⚠️ This package has been created years ago and has never been modernized. Its usage is restricted to concrete types (e.g.
Vector{Float64}
instead ofAbstractVector{<:Real}
). The API is inconsistent and sometimes confusing. MLJ.jl wraps NaiveBayes.jl, fixing some of these issues, but ghosts of the past still show up. You have been warned!
Naive Bayes classifier. Currently 3 types of NB are supported:
examples/nums.jl
for usage.examples/iris.jl
for usage.Dict{Symbol/AstractString, Vector}
rather than a Matrix
. Also, discrete features must be integers while continuous features must be floats. If all features are continuous Matrix
input is supported.Since GaussianNB
models multivariate distribution, it's not really a "naive" classifier (i.e. no independence assumption is made), so the name may change in the future.
As a subproduct, this package also provides a DataStats
type that may be used for incremental calculation of common data statistics such as mean and covariance matrix. See test/datastatstest.jl
for a usage example.
Continuous and discrete features as Dict{Symbol, Vector}}
f_c1 = randn(10)
f_c2 = randn(10)
f_d1 = rand(1:5, 10)
f_d2 = rand(3:7, 10)
training_features_continuous = Dict{Symbol, Vector{Float64}}(:c1=>f_c1, :c2=>f_c2)
training_features_discrete = Dict{Symbol, Vector{Int}}(:d1=>f_d1, :d2=>f_d2) #discrete features as Int64
labels = rand(1:3, 10)
hybrid_model = HybridNB(labels)
# train the model
fit(hybrid_model, training_features_continuous, training_features_discrete, labels)
# predict the classification for new events (points): features_c, features_d
features_c = Dict{Symbol, Vector{Float64}}(:c1=>randn(10), :c2=>randn(10))
features_d = Dict{Symbol, Vector{Int}}(:d1=>rand(1:5, 10), :d2=>rand(3:7, 10))
y = predict(hybrid_model, features_c, features_d)
Continuous features only as a Matrix
X_train = randn(3,400);
X_classify = randn(3,10)
hybrid_model = HybridNB(labels) # the number of discrete features is 0 so it's not needed
fit(hybrid_model, X_train, labels)
y = predict(hybrid_model, X_classify)
Continuous and discrete features as a Matrix{Float}
#X is a matrix of features
# the first 3 rows are continuous
training_features_continuous = restructure_matrix(X[1:3, :])
# the last 2 rows are discrete and must be integers
training_features_discrete = map(Int, restructure_matrix(X[4:5, :]))
# train the model
hybrid_model = train(HybridNB, training_features_continuous, training_features_discrete, labels)
# predict the classification for new events (points): features_c, features_d
y = predict(hybrid_model, features_c, features_d)
It is useful to train a model once and then use it for prediction many times later. For example, train your classifier on a local machine and then use it on a cluster to classify points in parallel.
There is support for writing HybridNB
models to HDF5 files via the methods write_model
and load_model
. This is useful for interacting with other programs/languages. If the model file is going to be read only in Julia it is easier to use JLD.jl for saving and loading the file.