Ranking Restaurant Services using Cluster Analysis of Yelp Dataset

Data

rev_Yelp
https://uofi.box.com/s/6y1y9p39env2rhzpnggm47mzvxga3xma

How to run

To generate the desired output, follow these steps:

1. Place rev_Yelp and business.py in the same folder and run the script. The resulting output should include a k-means graph.
2. Place rev_Yelp and feature_importance.py in the same folder and simply run. The resulting output should be a feature importance graph contain three sub graphs.

MEMBERS

Hanhui Song ([email protected]) MCS student Department of Computer Science
Mingyi Lai ([email protected]) MS student Department of Industrial and System Engineering
Shun-Hsiang Hsu ([email protected]) Ph.D. student Department of Civil and Environmental Engineering

INTRODUCTION

This project aims to use cluster analysis of the Yelp dataset to obtain the preferred services of the active users and rank restaurant services for providing recommendations. To that end, our clustering methods will take only the numerical and categorical data (e.g. the review count and category of a restaurant) in the Yelp dataset as the inputs to analyze the relationship between the active users and their reviews for ranking restaurant services. Based on the results, customized rules for different users will be developed to provide recommendations.

DATASET

The Yelp dataset consists of reviews, businesses, users, tips, and check-in data with a total size of 9 GB, including about 700k reviews, 1M+ users, and 150k+ restaurants. In addition to the categorical and numerical data, the text data of review contents are also included.

METHODS

This project will first discuss which columns might be important to rank services and conduct data cleaning to the selected columns (e.g., handling missing data). With the columns, different clustering methods will be developed and compared to obtain the relationship between the active users and their reviews. Then, the recommendations based on the relationships will be presented. For feasibility, our objective and research scope are similar to the previous research [1], where K-means and DBSCAN were adopted to infer electricity behavior.

K-means

1. Place the data generated from data_comb.py in rev_Yelp.
2. Run the K-means-Final.py, there are 8 results. From the top to the bottom they are: normal sum data using fill NaN with 0 method, standardized sum data using fill NaN with 0 method, normal time data using fill NaN with 0 method, standardized time data using fill NaN with 0 method, normal sum data using delete NaN data method, standardized sum data using delete NaN data method, normal time data using delete NaN data method, standardized time data using delete NaN data method.
3. You can set the number of cluster for each results and see how the results change.

SCHEDULES

• Before the midpoint (3/21), data preprocessing and preliminary clustering results using baseline approaches (K-means and DBSCAN) should be completed.
• [3/21-4/4] Review and develop new methods for comparison.
• [4/4-4/18] Rank restaurant services for providing recommendation.
• [4/18-5/2] Write final report.

REFERENCES

[1] Liping Zhang, Song Deng, and Shiyue Li. 2017. Analysis of power consumer behavior based on the complementation of K-means and DBSCAN. In 2017 IEEE Conference on Energy Internet and Energy System Integration (EI2). IEEE, 1–5.