compute cosine similarity for vertex pairs (#4482)

compute cosine similarity for vertex pairs

Authors:
  - Naim (https://github.com/naimnv)

Approvers:
  - Seunghwa Kang (https://github.com/seunghwak)

URL: #4482

Author: naimnv

cpp/CMakeLists.txt

@@ -193,9 +193,11 @@ set(CUGRAPH_SOURCES
     src/link_prediction/jaccard_sg.cu
     src/link_prediction/sorensen_sg.cu
     src/link_prediction/overlap_sg.cu
+    src/link_prediction/cosine_sg.cu
     src/link_prediction/jaccard_mg.cu
     src/link_prediction/sorensen_mg.cu
     src/link_prediction/overlap_mg.cu
+    src/link_prediction/cosine_mg.cu
     src/layout/legacy/force_atlas2.cu
     src/converters/legacy/COOtoCSR.cu
     src/community/legacy/spectral_clustering.cu

cpp/include/cugraph/algorithms.hpp

@@ -2082,6 +2082,37 @@ rmm::device_uvector<weight_t> jaccard_coefficients(
   std::tuple<raft::device_span<vertex_t const>, raft::device_span<vertex_t const>> vertex_pairs,
   bool do_expensive_check = false);
 
+/**
+ * @brief     Compute Cosine similarity coefficient
+ *
+ * Similarity is computed for every pair of vertices specified. Note that
+ * similarity algorithms expect a symmetric graph.
+ *
+ * @throws                 cugraph::logic_error when an error occurs.
+ *
+ * @tparam vertex_t Type of vertex identifiers. Needs to be an integral type.
+ * @tparam edge_t Type of edge identifiers. Needs to be an integral type.
+ * @tparam weight_t Type of edge weights. Needs to be a floating point type.
+ * @tparam multi_gpu Flag indicating whether template instantiation should target single-GPU (false)
+ * @param handle RAFT handle object to encapsulate resources (e.g. CUDA stream, communicator, and
+ * handles to various CUDA libraries) to run graph algorithms.
+ * @param graph_view Graph view object.
+ * @param edge_weight_view Optional view object holding edge weights for @p graph_view. If @p
+ * edge_weight_view.has_value() == true, use the weights associated with the graph. If false, assume
+ * a weight of 1 for all edges.
+ * @param vertex_pairs tuple of device spans defining the vertex pairs to compute similarity for
+ * In a multi-gpu context each vertex pair should be local to this GPU.
+ * @param do_expensive_check A flag to run expensive checks for input arguments (if set to `true`).
+ * @return similarity coefficient for the corresponding @p vertex_pairs
+ */
+template <typename vertex_t, typename edge_t, typename weight_t, bool multi_gpu>
+rmm::device_uvector<weight_t> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<vertex_t, edge_t, false, multi_gpu> const& graph_view,
+  std::optional<edge_property_view_t<edge_t, weight_t const*>> edge_weight_view,
+  std::tuple<raft::device_span<vertex_t const>, raft::device_span<vertex_t const>> vertex_pairs,
+  bool do_expensive_check = false);
+
 /**
  * @brief     Compute Sorensen similarity coefficient
  *
@@ -2202,6 +2233,62 @@ std::
     std::optional<size_t> topk,
     bool do_expensive_check = false);
 
+/**
+ * @brief     Compute Consine all pairs similarity coefficient
+ *
+ * Similarity is computed for all pairs of vertices.  Note that in a sparse
+ * graph, many of the vertex pairs will have a score of zero.  We actually
+ * compute similarity only for vertices that are two hop neighbors within
+ * the graph, since vertices that are not two hop neighbors will have
+ * a score of 0.
+ *
+ * If @p vertices is specified we will compute similarity on two hop
+ * neighbors the @p vertices.  If @p vertices is not specified it will
+ * compute similarity on all two hop neighbors in the graph.
+ *
+ * If @p topk is specified only the top @p topk scoring vertex pairs
+ * will be returned, if not specified then scores for all computed vertex pairs
+ * will be returned.
+ *
+ * Note the list of two hop neighbors in the entire graph might be a large
+ * number of vertex pairs.  If the graph is dense enough it could be as large
+ * as the the number of vertices squared, which might run out of memory.
+ *
+ * @throws                 cugraph::logic_error when an error occurs.
+ *
+ * @tparam vertex_t Type of vertex identifiers. Needs to be an integral type.
+ * @tparam edge_t Type of edge identifiers. Needs to be an integral type.
+ * @tparam weight_t Type of edge weights. Needs to be a floating point type.
+ * @tparam multi_gpu Flag indicating whether template instantiation should target single-GPU (false)
+ * @param handle RAFT handle object to encapsulate resources (e.g. CUDA stream, communicator, and
+ * handles to various CUDA libraries) to run graph algorithms.
+ * @param graph_view Graph view object.
+ * @param edge_weight_view Optional view object holding edge weights for @p graph_view. If @p
+ * edge_weight_view.has_value() == true, use the weights associated with the graph. If false, assume
+ * a weight of 1 for all edges.
+ * @param vertices optional device span defining the seed vertices. In a multi-gpu context the
+ * vertices should be local to this GPU.
+ * @param topk optional specification of the how many of the top scoring vertex pairs should be
+ * returned
+ * @param do_expensive_check A flag to run expensive checks for input arguments (if set to `true`).
+ * @return tuple containing three device vectors (v1, v2, score) of the same length.  Corresponding
+ * elements in the vectors identify a result, v1 identifying a vertex in the graph, v2 identifying
+ * one of v1's two hop neighors, and the score identifying the similarity score between v1 and v2.
+ * If @p topk was specified then the vectors will be no longer than @p topk elements.  In a
+ * multi-gpu context, if @p topk is specified all results will return on GPU rank 0, otherwise they
+ * will be returned on the local GPU for vertex v1.
+ */
+template <typename vertex_t, typename edge_t, typename weight_t, bool multi_gpu>
+std::
+  tuple<rmm::device_uvector<vertex_t>, rmm::device_uvector<vertex_t>, rmm::device_uvector<weight_t>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<vertex_t, edge_t, false, multi_gpu> const& graph_view,
+    std::optional<edge_property_view_t<edge_t, weight_t const*>> edge_weight_view,
+    std::optional<raft::device_span<vertex_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check = false);
+
 /**
  * @brief     Compute Sorensen similarity coefficient
  *

cpp/src/link_prediction/cosine_mg.cu

@@ -0,0 +1,122 @@
+/*
+ * Copyright (c) 2022-2024, NVIDIA CORPORATION.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "link_prediction/cosine_similarity_impl.cuh"
+
+namespace cugraph {
+
+template rmm::device_uvector<float> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int32_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int32_t, float const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<float> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int64_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<float> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int64_t, int64_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+  std::tuple<raft::device_span<int64_t const>, raft::device_span<int64_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<double> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int32_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int32_t, double const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<double> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int64_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<double> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int64_t, int64_t, false, true> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+  std::tuple<raft::device_span<int64_t const>, raft::device_span<int64_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<float>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int32_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int32_t, float const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<float>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int64_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int64_t>, rmm::device_uvector<int64_t>, rmm::device_uvector<float>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int64_t, int64_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+    std::optional<raft::device_span<int64_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<double>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int32_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int32_t, double const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<double>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int64_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int64_t>, rmm::device_uvector<int64_t>, rmm::device_uvector<double>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int64_t, int64_t, false, true> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+    std::optional<raft::device_span<int64_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+}  // namespace cugraph

cpp/src/link_prediction/cosine_sg.cu

@@ -0,0 +1,122 @@
+/*
+ * Copyright (c) 2022-2024, NVIDIA CORPORATION.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include "link_prediction/cosine_similarity_impl.cuh"
+
+namespace cugraph {
+
+template rmm::device_uvector<float> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int32_t, false, false> const& graph_view,
+  std::optional<edge_property_view_t<int32_t, float const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<float> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int64_t, false, false> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<float> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int64_t, int64_t, false, false> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+  std::tuple<raft::device_span<int64_t const>, raft::device_span<int64_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<double> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int32_t, false, false> const& graph_view,
+  std::optional<edge_property_view_t<int32_t, double const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<double> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int32_t, int64_t, false, false> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+  std::tuple<raft::device_span<int32_t const>, raft::device_span<int32_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template rmm::device_uvector<double> cosine_similarity_coefficients(
+  raft::handle_t const& handle,
+  graph_view_t<int64_t, int64_t, false, false> const& graph_view,
+  std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+  std::tuple<raft::device_span<int64_t const>, raft::device_span<int64_t const>> vertex_pairs,
+  bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<float>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int32_t, false, false> const& graph_view,
+    std::optional<edge_property_view_t<int32_t, float const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<float>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int64_t, false, false> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int64_t>, rmm::device_uvector<int64_t>, rmm::device_uvector<float>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int64_t, int64_t, false, false> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, float const*>> edge_weight_view,
+    std::optional<raft::device_span<int64_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<double>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int32_t, false, false> const& graph_view,
+    std::optional<edge_property_view_t<int32_t, double const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int32_t>, rmm::device_uvector<int32_t>, rmm::device_uvector<double>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int32_t, int64_t, false, false> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+    std::optional<raft::device_span<int32_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+template std::
+  tuple<rmm::device_uvector<int64_t>, rmm::device_uvector<int64_t>, rmm::device_uvector<double>>
+  cosine_similarity_all_pairs_coefficients(
+    raft::handle_t const& handle,
+    graph_view_t<int64_t, int64_t, false, false> const& graph_view,
+    std::optional<edge_property_view_t<int64_t, double const*>> edge_weight_view,
+    std::optional<raft::device_span<int64_t const>> vertices,
+    std::optional<size_t> topk,
+    bool do_expensive_check);
+
+}  // namespace cugraph