[03] Reproduce Results
You can follow anlaysis/synthetic_data_analysis.ipynb to reproduce our works in synthetic simulation analysis
to replicate Figure 1 and Figure 2 in section 6.2 of our paper.
Block [22] generates Figure 1 and block [21] generates Figure 2.
Details
python/gen_simulation_dat.py provides a package for generating simulation data.
In block [3], we use functions beta_gaussian_process and get_game_matrix_list to generate toy example with N objects and T time points for a sanity check on time-varying Bradley-Terry model. beta_gaussian_process samples ground-true parameter beta_gp from a Gaussian process, and then get_game_matrix_list simulates tn comparisons between each pair of objects
We compare time-varying Bradley-Terry model to two traditional methods, win rates and separate Bradley-Terry models at different time points.
In block [7], we get win rates for the objects.
In block [8], we use pgd_l2_sq with l_penalty=0 to run separate Bradley-Terry models at different time points.
In block [10], we use cv_utils.loocv to cross-validate on smoothness penalty parameter and get a cross-validated estimate
beta_l2sq_cv.
In block [12], we get a plot of cross-validation loss versus .
In block [13] and [14], we fit our model with -penalty by setting a specific value of
. One can uncomment the code in the two blocks to run LOOCV.
We can download the nflWAR data by simply running
make get_nfl_filt_dataYou can follow anlaysis/analysis_nfl.ipynb to reproduce our works in synthetic simulation analysis
to replicate Table 1 in section 7 of our paper.
Block [12] generates Table 1 and Block [16] gets average differences from ELO to our model for 2009-2015 seasons.