How to get different results running sklearn's MeanShift in a single program? (Python3)

Question

I ran into a quirk with sklearn's MeanShift that I don't know how to get around.

MeanShift doesn't predictably give the same results on every run, so I wanted to run it multiple times within one program, but the results are the same on every iteration. They are different across program runs, but the same over every loop.

Minimal code sans convert_non_numerical_data(...).

import numpy as np
import pandas as pd
from sklearn.cluster import MeanShift
from sklearn import preprocessing

# EDIT: added on request 
def convert_non_numerical_data_to_numerical(data_frame):
    columns = data_frame.columns.values

    for col in columns:
        text_digit_vals = {}

        if data_frame[col].dtype in [np.int64, np.float64]:
            continue

        x = 0
        for unique in set(data_frame[col].values):
            if unique not in text_digit_vals:
                text_digit_vals[unique] = x
                x += 1

        def convert_to_int(val):
            return text_digit_vals[val]
        data_frame[col] = list(map(convert_to_int, data_frame[col]))

df = pd.read_csv("test_data/titanic.csv")
df.fillna(0, inplace=True)  # bad idea to shove in 0?
convert_non_numerical_data_to_numerical(df)

y = np.array(df["survived"])
X = np.array(df.drop(["survived"], 1))
X = preprocessing.scale(X)

run_data = []
for run_count in range(3):
    clf = MeanShift()
    clf.fit(X)
    run_data.append(clf.labels_.copy())
    clf = None

run_0 = run_data[0]
for other_run in run_data[1:]:
    if np.array_equiv(run_0, other_run):
        print("same")
    else:
        print("different")

Output:

same
same

Any idea what is causing this, or at least how to get around it?

Note:

• Data set from pythonprogramming.net

Answer 1

Have you looked at smoothing your dataset with any other methods? What hyper-parameter tuning have you done? Also how did you convert your non_numerical data, just wondering if you can provide pastbin?

Answer 2

A couple of work arounds: 1- Set bin_seeding to True, and/or using random seeds at each iteration. 2- Shuffling the data at each iteration.