Forecasting the future of artificial intelligence

Forecasting the future of artificial intelligence with machine learning-based link prediction in an exponentially growing knowledge network
Mario Krenn, Lorenzo Buffoni, Bruno Coutinho, Sagi Eppel, Jacob Gates Foster, Andrew Gritsevskiy, Harlin Lee, Yichao Lu, Joao P. Moutinho, Nima Sanjabi, Rishi Sonthalia, Ngoc Mai Tran, Francisco Valente, Yangxinyu Xie, Rose Yu, Michael Kopp

This page collects results for an AI benchmark for link prediction in exponentially growing knowledge networks. It is a follow-up of the IEEE BigData Science4Cast competition 2021, described in our Nature Machine Intelligence paper.

Files

• create_data.py: A simple python file for creating the datasets SemanticGraph_delta_N_cutoff_M_minedge_P.pkl from the full semantic network all_edges.pkl.
• evaluate_model.py: Runs my simple baseline model on all datasets
• simple_model.py: My baseline, containing 15 predefined properties that are computed for each unconnected pair. Same model as in the competition.
• utils.py: Contains useful functions, such as the creation of datasets from the full semantic network (unbiased for test-set, and biased [i.e. same number of positive and negative solutions] for training if desired), and AUC computation.

Datasets can be downloaded via zenodo.org (file names: SemanticGraph_delta_N_cutoff_M_minedge_P.pkl).

Running the code

1. Create the required environment (see environmen.yaml and pyproject.toml). Requirements: pytorch, numpy, networkx, scipy, matplotlib.
2. Download the dataset files from zenodo.org and store them in the same directory as the python files.
3. Run python evaluate_model.py. This runs the baseline model M6 on all 18 datasets. Each dataset takes roughly 2h on a standard notebook (i.e. a total of 36 hours).
4. The python code generates a log file called logs_SemanticGraph_delta_M_cutoff_N_minedge_K.pkl.txt (M,N,K are numbers). One example (for M=1, N=25, K=1) of such an expected output file is stored here. Similar files will be generated for all other 17 datasets.

Degree Evolution

Result of Yichao's Model (M1)

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9330	0.9252	0.9248
delta=3	0.9172	0.9191	0.9096
delta=5	0.8960	0.8987	0.8935

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9926	0.9945	0.9982
delta=3	0.9853	0.9965	0.9949
delta=5	0.9793	0.9893	0.9990

Result of Team HashBrown's Model (M2)

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9147	0.9175	0.9156
delta=3	0.8953	0.8977	0.8949
delta=5	0.8610	0.8645	0.8630

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9900	0.9876	0.9944
delta=3	0.9786	0.9861	0.9867
delta=5	0.9595	0.9689	0.9692

Result of Nima Sanjabi's Model (M3)

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.8979	0.8980	0.9010
delta=3	0.8830	0.8823	0.8823
delta=5	0.8489	0.8433	0.8409

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9496	0.9687	0.9481
delta=3	0.9652	0.9765	0.9788
delta=5	0.9480	0.9538	0.9488

Results of Bacalhink Team (M4)

Preferential Attachment (M4A)

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.8838	0.8862	0.8836
delta=3	0.8695	0.8673	0.8628
delta=5	0.8422	0.8359	0.8300

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9754	0.9649	0.9789
delta=3	0.9590	0.9620	0.9646
delta=5	0.9380	0.9442	0.9386

Common Neighbours (M4B)

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.8942	0.9016	0.9009
delta=3	0.8476	0.8761	0.8783
delta=5	0.7677	0.8266	0.8345

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9369	0.9771	0.9889
delta=3	0.9247	0.9760	0.9786
delta=5	0.8658	0.9520	0.9526

Result of Francisco Valente's Model (M5)

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1
delta=3	0.8467	0.8490	0.8335
delta=5	0.7897	0.8023	0.8004

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1			0.9819
delta=3	0.9420	0.9562	0.9461
delta=5	0.8914	0.9262	0.9150

Result of MK's Baseline (M6)

• Prediction from Year (2021-delta,2021), with delta=[1,3,5]
• Minimal Vertex Degree: cutoff=[0,5,25]
• Prediction from unconnceted to edge_weight=[1,3] edges

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.8520	0.8526	0.8512
delta=3	0.8411	0.8379	0.8317
delta=5	0.8201	0.8093	0.8045

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9473	0.9317	0.9490
delta=3	0.9408	0.9465	0.9296
delta=5	0.9055	0.9160	0.9030

Resuts of Andrew Gritsevskiy -- Feature embedding method (M7)

Node2Vec (M7A)

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.8768	0.8558	0.8467
delta=3	0.8361	0.5039	0.5127
delta=5	0.8755	0.6106	0.6026

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9258	0.9624	0.9891
delta=3	0.8648	0.5025	0.5402
delta=5	0.8573	0.6133	0.6423

ProNE (M7B)

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.8354	0.8538	0.7375
delta=3	0.8210	0.7043	0.7763
delta=5	0.7383	0.7063	0.6872

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9952	0.9898	0.9989
delta=3	0.8844	0.9817	0.9862
delta=5	0.8586	0.8609	0.8251

Result of Team Harlin -- Transformer Method (M8)

Area under the Curve (AUC) for prediction of new edge_weights of 1

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.8232	0.8253	0.8321
delta=3	0.7418	0.7659	0.7435
delta=5	0.6980	0.7023	0.6743

Area under the Curve (AUC) for prediction of new edge_weights of 3

	cutoff=0	cutoff=5	cutoff=25
delta=1	0.9407	0.9373	0.9636
delta=3	0.8518	0.8804	0.8754
delta=5	0.7365	0.7977	0.7467