1. RFID Data Aggregation Dritan Bleco, Yannis Kotidis Department of Informatics Athens University of Economics and Business

2. Yannis Kotidis Outline Introduction Temporal Aggregation Basic Temporal Aggregation - BTA Lossy Temporal Aggregation - LTA Spatial Aggregation Evaluation Conclusions

3. Yannis Kotidis Radio Frequency Identification (RFID) Use radio-frequency waves to transfer data between a reader and a movable item to identify, categorize, track... Does not require physical sight or contact between reader/scanner and the tagged item

4. Yannis Kotidis Variations Active/Passive Tags Memory Size (16bits- KBs) Memory Type Read Only, WORM, Read/Write Frequency 125KHz - 5.8 GHz Physical Dimensions Thumbnail to Brick sizes Price (few cents-hundred euros)

5. Yannis Kotidis Existing Applications Animal/livestock tracking Postal services (routing and sorting) Libraries Toll collection Warehousing Supply chain management …

6. Yannis Kotidis RFID System Architecture RFID Reader tag RFID Reader tag Edgeware (on-site) filtering, cleaning, aggregation Raw RFID Data Stream Aggregated RFID Data Stream Middleware (remote) IT Applications tag Level-0 Level-1 Level-2 Level-3 EPC Code (tag) Time (reader) Location (reader)

7. Yannis Kotidis Simple RFID data stream model Assume streaming records with schema EPC code: EPC i (Discrete) Time: t i Location: loc i

8. Yannis Kotidis Basic Temporal Aggregation (BTA) Collate consecutive reports of the same tag Reader (EPC i,loc i,t start ) (EPC i,loc i,t start+1 ) (EPC i,loc i,t start+2 ) … (EPC i,loc i,t end ) t start t end Raw stream (EPC i,loc i,t start,t end ) Aggregated stream EPC i loc i

9. Yannis Kotidis Problems with BTA RFID readers often drop observations e.g. due to collisions Up to 30% loss is not uncommon [Jeffery2006] Objects are often moved within the facility Multiple BTA records Reduction depends on data characteristics Need an application-controllable reduction framework OLAP analysis does not require precise knowledge!

10. Yannis Kotidis Lossy Temporal Aggregation (LTA) LTA record format: (EPC,loc,t start,t end,p) Tag may be partially present during the interval Value denotes the fraction of times that the tag was observed during the interval BTA: p=1 (implied) LTA: 0<p≤1 Allow us to control the size of the aggregated stream or the level of accuracy

11. Yannis Kotidis Types of Error in LTA t start t end X=epochs when tag was reported in [t start,t end ] Y=epochs when tag was not reported in [t start,t end ] p = X / (X+Y) Tag spotted but not reported Tag spotted but reported with probability p instead of 1 Tag not spotted but nevertheless reported with probability p selected LTA interval

12. Yannis Kotidis Problem Formulation Compute best B-tuple LTA representation such that cumulative error (including both false negative and false positive error types) is minimized Cumulative Error = 2*X*Y/(X+Y) 2 Other error metrics? Dual problem also interesting

13. Yannis Kotidis Helpful Observations 1. Selected end-points t start,t end must contain appearance of a tag 2. Should not break consecutive observations Bad choice due to (1) Bad choice due to (2) Thus, we can first apply BTA and afterwards LTA

14. Yannis Kotidis Linear Algorithm Goal: generate B LTA records Input: n BTA records Example: Reduce stream from 8 to B=4 records BTA Interval LTA Interval

15. Yannis Kotidis Greedy Algorithm Iteratively merge intervals Select best candidate at each step Stop when left with exactly B intervals

16. Yannis Kotidis Optimal LTA Dynamic Programming formulation E(i,k): error of best k LTA representation for first i BTA intervals err(a,b): error for single LTA record encoding intervals a, a+1, … b k-1 LTA intervals 12jj+1i 1 LTA interval E(i,k)=min(E(j,k-1)+err(j+1,i)) j<i BTA intervals:

17. Yannis Kotidis Spatial Aggregation Tags often move in batches Common in supply- chain/distribution networks Idea: create surrogate EPC codes to replace multiple tags packaged together Proposed in [Gonzales et all ICDE 2006] Note: Do not know in advance how items are grouped Surrogate codes do not imply physical grouping

18. Yannis Kotidis Example G1: I1 I2 Surrogate Group codes I1 L1 T1 T5.78 I2 L1 T1 T5.69 I3 L1 T2 T5.90 I1 L2 T12 T22.67 I2 L2 T12 T22.62 I4 L2 T12 T22.66 LTA stream These items are observed at the same interval/location

19. Yannis Kotidis Example G1L1 T1 T5.69 I3 L1 T2 T5.90 I1 L2 T12 T22.67 I2 L2 T12 T22.62 I4 L2 T12 T22.66 LTA stream New record replaces both entries G1: I1 I2 G2: G1 I4 Surrogate Group codes More tags spotted together

20. Yannis Kotidis Resulting Tables G1L1 T1 T5.69 I3 L1 T2 T5.90 G2 L2 T12 T22.62 Reduced stream G1: I1 I2 G2: G1 I4 Surrogate Group codes I1 L1 T1 T5.78 I2 L1 T1 T5.69 I3 L1 T2 T5.90 I1 L2 T12 T22.67 I2 L2 T12 T22.62 I4 L2 T12 T22.66 LTA stream

21. Yannis Kotidis Experiments Used RFID traces from 2008 Hope Conference in New York Sampled data at 30sec intervals 1.9Million records Reduced to 423K records via BTA

22. Yannis Kotidis Accuracy (LTA) Picked tag with most intervals (569) Vary number of requested LTA-tuples (B)

23. Yannis Kotidis Execution Times (LTA)

24. Yannis Kotidis Notes on Spatial Aggregation Input: 434K BTA records Output 77K surrogate group ids 39% space reduction (accounting surrogates) 3.3secs (1.83GHz Core Duo with 1GB)

25. Yannis Kotidis Different Choices LosslessLossy 3:1

26. Yannis Kotidis Conclusions Explored different aggregation schemes Exploit temporal and spatial correlations Schemes reduce size of RFID stream in a user- controllable manner All algorithms are fairly fast Greedy is orders of magnitude faster than OptimalDP with practically identical performance More schemes possible Ex: spatial with fuzzy groups Other error metrics, dual problem

27. Yannis Kotidis Thank you, Questions?