diff --git a/src/model/TEXTGCN.cpp b/src/model/TEXTGCN.cpp
index 532e5fe..567e1b1 100644
--- a/src/model/TEXTGCN.cpp
+++ b/src/model/TEXTGCN.cpp
@@ -15,16 +15,34 @@ void TEXTGCN::SaveWeights(string model_name){
                 counter++;
         model << counter << " " << dim << endl;
 
+        vector<double> w_avg;
+        w_avg.resize(dim, 0.0);
+
+        long temp_vid;
         for (long vid=0; vid!=pnet.MAX_vid; vid++)
         {
             if (pnet.field[vid].fields[0]==1)
                 continue;
             
             model << pnet.vertex_hash.keys[vid];
-            
             if (pnet.field[vid].fields[0]==0)
+            {
+                for (int b=0; b<pnet.vertex[vid].branch; b++)
+                {
+                    temp_vid = pnet.context[pnet.vertex[vid].offset + b].vid;
+                    for (int d=0; d<dim; ++d)
+                    {
+                        w_avg[d] += w_vertex[temp_vid][d];
+                    }
+                }
                 for (int d=0; d<dim; ++d)
-                    model << " " << w_context[vid][d];
+                {
+                    model << " " << w_avg[d];
+                    w_avg[d] = 0.0;
+                }
+                //for (int d=0; d<dim; ++d)
+                //    model << " " << w_context[vid][d];
+            }
             if (pnet.field[vid].fields[0]==2)
                 for (int d=0; d<dim; ++d)
                     model << " " << w_vertex[vid][d];
diff --git a/src/proNet.cpp b/src/proNet.cpp
index f13a013..3d6fcef 100644
--- a/src/proNet.cpp
+++ b/src/proNet.cpp
@@ -2739,23 +2739,42 @@ void proNet::UpdateCBOWdev(vector< vector<double> >& w_vertex, vector< vector<do
 
 void proNet::UpdateCBOW(vector< vector<double> >& w_vertex, vector< vector<double> >& w_context, long vertex, long context, int dimension, double reg, int walk_steps, int negative_samples, double alpha){
 
-    vector<long> vertices;
-    vector<double> w_avg, back_err;
+    vector<long> vertex_contexts, context_contexts;
+    vector<double> w_avg, c_avg, back_err, back_c_err;
     w_avg.resize(dimension, 0.0);
+    c_avg.resize(dimension, 0.0);
     back_err.resize(dimension, 0.0);
-    long vertex_context;
+    back_c_err.resize(dimension, 0.0);
+    long vertex_context, context_context;
 
-    //vertices.push_back(vertex);
+    // context-context (user-event)
+    for (int i=0; i!=walk_steps; ++i)
+    {
+        context_context = TargetSample(context);
+        if (context_context==-1) break;
+        context_contexts.push_back(context_context);
+    }
+
+    // vertex-context (event-word)
     for (int i=0; i!=walk_steps; ++i)
     {
         vertex_context = TargetSample(vertex);
         if (vertex_context==-1) break;
-        vertices.push_back(vertex_context);
+        vertex_contexts.push_back(vertex_context);
     }
 
-    //double decay = 1.0;
     vector<double>* w_ptr;
-    for (auto v: vertices)
+    // context-context (user-event)
+    for (auto c: context_contexts)
+    {
+        w_ptr = &w_vertex[c];
+        for (int d=0; d!=dimension;++d)
+        {
+            c_avg[d] += (*w_ptr)[d];
+        }
+    }
+    // vertex-context (event-word)
+    for (auto v: vertex_contexts)
     {
         w_ptr = &w_vertex[v];
         for (int d=0; d!=dimension;++d)
@@ -2776,21 +2795,73 @@ void proNet::UpdateCBOW(vector< vector<double> >& w_vertex, vector< vector<doubl
     
     // positive training
     label = 1.0;
-    Opt_SigmoidRegSGD(w_avg, w_context[context], label, alpha, reg, back_err, w_context[context]);
+    Opt_SigmoidRegSGD(w_avg, c_avg, label, alpha, reg, back_err, back_c_err);
     //Opt_SGD(w_vertex[vertex], w_context[context], label, alpha, reg, back_err, w_context[context]);
 
+    // update context_contexts (user-event)
+    for (auto c: context_contexts)
+    {
+        w_ptr = &w_vertex[c];
+        for (int d=0; d!=dimension;++d)
+        {
+            (*w_ptr)[d] += back_c_err[d];
+        }
+    }
+    for (int d=0; d!=dimension;++d)
+    {
+        c_avg[d] = 0.0;
+        back_c_err[d] = 0.0;
+    }
+   
     // negative sampling
     label = 0.0;
     for (int neg=0; neg!=negative_samples; ++neg)
     {
-        neg_context = random_gen(0, MAX_vid);
-        while(field[neg_context].fields[0]!=0)
-            neg_context = random_gen(0, MAX_vid);
-        Opt_SigmoidRegSGD(w_avg, w_context[neg_context], label, alpha, reg, back_err, w_context[neg_context]);
+        //neg_context = random_gen(0, MAX_vid);
+        //while(field[neg_context].fields[0]!=0)
+        //    neg_context = random_gen(0, MAX_vid);
+        
+        // sample neg context-context (user-event)
+        context_contexts.clear();
+        for (int i=0; i!=walk_steps; ++i)
+        {
+            context_context = random_gen(0, MAX_vid);
+            while(field[context_context].fields[0]!=1) // event
+                context_context = random_gen(0, MAX_vid);
+            context_contexts.push_back(context_context);
+        }
+        // context-context (user-event)
+        for (auto c: context_contexts)
+        {
+            w_ptr = &w_vertex[c];
+            for (int d=0; d!=dimension;++d)
+            {
+                c_avg[d] += (*w_ptr)[d];
+            }
+        }
+
+        Opt_SigmoidRegSGD(w_avg, c_avg, label, alpha, reg, back_err, back_c_err);
         //Opt_SGD(w_vertex[vertex], w_context[context], label, alpha, reg, back_err, w_context[context]);
+
+        // update context_contexts (user-event)
+        for (auto c: context_contexts)
+        {
+            w_ptr = &w_vertex[c];
+            for (int d=0; d!=dimension;++d)
+            {
+                (*w_ptr)[d] += back_c_err[d];
+            }
+        }
+        for (int d=0; d!=dimension;++d)
+        {
+            c_avg[d] = 0.0;
+            back_c_err[d] = 0.0;
+        }
+
     }
 
-    for (auto v: vertices)
+    // update context_contexts (event-words)
+    for (auto v: vertex_contexts)
     {
         w_ptr = &w_vertex[v];
         for (int d=0; d!=dimension;++d)