def gen_samples(x_indexes, cust_num_protol_nums, feats, labels):
'''
将原有数据记录转为Samples格式
'''
samples = Samples(x_indexes)
for feat, label in zip(feats, labels):
cust_num = feat[0]
protol_nums = cust_num_protol_nums[cust_num]
X = feat[1]
y = label[1]
sample = Sample(cust_num, protol_nums, X, y)
samples.append(sample)
return samples
def analyze_samples(trn_samples, tst_samples, n, filename):
'''
分析样本中逾期客户的风险点
Args:
trn_samples (Samples): 训练样本
tst_samples (Samples): 测试样本
n (int): 取top n个重要特征
filename (str): 结果文件
'''
x_importances = get_x_importances(trn_samples)
top_n_xs = sorted(x_importances.keys(), key=lambda item: x_importances[item])[: n] # 取top n个重要特征
x_indexes = trn_samples.get_x_indexes()
tst_pos_samples = Samples(x_indexes, [sample for sample in tst_samples if sample.get_y_pred() == cf.OVERDUE]) # 逾期客户
trn_neg_samples = Samples(x_indexes, [sample for sample in trn_samples if sample.get_y() == cf.NON_OVERDUE]) # 正常客户
trn_neg_avg_X = trn_neg_samples.get_avg_X()
trn_neg_std_X = trn_neg_samples.get_std_X()
with open(filename, 'w') as outfile:
d = {'samples': {'sample': []}}
for sample in tst_pos_samples:
s = {}
s['cust_num'] = sample.get_cust_num()
s['protol_nums'] = {'protol_num': sample.get_protol_nums()}
s['y_pred'] = sample.get_y_pred()
X = sample.get_X()
ratios = []
for name in top_n_xs:
index = x_indexes[name]
x = X[index]
avg_x = trn_neg_avg_X[index]
std_x = trn_neg_std_X[index]
ratio = abs(x - avg_x) / std_x
ratios.append([name, x, avg_x, std_x, ratio])
ratios.sort(key=lambda item: item[4], reverse=True)
s['X'] = {'x': []}
for name, x, avg_x, std_x, ratio in ratios:
s['X']['x'].append({'ratio': str(ratio), 'val': str(x), 'avg_val': str(avg_x), 'std_val': str(std_x), '@name': name.decode('utf-8')})
d['samples']['sample'].append(s)
xml = xmltodict.unparse(d, pretty=True).encode('utf-8')
outfile.write(xml)
