the sensors are inactive.
-
Clean the sensors which are always inactive
- If the site is not power-on yet, it will not be counted as inactive
- Only after sensor(s)(maybe just a few of them)actived, start to count inactive missing data
-
Remove the outliers with Turkey's fences and replace with min/max value
What is outliers In statistics, an outlier is an observation point that is distant from other observations. An outlier may be due to variability in the measurement or it may indicate experimental error; the latter are sometimes excluded from the data set.[3] Outliers can occur by chance in any distribution, but they often indicate either measurement error or that the population has a heavy-tailed distribution. In the former case one wishes to discard them or use statistics that are robust to outliers, while in the latter case they indicate that the distribution has high skewness and that one should be very cautious in using tools or intuitions that assume a normal distribution. A frequent cause of outliers is a mixture of two distributions, which may be two distinct sub-populations, or may indicate 'correct trial' versus 'measurement error'; this is modeled by a mixture model. Output Two ways to indicate a data point is an outlier - Real-valued outlier score, higher values of the score make the point more like an outlier - Binary label binary value yes or no for an data point to be outlier -
identify outliers by using Turkey's fences, aka inter quartile range
Q1 = First QuartileQ3 = Third QuartileInter-quartile Range (IQR) = Q3 - Q1Lower Outlier Boundary = Q1 - 3 * IQRUpper Outlier Boundary = Q3 + 3 * IQR -
identify outliers by using a sliding windows W holds last W-1 values Moving windows through data from the beginning
- If the inter quartile range becomes biggest ever seen,here comes a outliers : replace it with min or max
- If the new value is NaN, it is also an outlier : replace it with average
- min/max/average = min/max/average (previous W-1 values)
-
moving window average to smooth out short-term fluctuations and highlight longer-term trends or cycles
The SMA is the most straightforward calculation, the average over a chosen time period. The main advantage of the SMA is that it offers a smoothed line, less prone to whipsawing up and down in response to slight, temporary price swings back and forth. Therefore, it provides a more stable level indicating support or resistance. The SMA's weakness is that it is slower to respond to rapid changes that often occur at market reversal points. The SMA is often favored by analysts operating on longer time frames, such as daily or weekly charts. -
refill the NaN with average of the whole series values
-
Linear fit
In statistics, linear regression is a linear approach for modeling the relationship between a scalar dependent variable y and one or more explanatory variables (or independent variables) denoted X.``` -
Visualize one day temperature data after processes mentioned above
Can we retrieve outdoor weather through API ?
Openweathermap for real-time data
But this response is only for the real time request.
Worldweatheronline for history data
Both of APIs response :
| Temperature | Wind | Humidity | Pressure | Cloud... |
|---|
What about the accuracy between data retrieved from API and sensors?
Notice API from worldweatheronline does not provide longer than 32days data
Conclusion : Yes we can retrieve both realtime and history,but the accuracy is not pretty enough
-
Assuming the indoor temperature should rise as the day time passing by. We do not put human activity or others into the consideration, for now
-
Identify patten by peak time:
-
Intuitively while observing the temperature peak for different rooms:
- East: the peak temperature mostly arrives at the early day
- West: the peak at the late of the day
- South: the peak should be at the mid-noon or later
- Rest: room facing north/music room/computer lab/basement room will have relatively low variation and average of the temperature
If we put cloud cover persentage with orientation of the room.
Observe the time difference for indoor vs outdoor reach daily peak temperature
-
Predicting orientation by using peak temperature :
-
Using RESTful API to retrieve the time : [sunrise , noon ,and sunset],unit: hour.
-
Only check the temperature during the daytime between [sunrise,sunset]
-
pick the hottest time and put that time:hour into [list_peak_at_hour]
-
Orientation = sum( [list_peak_at_hour] /24hour *360degree)/length([list_peak_at_hour] ) (unit:degree)
-
Simply match Orientation into :
- 0-90 degree: North-East
- 90-180 degree: South-East
- 180-270 degree: South-West
- 270-360 degree: North-West
(./image/comp1.jpg)
Exmaple:8ο Γυμνάσιο Πατρών, Class 1 id fb8, the classroom reach the highest temperature at hour: peak_at_hour_list = [ 18, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 18, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 17, 16, 16, 16, 16, 18, 17, 16, 16, 17, 16, 17, 16, 16, 17, 16, 18, 17, 17, 17, 16, 16, 17, 17, 16, 17,17, 16, 16, 17, 17, 17, 16, 17, 17, 18, 17, 17, 16, 16, 17, 17, 17, 17,17, 17, 16, 17, 17, 16, 17, 15, 16, 17, 16, 16, 16, 16, 16, 16, 17, 17, 16, 17, 16, 17, 17, 15, 16, 17, 17, 17, 16, 16, 16] in total: [(16 o'clock, 33times), (18 o'clock, 5times),(15 o'clock, 2times), (17 o'clock, 60times)] Orientation = sum(16/24*360*26+...+15/24*360*2+17/24*360*60)/Length(peak_at_hour_list) = 250.2 degree So we got the orientation is 250.2 degrees which looks like south-west.
But ... if there are also a lot of peaks happened in the morning ClasslB2 id :0x317. This room gets enough exposed in the sunshine/high temperature during the morning and also likely it facing to the south-east in total: (14 o'clock, 5times), (16 o'clock, 3times), (10 o'clock, 3times), (12 o'clock, 3times), (11 o'clock, 3times), (15 o'clock, 3times),(13 o'clock, 2times), (18 o'clock, 2times), (17 o'clock, 2times), (7 o'clock, 1times), (19 o'clock, 1times) Orientation = 205.7 degree,have peak temperature in the mornings at [7, 10, 11], takes 25.0% in total we got something more like heading to the east , or south east
We take a close look at the distribution of site rooms in different orientation
The south-east are in lower temperature compared with south and south-west room
Category by site :
SITE ID Orientation Prediction Correct 144024 60% 144243 50% 155851 20% 155865 50% 19640 0% 27827 25% 144242 0% 155877 33% 159705 28% 155849 0% 157185 44% REST (no data) Category by the room orientation :
Orientation Orientation Prediction Correct comment N - E 0% 0/12 E - S 35% 6/17 S - W 56% 9/16 W 0% 0/1 W - N 0% 0/11 -
-
The cloudy coverage impact on the indoor temperature:
-
-
Identify deviation by slope = delta(temperature)/delta(time):
-
Slope for one room: first picture is the data after ETL, second is the slop for the data If we are focusing on detecting sudden change of the indoor temperature, this slope plot could provide fast and efficient way, such as the fluctuate exists on Sep-23 on the top plot, a peak matching to this change is observed at the bottom plot as well. We can use the similar algorithm: searching the "outliers" for slope to detect the fluctuate.
-
And put all the classrooms from one site together(ETL data on top and slope on bottom) we can detect which room behavior abnormal : the room in red during the day time and room in orange in the night time.
-
From Wikipedia
Thermal comfort
Thermal comfort is the condition of mind that expresses satisfaction with the thermal environment
and is assessed by subjective evaluation (ANSI/ASHRAE Standard 55).
ANSI/ASHRAE Standard 55
(Thermal Environmental Conditions for Human Occupancy) is a standard
that provides minimum requirements for acceptable thermal indoor environments.
The purpose of the standard is to specify the combinations of indoor thermal environmental factors
and personal factors that will produce thermal environmental conditions
acceptable to a majority of the occupants within the space
The standard addresses the four primary environmental factors
(temperature, thermal radiation, humidity, and air speed)
and two personal factors (activity and clothing) that affect thermal comfort.
It is applicable for healthy adults at atmospheric pressures in altitudes up to (or equivalent to) 3,000 m (9,800 ft),
and for indoor spaces designed for occupancy of at least 15 minutes.
Comfort zone
Refers to the combinations of air temperature, mean radiant temperature (tr),
and humidity that are predicted to be an acceptable thermal environment at particular values of
air speed, metabolic rate, and clothing insulation (Icl)
Intuitively, we want the temperature indoor in the certain range like [18,24]
during Monday to Friday, from 8:00 to 18:00
Obviously, the truth is not always what we wish for
Tool: CBE Thermal Comfort Tool for ASHRAE-55 How to use:
By choosing the Adaptive method at the very top of the user interface,
the chart changes and the input variables include air temperature, mean radiant temperature and prevailing mean outdoor temperature.
This is because the personal factors and humidity are not significant in this method since adaptation is considered, and the only variable is the outdoor temperature.
See above for explanation of the first two variables, air and mean radiant temperature.
Prevailing mean outdoor temperature
Here you can type the outdoor temperature averaged as explained on the standard.
See the Wikipedia link for a brief explanation.
Changing this variable makes the dot representing the current condition move horizontally.
The meaning of this chart is that certain conditions of indoor-outdoor temperature fall inside the comfort zone,
which in this case is static.
-
Sample period 2017.Sep.05-2017.Nov.04 , weekday: Monday-Friday, Time: 8:00-16:00, week of year: 36-44
-
In daytime, this room is comforable for (n*8hours) , 0 < n < 1
-
All day comfortable = 1 all day not comfortable = 0
-
The following two pictures are comfortness ratio per day is based on hourly temperature from site ID 27827 8ο Γυμνάσιο Πατρών
- use "Site-Temperature" sensor as outdoor temperature source
- Use Worldweatheronline API as outdoor temperature source
- The difference might because of the not accuracy from "Site-Temperature"
- use "Site-Temperature" sensor as outdoor temperature source
-
The following two pictures are comfortness ratio per day is based on hourly temperature from site ID 144242,1ο Γυμνάσιο Ν. Φιλαδέλφειας
- use libelium sensor as outdoor temperature source
- Use Worldweatheronline API as outdoor temperature.
- The difference might because of the "complete nonsense data" from this libelium sensor
- use libelium sensor as outdoor temperature source
-
The following two pictures are comfortness ratio per day is based on hourly temperature from site ID 19640,Γυμνάσιο Πενταβρύσου Καστοριάς - use libelium sensor as outdoor temperature source
- Use Worldweatheronline API as outdoor temperature.
- Use Worldweatheronline API as outdoor temperature.
-
-
Retrieve the data by using APIs on https://api.sparkworks.net/swagger-ui.html
Demo for "POST /v1/resource/query/timerange" :
Command line:
curl -X POST --header 'Content-Type: application/json' --header 'Accept: application/json' --header 'Authorization: Bearer cd885cf5-7fca-4be8-b32e-97225da6763f' -d '{
"queries": [
{
"from": 1498867200000,
"granularity": "day",
"resourceID": 156972,
"resultLimit": 0,
"to": 1500076799000
}
]
}' 'https://api.sparkworks.net/v1/resource/query/timerange'
Reponse body:
{
"results": {
"{\"resourceID\":156972,\"resourceURI\":\"gaia-ea/room-1/temp\",\"from\":1498867200000,\"to\":1500076799000,\"granularity\":\"day\",\"resultLimit\":0}": {
"average": 31.995042261495424,
"summary": 479.92563392243136,
"data": [
{
"timestamp": 1498856400000,
"reading": 34.19535065107274
},
{
"timestamp": 1498942800000,
"reading": 35.618584889499054
},
....
{
"timestamp": 1499979600000,
"reading": 25.807027188020786
},
{
"timestamp": 1500066000000,
"reading": 28.399119190883642
}
]
}
}
}
-
when we request data within the time range, for different granularity , the response time stamps are different
- 5min : it's the code running time, not the fixed data timestamp as below
- 1hour : beginning of every hour
- 1day : 21:00 for each day
- 1month : at 21:00 last day of the month
Solution : use ".toInstant().toEpochMilli()/ 300000*300000" to change the time into 5mins interval in one hour. So we will have 5', 10',15',...55' for the record timestamp
-
Cassandra do not have connector as data source for flink
Solution
reference : CassandraConnectorITCase
ClusterBuilder cb = new ClusterBuilder() { @Override public Cluster buildCluster(Cluster.Builder builder) { return builder.addContactPoint("127.0.0.1").build(); } }; String query = "SELECT ResourceID,Reading FROM gaia.reading_data WHERE ResourceID=155873"; InputFormat<Tuple2<Integer, Float>, InputSplit> source = new CassandraInputFormat<>(query, cb); source.configure(null); source.open(null); List<Tuple2<Integer, Float>> result = new ArrayList<>(); while (!source.reachedEnd()) result.add(source.nextRecord(new Tuple2<Integer, Float>())); source.close(); -
Can't read data through API:
Resource historical data resource ID: 90946 from : 2017-08-24T09:25:49.323Z to 2017-08-31T09:25:49.080Z with steps per hour [0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0] However the real data is not all-zeros in https://console.sparkworks.net/resource/view/90946
We will use the data extracted from API for further process , the console data are only for reference.
-
Data is missed from time to time:
Resource historical data resource ID: 155918 from : 2017-08-24T09:25:49.323Z to 2017-08-31T09:25:49.080Z with steps per hour [32.34, 32.34, 32.34, 32.473637, 33.4425, 34.259167, 34.75733, 34.365334, 33.32, 32.764668, 32.570587, 32.36722, 32.3155, 32.3068, 32.241306, 31.868149, 31.868149, 31.85, 31.838118, 31.808867, 31.826338, 31.822662, 31.832302, 31.852188, 31.842134, 31.841246, 31.85319, 32.140522, 33.013374, 33.919827, 34.664608, 34.72307, 32.34, 32.34, 32.34, 31.85, 31.85, 32.34, 31.85, 31.85, 31.85, 31.85, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 31.367912, 31.374935, 31.410751, 31.416487, 31.51673, 31.94285, 32.77366, 33.64148, 34.232605, 33.768044, 32.68922, 32.30267, 32.140594, 31.958946, 31.87191, 31.852951, 31.830269, 31.64, 31.46889, 31.456898, 31.3657, 31.385254, 31.381018, 31.41322, 31.534771, 31.668463, 31.79291, 31.836, 31.869188, 32.38108, 33.044456, 33.666306, 33.827496, 32.96945, 32.38393, 32.23478, 32.038074, 31.907423, 31.856163, 31.844433, 31.838556, 31.69923, 31.6932, 31.511343, 31.412098, 31.391008, 31.380096, 31.443954, 31.619139, 31.776262, 31.830036, 31.832235, 31.799591, 31.84828, 32.09569, 32.928936, 33.33732, 33.126163, 32.188957, 31.83223, 31.623442, 31.422161, 31.361742, 31.356749, 31.3123, 30.906296, 30.886333, 30.878448, 30.859062, 30.843264, 30.859325, 30.861336, 30.849495, 30.85837, 30.85697]
- The building data genome project
- Big data computing
- Naked Statistics: Stripping the Dread from the Data by Charles Wheelan
- Real Time sensor status on GAIA
Professor Ioannis Chatzigiannakis
Professor Aris Anagnostopoulos
Since the first day I stepped into Sapienza , you are the ones lead me to “find true love with data mining”. Till the last six months, you not only gave me “fish” but also taught me “how to fish” with all your supports and encourage. That is beyond anything I could expect but I am so blessed to have. Without you, I won’t be standing here today.