Imbalanced data set is serious problem in classification. It is caused by skewed distribution of data between classes. Most of standard algorithms assume or expect balanced class distribution or equal misclassification cost. Therefore, when presented with large imblanced data sets, these algorithms fail to properly represent distributive characteristics of data. The best of my knowledge, there are two approaches can be done, first on data level and second on algorithm level. On data level, most of studies applied resampling techniques to get the balanced distribution. You can undersample the majority class, oversample the minority class or both of them coincidely. Synthetic minority over sampling technique (SMOTE) is a familiar technique you can use. On algorithm level, you can apply boosting algorithm or adjust misclassification cost. Boosting algorithm is you construct strong classifier from some weak classifiers (baseline classifiers). You can use SVM and KNN as baseline classifiers.

Check out http://www.ele.uri.edu/faculty/he/PDFfiles/ImbalancedLearning.pdf

There are many different approaches. An interesting option is SMOTE.

Best Matthias

Before applying any classification techniques, you should do outliers analysis. 

Box-plot analysis is one best way to identify outliers and remove effective outliers. 

You can try to apply any of well known resampling techniques (ROS, RUS, SMOTE or ROSE). You can use also 'unbalanced' R library. It implements a lot of functions to deal with imbalanced data. For example, ubRacing method automatically selects the best technique to re-balance your specific data.

Evidently, general purpose preprocessing techiques (outlier analysis, noise detection, feature selection, etc) should also be applied.

Imbalanced data set is serious problem in classification. It is caused by skewed distribution of data between classes. Most of standard algorithms assume or expect balanced class distribution or equal misclassification cost. Therefore, when presented with large imblanced data sets, these algorithms fail to properly represent distributive characteristics of data. The best of my knowledge, there are two approaches can be done, first on data level and second on algorithm level. On data level, most of studies applied resampling techniques to get the balanced distribution. You can undersample the majority class, oversample the minority class or both of them coincidely. Synthetic minority over sampling technique (SMOTE) is a familiar technique you can use. On algorithm level, you can apply boosting algorithm or adjust misclassification cost. Boosting algorithm is you construct strong classifier from some weak classifiers (baseline classifiers). You can use SVM and KNN as baseline classifiers.

Random Forests (R package randomForest) are known to deal quite robustely with imbalanced data sets. Consider this as an alternative to SVN and KNN. You should also think about what is the right error measure for your task. Plain error rate is probably not the best choice. You might consider mean per-class error rate or another weighted error rate with specific weights for each element of the confusion matrix. randomForest offers the options classwt and cutoff, which reflect the importance given to each class during training and prediction. Both options need however careful tuning to get best results.

SMOTE technique gives desired solution for imbalanced data. I have used in my M.Sc. thesis.

I think that SVM is adequate for unbalanced data and it doesn't need preprocessing steps to achieve such classification

How much imbalanced? Considering how they work, both SVM and KNN should do well with imbalanced data, to some extent.

Better testing first? Keep in mind, however, not to use the accuracy as a metric for the performance, since the baseline model (predictions on the majority class) may give you very high percentages. F1 may be a better choice.

If some data preparation needed, you may consider over-sampling the minority clas(es) or under-sample the majority class, or, as suggested above, find some cost-sensitive classification algorithms (or implementations of those you want to use).

Regards,

Using SMOTE is a better option...

You can try my new R package "BayesRandomForest", it handles imbalance data via bootstrap prior technique.

Among preprocessing (data resampling) strategies, researchers often invest in oversampling methods since they are capable of balancing class distributions without ruling out potentially important examples. Over an extensive comparison of oversampling algorithms, the best seem to possess 3 key characteristics: cluster-based oversampling, adaptive weighting of minority examples and cleaning procedures (e.g. SMOTE, SMOTE-ENN, SMOTE-TL, MWMOTE). The following paper may be useful, as it performs a thorough empirical comparison

of well-established oversampling algorithms, focusing on their behaviour and intrinsic characteristics:

Article Cross-Validation for Imbalanced Datasets: Avoiding Overoptim...