Business Recommendation using Topic-Modelling on Yelp Data Set

Objective :

• Based on Yelp Data Set of User Reviews for Restaurants. Business value for this project are:
  • To provide recommendation on topics a Business Owner can improve in their restaurant.
  • To provide recommendation on topics for a new Restaurant.
• This project involved parsing raw customer reviews data with python's NLTK tokenizing words, POS tagging, filtering out only NN, NNS (noun singular/plurals) and removing stop words .
• Created Corpus documents with tokenized words
• Apply LDA algorithm on the documents and create a set of topics and the distribution of words in them.
• Predict new topics based on trained model.

Dataset

1. Please download and extract Yelp Dataset Challenge data, yelp_dataset_challenge_academic_dataset from the following link: Yelp Data Set
2. Put this data dolder in the dataset folder in the same directory as Topic_Modelling.
3. Install MongoDB in your system from the following link : Install MongoDB
4. Download the MongoDB data files from the following google drive link: MongoDB
5. Put this folder in /var/lib

Python Packages:

Python packages to be installed for running the program:

1. Gensim
2. Nltk
3. Pandas
4. Numpy
5. Matplotlib
6. Scipy
7. pymongo

File Information:

Relevant File Informations:

1. __init__.py - To treat the current directory as package.
2. db_objects.py - Factory class to return database collections.
3. filter_review.py - Filters the reviews from JSON file and stores into MongoDB REVIEW collections.
4. make_json_serializable.py - Creates serializable JSON objects from the MongoDB collection to write JSON data to JSON file.
5. processing.py - Processes the reviews by tokenizing, removing stop words, lemmatizing and POS tagging and stores it into MongoDB CORPUS collection.
6. settings.py - Contains database constants, connection strings, collection names, CORPUS disctionary object names.
7. show_db_information.py - Displays the database names currently present in the MongoDB database, collections present and collection objects informations.
8. training.py - Creates CORPUS documents and train the LDA Model with documents as input.
9. klDivergence.py - Creates KL Divergence graph to find out optimal number of topics for LDA.

Project Setup:

1. We have filtered out review data based on “Pittsburgh” city (total 61849 reviews after filtering) and trained our LDA model on this data. It took a total of 7.91 Hrs to preprocess (tokenization, removing stop words, lemmatizing,POS tagging) the reviews on running 4 processes on 4 different cores on Dell, Intel i5 processor.
2. The review data are present in the REVIEWS_COLLECTION, while the corpus document data are present in the CORPUS_COLLECTION.

Steps To Run:

1. Run transform.py file to extract the reviews data for pittsburgh area into a new json file(review_json_file_pittsburgh_restaurant.json).
2. Run kldivergence.py to plot the symmetric KL divergence vs number of topic to find the optimal number of topics for LDA.
3. Run training.py to train the LDA model and populate the optimal number of topics found by KL Divergence method in the previous step.
4. Run display.py to check the results based on business ID provided.

Project WorkFlow:

Results:

The topics that would be output from the program

(0, u'0.093*pizza + 0.023*sauce + 0.015*slice + 0.015*place + 0.014*salad + 0.014*tomato + 0.013*cheese + 0.012*crust + 0.012*delivery + 0.012*bread')
(1, u'0.017*meal + 0.016*restaurant + 0.015*dinner + 0.014*flavor + 0.014*dessert + 0.013*menu + 0.011*plate + 0.011*pork + 0.011*meat + 0.010*potato')
(2, u'0.031*taco + 0.023*place + 0.021*breakfast + 0.020*coffee + 0.019*egg + 0.019*chip + 0.018*brunch + 0.010*time + 0.009*potato + 0.009*day')
(3, u'0.037*place + 0.026*time + 0.022*order + 0.021*burger + 0.012*sandwich + 0.012*don + 0.012*lunch + 0.011*fry + 0.011*service + 0.010*people')
(4, u'0.029*place + 0.026*chicken + 0.021*restaurant + 0.019*soup + 0.018*spicy + 0.016*rice + 0.016*roll + 0.013*sushi + 0.013*sauce + 0.013*service')
(5, u'0.036*place + 0.030*bar + 0.024*beer + 0.022*service + 0.021*time + 0.018*night + 0.013*drink + 0.013*restaurant + 0.012*selection + 0.012*menu')

This can be summarised as the following chart:

The following table shows the average rating for the topics of business id -

SsGNAc9U-aKPZccnaDtFkA

The above results infer that the restaurant needs to improve on Breakfast.

References:

1. http://mlwave.com/tutorial-online-lda-with-vowpal-wabbit/
2. http://xmodulo.com/how-to-find-number-of-cpu-cores-on.html
3. http://stackoverflow.com/questions/20886565/python-using-multiprocessing-process-with-a-maximum-number-of-simultaneous-pro

• LDA Algorithm Explanation:

1. http://blog.echen.me/2011/08/22/introduction-to-latent-dirichlet-allocation/
2. https://wellecks.wordpress.com/2014/09/03/these-are-your-tweets-on-lda-part-i/
3. http://stackoverflow.com/questions/10624760/latent-dirichlet-allocation-solution-example
4. http://obphio.us/pdfs/lda_tutorial.pdf
5. http://www.vladsandulescu.com/topic-prediction-lda-user-reviews/
6. https://www.quora.com/What-is-a-good-explanation-of-Latent-Dirichlet-Allocation