The self-flow in taxi flow are removed when generating embedding

[thekingofkings]

The taxi-flow function sets self-flow f_{i,i} to 0. This is required in the regression-based method.

However, when learning graph embedding (#27), self-flow is an important feature that cannot be omitted.

[thekingofkings]

With self-flow, the best setting is

1. Consider 8 neighbors
2. Vector size is 10 (5 from 1st order, 5 from 2nd order neighbors)
3. train on 20 millions graph samples
4. Number of negative sample is 5

[thekingofkings]

The leave one out error is much lower. In year 2010 data,

MAE: 465.89, MRE: 0.3477