Group trajectories according to queries, hold 30% of trajectories conform to each query for test, and use all other trajectories for training.
Given a query $x$, the nearest neighbour heuristic make a prediction by choosing one from the set of trajectories that conform to $x$ in training set, there's a number of strategies to make a choice, e.g., choose uniformly at random, choose the one with the most support etc.
The conjecture is that the nearest neighbour heuristic will probably beat all other sophisticated methods (e.g., ranking, ssvm, memm) in this evaluation protocol.
We hope the results can provide helpful information for designing an evaluation protocol.
cdawei
commented
7 years ago
Group trajectories according to queries, hold 30% of trajectories conform to each query for test, and use all other trajectories for training.
Given a query $x$, the nearest neighbour heuristic make a prediction by choosing one from the set of trajectories that conform to $x$ in training set, there's a number of strategies to make a choice, e.g., choose uniformly at random, choose the one with the most support etc.
The conjecture is that the nearest neighbour heuristic will probably beat all other sophisticated methods (e.g., ranking, ssvm, memm) in this evaluation protocol.
We hope the results can provide helpful information for designing an evaluation protocol.