Techniques and Challenges for Trajectory Prediction Slides credit: - PowerPoint PPT Presentation

Where Next? Data Mining Techniques and Challenges for Trajectory Prediction Slides credit: Layla Pournajaf

o Navigational services. o Traffic management. o Location-based advertising. Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

o Destination prediction o Path prediction with known destination o Path prediction with unknown destination o Similar to predicting next N locations Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

Raw Trajectories Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

Raw Trajectories Preprocessed Trajectories Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

Prediction Model Raw Trajectories Preprocessed Trajectories Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

Source: www.openstreetmap.org

Real-world data include raw trajectories of continuous GPS coordinates which are noisy and inaccurate! Source: www.openstreetmap.org

Raw Trajectories Preprocessed Trajectories Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

o Discretizing Time o 30 seconds, one hour o Temporal Representation o Location-series o Fixed-interval time-location series o Variable-interval time-location series Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

o Discretizing Location o Grid-based (uniform vs hierarchical) o Mining Frequent Regions  Clustering  Semantic-based Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

Map of Beijing with 30 × 30 grid overlay: Each cell ≈ 1.78km 2 Source: Xue, Andy Yuan, et al. "Destination prediction by sub-trajectory synthesis and privacy protection against such prediction." ICDE 2013.

o Clustering o DBScan o Hierarchical Clustering o Semantic-based o Using points of interests Source: Lei, Po-Ruey, et al. "Exploring spatial-temporal trajectory model for location prediction." MDM 2011.

Prediction Models Raw Trajectories Preprocessed Trajectories Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

 Personalized / Individual-based: o Utilize only the history of one object to predict its future locations  General: o Utilize the history of all objects to predict future locations

 Model-based ( formulate the movement of moving objects using mathematical models) o Markov Models o Hidden Markov Models (Zhou et. al., ACM SIGKDD 2013) o Recursive Motion Function ( Y . Tao et. al., ACM SIGMOD 2004) o Deep learning models  Pattern-based ( exploit pattern mining algorithms for prediction) o Sequential Pattern Mining ( G. Yavas et. al., DKE 2005) o Trajectory Pattern Mining  Hybrid o Recursive Motion Function + Sequential Pattern Mining (H. Jeung et. al., ICDE 2008)

Source: Xue, Andy Yuan, et al. "Destination prediction by sub-trajectory synthesis and privacy protection against such prediction." ICDE 2013.

2 𝑞 45 = 3 1 𝑞 56 = 3 Source: Xue, Andy Yuan, et al. "Destination prediction by sub-trajectory synthesis and privacy protection against such prediction." ICDE 2013.

Source: Xue, Andy Yuan, et al. "Destination prediction by sub-trajectory synthesis and privacy protection against such prediction." ICDE 2013.

Partial Trajectory: < 𝑠 1 , 𝑢 1 > , < 𝑠 2 , 𝑢 2 >, …., < 𝑠 𝑑 , 𝑢 𝑑 > <?, 𝑢 𝑑 +1 > Prediction: • Having a partial trajectory (discretized) including the current region 𝑠 𝑑 , find the most probable region at time point 𝑢 𝑑 +1 arg max P ( 𝑆 𝑑 +1 = 𝑠 𝑑 +1 | 𝑠 1 , … 𝑠 𝑑 ) 𝑠 𝑑 +1

Embedding Higher-Order Chains • Each new state depends on fixed-length window of preceding state values • We can represent this as a first-order model via state augmentation : (N 2 augmented states)

Semi-Lazy Hidden Markov Approach ( SIGKDD ‘13) • Find similar trajectories from historical trajectories (reference objects) • Build a hidden Markov Model on the fly (vs. eager or lazy approach) • Self-correcting continuous prediction (real time) • Refine prediction model • Adjust weights for reference objects

 Model-based ( formulate the movement of moving objects using mathematical models) o Markov Models o Hidden Markov Models (Zhou et. al., ACM SIGKDD 2013) o Recursive Motion Function ( Y . Tao et. al., ACM SIGMOD 2004) o Deep learning models  Pattern-based ( exploit pattern mining algorithms for prediction) o Sequential Pattern Mining ( G. Yavas et. al., DKE 2005) o Trajectory Pattern Mining ( Monreale et al ACM SIGKDD 2009)  Hybrid o Recursive Motion Function + Sequential Pattern Mining (H. Jeung et. al., ICDE 2008)

1. Preprocess raw trajectories and extract frequent sequential patterns (T-Pattern) Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

1. Preprocess raw trajectories and extract frequent sequential patterns (T-Pattern) 2. Build a Prefix Tree (T-Pattern Tree) Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

1. Preprocess raw trajectories and extract frequent sequential patterns (T-Pattern) 2. Build a Prefix Tree (T-Pattern Tree) 3. Predict Next Location Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

< 𝑦 1 , 𝑧 1 , 𝑢 1 > , < 𝑦 2 , 𝑧 2 , 𝑢 2 >, …., < 𝑦 𝑜 , 𝑧 𝑜 , 𝑢 𝑜 > Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

< 𝑦 1 , 𝑧 1 , 𝑢 1 > , < 𝑦 2 , 𝑧 2 , 𝑢 2 >, …., < 𝑦 𝑜 , 𝑧 𝑜 , 𝑢 𝑜 > • Two points match if one falls within a spatial neighborhood N() of the other • Two transition times match if their temporal difference is ≤ τ Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

< 𝑦 1 , 𝑧 1 , 𝑢 1 > , < 𝑦 2 , 𝑧 2 , 𝑢 2 >, …., < 𝑦 𝑜 , 𝑧 𝑜 , 𝑢 𝑜 > • Two points match if one falls within a spatial neighborhood N() of the other • Two transition times match if their temporal difference is ≤ τ • Calculate support for each T-pattern Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

Generating all association rules from each T-pattern and using them to build a classifier is too expensive. α 3 α 1 α 2 T-Pattern R 1 R 2 R 3 R 4 Rules R 1 R 2 R 3 R 4 R 1 R 2 R 3 R 4 R 1 R 2 R 3 R 4 Source: A. Monreale, F. Pinelli, R. Trasarti, F. Giannotti. WhereNext: a Location Predictor on Trajectory Pattern Mining . KDD 2009

To avoid the rules generation, the T-Pattern set is organized as a prefix tree. For Each node v • Id identifies the node v • Region is a spatial component of the T-Pattern • Support is the support of the T-pattern For Each edge j [a,b] correspond to the time interval α n of the T-Pattern

Three steps: 1. Search for best match 2. Candidate generation 3. Make predictions Best Match Prediction

Three steps: 1. Search for best match 2. Candidate generation 3. Make predictions The Best Match is the path having:  the maximum path score using the time and location matching and support  at least one admissible prediction.

o Prediction errors (distance and time) o Prediction accuracy (precision and recall) o Prediction rate