Yelp is a regional directory platform and review site with social networking tools. The website provides crowd-sourced reviews for local businesses (spas, restaurants, department stores, pubs, home-local services, shops, cars). This helps users to give business ratings and reviews. Normally, the review is a brief text composed of a few lines of about 100 words describing various user experiences with respect to various dimensions. This offers business owners the opportunity to improve their products and customers to choose among the best available sector.
The management might not have enough time to go through each and every review. If valuable information and insights can be provided to them at a glance, it can be really helpful and time-saving. And not only for the management but also for the customers who are trying to know more about the restaurant and need some help in ordering or selecting the restaurants. After all, in today’s world, every person prefers to read reviews and feedbacks before they take a decision. In our project, we have used Natural Language Processing and Machine Learning to achieve these business and customer goals. We have focussed on Sentiment Analysis, Topic Modelling, Data Analysis, and Classification for rating prediction.
- The dataset is available at https://www.yelp.com/dataset. It contains 6,685,900 reviews for 192,609 businesses based in 10 metropolitan areas.
- As a part of the data preparation and cleaning process, corrected the datatypes of every column in the dataset and dropped all the unnecessary columns.
- Wanted to focus on only one city which had the maximum no of reviews for restaurants.
- In this case, it was Las Vegas. Hence, filtered the dataset records to keep only reviews for restaurants in Las Vegas. The resultant dataset had 1,242,711 reviews for 6450 restaurant businesses.
Interpretation: Restaurants are the most reviewed business category followed by shopping, food and home services. It is always better to focus on one category for analysis in order to get relevant insights from data. Hence, reviews were filtered in the dataset to keep only restaurant category reviews.
Interpretation: Las Vegas has the maximum number of reviews. The reviews were filtered to keep the Las Vegas restaurant reviews.
Interpretation: From April 2016 to December 2018, it can be observed how the number of reviews for each restaurant varies. For example, for Nacho Daddy the reviews dropped from 100 in May 2016 to 20 in June 2016. It stayed the same for quite some time until Sept 2016. This information can be useful for the management to evaluate the possible reasons behind the drop and gain in number of reviews and when it happened.
Now, after cleaning the dataset, sentimental analysis was applied on the data to achieve analysis aim on Las vegas's top three restaurants.
Interpretations: It can be observed how the sentiment varies for each review for different restaurants. It helps us to understand how many reviews were positive and negative. For example, Hash House A Go Go had the maximum number of reviews with -0.1 polarity (negative). Most of the reviews for each restaurant were positive and were between 0.1 and 0.6.
Interpretations: It can be observed that for all the three carriers, the number of positive reviews is more than negative reviews. But as the number of reviews for a restaurant increases, the proportion of negative reviews also increases. Hash House A Go Go has maximum reviews and more negative reviews compared to others.
Interpretations: The frequent words in positive and negative reviews are quite similar. It’s good to have this as the starting point. We can see that reviewers talked a lot about the portion size in positive and service in negative reviews for Hash House A Go Go.
After separating the negative from the positive reviews, we built topic models using TF-IDF(Term Frequency - Inverse Document Frequency) and LDA(Latent Dirichlet Allocation) from the gensim package to understand what reviewers are talking about in the reviews.
From the results, it can be observed the following for each restaurant:
- Bacchanal Buffet: Reviewers probably liked the quality of seafood and buffet and were dissatisfied with the manager, service and wait time.
- Nacho daddy: Reviewers praised margarita pizza, location, portion size, and breakfast but left negative reviews for bartenders, tacos, burritos, and servers.
- Hash House A Go Go: Reviewers left positive feedback for pancakes, vega, and portion size but were not happy with biscuits, benedict, manager, waiter, and service.
There can be times when a user leaves a review and no rating for the restaurant. To predict the rating that the restaurant would have received based on each review we have used three different models - Naive Bayes, Random Forest, and Neural Network. We partitioned the dataset into training(70%) and testing(30%) dataset. Using the TF-IDF vectorizer, we transformed the clean preprocessed reviews text and used it as an input for all the three models.