W55699 / DEGC

Code for our paper 'Dynamically Expandable Graph Convolution for Streaming Recommendation' accepted by the Web Conference (WWW) 2023.

Dynamically Expandable Graph Convolution for Streaming Recommendation.

Note: We will continue to update and improve the code and related docs.

Environment requirement

See requirement.txt

Log file directory

-log-files sets where the results logs are saved

-load_save_path_prefix sets top-level directory for saved models

log_folder sets the sub-directory for saved models

Command examples

• Training baselines

Fine-tune: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg Finetune -de 0 -e 25 -train_mode sep -log_folder test -log test_Finetune -save_cp b0_100e -patience 2 -lr 1e-3

TWP: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg TWP -de 1 -e 25 -train_mode sep -log_folder test -log test_TWP -save_cp b0_100e -mse 100 -local_distill 1e4 -local_mode LSP_s -patience 2 -lr 1e-3

GraphSAIL: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg GraphSAIL -de 0 -e 25 -train_mode sep -log_folder test -log test_GraphSAIL -save_cp b0_100e -mse 100 -local_distill 1e4 -local_mode LSP_s -global_distill 1e4 -global_k 50,50 -global_tau 1 -patience 2 -lr 1e-3

Inverse Degree Sampling: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg Inverse_Sampling -de 2 -e 20 -train_mode sep -log_folder test -log test_Inverse_Sampling -save_cp b0_100e -rs full -union_mode snu -replay_ratio 0.2 -sampling_mode inverse_deg -patience 2 -lr 1e-3

SGCT: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg SGCT -de 3 -e 25 -train_mode sep -log_folder test -log test_SGCT -save_cp b0_100e -layer_wise 0 -contrastive_mode 'Single' -lambda_contrastive 1000,0,0 -con_positive 15 -con_ratios 2,1,2,0,0,0,0 -patience 2 -lr 1e-3

MGCT: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg MGCT -de 0 -e 25 -train_mode sep -log_folder test -log test_MGCT -save_cp b0_100e -layer_wise 0 -contrastive_mode 'Multi' -lambda_contrastive 100,0,0 -con_positive 15 -con_ratios 2,1,2,1,1,1,1 -patience 2 -lr 1e-3

LWC-KD: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg LWC_KD -de 0 -e 25 -train_mode sep -log_folder test -log test_LWC_KD -save_cp b0_100e -layer_wise 1 -contrastive_mode 'Multi' -lambda_contrastive 100,100,1000 -con_positive 15 -con_ratios 2,1,2,1,1,1,1 -patience 2 -lr 1e-3

ContinualGNN: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg ContinualGNN -de 0 -e 25 -train_mode sep -log_folder test -log test_ContinualGNN -save_cp b0_100e -mse 100 -local_distill 1e4 -local_mode LSP_s -rs full -union_mode snu -replay_ratio 0.2 -sampling_mode uniform -patience 2 -lr 1e-3 -first_segment_time 18 -last_segment_time 48

Uniform Experience Replay: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg uniform_sampling -de 0 -e 20 -train_mode sep -log_folder test -log test_uniform_sampling -save_cp b0_100e -rs full -union_mode snu -replay_ratio 0.2 -sampling_mode uniform -patience 2 -lr 1e-3

Full Batch Replay: python -u run_baselines_segments.py -d Taobao2014 -bm MGCCF -alg Full_batch -de 0 -e 20 -train_mode acc -log_folder test -log test_Full_batch -save_cp b0_100e -full_batch -patience 2 -lr 1e-3

• Training DEGC

DEGC+Finetune: python -u run_DEGC.py -d Taobao2014 -bm MGCCF -alg DEGC+Finetune -de 0 -e 25 -train_mode sep -log_folder test -log test_DEGC+Finetune -save_cp b0_100e -patience 2 -lr 1e-3

DEGC+LWC-KD: python -u run_DEGC.py -d Taobao2014 -bm MGCCF -alg DEGC+LWC_KD -de 0 -e 25 -train_mode sep -log_folder test -log test_DEGC+LWC_KD -save_cp b0_100e -layer_wise 1 -contrastive_mode 'Multi' -lambda_contrastive 100,100,1000 -con_positive 15 -con_ratios 2,1,2,1,1,1,1 -patience 2 -lr 1e-3