Updates to IMDB-based graph-NSL tutorial.

- use 2 neighbors for graph regularization (demonstrates use of > 1 nbr)
- increase # training epochs from 4 to 10 (reduces variance and instability)
- recompute accuracy values for base and graph-regularized model for various supervision ratio values for both the Bi-LSTM and feed-forward NN architectures.

PiperOrigin-RevId: 266485487

Author: Neural-Link Team

g3doc/tutorials/graph_keras_lstm_imdb.ipynb

@@ -743,13 +743,13 @@
         "    ### neural graph learning parameters\n",
         "    self.distance_type = nsl.configs.DistanceType.L2\n",
         "    self.graph_regularization_multiplier = 0.1\n",
-        "    self.num_neighbors = 1\n",
+        "    self.num_neighbors = 2\n",
         "    ### model architecture\n",
         "    self.num_embedding_dims = 16\n",
         "    self.num_lstm_dims = 64\n",
         "    self.num_fc_units = 64\n",
         "    ### training parameters\n",
-        "    self.train_epochs = 4\n",
+        "    self.train_epochs = 10\n",
         "    self.batch_size = 128\n",
         "    ### eval parameters\n",
         "    self.eval_steps = None  # All instances in the test set are evaluated.\n",
@@ -1459,11 +1459,12 @@
         "# Accuracy values for both the Bi-LSTM model and the feed forward NN model have\n",
         "# been precomputed for the following supervision ratios.\n",
         "\n",
-        "supervision_ratios = [0.3, 0.15, 0.05, 0.03, 0.01]\n",
+        "supervision_ratios = [0.3, 0.15, 0.05, 0.03, 0.02, 0.01, 0.005]\n",
         "\n",
         "model_tags = ['Bi-LSTM model', 'Feed Forward NN model']\n",
-        "base_model_accs = [[85, 85, 62, 58, 50], [85, 79, 61, 53, 50]]\n",
-        "graph_reg_model_accs = [[85, 84, 76, 63, 51], [85, 79, 73, 62, 50]]\n",
+        "base_model_accs = [[84, 84, 83, 80, 65, 52, 50], [87, 86, 76, 74, 67, 52, 51]]\n",
+        "graph_reg_model_accs = [[84, 84, 83, 83, 65, 63, 50],\n",
+        "                        [87, 86, 80, 75, 67, 52, 50]]\n",
         "\n",
         "plt.clf()  # clear figure\n",
         "\n",
@@ -1498,12 +1499,12 @@
         "It can be observed that as the superivision ratio decreases, model accuracy also\n",
         "decreases. This is true for both the base model and for the graph-regularized\n",
         "model, regardless of the model architecture used. However, notice that the\n",
-        "graph-regularized model is consistenly better than the base model -- sometimes\n",
-        "by as much as 15% -- and further, as the supervision ratio decreases, the\n",
-        "decrease in accuracy is much less for the graph-regularized model than the base\n",
-        "model. This is primarily because of semi-supervised learning for the\n",
-        "graph-regularized model, where structural similarity among training samples is\n",
-        "used in addition to the training samples themselves."
+        "graph-regularized model performs better than the base model for both the\n",
+        "architectures. In particular, for the Bi-LSTM model, when the supervision ratio\n",
+        "is 0.01, the accuracy of the graph-regularized model is **~20%** higher than\n",
+        "that of the base model. This is primarily because of semi-supervised learning\n",
+        "for the graph-regularized model, where structural similarity among training\n",
+        "samples is used in addition to the training samples themselves."
       ]
     },
     {