1{
2  "op": "replace",
3  "filename": "scikit-learn/custom_clustering.py",
4  "start_line": 36,
5  "end_line": 109,
6  "content": "# ================================================================\n# EDITABLE -- agent modifies this section (lines 36 to 109)\n# ================================================================\n\n\nclass CustomClustering(BaseEstimator, ClusterMixin):\n    \"\"\"Multi-Scale Density-Adaptive Spectral Clustering (MSDASC).\n\n    Extends self-tuning spectral clustering (Zelnik-Manor & Perona, 2004)\n    with a novel multi-scale kernel bandwidth averaging scheme. Each point's\n    affinity kernel adapts to local density at multiple neighborhood scales,\n    and the resulting affinities are averaged to capture both fine-grained\n    and coarse cluster structure robustly across diverse geometries.\n    \"\"\"\n    def __init__(self, n_clusters=None, random_state=42):\n        self.n_clusters = n_clusters\n        self.random_state = random_state\n        self.labels_ = None\n\n    def fit(self, X):\n        from sklearn.decomposition import PCA\n        from sklearn.cluster import KMeans\n        from sklearn.neighbors import NearestNeighbors\n        from scipy.sparse.linalg import eigsh\n        from scipy.sparse import csr_matrix\n        n, d = X.shape\n        k = self.n_clusters if self.n_clusters is not None else 8\n        # Adaptive PCA for high-dimensional data\n        Xw = X\n        if d > 15:\n            nc = min(n - 1, d, max(k * 5, 50))\n            Xw = PCA(n_components=nc, random_state=self.random_state).fit_transform(X)\n        # Build k-NN graph\n        nn_k = min(max(15, int(np.sqrt(n))), n - 1)\n        dists, idxs = NearestNeighbors(n_neighbors=nn_k + 1).fit(Xw).kneighbors(Xw)\n        ri = np.repeat(np.arange(n), nn_k)\n        ci = idxs[:, 1:].ravel()\n        di = dists[:, 1:].ravel()\n        # Novel: Multi-scale self-tuning Gaussian kernel averaging\n        # Average affinities over multiple local bandwidth scales to\n        # robustly capture structure at different resolutions\n        scales = [s for s in [3, 5, 7, 10, 15, 20] if s < dists.shape[1]]\n        Wsum = None\n        for s in scales:\n            sig = np.maximum(dists[:, s], 1e-10)\n            w = np.exp(-di**2 / (np.repeat(sig, nn_k) * sig[ci] + 1e-10))\n            Ws = csr_matrix((w, (ri, ci)), shape=(n, n))\n            Ws = (Ws + Ws.T) / 2.0\n            Wsum = Ws if Wsum is None else Wsum + Ws\n        W = Wsum * (1.0 / len(scales))\n        # Normalized spectral decomposition (Ng-Jordan-Weiss)\n        deg = np.array(W.sum(axis=1)).ravel()\n        dinv = np.where(deg > 0, 1.0 / np.sqrt(deg), 0.0)\n        Dsp = csr_matrix((dinv, (range(n), range(n))), shape=(n, n))\n        Lsym = Dsp @ W @ Dsp\n        try:\n            _, V = eigsh(Lsym, k=k, which='LM')\n        except Exception:\n            e, Va = np.linalg.eigh(Lsym.toarray())\n            V = Va[:, np.argsort(-e)[:k]]\n        V = V / np.maximum(np.linalg.norm(V, axis=1, keepdims=True), 1e-10)\n        self.labels_ = KMeans(\n            n_clusters=k, random_state=self.random_state, n_init=30, max_iter=500\n        ).fit_predict(V)\n        return self\n\n    def predict(self, X):\n        if self.labels_ is None:\n            self.fit(X)\n        return self.labels_\n\n\ndef custom_distance(x, y):\n    return np.sqrt(np.sum((x - y) ** 2))\n\n"
7}

1{
2  "op": "replace",
3  "filename": "scikit-learn/custom_clustering.py",
4  "start_line": 36,
5  "end_line": 110,
6  "content": "# ================================================================\n# EDITABLE -- agent modifies this section (lines 36 to 109)\n# ================================================================\n\n\nclass CustomClustering(BaseEstimator, ClusterMixin):\n    \"\"\"SNN-Weighted Multi-Scale Spectral Clustering (SNN-MSSC).\n\n    Novel algorithm that combines three complementary affinity signals:\n    1. Self-tuning Gaussian kernel (Zelnik-Manor & Perona): local density\n       adaptation via per-point bandwidth sigma_i = d(x_i, nn_k(x_i))\n    2. Multi-scale averaging: affinities from multiple bandwidth scales\n       are averaged to robustly capture structure at different resolutions\n    3. Shared Nearest Neighbor (SNN) weighting: edges are weighted by the\n       fraction of shared k-nearest neighbors, providing topology-aware\n       affinity that is robust in high-dimensional spaces\n\n    The product W = SNN_weight * mean_scale(Gaussian) ensures that only\n    pairs that are both geometrically close AND topologically connected\n    receive high affinity, improving cluster separation.\n    \"\"\"\n    def __init__(self, n_clusters=None, random_state=42):\n        self.n_clusters = n_clusters\n        self.random_state = random_state\n        self.labels_ = None\n\n    def fit(self, X):\n        from sklearn.decomposition import PCA\n        from sklearn.cluster import KMeans\n        from sklearn.neighbors import NearestNeighbors\n        from scipy.sparse.linalg import eigsh\n        from scipy.sparse import csr_matrix\n        n, d = X.shape\n        k = self.n_clusters if self.n_clusters is not None else 8\n        # Adaptive PCA for high-dimensional data (reduce noise, not structure)\n        Xw = X\n        if d > 15:\n            nc = min(n - 1, d, max(k * 4, 40))\n            Xw = PCA(n_components=nc, random_state=self.random_state).fit_transform(X)\n        # Build k-NN graph\n        nn_k = min(max(15, int(np.sqrt(n))), n - 1)\n        dists, idxs = NearestNeighbors(n_neighbors=nn_k + 1).fit(Xw).kneighbors(Xw)\n        ri = np.repeat(np.arange(n), nn_k)\n        ci = idxs[:, 1:].ravel()\n        di = dists[:, 1:].ravel()\n        # SNN: count shared neighbors via sparse matrix product\n        B = csr_matrix((np.ones(len(ri)), (ri, ci)), shape=(n, n))\n        SNN = B @ B.T\n        snn_raw = np.array(SNN[ri, ci]).ravel()\n        snn_w = (snn_raw + 1.0) / (nn_k + 1.0)  # Smoothed, in (0, 1]\n        # Multi-scale self-tuning Gaussian kernel, weighted by SNN\n        scales = [s for s in [3, 5, 7, 10, 15, 20] if s < dists.shape[1]]\n        Wsum = None\n        for s in scales:\n            sig = np.maximum(dists[:, s], 1e-10)\n            g = np.exp(-di**2 / (np.repeat(sig, nn_k) * sig[ci] + 1e-10))\n            w = snn_w * g  # SNN-weighted Gaussian affinity\n            Ws = csr_matrix((w, (ri, ci)), shape=(n, n))\n            Ws = (Ws + Ws.T) / 2.0\n            Wsum = Ws if Wsum is None else Wsum + Ws\n        W = Wsum * (1.0 / len(scales))\n        # Normalized spectral decomposition (Ng-Jordan-Weiss)\n        deg = np.array(W.sum(axis=1)).ravel()\n        dinv = np.where(deg > 0, 1.0 / np.sqrt(deg), 0.0)\n        Dsp = csr_matrix((dinv, (range(n), range(n))), shape=(n, n))\n        Lsym = Dsp @ W @ Dsp\n        try:\n            _, V = eigsh(Lsym, k=k, which='LM')\n        except Exception:\n            e, Va = np.linalg.eigh(Lsym.toarray())\n            V = Va[:, np.argsort(-e)[:k]]\n        V = V / np.maximum(np.linalg.norm(V, axis=1, keepdims=True), 1e-10)\n        self.labels_ = KMeans(\n            n_clusters=k, random_state=self.random_state, n_init=50, max_iter=500\n        ).fit_predict(V)\n        return self\n\n    def predict(self, X):\n        if self.labels_ is None:\n            self.fit(X)\n        return self.labels_\n\n\ndef custom_distance(x, y):\n    return np.sqrt(np.sum((x - y) ** 2))\n\n"
7}

1{
2  "n": 2
3}