I have a question about Dynotears about causalnex HOT 11 CLOSED

Did you have a look at
from causalnex.structure.transformers import DynamicDataTransformer?

Its a util to get the data in the right shape.

from causalnex.

@qbphilip
Thank you for your comment.
I have a look at DynamicDataTransformer.
It transforms data like g1_lag0 g2_lag0 g1_lag1 g2_lag2.... g1_lagp g2_lagp.
However from_pandas_dynamic function contains DynamicDataTransformer.
Could you give me a simple example code and data for dynotears?

from causalnex.

Hi!
Thanks for the interest in causalnex.

if you provide a pandas dataset whose index represent timestamps, from_pandas_dynamic should work. That is, the row 0 is the timestamp zero, row 1 the first timestamp, and so on... We do not need to create the "lagged" columns because from_pandas_dynamic creates it for you.

I am printing below some code of a simulated dataset. Maybe the API changed a little because this code as many months old:

#### GET A SIMULATED TIME SERIES ####
import warnings
import pandas as pd
import numpy as np
from sklearn import preprocessing
import seaborn as sns
# silence warnings
warnings.filterwarnings("ignore")

from causalnex.structure.data_generators import gen_stationary_dyn_net_and_df

# Obtain simulated structure (g), dataset sampled from g and list of intra- and inter-slice node names
g, df, intra_nodes, inter_nodes = gen_stationary_dyn_net_and_df(
    num_nodes = 10,
    n_samples = 10000,
    p = 1,
    degree_intra = 3,
    degree_inter = 2,
    graph_type_intra = 'erdos-renyi',
    graph_type_inter = 'erdos-renyi',
    w_min_intra = 0.3,
    w_max_intra = 2,
    w_min_inter = 0.3,
    w_max_inter = 0.5,
    w_decay = 1.0,
    sem_type = 'linear-gauss',
    noise_scale = 1,
    max_data_gen_trials = 1000,
)

df = df[intra_nodes]
df.columns = [el.split('_')[0] for el in df.columns]
df.head() # <<--- this is a time series data. each row is a timestamp 

Then you can call dynotears:

from causalnex.structure.dynotears import from_pandas_dynamic
g_learnt = from_pandas_dynamic(df,1,lambda_w=.1,lambda_a=.1,w_threshold=.1)
g_learnt

then, to see the graph you can do:

from copy import deepcopy
g_learnt_2 = deepcopy(g_learnt)
g_learnt_2.remove_edges_below_threshold(.1)
from causalnex.plots import plot_structure
from IPython.display import Image
viz = plot_structure(g_learnt_2.get_largest_subgraph())
f='dbn_learnt.jpg'
viz.draw(f)
Image(f)

from causalnex.

each variable has a _lagX on it. this represents the variable shiftted of one timestamp. for example,
the edge 1_lag1 --> 5_lag0 indicates that variable 5 in a time $t$ will be affected by the value of the variable 1 at $t-1$.

from causalnex.

Let me know if it helps :)

from causalnex.

Hi @GabrielAzevedoFerreiraQB ,

I have tried to execute your code from #74 (comment)

When running second snippet (with g_learnt) I am getting this error in Jupyter notebook (Python 3.7.6):

`~\anaconda3\lib\site-packages\causalnex\structure\transformers.py in _check_input_from_pandas(self, time_series)
203
204 if t.index.dtype != int:
--> 205 raise TypeError("Index must be integers")
206
207 if self.columns is not None:

TypeError: Index must be integers`

I get the same error when I initially tried to execute from_pandas_dynamic on my own dataframe.

Do you know what can be the problem?

cc: @qbphilip
Thanks!

from causalnex.

hmm, somehow the returned index are not integers. Maybe there was a change in the generators, and the indexes are not integers anymore

The index of your dataframes must represent the sampling time of your time series. (i.e. x_1, x_2, x_3...) (this is very important)

I suggest taking a look at df.index to make sure that they are (1) integer, (2) in increasing order and (3) that the indexes increases 1 by one.

from causalnex.

Hi I can execute the code at comment 74 #74 (comment),
Is there a way to impose tabu_edges between lag1 and lag0 variables?
How should we use the learned structure to instantiate a DBN?

from causalnex.

A more robust integer type checking has been implemented in this commit and will be available in the next CausalNex release.

from causalnex.

Hi,

How do I access the adjacency matrix of the "true" graph from the above code ?

from causalnex.

Hi @donaldRwilliams,

My understanding is that the StructureModel object produced by from_pandas_dynamic extends (inherits) the networkx.DiGraph class (see here). So in principle you can extract the adjacency matrix using networkx's functionalities such as networkx.adjancency_matrix. Alternatively, you can write your own code to extract the matrix by iterating over g_learnt.nodes() or g_learnt.edges().

Hope this helps. Thanks.

from causalnex.