02_graph_kernels_solution.ipynb

{"nbformat":4,"nbformat_minor":0,"metadata":{"colab":{"provenance":[]},"kernelspec":{"name":"python3","display_name":"Python 3"},"language_info":{"name":"python"}},"cells":[{"cell_type":"markdown","source":["# Graph kernels"],"metadata":{"id":"TzJ5HxzLePkc"}},{"cell_type":"markdown","source":["## Setup"],"metadata":{"id":"cPoYFsfYebYG"}},{"cell_type":"markdown","source":["Install libraries:\n","- standard data science stack (Numpy, Pandas, Scikit-learn)\n","- graph and graph kernels libraries (NetworkX, GraKeL)\n"],"metadata":{"id":"tWD4nzvqexRR"}},{"cell_type":"code","source":["!pip install numpy pandas scikit-learn grakel"],"metadata":{"id":"I6U3WCeUeaw4"},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":["## Data loading"],"metadata":{"id":"P_P7e3WCeWoA"}},{"cell_type":"markdown","source":["We will use **PROTEINS** dataset:\n","- bioinformatics proteins dataset\n","- each graph represents a protein: enzyme or non-enzyme\n","- nodes are secondary structure elements (SSE)\n","- there is an edge if nodes are neighbors along the amino acid sequence\n","or one of three nearest neighbors in space\n","- statistics:\n","  - 1113 graphs\n","  - 2 classes\n","  - avg # nodes: 39.1\n","  - avg # edges: 72.9\n","  - discrete node labels - 3 SSE types\n","  - continous node features - various chemical properties\n","\n","We won't be using continous node features $X_n$. However, we can easily use node labels as starting labels for WL-kernel.\n","\n","Dataset hosted by TU Dortmund University at [TUDataset](https://chrsmrrs.github.io/datasets/docs/datasets/)."],"metadata":{"id":"ptQpmpa4gMCV"}},{"cell_type":"code","source":["!wget -P data -r \"https://www.chrsmrrs.com/graphkerneldatasets/PROTEINS.zip\" -O PROTEINS.zip"],"metadata":{"id":"ppSSP0r8dDR-"},"execution_count":null,"outputs":[]},{"cell_type":"code","execution_count":null,"metadata":{"id":"Qkmr_Dv9eII4"},"outputs":[],"source":["import numpy as np\n","import pandas as pd\n","\n","from grakel.datasets import fetch_dataset\n","\n","\n","dataset = fetch_dataset(\n","    data_home=\".\",  # dataset location\n","    name=\"PROTEINS\",\n","    with_classes=True,\n","    as_graphs=True,\n",")\n","graphs = dataset.data\n","y = dataset.target\n","y = y - 1  # remap {1, 2} to {0, 1}\n","\n","print(f\"Number of classes: {len(np.unique(y))}\")\n","print()\n","pd.Series(y).value_counts().plot.bar(title=\"Class distribution\")"]},{"cell_type":"markdown","source":["## Weisfeiler-Lehman kernel classification"],"metadata":{"id":"SGG3dPLbnDBu"}},{"cell_type":"markdown","source":["1. Calculate kernel matrix for training and test data\n","2. Classify with kernel SVM\n","\n","We will use code inspired by [GraKeL MUTAG dataset tutorial](https://ysig.github.io/GraKeL/0.1a8/auto_examples/shortest_path.html)."],"metadata":{"id":"Jq9F_49Vn1qp"}},{"cell_type":"code","source":["import numpy as np\n","from grakel import Graph\n","from grakel.kernels import WeisfeilerLehman\n","from sklearn.metrics import accuracy_score\n","from sklearn.model_selection import cross_val_score, train_test_split\n","from sklearn.pipeline import make_pipeline\n","from sklearn.svm import SVC\n","\n","\n","def train_wl_kernel_svm(graphs: list[Graph], y: np.ndarray) -> None:\n","    # create kernel\n","    kernel = WeisfeilerLehman(n_jobs=-1, normalize=True, n_iter=5)\n","\n","    # create SVM with precomputed kernel\n","    clf = SVC(\n","        kernel=\"precomputed\",\n","        cache_size=1024,\n","        class_weight=\"balanced\",\n","        random_state=0\n","    )\n","\n","    # make pipeline\n","    pipeline = make_pipeline(kernel, clf, verbose=True)\n","\n","    # testing with 10-fold CV\n","    scores = cross_val_score(\n","        estimator=pipeline,\n","        X=graphs,\n","        y=y,\n","        scoring=\"accuracy\",\n","        cv=10,\n","    )\n","    test_acc = np.mean(scores)\n","    print(f\"Test accuracy: {100 * test_acc:.2f}\")\n"],"metadata":{"id":"jWmQL_66PK3L"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":["train_wl_kernel_svm(graphs, y)"],"metadata":{"id":"w2xTTrYxQRGQ"},"execution_count":null,"outputs":[]},{"cell_type":"code","source":[],"metadata":{"id":"vEgtttw8kJlS"},"execution_count":null,"outputs":[]}]}