Scikit-Learn SVR Prediction Always Gives the Same Value

Question

I'm about to predict IMDB score (film rate) using Support Vector Regression in Scikit-Learn. The problem is it always gives the same prediction result for every input.

When i predict using data training, it gives various result. But when using data testing, it always gives the same value.

Data training prediction:

Data testing prediction:

Here is the link for dataset: IMDB 5000 Movie Dataset

My codes:

import matplotlib.pyplot as plt
%matplotlib inline
import pandas as pd
import numpy as np
import seaborn as sb
from sklearn import metrics as met


df = pd.read_csv("movie_metadata.csv")
df.head()


original = df.shape[0]
df = df.drop_duplicates(["movie_title"])
notDuplicated = df.shape[0]
df.reset_index(drop = True, inplace = True)
print(original, notDuplicated)


df["num_critic_for_reviews"].fillna(0, inplace = True)
df["num_critic_for_reviews"] = df["num_critic_for_reviews"].astype("int")

df["director_facebook_likes"].fillna(0, inplace = True)
df["director_facebook_likes"] = df["director_facebook_likes"].astype("int")

df["actor_3_facebook_likes"].fillna(0, inplace = True)
df["actor_3_facebook_likes"] = df["actor_3_facebook_likes"].astype(np.int64)

df["actor_2_facebook_likes"].fillna(0, inplace = True)
df["actor_2_facebook_likes"] = df["actor_2_facebook_likes"].astype(np.int64)

df["actor_1_facebook_likes"].fillna(0, inplace = True)
df["actor_1_facebook_likes"] = df["actor_1_facebook_likes"].astype(np.int64)

df["movie_facebook_likes"].fillna(0, inplace = True)
df["movie_facebook_likes"] = df["movie_facebook_likes"].astype(np.int64)

df["content_rating"].fillna("Not Rated", inplace = True)
df["content_rating"].replace('-', "Not Rated", inplace = True)
df["content_rating"] = df["content_rating"].astype("str")

df["imdb_score"].fillna(0.0, inplace = True)

df["title_year"].fillna(0, inplace = True)
df["title_year"].replace("NA", 0, inplace = True)
df["title_year"] = df["title_year"].astype("int")

df["genres"].fillna("", inplace = True)
df["genres"] = df["genres"].astype("str")


df2 = df[df["title_year"] >= 1980]
df2.reset_index(drop = True, inplace = True)

nRow = len(df2)
print("Number of data:", nRow)
nTrain = np.int64(np.floor(0.7 * nRow))
nTest = nRow - nTrain
print("Number of data training (70%):", nTrain, "\nNumber of data testing (30%):", nTest)

dataTraining = df2[0:nTrain]
dataTesting = df2[nTrain:nRow]
dataTraining.reset_index(drop = True, inplace = True)
dataTesting.reset_index(drop = True, inplace = True)


xTrain = dataTraining[["num_critic_for_reviews", "director_facebook_likes", "actor_3_facebook_likes", "actor_2_facebook_likes", "actor_1_facebook_likes", "movie_facebook_likes"]]
yTrain = dataTraining["imdb_score"]

xTest = dataTesting[["num_critic_for_reviews", "director_facebook_likes", "actor_3_facebook_likes", "actor_2_facebook_likes", "actor_1_facebook_likes", "movie_facebook_likes"]]
yTest = dataTesting["imdb_score"]

movieTitle = dataTesting["movie_title"].reset_index(drop = True)


from sklearn.svm import SVR

svrModel = SVR(kernel = "rbf", C = 1e3, gamma = 0.1, epsilon = 0.1)
svrModel.fit(xTrain,yTrain)


predicted = svrModel.predict(xTest)
[print(movieTitle[i], ":", predicted[i]) for i in range(10)]

Answer 1

Change gamma from 0.1 to 1e-8 while keeping everything else the same.

When gamma is set to 0.1, the number of unique predictions is 8 and they're all close to 6.37. When gamma is set to 1e-8, 1366 unique predictions are outputted (xTest contains 1368 total samples).

Why does gamma matter?

Intuitively, the gamma parameter defines how far the influence of a single training example reaches, with low values meaning ‘far’ and high values meaning ‘close’. The gamma parameters can be seen as the inverse of the radius of influence of samples selected by the model as support vectors.

There's a deeper explanation and example at RBF SVM Parameters.

There's also a similar explanation here: Output of Scikit SVM in multiclass classification always gives same label

Personally, I would use GridSearchCV at the bottom of your script instead. Here's an example for finding an ideal gamma an C value:

from sklearn.svm import SVR
from sklearn.model_selection import GridSearchCV

#svrModel = SVR(kernel = "rbf", C = 1e3, gamma = 1e-8, epsilon = 0.1)
#svrModel.fit(xTrain,yTrain)


#predicted = svrModel.predict(xTest)
#[print(movieTitle[i], ":", predicted[i]) for i in range(10)]

#print('Unique predictions:', np.unique(predicted))

parameters = {
    "kernel": ["rbf"],
    "C": [1,10,10,100,1000],
    "gamma": [1e-8, 1e-7, 1e-6, 1e-5, 1e-4, 1e-3, 1e-2, 1e-1]
    }

grid = GridSearchCV(SVR(), parameters, cv=5, verbose=2)
grid.fit(xTrain, yTrain)