1. KARI LAASONEN BASIC RESEARCH UNIT, HELSINKI INSTITUTE FOR INFORMATION TECHNOLOGY DEPARTMENT OF COMPUTER SCIENCE, UNIVERSITY OF HELSINKI WORKSHOP ON CONTEXT-AWARENESS FOR PROACTIVE SYSTEMS 2005 Route Prediction form Cellular Data

2. 1. Introduction 2. Problem Description 3. Prediction Algorithm  3.1 Route Composition  3.2 Route Similarity  3.3 Making Predictions 4 Evaluation4 Evaluation Comments Outline 2

3. Location awareness plays a large role in ubiquitous computing. Several applications relied on knowing or predicting the location of the user.  Not merely to a known location, but to accurately predict human movement  Early-reminder system  Traffic planning We present an algorithm for predicting movement from cell-based location data.  to learn places that are personally important to that user,  To predict the place the user is moving to. 1. Introduction-1 3

4. Existing approaches to learning important locations and predicting routes rely on GPS data such as [4,2]  GPS can be problematic in urban areas  Privacy. The contribution of the present paper  an enhanced algorithm for predicting routes.  The algorithm analyzes whole paths using string processing techniques, instead of relying on the short path fragments of the earlier paper. 1. Introduction-2 4 2. Harrington, A., Cahill, V.: Route Proling--Putting Context to Work. In: 2004 ACM Symposium on Applied Computing SAC'04, ACM Press(2004) 1567-1573 4. Marmasse, N., Schmandt, C.: A User-centered Location Model. Personal and Ubiquitous Computing 6 (2002) 318-321

5. Problem  A GSM phone communicates with a base station.  over the air  several base stations signal reaches the phone. Select the station which has the strongest signal How about the signal strengths are equal?  A cell is the area covered by a single base station  we say the phone is in cell the phone is in the area of the corresponding base station.  overlapping each other  A physical location does not one-to-one to cells 2 Problem Description 2.1 Locations and Bases 5

6. 6 We can visualize the data by making a graph, the vertices are the observed cells, edge (c i, c j ) is a transition from c i to c j.

7. This graph shows both  daily commute from home (“Vuosaari”) to work  from home to downtown Helsinki.  does not include transitions in the opposite direction. A location is either a cell cluster or a single cell. A location is promoted to a base the time spent there as a portion of the total time the software run goes above a certain threshold. Locations  we can reliably detect the user entering and leaving them.  are important to the user are known as bases. Fig. 1 7

8. the most important consequence of using cell- based location data is that  lack the physical topology of the cell network.  includes the correspondence between cells and physical locations, and also all indications of direction. cell sequence: ABA?  the user visited B and came back.  Or cell A was just briefly shadowed by B. Looking at the immediate context is all but useless. 2.2 Route Prediction 8

9. The first is to examine the local context of recent cells [3].  Suppose in cell c and the have been h1, h2…  prepare strings h k h k-1 …h 1 c, variable k.  matched against a database of stored fragments.  Based on the matches found, and the bases reached from c, we get probabilities for the next base. Two basic approaches to the problem-1 9 3. Laasonen, K., Raento, M., Toivonen, H.: Adaptive On-device Location Recognition. In Pervasive Computing: Second International Conference, LNCS 3001, Springer Verlag (2004), 287{304

10. The second approach, in this paper  entire routes between two bases.  To learn all different physical routes as strings of cell identifiers.  Whenever the user completes a route r between bases a and b  if an existing route between a and b is similar to r.  the two routes are merged together.  To make a prediction  the user has left base a, we have a set of possible routes and their destinations b.  We now use a recent history h of cells and find the route that exhibits the largest similarity to h. Two basic approaches to the problem-2 10

11. 11 route clustering to the data of Fig. 1.  the two most frequently traveled are shown  the routes actually traveled in the physical world.

12. There are three phases in the algorithm.  Phase 1, the user leaves a base enters a cell c,  prepares for a new route prediction task.  Phase 2, at each cell transition, we make a prediction, which is a set of pairs (b, p),  b is a possible future base  p the probability of the user going  Phase 3, when the user arrives at a base, the entire route a, c 1,…, c n, b * is used to make better subsequent predictions. 3. Prediction Algorithm 12

13. For each pair of bases a and b we maintain a set of routes R ab. When the user arrives at base b a new route t = ac 1 … c n b is added to the database  If the maximum similarity of t against all occurs with some r = r max and is greater than a threshold value.  Then t is merged with route r max.  falls below threshold value for all existing routes,  add t to R ab, the set of (distinct) routes between a and b. 3.1. Route Composition 13

14. the paths as strings of cell identifiers or “letters.”  give each letter in both strings a position (value), [0; 1]  the initial value assign to i th letter is v(xi) = (i -1)/(n -1). For example, The merged string is thus “tw(ir)les",  i and r share the same position.  some cells do not necessarily have fixed order to them. the average value  v(t) = 0; v(w) = 1/6 ; v(i) = v(r) = 1/3 ; v(l) = 2/3 ; v(e) = 5/6; v(s) = 1: To handle cyclic paths 14

15. The similarity function, sim(r, t)  r is a composite route between two bases.  t be a complete path. Jaccard measure J = n rt /(n r + n t - n rt ),  n r and n t are the number of elements in r and t,  n rt is the number that is in both.  symmetric, but ignores direction, so a string is equivalent to its reverse. 3.2. Route Similarity 15

16. Inclusion similarity, I  is similar to J but asymmetric let I(r, t) = T/|t|,  T is the number of elements in t that are found, in-order, in r. For example,  I(abcdef; acbdg) = 3/5;  letters `a' and `c' are in order,  but `b' and `c' have been exchanged. Inclusion similarity 16

17. Prediction  By the most recent h = c k-m … c k, not all c 1 … c k  detect faster and more efficient. Route matching has produced a set S of possible reachable bases when starting from base a.  Making a prediction entails computing for each candidate base b S the similarity  the largest similarity of the cell history against all routes leading to b  equal similarities, choose by additional context variables.  time of day, weekday and cell frequency 3.3. Making Predictions 17

18. The data was collected for six months in 2003  With the Context Phone software on a Nokia 7650 phone.  three volunteer users  both at work and at leisure. The baseline algorithm is the fragment-based method [3], which was tested with several window sizes k. The resulting prediction was then compared to the actual base. 4. Evaluation 18 3. Laasonen, K., Raento, M., Toivonen, H.: Adaptive On-device Location Recognition. In Pervasive Computing: Second International Conference, LNCS 3001, Springer Verlag (2004), 287{304

19. A prediction is correct  matches the actual next base and larger than the threshold value = 0.3. A low correct prediction  is correct, but probability is less than the threshold,  or the second-best prediction is correct with nearly equal probability (e.g., p1 = 0:55 and p2 = 0:44),  or the fork point was predicted correctly. A low fail prediction  was wrong, but the probability was also low. A fail -type prediction  wrong, or no prediction at all. Fig. 3 19

20. 20 The F2 and F4 are the fragment method with a window size of 2 and 4, respectively. The C denotes the route prediction using the normal context database, which maintains a time distribution for all intermediate route cells; The reduce model C’ has a time distribution only for the starting times.

21. Fig. 4 21 Fig. 4 Comparison of the memory consumption of the algorithms. Accuracy  models C and C’ are very similar  the latter uses much less memory But even model C consumes less memory than any fragment-based method.

22. Route predictions are based on approximate string matching techniques. 在當時用 GSM 系統進行研究，就我們的研究可以用 WiMAX 進行研究 用 base station 來進行路徑的紀錄及預測，在準確度 上是否會不足 這篇 paper 的做法是需要有歷史紀錄的支持才能進行 預測，對不曾前往的地點則無法預測。 Comments 22

23. Algorithm 23 3. Laasonen, K., Raento, M., Toivonen, H.: Adaptive On-device Location Recognition. In Pervasive Computing: Second International Conference, LNCS 3001, Springer Verlag (2004), 287{304