business_reviews

Business Reviews

Multi-input Deep Learning model to predict ratings of business reviews

Background

In the text classification tasks, we make use of multiple input data types - textual as well as the non-textual information or meta-data to classify the text which can improve accuracy of statistical model. For example, gender, nationalities, etc.

Goal

To create a text classification system that classifies user reviews regarding different businesses into “good”, “bad”, and “average” using a multiple inputs model based on the text of the review along with the associated meta data.

Process

• Use Keras Functional API since it supports multiple inputs and multiple output models.
• Create a deep learning model
  • Accept multiple inputs
  • Concatenate the two outputs
  • Performing classification using the aggregated input.

Dependencies

• Pandas
• Seaborn
• Numpy
• Tensorflow
• Keras
• Scikit-Learn
• Matplotlib

Dataset

Reviews of different business (yelp_review): https://www.kaggle.com/likarajo/yelp-business-reviews

Data Analysis and preparation

• The dataset contains a column Stars that contains ratings for different businesses.
  • Have values between 1 and 5.
• Convert the numerical stars values for the reviews into categorical ones.
  • Add a new column rating to our dataset.
  • Stars = 1 => rating = bad
  • Stars = 2 or 3 => rating = average
  • Stars = 4 or 5 => rating = good

Data Preprocessing

• Remove punctuations and numbers
• Remove single characters
• Replace multiple spaces with single space

Text-only Model

• Convert text label to one-hot encoded vector
  • covert text to integer labels
  • convert the labels to categories
• Split data into training and test sets
• Create Word embeddings
  • convert text into token sequences
  • pad each sequence to equal length
  • load pre-trained GloVe file with word vectors from the Stanford Core NLP glove file
  • create embedding matrix using the word vectors from the GloVe file
• Create the model using Keras functional API
  • One Embedding (input) layer
  • One LSTM layer
    • 128 neurons
  • One Dense (output) layer
    • 3 neurons (for 3 outputs bad, average, good)
    • Activation function = Softmax
  • Compilation
    • Loss function = categorical cross-entropy
    • Optimization fucntion = Adam
    • Metric = Accuracy
• Train the model
  • Train data = 80%
  • Validation data = 20%
  • Epochs = 10
• Evaluate the model
  • Test score
  • Test accuracy

---

Metadata-only Model

• Convert text label to one-hot encoded vector
  • covert text to integer labels
  • convert the labels to categories
• Split data into training and test sets
• Create the model using Keras functional API
  • input layer
    • 3 metadata
  • two hidden dense layers
    • neurons = 10
    • activation = Relu
  • output dense layer
    • neurons = 3 for 3 outputs (bad, average, good)
    • activation = Softmax
  • Compilation
    • Loss function = categorical cross-entropy
    • Optimization fucntion = Adam
    • Metric = Accuracy
• Train the model
  • Train data = 80%
  • Validation data = 20%
  • Epochs = 10
• Evaluate the model
  • Test score
  • Test accuracy
    NOTE: The accuracy can be increased by increasing the number of dense layers or by increasing the number of epochs.

---

Model with multiple inputs

• Input is the combination of both textual information and meta information
• Built with Keras Functional API
• Create two sub-models

Sub-model 1

• Input = textual: review text
• Embedding layer
• LSTM layer: 128 neurons
• Output = Dense: 3 neurons (bad, average, good)

Sub-model 2

• Input = meta info: useful, funny, cool
• Dense layer 1: 10 neurons
• Dense layer 2: 10 neurons
• Output = Dense: 3 neurons (bad, average, good)

Concatenate sub-model outputs

• Input = Sub-model 1 output + Sub-model 2 output
• Dense layer: 10 neurons
• Output = Dense: 3 neurons (bad, average, good)

---

Conclusion and Improvements

• If the differences for loss and accuracy values is minimal between the training and test sets, then the model is not overfitting.
• We only used 10,000 rveiew records due to hardware constraint. Better performance can be achieved by training the model on a higher number of records.
• More LSTM and dense layers can be added to the model. If the model overfits, then layers can be dropped out as necessary.
• The optimizer function can be changed.
• The model can be trained with higher number of epochs.