Intel® oneAPI Data Analytics Library Developer Guide and Reference
A newer version of this document is available. Customers should click here to go to the newest version.
k-Nearest Neighbors Classification and Search (k-NN)
k-NN classification and search algorithms are based on finding the k nearest observations to the training set. For classification, the problem is to infer the class of a new feature vector by computing the majority vote of its k nearest observations from the training set. For search, the problem is to infer k nearest observations from the training set to a new feature vector. The nearest observations are computed based on the chosen distance metric.
Operation |
Computational methods |
Programming Interface |
|||
Mathematical formulation
Training
Classification
Let 
be the training set of p-dimensional feature vectors, let 
be the set of class labels, where 
, 
, and C is the number of classes. Given X, Y, and the number of nearest neighbors k, the problem is to build a model that allows distance computation between the feature vectors in training and inference sets at the inference stage.
Search
Let be the training set of p-dimensional feature vectors. Given X and the number of nearest neighbors k, the problem is to build a model that allows distance computation between the feature vectors in training and inference sets at the inference stage.
Training method: brute-force
The training operation produces the model that stores all the feature vectors from the initial training set X.
Training method: k-d tree
The training operation builds a k-d tree that partitions the training set X (for more details, see k-d Tree).
Inference
Classification
Let 
be the inference set of p-dimensional feature vectors. Given 
, the model produced at the training stage, and the number of nearest neighbors k, the problem is to predict the label 
from the Y set for each 
, 
, by performing the following steps:
Identify the set
of k feature vectors in the training set that are nearest to with respect to the Euclidean distance, which is chosen by default. The distance can be customized with the predefined set of pairwise distances: Minkowski distances with fractional degree (including Euclidean distance), Chebyshev distance, and Cosine distance.Estimate the conditional probability for the l-th class as the fraction of vectors in
whose labels 
are equal to l:
Predict the class that has the highest probability for the feature vector
:
Search
Let be the inference set of p-dimensional feature vectors. Given , the model produced at the training stage, and the number of nearest neighbors k:
Identify the set
of k feature vectors in the training set that are nearest to with respect to the Euclidean distance, which is chosen by default. The distance can be customized with the predefined set of pairwise distances: Minkowski distances with fractional degree (including Euclidean distance), Chebyshev distance, and Cosine distance.
Inference method: brute-force
Brute-force inference method determines the set of the nearest feature vectors by iterating over all the pairs 
in the implementation defined order, 
, 
.
Inference method: k-d tree
K-d tree inference method traverses the k-d tree to find feature vectors associated with a leaf node that are closest to , 
. The set 
of the currently known nearest k neighbors is progressively updated during the tree traversal. The search algorithm limits exploration of the nodes for which the distance between the and respective part of the feature space is not less than the distance between and the most distant feature vector from . Once tree traversal is finished, 
.
Programming Interface
Refer to API Reference: k-Nearest Neighbors Classification and Search.
Usage example
Training
knn::model<> run_training(const table& data,
const table& labels) {
const std::int64_t class_count = 10;
const std::int64_t neighbor_count = 5;
const auto knn_desc = knn::descriptor<float>{class_count, neighbor_count};
const auto result = train(knn_desc, data, labels);
return result.get_model();
}
Inference
table run_inference(const knn::model<>& model,
const table& new_data) {
const std::int64_t class_count = 10;
const std::int64_t neighbor_count = 5;
const auto knn_desc = knn::descriptor<float>{class_count, neighbor_count};
const auto result = infer(knn_desc, model, new_data);
print_table("labels", result.get_labels());
}
Examples
oneAPI DPC++
Batch Processing:
dpc_knn_cls_brute_force_dense_batch.cpp
oneAPI C++
Batch Processing:
cpp_knn_cls_brute_force_dense_batch.cpp
cpp_knn_cls_kd_tree_dense_batch.cpp
cpp_knn_search_brute_force_dense_batch.cpp
Python* with DPC++ support
Batch Processing: