Towards Privacy-Aware Causal Structure Learning in Federated Setting

Usage

"FedPC_discrete.m" (for discrete datasets) and "FedPC_continues" (for continuous datasets) are two main scripts.

Note that the current code has only been debugged on Matlab (2018a) with a 64-bit Windows system.

'layer_dis.m'(for discrete datasets) and 'layer_con.m'(for continues datasets) is implemented for the layer-wise strategy.

In 'layer_dis.m' and 'layer_con.m', the inputs and outputs are the same, shown below:

function [DAG,sep] = layer_dis(data,alpha,G,sep,ord) / function [DAG,sep] = layer_con(data,alpha,G,sep,ord)

INPUT:

data is the data matrix

alpha is the significance level

G is the learned skeleton; for the first time, a fully connected matrix

sep is the separation set

ord is the length of the separation set

OUTPUT:

DAG is incomplete DAG under the length of separation set

sep is the separation set identified

'orient_dis.m'(for discrete datasets) and 'orient_con.m' (for continuous datasets) are to orient edges.

In 'orient_dis.m' and 'orient_con.m', the inputs and outputs are the same, shown below:

function [DAG] = orient_dis(G,dataset,clients) / function [DAG] = orient_con(G,dataset,clients)

INPUT:

G is the skeleton of DAG

dataset is the data 

clients is the number of clients

OUTPUT:

DAG is the output after orienting

'random_partition.m' is to generate the random number of clients across clients.

function [partition] = random_partition(data_samples,clients)

INPUT:

data_samples is the number of all data samples 

clients is the number of clients

OUTPUT:

partition is the distribution of data samples on clients 

'layer_federate.m' is the process federating parameters at the server.

function [fed_ske,fed_sep] = layer_federate(ske_set,sep_set,clients,ratio)

INPUT:

ske_set is the data matrix set learning from clients

sep_set is the separation set learning from clients

clients is the number of clients

ratio is the voting ratio to determine the edges

OUTPUT:

fed_ske is DAG after voting

fed_sep is the union of nodes in separation sets 

'max_p.m' and 'min_r.m' is to find out the separation set with max p-value.

function [maxsep] = max_p(x,y,candidate,dataset,clients)

INPUT:

x,y is 2 nodes

candidate is candidates for separation set 

dataset is data across clients

clients is the number of clients

OUTPUT:

maxsep is the separation set between 2 nodes with max p-value

function [minsep] = min_r(x,y,candidate,dataset,clients)

INPUT:

x,y is 2 nodes

candidate is candidates for separation set 

dataset is data across clients

clients is the number of clients

OUTPUT:

minsep is the seperation set between 2 nodes with min r-value

References

[1] Huang J, Guo X, Yu K, et al. Towards Privacy-Aware Causal Structure Learning in Federated Setting[J]. IEEE Transactions on Big Data, 2023.