Overlapping Spaces for Compact

Graph Representations: Code

Requirements:

TF 2.0, 512GB Memory (in fact, we did not measure memory consumption, but it was trained on a server with 512GB :) ).

Section to notebook mapping:

Our new WLA6 dataset

• /datasets/wla6.edges and /datasets/wla6.nodes

5.1 distortion experiment

• /new_graph_exp/more_graph_exp_csphds.ipynb
• /new_graph_exp/more_graph_exp_facebook.ipynb
• /new_graph_exp/more_graph_exp.ipynb (EuCore dataset)
• /new_graph_exp/more_graph_exp_power.ipynb
• /new_graph_exp/more_graph_exp_ucsa312.ipynb
• /new_graph_exp/more_graph_exp_wips.ipynb (all wips exps)
• /new_graph_exp/more_graph_exp_wla6.ipynb
• /hdssm_story/offline_distortion_all_sphere_fix.ipynb (old experiments without all metrics, but in single file)

5.1 map experiment

• /new_graph_exp/more_graph_exp_csphds.ipynb
• /new_graph_exp/more_graph_exp_facebook.ipynb
• /new_graph_exp/more_graph_exp.ipynb (EuCore dataset)
• /new_graph_exp/more_graph_exp_power.ipynb
• /new_graph_exp/more_graph_exp_ucsa312.ipynb
• /new_graph_exp/more_graph_exp_wips.ipynb (all wips exps)
• /new_graph_exp/more_graph_exp_wla6.ipynb
• /hdssm_story/offline_ranking_all_sphere_fix.ipynb (old experiments without all metrics, but in single file)

5.2 dssm tiny experiment

• /new_graph_exp/dssm-exps-tiny-exdot.ipynb
• /hdssm_story/dssm-exps-tiny-more-metrics.ipynb

5.2 dssm large experiment

• /new_graph_exp/dssm-exps-large-exdot.ipynb
• /hdssm_story/dssm-exps-large.ipynb

(there is no data for dssm exp due to privacy)

5.3 synthetic bipartite graph

• /hdssm_story/strange_graph_exp.ipynb
• /hdssm_story/strange_graph_exp_exdot_and_hs.ipynb
• /hdssm_story/strange_graph_exp_wips.ipynb

graph itself

• /datasets/bg_20_700_0.05.edges

Additional:

different conversion from distance to probability

• /hdssm_story/offline_ranking_1_div_(no_exp)_dist_exp.ipynb
• /hdssm_story/offline_ranking_1_div_dist_exp.ipynb

weights analysis

• /hdssm_story/reviews_osweights_distortion.ipynb

overlapping spaces' depth analyis

• /hdssm_story/reviews_ttl_distortion.ipynb (t = 2)
• /hdssm_story/reviews_ttl_distortion-ttl3.ipynb (t = 3)
• /hdssm_story/reviews_ttl_distortion-ttl3_sq.ipynb (t = 3)

Some hints:

• S10 = spherical space with simple parametrization
• hS10 = spherical space with hyperspherical parametrization
• tighten spaces = overlaying spaces (early name)
• triple trainable = universal signature (4.1, also early name)
• triple_trainable_lx = overlaying space $$O_{l1}, (t = x)$$
• triple_trainable_lx_sq = overlaying space $$O_{l2}, (t = x)$$
• triple_trainable_lx_sq_ex = OS-Mixed $$O_{mix-l2}, (t = x)$$