Modeling Knowledge Graphs with Composite Reasoning

This is the code of paper Modeling Knowledge Graphs with Composite Reasoning.

Wanyun Cui, Linqiu Zhang. AAAI 2024

1. Results

The results of CompilE_D and CompilE_N on WN18RR, FB15k237, UMLS and KINSHIP are as follows.

2. Reproduce the Results

To reproduce the above results, download the pkl files for the four datasets here.

Move those pkl files to the current directory, and run the following commands.

#################################### WN18RR ####################################
# CompilE_N
python learn.py --dataset WN18RR --model ComplEx --rank 2000 \
--optimizer Adagrad --learning_rate 1e-1 --batch_size 100 --use_N3 --use_N3_weight 0.1 \
--save_path trained_models/wn/dura \
--max_epochs 100 --valid 2 --data_path data \
--regularizer composite --mode_list "hrt-t-hrt;hrt-h-hrt" \
--n_pos 10 --w3 0.6 --fully_train

# CompilE_D
python learn.py --dataset WN18RR --model ComplEx --rank 2000 \
--optimizer Adagrad --learning_rate 1e-1 --batch_size 100 --use_DURA_W --use_DURA_W_weight 0.1 \
--save_path trained_models/wn/dura \
--max_epochs 100 --valid 2 --data_path data \
--regularizer composite --mode_list "hrt-t-hrt;hrt-h-hrt" \
--n_pos 10 --w3 0.4 --fully_train --do_ce_weight

#################################### FB237 ####################################
# CompilE_N
python learn.py --dataset FB237 --model ComplEx --rank 2000 \
--optimizer Adagrad --learning_rate 1e-1 --batch_size 100 --use_N3 --use_N3_weight 0.05 \
--save_path trained_models/umls/dura \
--max_epochs 100 --valid 2 --data_path data \
--regularizer composite --mode_list "hrt-hr-rt;hrt-rt-hr" \
--n_pos 10 --w2 0.2 --fully_train


# CompilE_D
python learn.py --dataset FB237 --model ComplEx --rank 2000 \
--optimizer Adagrad --learning_rate 1e-1 --batch_size 200 --use_DURA_W --use_DURA_W_weight 0.05 \
--save_path trained_models/umls/dura \
--max_epochs 100 --valid 2 --data_path data \
--regularizer composite --mode_list "hrt-hr-rt;hrt-rt-hr" \
--n_pos 10 --w2 0.1 --fully_train


#################################### UMLS ####################################
# CompilE_N
python learn.py --dataset umls --model ComplEx --rank 2000 \
--optimizer Adagrad --learning_rate 1e-1 --batch_size 100 --use_N3 --use_N3_weight 0.005 \
--save_path trained_models/umls/N3 \
--max_epochs 100 --valid 2 --data_path data \
--regularizer composite --mode_list "hrt-hr-rt;hrt-rt-hr" \
--n_pos 10 --w2 1.0 --fully_train

# CompilE_D
python learn.py --dataset umls --model ComplEx --rank 2000 \
--optimizer Adagrad --learning_rate 5e-2 --batch_size 100 --use_DURA_W --use_DURA_W_weight 0.001 \
--save_path trained_models/umls/dura \
--max_epochs 100 --valid 2 --data_path data \
--regularizer composite --mode_list "hrt-hr-rt;hrt-rt-hr" \
--n_pos 10 --w2 0.8 --fully_train


#################################### KINSHIP ####################################
# CompilE_N
python learn.py --dataset kinship --model ComplEx --rank 2000 \
--optimizer Adagrad --learning_rate 1e-1 --batch_size 100 --use_N3 --use_N3_weight 0.05 \
--save_path trained_models/kinship/N3 \
--max_epochs 100 --valid 2 --data_path data \
--regularizer composite --mode_list "hrt-rt-hrt;hrt-hr-hrt;hrt-t-hrt;hrt-h-hrt;hrt-hr-t;hrt-rt-h" \
--n_pos 10 --w1 0.4 --w3 0.4 --fully_train

# CompilE_D
python learn.py --dataset kinship --model ComplEx --rank 2000 \
--optimizer Adagrad --learning_rate 5e-2 --batch_size 100 --use_DURA_W --use_DURA_W_weight 0.005 \
--save_path trained_models/kinship/dura \
--max_epochs 100 --valid 2 --data_path data \
--regularizer composite --mode_list "hrt-hr-rt;hrt-rt-hrt" \
--n_pos 10 --w2 1.0 --fully_train

Acknowledgement

We refer to the code of kbc and DURA. Thanks for their contributions.