Below text-classification model gives accuracy of 0.77 only on one dataset and 0.99 on spam-ham dataset? What should I do to increase with my dataset?

Question

from keras.models import Model
from keras.layers import Input, Dense, Dropout, Embedding, Conv1D, MaxPooling1D, Flatten, Bidirectional, GRU, Concatenate, Lambda, Multiply, Permute, RepeatVector,dot



text_input = tf.keras.layers.Input(shape=(), dtype=tf.string)
encoder_inputs = preprocessor(text_input)
outputs = encoder(encoder_inputs)
#pooled_output = outputs["pooled_output"]      # [batch_size, 768].

sequence_output = outputs["sequence_output"]
dropout_layer = Dropout(0.3)(sequence_output)  
# add BiGRU layer with attention mechanism
bigru_output= Bidirectional(GRU(units=64,activation='tanh',return_sequences=True))(dropout_layer)


# Add a CNN layer
conv_layer1 = Conv1D(filters=128, kernel_size=2, activation='relu',padding="same")(bigru_output)
conv_layer2 = Conv1D(filters=128, kernel_size=3, activation='relu',padding="same")(bigru_output)
conv_layer3 = Conv1D(filters=128, kernel_size=4, activation='relu',padding="same")(bigru_output)
# max_pool_layer = MaxPooling1D(pool_size=2)(conv_layer)
conv_layer= tf.keras.layers.Concatenate()([conv_layer1,conv_layer2,conv_layer3])
# Add a dropout layer after the CNN layer
conv_layer = Dropout(0.3)(conv_layer)
# Map each cnn output vector to a unique context vector using a Dense layer
context_vectors = Dense(128, activation='tanh')(conv_layer)

# Define a function to compute attention scores
def compute_attention_score(context_vector, query_vector):
    """
    Computes the attention score between a context vector and a query vector.
    """
    score = dot([context_vector, query_vector], axes=[1, 1])
    score = Activation('softmax')(score)
    return score

# Compute attention scores for each context vector using a lambda function
attention_scores = Lambda(lambda x: compute_attention_score(x[0], x[1]))([context_vectors, bigru_output])

# Compute the weighted sum of the context vectors using the attention scores
weighted_context_vectors = Lambda(lambda x: dot([x[0], x[1]], axes=[1, 1]))([attention_scores, context_vectors])

# Concatenate the weighted context vectors with the BiGRU output vector
attention_output = Lambda(lambda x: tf.concat([x[0], x[1]], axis=-1))([bigru_output, weighted_context_vectors])


# Add max pooling layer
max_pool_layer = MaxPooling1D(pool_size=2)(attention_output)

# Flatten and add dense layer for final output
flatten_layer = Flatten()(attention_output)
output_layer = Dense(units=1, activation='sigmoid')(flatten_layer)


# define the model
model = Model(name="BBRCA",inputs=text_input, outputs=output_layer)

Answer 1

Your question can be given a general answer along the lines of the comment from @Erwan . Scoring highly on a chosen metric (and as he mentioned accuracy is only one type and possibly not the best one) is affected by a myriad of considerations. Trying to compare the results of a metric on completely distinct types of problems does not have any meaning

• The datasets are different
• The objectives to be measured are different

There should be no expectation of results on the one hand having any carry-over meaning to the other. Even if you had 99% precisely on both datasets it would not indicate that the ham-spam results were indicative of likely statistical strength on the second problem.