How to avoid float values in regression models

Question

I am trying to predict wine quality (ranges from 1 to 10) using regression models such as linear,SGDRegressor, ridge,lasso.

dataset:http://archive.ics.uci.edu/ml/machine-learning-databases/wine-quality/winequality-white.csv

Independent values:volatile acidity,residual sugar,free sulfur dioxide,total sulfur dioxide,alchohol Dependent:Quality

Linear model

regr = linear_model.LinearRegression(n_jobs=3)
regr.fit(x_train, y_train)
predicted = regr.predict(x_test)

predicted values for LinearRegression array([ 5.33560542, 5.47347404, 6.09337194, ..., 5.67566813, 5.43609198, 6.08189 ])

predicted values are in float instead of (1,2,3...10) I tried to round predicted values using numpy

predicted = np.round(regr.predict(x_test))` but my accuracy gone down with this attempt.

SGDRegressor model.

from sklearn import linear_model
np.random.seed(0)
clf = linear_model.SGDRegressor()
clf.fit(x_train, y_train)
redicted = np.floor(clf.predict(x_test))

predicted output values for SGDRegressor:

array([ -2.77685458e+12,   3.26826414e+12,   4.18655713e+11, ...,
     4.72375220e+12,  -7.08866307e+11,   3.95571514e+12])

Here I am unable to convert the output values into integers.

Could someone please let me know the best way to predict the wine quality using these regression models.

Have you normalised the data between 0 and 1? or sometimes depending on the regression between -1 and 1 — pbu, Jun 18 '15 at 12:50
As a part of academic assignment, we have to use both classifications(to classify wine based upon the quality) and regression models (to predict the quality of the wine) @Chung-YenHung do you think is there any alternative or am I missing any other measures? — Praneeth, Jun 18 '15 at 13:15
@pbu I haven't normalised the data. do you think normalising the data results in the output quality to integer form? — Praneeth, Jun 18 '15 at 13:20
You can reformulate the "regression" problem of wine grade prediction as a 45-way classification problem, see e.g., http://stackoverflow.com/questions/9041753/multi-class-classification-in-libsvm but that might be more expensive than living with float predictions. **But** what do you mean your prediction error went down? When your target values are integers, and you compare them against predictions that are int/float, you *must* use different success metrics! — Ahmed Fasih, Jun 18 '15 at 13:43
The SGDRegressor result seems totally worthless, why is it predicting negative and ~1e12 for wine scores?! I haven't looked at its documentation, are you sure you don't need to pass in some parameters? Or, *maybe* normalizing your input features to zero-mean-unit-variance (or something like that) will help this. — Ahmed Fasih, Jun 18 '15 at 13:47

score 4 · Accepted Answer · answered Jun 18 '15 at 14:42

You are doing a regression and therefore the output is continuous in nature.

The thing you should note is that your mini-project on predicting wine quality is not a classification problem. The response variable y, the wine quality, has intrinsic order which means a score of 6 is strictly better than a score of 5. It is NOT categorical variable where different numbers just represent different groups where groups are non-comparable.

How to avoid float values in regression models

1 Answers1