sequenzo 0.1.24__cp311-cp311-macosx_10_9_x86_64.whl

sequenzo/big_data/clara/clara.py

@@ -0,0 +1,474 @@
+"""
+@Author  : 李欣怡
+@File    : clara.py
+@Time    : 2024/12/27 12:04
+@Desc    : 
+"""
+
+import gc
+import os
+from contextlib import redirect_stdout
+import warnings
+
+from joblib import Parallel, delayed
+from sequenzo.clustering.sequenzo_fastcluster.fastcluster import linkage
+from scipy.special import comb
+from itertools import product
+
+from sequenzo.big_data.clara.utils.aggregatecases import *
+from sequenzo.big_data.clara.utils.davies_bouldin import *
+from sequenzo.clustering.KMedoids import *
+from sequenzo.big_data.clara.utils.get_weighted_diss import *
+
+from sequenzo.define_sequence_data import SequenceData
+from sequenzo.dissimilarity_measures.get_distance_matrix import get_distance_matrix
+
+
+def adjustedRandIndex(x, y=None):
+    if isinstance(x, np.ndarray):
+        x = np.array(x)
+        y = np.array(y)
+        if len(x) != len(y):
+            raise ValueError("Arguments must be vectors of the same length")
+
+        tab = pd.crosstab(x, y)
+    else:
+        tab = x
+
+    if tab.shape == (1, 1):
+        return 1
+
+    # 计算 ARI 的四个部分：a, b, c, d
+    a = np.sum(comb(tab.to_numpy(), 2))  # 选择每对组合的组合数
+    b = np.sum(comb(np.sum(tab.to_numpy(), axis=1), 2)) - a
+    c = np.sum(comb(np.sum(tab.to_numpy(), axis=0), 2)) - a
+    d = comb(np.sum(tab.to_numpy()), 2) - a - b - c
+
+    ARI = (a - (a + b) * (a + c) / (a + b + c + d)) / ((a + b + a + c) / 2 - (a + b) * (a + c) / (a + b + c + d))
+    return ARI
+
+
+def jaccardCoef(tab):
+    if tab.shape == (1, 1):
+        return 1
+
+    # 计算交集（n11）和并集（n01 和 n10）
+    n11 = np.sum(tab.to_numpy() ** 2)  # 交集
+    n01 = np.sum(np.sum(tab.to_numpy(), axis=0) ** 2)  # 列和的平方
+    n10 = np.sum(np.sum(tab.to_numpy(), axis=1) ** 2)  # 行和的平方
+
+    return n11 / (n01 + n10 - n11)
+
+
+def clara(seqdata, R=100, kvals=None, sample_size=None, method="crisp", dist_args=None,
+          criteria=["distance"], stability=False, max_dist=None):
+
+    # ==================
+    # Parameter checking
+    # ==================
+    if kvals is None:
+        kvals = range(2, 11)
+
+    if sample_size is None:
+        sample_size = 40 + 2 * max(kvals)
+
+    print("[>] Starting generalized CLARA for sequence analysis.")
+
+    # Check for input data type (should be a sequence object)
+    if not isinstance(seqdata, SequenceData):
+        raise ValueError("[!] 'seqdata' should be SequenceData, check the input format.")
+
+    if max(kvals) > sample_size:
+        raise ValueError("[!] More clusters than the size of the sample requested.")
+
+    allmethods = ["crisp"]
+    if method.lower() not in [m.lower() for m in allmethods]:
+        raise ValueError(f"[!] Unknown method {method}. Please specify one of the following: {', '.join(allmethods)}")
+
+    if method.lower() == "representativeness" and max_dist is None:
+        raise ValueError("[!] You need to set max.dist when using representativeness method.")
+
+    allcriteria = ["distance", "db", "xb", "pbm", "ams"]
+    criteria = [c.lower() for c in criteria]
+    if not all(c in allcriteria for c in criteria):
+        raise ValueError(
+            f"[!] Unknown criteria among {', '.join(criteria)}. Please specify at least one among {', '.join(allcriteria)}.")
+
+    if dist_args is None:
+        raise ValueError("[!] You need to set the 'dist_args' for get_distance_matrix function.")
+
+    print(f"[>] Using {method} clustering optimizing the following criterion: {', '.join(criteria)}.")
+
+    # FIXME : Add coherance check between method and criteria
+
+    # ===========
+    # Aggregation
+    # ===========
+    number_seq = len(seqdata.seqdata)
+    print(f"  - Aggregating {number_seq} sequences...")
+
+    ac = DataFrameAggregator().aggregate(seqdata.seqdata)
+    agseqdata = seqdata.seqdata.iloc[ac['aggIndex'], :]
+    # agseqdata.attrs['weights'] = None
+    ac['probs'] = ac['aggWeights'] / number_seq
+    print(f"  - OK ({len(ac['aggWeights'])} unique cases).")
+
+    # Memory cleanup before parallel computation
+    gc.collect()
+    print("[>] Starting iterations...")
+
+    def calc_pam_iter(circle, agseqdata, sample_size, kvals, ac):
+        # Sampling with replacement allows the process to proceed normally
+        # even when the sample size exceeds the dataset size, as samples can be repeatedly drawn."
+        mysample = np.random.choice(len(agseqdata), size=sample_size, p=ac['probs'], replace=True)
+        mysample = pd.DataFrame({'id': mysample})
+
+        # Re-aggregate!
+        ac2 = DataFrameAggregator().aggregate(mysample)
+        data_subset = agseqdata.iloc[mysample.iloc[ac2['aggIndex'], 0], :]
+
+        with open(os.devnull, 'w') as fnull:
+            with redirect_stdout(fnull):
+                states = np.arange(1, len(seqdata.states)).tolist()
+                data_subset = SequenceData(data_subset,
+                                           time=seqdata.time,
+                                           states=states)
+                dist_args['seqdata'] = data_subset
+                diss = get_distance_matrix(opts=dist_args)
+
+        diss = diss.values
+        _diss = diss.copy()
+        _diss = get_weighted_diss(_diss, ac2['aggWeights'])
+        hc = linkage(_diss, method='ward')
+        del _diss
+
+        # For each number of clusters
+        allclust = []
+
+        for k in kvals:
+            # Weighted PAM clustering on subsample
+            # TODO : hc 已经是选好的中心点了，为什么初始化 clusterid 的时候要用 -1 呢？
+            #  因为没有必要啊，直接用原来的不好吗？尤其在没有进入 if 分支的情况下，这样处理也能避免 -1 的数据访问越界。所以为什么要初始化为-1呢？
+            clustering = KMedoids(diss=diss, k=k, cluster_only=True, initialclust=hc, weights=ac2['aggWeights'], verbose=False)
+            medoids = mysample.iloc[ac2['aggIndex'][np.unique(clustering)], :]
+            medoids = medoids.to_numpy().flatten()
+
+            del clustering
+
+            # =====================================================
+            # Compute Distances Between All Sequence to the Medoids
+            # =====================================================
+            refseq = [list(range(0, len(agseqdata))), medoids.tolist()]
+            with open(os.devnull, 'w') as fnull:
+                with redirect_stdout(fnull):
+                    states = np.arange(1, len(seqdata.states)).tolist()
+                    agseqdata = SequenceData(agseqdata,
+                                             time=seqdata.time,
+                                             states=states)
+                    dist_args['seqdata'] = agseqdata
+                    dist_args['refseq'] = refseq
+                    diss2 = get_distance_matrix(opts=dist_args)
+                    del dist_args['refseq']
+                    agseqdata = agseqdata.seqdata   # Restore scene
+
+            # Compute two minimal distances are used for silhouette width
+            # and other criterions
+            diss2 = diss2.to_numpy()
+            alphabeta = np.array([np.sort(row)[:2] for row in diss2])
+            sil = (alphabeta[:, 1] - alphabeta[:, 0]) / np.maximum(alphabeta[:, 1], alphabeta[:, 0])
+
+            # Allocate to clusters
+            memb = np.argmin(diss2, axis=1)     # Each data point is assigned to its nearest cluster
+
+            mean_diss = np.sum(alphabeta[:, 0] * ac['probs'])
+
+            warnings.filterwarnings('ignore', category=RuntimeWarning)  # The ÷0 case is ignored
+            db = davies_bouldin_internal(diss=diss2, clustering=memb, medoids=medoids, weights=ac['aggWeights'])['db']
+            warnings.resetwarnings()
+            pbm = ((1 / len(medoids)) * (np.max(diss2[medoids]) / mean_diss)) ** 2
+            ams = np.sum(sil * ac['probs'])
+
+            distmed = diss2[medoids, :]
+            distmed_flat = distmed[np.triu_indices_from(distmed, k=1)]  # Take the upper triangular part
+            minsep = np.min(distmed_flat)
+
+            xb = mean_diss / minsep
+
+            del alphabeta
+            del sil
+            del diss2
+            del distmed
+            del minsep
+
+            allclust.append({
+                'mean_diss': mean_diss,
+                'db': db,
+                'pbm': pbm,
+                'ams': ams,
+                'xb': xb,
+                'clustering': memb,
+                'medoids': medoids
+            })
+
+        del diss
+        gc.collect()
+
+        return allclust
+
+    # Compute in parallel using joblib
+    # the output example of `results`:
+    #         results[0] = all iter1's = [{k=2's}, {k=3's}, ... , {k=10's}]
+    #         results[1] = all iter2's = [{k=2's}, {k=3's}, ... , {k=10's}]
+    results = Parallel(n_jobs=-1)(
+        delayed(calc_pam_iter)(circle=i, agseqdata=agseqdata, sample_size=sample_size, kvals=kvals, ac=ac) for i in range(R))
+    # results = []
+    # for i in range(R):
+    #     res = calc_pam_iter(circle=i,
+    #                         agseqdata=agseqdata,
+    #                         sample_size=sample_size,
+    #                         kvals=kvals,
+    #                         ac=ac)
+    #     results.append(res)
+
+    print("  - Done.")
+    print("[>] Aggregating iterations for each k values...")
+
+    # aggregated output example :
+    #         data[0] = all k=2's = [{when iter1, k=2's}, {when iter2, k=2's}, ... , {when iter100, k=2's}]
+    #         data[1] = all k=3's = [{when iter1, k=3's}, {when iter2, k=3's}, ... , {when iter100, k=3's}]
+    collected_data = [[] for _ in kvals]
+    for iter_result in results:
+        k = 0
+        for item in iter_result:
+            collected_data[k].append(item)
+            k += 1
+
+    kvalscriteria = list(product(range(len(kvals)), criteria))
+    kret = []
+    for item in kvalscriteria:
+        k = item[0]
+        _criteria = item[1]
+
+        mean_all_diss = [d['mean_diss'] for d in collected_data[k]]
+        db_all = [d['db'] for d in collected_data[k]]
+        pbm_all = [d['pbm'] for d in collected_data[k]]
+        ams_all = [d['ams'] for d in collected_data[k]]
+        xb_all = [d['xb'] for d in collected_data[k]]
+        clustering_all_diss = [d['clustering'] for d in collected_data[k]]
+        med_all_diss = [d['medoids'] for d in collected_data[k]]
+
+        # Find best clustering
+        objective = {
+            "distance": mean_all_diss,
+            "pbm": pbm_all,
+            "db": db_all,
+            "ams": ams_all,
+            "xb": xb_all
+        }
+        objective = objective[_criteria]
+        best = np.argmax(objective) if _criteria in ["ams", "pbm"] else np.argmin(objective)
+
+        # Compute clustering stability of the best partition
+        if stability:
+            def process_task(j, clustering_all_diss, ac, best):
+                df = pd.DataFrame({
+                    'clustering_j': clustering_all_diss[j],        # The J-TH cluster
+                    'clustering_best': clustering_all_diss[best],  # The best-TH clustering
+                    'aggWeights': ac['aggWeights']
+                })
+                tab = df.groupby(['clustering_j', 'clustering_best'])['aggWeights'].sum().unstack(fill_value=0)
+
+                val = [adjustedRandIndex(tab), jaccardCoef(tab)]
+                return val
+
+            arilist = []
+
+            if method in ["noise", "fuzzy"]:
+                for j in range(R):
+                    val = process_task(j, clustering_all_diss, ac, best)
+                    arilist.append(val)
+            else:
+                arilist = Parallel(n_jobs=-1)(
+                    delayed(process_task)(j, clustering_all_diss, ac, best) for j in range(R))
+
+            arimatrix = np.vstack(arilist)
+            arimatrix = pd.DataFrame(arimatrix, columns=["ARI", "JC"])
+            ari08 = np.sum(arimatrix.iloc[:, 0] >= 0.8)
+            jc08 = np.sum(arimatrix.iloc[:, 1] >= 0.8)
+
+        else:
+            arimatrix = np.nan
+            ari08 = np.nan
+            jc08 = np.nan
+
+        _clustering = clustering_all_diss[best]
+
+        disagclust = np.full(seqdata.seqdata.shape[0], -1)
+        for i, index in enumerate(ac["disaggIndex"]):
+            disagclust[i] = _clustering[index] + 1      # 1-based index for clusters
+
+        evol_diss = np.maximum.accumulate(objective) if _criteria in ["ams", "pbm"] else np.minimum.accumulate(objective)
+
+        # Store the best solution and evaluations of the others
+        bestcluster = {
+            "medoids": ac["aggIndex"][med_all_diss[best]],
+            "clustering": disagclust,
+            "evol_diss": evol_diss,
+            "iter_objective": objective,
+            "objective": objective[best],
+            "iteration": best,
+            "arimatrix": arimatrix,
+            "criteria": _criteria,
+            "method": method,
+            "avg_dist": mean_all_diss[best],
+            "pbm": pbm_all[best],
+            "db": db_all[best],
+            "xb": xb_all[best],
+            "ams": ams_all[best],
+            "ari08": ari08,
+            "jc08": jc08,
+            "R": R,
+            "k": k
+        }
+
+        # Store computed cluster quality
+        kresult = {
+            "k": k+2,
+            "criteria": criteria,
+            "stats": [bestcluster["avg_dist"], bestcluster["pbm"], bestcluster["db"], bestcluster["xb"],
+                      bestcluster["ams"], bestcluster["ari08"], bestcluster["jc08"], best],
+            "bestcluster": bestcluster
+        }
+
+        kret.append(kresult)
+
+    def claraObj(kretlines, method, kvals, kret, seqdata):
+        clustering = np.full((seqdata.seqdata.shape[0], len(kvals)), -1)
+        clustering = pd.DataFrame(clustering)
+        clustering.columns = [f"Cluster {val}" for val in kvals]
+        clustering.index = seqdata.ids
+
+        ret = {
+            "kvals": kvals,
+            "clara": {},
+            "clustering": clustering,
+            "stats": np.full((len(kvals), 8), -1, dtype=float)
+        }
+
+        for i in kretlines:
+            k = kret[i]['k'] - 2    # start from 0, not 2
+            ret['stats'][k, :] = np.array(kret[i]['stats'])
+            ret['clara'][k] = kret[i]['bestcluster']
+
+            ret['clustering'].iloc[:, k] = kret[i]['bestcluster']['clustering']
+
+        ret['stats'] = pd.DataFrame(ret['stats'],
+                                    columns=["Avg dist", "PBM", "DB", "XB", "AMS", "ARI>0.8", "JC>0.8", "Best iter"])
+        ret['stats'].insert(0, "Number of Clusters", [f"Cluster {k}" for k in kvals])
+        ret['stats']["k_num"] = kvals
+
+        return ret
+
+    if len(criteria) > 1:
+        ret = {
+            'param': {
+                'criteria': criteria,
+                'pam_combine': False,
+                'all_criterias': criteria,
+                'kvals': kvals,
+                'method': method,
+                'stability': stability
+            }
+        }
+
+        for meth in criteria:
+            indices = np.where(np.array([tup[1] for tup in kvalscriteria]) == meth)[0]
+            ret[meth] = claraObj(kretlines=indices, method=method, kvals=kvals, kret=kret, seqdata=seqdata)
+
+        allstats = {}
+
+        for meth in criteria:
+            stats = pd.DataFrame(ret[meth]['stats'])
+            stats['criteria'] = meth
+
+            allstats[meth] = stats
+
+        ret['allstats'] = pd.concat(allstats.values(), ignore_index=False)
+    else:
+
+        ret = claraObj(kretlines=range(len(kvalscriteria)), method=method, kvals=kvals, kret=kret, seqdata=seqdata)
+
+    print("  - Done.")
+
+    return ret
+
+
+if __name__ == '__main__':
+    from sequenzo import *  # Social sequence analysis
+    import pandas as pd  # Import necesarry packages
+
+    # TODO : clara 返回的隶属矩阵要转置一下，因为plot_sequence_index里的参数id_group_df：cluster id 是行，id 是列
+
+    # ===============================
+    #             Sohee
+    # ===============================
+    # df = pd.read_csv('D:/college/research/QiQi/sequenzo/data_and_output/orignal data/sohee/sequence_data.csv')
+    # time_list = list(df.columns)[1:133]
+    # states = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
+    # # states = ['A', 'B', 'C', 'D', 'E', 'F', 'G']
+    # labels = ['FT+WC', 'FT+BC', 'PT+WC', 'PT+BC', 'U', 'OLF']
+    # sequence_data = SequenceData(df, time=time_list, time_type="age", states=states, labels=labels, id_col="PID")
+
+    # om.to_csv("D:/college/research/QiQi/sequenzo/files/sequenzo_Sohee_string_OM_TRATE.csv", index=True)
+
+    # ===============================
+    #             kass
+    # ===============================
+    # df = pd.read_csv('D:/college/research/QiQi/sequenzo/files/orignal data/kass/wide_civil_final_df.csv')
+    # time_list = list(df.columns)[1:]
+    # states = ['Extensive Warfare', 'Limited Violence', 'No Violence', 'Pervasive Warfare', 'Prolonged Warfare',
+    #           'Serious Violence', 'Serious Warfare', 'Sporadic Violence', 'Technological Warfare', 'Total Warfare']
+    # sequence_data = SequenceData(df, time=time_list, time_type="year", states=states, id_col="COUNTRY")
+
+    # ===============================
+    #             CO2
+    # ===============================
+    # df = pd.read_csv("D:/country_co2_emissions_missing.csv")
+    # time = list(df.columns)[1:]
+    # states = ['Very Low', 'Low', 'Middle', 'High', 'Very High']
+    # sequence_data = SequenceData(df, time_type="age", time=time, id_col="country", states=states)
+
+    # ===============================
+    #            detailed
+    # ===============================
+    # df = pd.read_csv("/Users/xinyi/Projects/sequenzo/sequenzo/data_and_output/sampled_data_sets/detailed_data/sampled_1000_data.csv")
+    # time = list(df.columns)[4:]
+    # states = ['data', 'data & intensive math', 'hardware', 'research', 'software', 'software & hardware', 'support & test']
+    # sequence_data = SequenceData(df[['worker_id', 'C1', 'C2', 'C3', 'C4', 'C5', 'C6', 'C7', 'C8', 'C9', 'C10']],
+    #                              time=time, id_col="worker_id", states=states)
+
+    # ===============================
+    #             broad
+    # ===============================
+    # df = pd.read_csv("D:/college/research/QiQi/sequenzo/data_and_output/sampled_data_sets/broad_data/sampled_1000_data.csv")
+    # time = list(df.columns)[4:]
+    # states = ['Non-computing', 'Non-technical computing', 'Technical computing']
+    # sequence_data = SequenceData(df[['worker_id', 'C1', 'C2', 'C3', 'C4', 'C5']],
+    #                              time_type="age", time=time, id_col="worker_id", states=states)
+
+    df = pd.read_csv("/Users/xinyi/Projects/sequenzo/sequenzo/data_and_output/orignal data/not_real_detailed_data/synthetic_detailed_U5_N1000.csv")
+    _time = list(df.columns)[2:]
+    states = ["Data", "Data science", "Hardware", "Research", "Software", "Support & test", "Systems & infrastructure"]
+    df = df[['id', '1', '2', '3', '4', '5', '6', '7', '8', '9', '10']]
+    sequence_data = SequenceData(df, time=_time, id_col="id", states=states)
+
+    result = clara(sequence_data,
+                   R=250,
+                   sample_size=500,
+                   kvals=range(2, 6),
+                   criteria=['distance'],
+                   dist_args={"method": "OM", "sm": "CONSTANT", "indel": 1},
+                   stability=True)
+
+    # print(result)
+    print(result['stats'])

sequenzo/big_data/clara/utils/__init__.py

@@ -0,0 +1,27 @@
+"""
+@Author  : 李欣怡
+@File    : __init__.py.py
+@Time    : 2025/2/28 00:30
+@Desc    : 
+"""
+from .aggregatecases import *
+from .davies_bouldin import *
+from .wfcmdd import *
+from sequenzo.clustering.KMedoids import KMedoids
+
+
+def _import_c_code():
+    """Lazily import the c_code module to avoid circular dependencies during installation"""
+    try:
+        from sequenzo.clustering import clustering_c_code
+        return clustering_c_code
+    except ImportError:
+        # If the C extension cannot be imported, return None
+        print(
+            "Warning: The C++ extension (c_code) could not be imported. Please ensure the extension module is compiled correctly.")
+        return None
+
+
+__all__ = [
+    'KMedoids'
+]

sequenzo/big_data/clara/utils/aggregatecases.py

@@ -0,0 +1,92 @@
+"""
+@Author  : 李欣怡
+@File    : aggregatecases.py
+@Time    : 2024/12/27 10:12
+@Desc    : 
+"""
+import pandas as pd
+import numpy as np
+
+
+class WcAggregateCases:
+    def aggregate(self, x, weights=None, **kwargs):
+        """
+        The appropriate aggregation method is invoked dynamically depending on the type of x
+        """
+        method_name = f"aggregate_{type(x).__name__}"
+        method = getattr(self, method_name, None)
+
+        if method is None:
+            raise NotImplementedError(f"No aggregation method for type {type(x).__name__}")
+
+        return method(x, weights, **kwargs)
+
+
+class WcAggregateCasesInternal:
+    def aggregate(self, x, weights=None):
+        x = pd.DataFrame(x)
+        lx = len(x)
+
+        if weights is None:
+            weights = np.ones(lx)
+
+        ids = x.apply(lambda row: "@@@WC_SEP@@".join(row.astype(str)), axis=1)
+
+        mcorr = [np.nan] * lx
+
+        def _compute_weight_each_group_and_sum(group):
+            first_element = group.iloc[0]
+
+            for idx in group:
+                mcorr[idx] = first_element
+            weighted_sum = np.sum(weights[group])
+            return [first_element, weighted_sum]
+
+        df = pd.DataFrame({
+            'index': range(0, lx),
+            'id': ids
+        })
+
+        grouped = df.groupby('id')['index'].apply(_compute_weight_each_group_and_sum)
+
+        agg_df = pd.DataFrame(grouped.tolist(), columns=['aggIndex', 'aggWeights'])
+
+        aggIndex = agg_df['aggIndex']
+        mcorr2 = [aggIndex[aggIndex == val].index[0] if val in aggIndex.values else -1 for val in mcorr]
+
+        ret = {
+            "aggIndex": agg_df['aggIndex'].values,
+            "aggWeights": agg_df['aggWeights'].values,
+            "disaggIndex": mcorr2,
+            "disaggWeights": weights
+        }
+
+        return ret
+
+
+class DataFrameAggregator(WcAggregateCases):
+    def aggregate_DataFrame(self, x, weights=None, **kwargs):
+        internal = WcAggregateCasesInternal()
+        return internal.aggregate(x, weights)
+
+
+class MatrixAggregator(WcAggregateCases):
+    def aggregate_ndarray(self, x, weights=None, **kwargs):
+        internal = WcAggregateCasesInternal()
+        return internal.aggregate(x, weights)
+
+
+class StsListAggregator(WcAggregateCases):
+    def aggregate_stslist(self, x, weights=None, weighted=True, **kwargs):
+        if weights is None and weighted:
+            weights = getattr(x, "weights", None)
+        internal = WcAggregateCasesInternal()
+        return internal.aggregate(x, weights)
+
+
+# Print function (for output)
+def print_wcAggregateCases(result):
+    print(f"Number of disaggregated cases: {len(result['disaggWeights'])}")
+    print(f"Number of aggregated cases: {len(result['aggWeights'])}")
+    print(f"Average aggregated cases: {len(result['disaggWeights']) / len(result['aggWeights'])}")
+    print(f"Average (weighted) aggregation: {np.mean(result['aggWeights'])}")

sequenzo/big_data/clara/utils/davies_bouldin.py

@@ -0,0 +1,91 @@
+"""
+@Author  : 李欣怡
+@File    : davies_bouldin.py
+@Time    : 2024/12/27 17:56
+@Desc    :
+    :param
+        diss : numpy 2D, 距离矩阵
+        clustering : numpy 1D, 每个数据点的隶属矩阵（一种可能：初始化时每个点构成一个簇，则每个数据点隶属自己）
+        medoids : numpy 1D, 簇的中心点
+"""
+import numpy as np
+
+
+def davies_bouldin_internal(diss, clustering, medoids, p=1, weights=None, medoidclust=False):
+    # If weights are not provided, use uniform weights
+    if weights is None:
+        weights = np.ones(diss.shape[0])
+
+    list_diam = np.zeros(len(medoids))
+
+    # Calculate the diameter for each medoid
+    for i in range(len(medoids)):
+        medi = medoids[i] if medoidclust else i
+        cond = (clustering == medi)
+
+        # Calculate the diameter (weighted distance)
+        list_diam[i] = (np.sum(weights[cond] * diss[cond, i] ** p) / np.sum(weights[cond])) ** (1 / p)
+
+    maximum = np.zeros(len(medoids))
+
+    # Calculate the maximum ratio for each medoid
+    for i in range(len(medoids)):
+        # Calculate the distance to other medoids
+        maximum2 = (list_diam[i] + list_diam) / diss[medoids[i], :]
+
+        # Take the maximum of the valid (finite) values
+        # ensure values for "same" medoids
+        maximum[i] = np.max(maximum2[np.isfinite(maximum2)])
+
+    # Calculate the final Davies-Bouldin index (average of maximum values)
+    final_db = np.mean(maximum)
+
+    return {'db': final_db, 'per_cluster': maximum}
+
+
+def fuzzy_davies_bouldin_internal(diss, memb, medoids, weights=None):
+    if weights is None:
+        weights = np.ones(diss.shape[0])
+
+    # R 中定义后未使用，用另一个值赋值了
+    list_diam = np.zeros(len(medoids))
+
+    # R 中只定义未使用
+    # n = np.sum(weights)
+
+    mw = memb * weights[:, None]
+    list_diam = np.sum(mw * diss, axis=0) / np.sum(mw, axis=0)
+
+    # 初始化一个数组来存储每个簇的最大值
+    maximum = np.zeros(len(medoids))
+
+    # 对每个簇计算其与其他簇的相似度
+    for i in range(len(medoids)):
+        maximum2 = (list_diam[i] + list_diam) / diss[medoids[i], :]
+
+        maximum[i] = np.max(maximum2[np.isfinite(maximum2)])
+
+    final_db = np.mean(maximum)
+
+    return {'db': final_db, 'per_cluster': maximum}
+
+
+def adjpbm_internal(diss, clustering, medoids, p=1, weights=None, medoidclust=False):
+    if weights is None:
+        weights = np.ones(diss.shape[0])
+
+    # Calculate internal distance
+    internaldist = [
+        (sum(weights[clustering == (medoids[i] if medoidclust else i)] * diss[
+            clustering == (medoids[i] if medoidclust else i), i] ** p) /
+         sum(weights[clustering == (medoids[i] if medoidclust else i)])) ** (1 / p)
+        for i in range(len(medoids))
+    ]
+
+    # Calculate the minimum separation distance between medoids
+    separation = np.nanmin(diss[medoids, :][:, medoids])
+
+    # Calculate pbm (probabilistic cluster separation)
+    pbm = (1 / len(medoids)) * (separation / np.sum(internaldist))
+
+    return pbm