fix(analytical): fix the louvain algorithm that the result is not consistent with #183

analytical_engine/apps/pregel/louvain/auxiliary.h

@@ -41,6 +41,8 @@ constexpr int phase_one_minor_step_0 = 0;
 constexpr int phase_one_minor_step_1 = 1;
 constexpr int phase_one_minor_step_2 = 2;
 
+constexpr double min_quality_improvement = 0.001;
+
 /**
  * The state of a vertex.
  */

analytical_engine/apps/pregel/louvain/louvain_app_base.h

@@ -16,6 +16,7 @@ limitations under the License.
 #ifndef ANALYTICAL_ENGINE_APPS_PREGEL_LOUVAIN_LOUVAIN_APP_BASE_H_
 #define ANALYTICAL_ENGINE_APPS_PREGEL_LOUVAIN_LOUVAIN_APP_BASE_H_
 
+#include <cmath>
 #include <memory>
 #include <string>
 #include <utility>
@@ -224,7 +225,9 @@ class LouvainAppBase
               actual_quality_aggregator);
       // after one pass if already decided halt, that means the pass yield no
       // changes, so we halt computation.
-      if (current_super_step <= 14 || actual_quality <= ctx.prev_quality()) {
+      if (current_super_step <= 14 ||
+          std::fabs(actual_quality - ctx.prev_quality()) <
+              min_quality_improvement) {
         // turn to sync community result
         ctx.compute_context().set_superstep(sync_result_step);
         syncCommunity(frag, ctx, messages);
@@ -273,13 +276,6 @@ class LouvainAppBase
       } else if (ctx.compute_context().superstep() == compress_community_step) {
         ctx.GetVertexState(v).is_alived_community = false;
       }
-
-      if (!ctx.compute_context().active(v)) {
-        std::vector<std::pair<vid_t, edata_t>> tmp_edges;
-        ctx.GetVertexState(v).fake_edges.swap(tmp_edges);
-        std::vector<vid_t> tmp_nodes;
-        ctx.GetVertexState(v).nodes_in_community.swap(tmp_nodes);
-      }
     });
 
     {
@@ -321,7 +317,7 @@ class LouvainAppBase
                                 int tid, vertex_t v) {
       const auto& member_list = ctx.vertex_state()[v].nodes_in_community;
       if (!member_list.empty()) {
-        auto community_id = frag.Gid2Oid(member_list.front());
+        auto community_id = frag.Gid2Oid(ctx.vertex_state()[v].community);
         // send community id to members
         for (const auto& member_gid : member_list) {
           auto fid = vid_parser.GetFid(member_gid);

python/graphscope/tests/unittest/test_app.py

@@ -235,7 +235,10 @@ def test_other_app_on_undirected_graph(
 
     # louvain
     ctx = louvain(p2p_project_undirected_graph, min_progress=50, progress_tries=2)
-    assert ctx is not None
+    df = ctx.to_dataframe({"node": "v.id", "r": "r"})
+    community_num = len(df["r"].unique())
+    print("community_num: ", community_num)  # it's not a fixed number
+
     # simple_path
     ctx = is_simple_path(p2p_project_undirected_graph, [1, 10])
     assert ctx is not None