Node feature selection

[cshjin]

• node feature descriptions

attr	type	comment
type	categorical	auxiliary / compute / transfer
is_clustered	int
ready	timestamp
pre_script_start	timestamp
pre_script_end	timestamp
submit	timestamp
execute_start	timestamp
execute_end	timestamp
post_script_start	timestamp
post_script_end	timestamp
wms_delay	float	= submit - ready
pre_script_delay	float	= pre_script_end - pre_script_start
queue_delay	float	= execute_start - submit
runtime	float	= execute_end - execute_start
post_script_delay	float	= post_script_end - post_script_start
stage_in_delay	float
stage_in_bytes	float
stage_out_delay	float
stage_out_bytes	float
kickstart_user	categorical	panorama
kickstart_site	categorical	local / condorpool
kickstart_hostname	categorical	master / worker-*
kickstart_transformations	str
kickstart_executables	str
kickstart_executables_argv	str
kickstart_executables_cpu_time	float
kickstart_status	int
kickstart_executables_exitcode	int

• exclude timestamps
• exclude long string
• handle missing values

[cshjin]

Done in https://github.com/PoSeiDon-Workflows/graph_nn_2/blob/hongwei_dev/psd_gnn/dataset.py

[cshjin]

• Normalization
  • MinMaxScalar vs StandardScalar
• Feature Selection
  • Auto feature selection from model

[cshjin]

• implement the column-wise normalization as a separate transform class
  • examples: https://pytorch-geometric.readthedocs.io/en/latest/_modules/torch_geometric/transforms/normalize_features.html
• try out transform function in pytorch_geometric, including split, toundirected and more.

[cshjin]

• check out new features in montage and predict_future_sales

[cshjin]

• fit_transform on train and transform on test
• remove the normalization in dataset.py