Final Presentation

Challenge Team

Salt baes

Objectives

• Elite Users
• Yelp has Elite users in its system, where people can nominate themselves/other users for Elite status
• Yelp then reviews these nominations through their “Elite Council”, with criteria unknown to us
• We would like to use the Yelp dataset to analyze Elite user trends
• Applications
• Provide Yelp a system that can automatically mark users as Elite and take away Elite status from users who have been slacking
• Help someone who wants to become an elite user

User Groups

• Elite
• Elite this year, and will be elite next year
• Non-Elite
• Non-elite this year, and will be non-elite next year
• Slackers
• Elite this year, and will be non-elite next year
• Potentials
• Non-elite this year, and will be elite next year

Elite, Non-elite detection

• Metadata Analysis
• Text Analysis
• Time Series Analysis

Metadata Analysis

How is the number of tips affiliated with a Yelp user’s elite status?

We looked for distinguishing characteristics in the tip-giving patterns among different user groups.

​

Elites and Non-Elites have defined traits, but Potentials and Slackers are difficult to define based solely on tip behavior.

Statistical Analysis

Significant at the 10% level

Significant at the 5% level

Significant at the 1% level

Text Analysis

Reviews reduced to 3 dimensions using T-SNE

• Vectorized reviews into both bag of words and TF-IDF (term frequency x inverse document frequency) format
• Looked into clustering of different types of users based on their vectorized reviews
• Unable to find distinct clusters from the different classes for both representations

Text Analysis

Non-Elite

Elite

Slacker

Potentials

Time Series Analysis

• Studied change in each user’s behavior over time
• Extracted both positive and negative anomalies
• Clustered pattern of user behavior
• Use this metric as a classification feature

Regression Models

SVC:

R^2 0.6695, MAE 0.0869, MSE 0.00809

Kernel Ridge:

R^2 0.6935 MAE 0.02598, MSE 0.00405

Decision Tree (depth = 3):

R^2 0.8711, MAE 0.02120, MSE 0.00143

• Each model is tuned by the grid search and 3-fold cross validation

​

​

​

​

Anomaly Detection

​

• Decision Tree model
• Detect the negative anomalies

​

• Kernel Ridge
• Extract anomalies without over-fitting.

Kernel Density Estimation

• Used local-minimum of the density graph as a “turning point”

Dynamic Time Warping Clustering

• Dynamically find the best alignment that minimizes overall euclidean distance.

​

• Spliced the data into 24-month period (2 years) and calculated DTW distance to cluster user pattern

​

Questions?

Subteam: Boss Llama

1st Topic: Survival Analysis

• Predict if a business is open on a certain date
• Features
• Average star rating
• Days between reviews
• Last activity
• Linear regression
• Achieved ~94% accuracy
• Unable to increase past 94%
• Unable to validate

2nd Topic: Competition analysis

• Key Question: Can we see competition in reviews?
• Performance Metric: Time taken to reach threshold count
• Goal:
• Construct a model of competition
1. Model the macroscopic trend of review intensity
2. Apply natural language processing to categorize businesses
3. Identify city centers or “clusters”
• Fit regression models to find the competition relationships

Features Considered

• 8 cities, 9857 Restaurants
• Review count and intensity
• Geographical information
• Survival information
• Is_open
• First Review Date
• Restaurants categories
• Topics in the Reviews
• Business Attributes

Fit a Poisson Process

1. Good model of activity or “arrivals”
1. Rate function has effect of “smoothing” the volatility
2. How to fit a rate function
1. Identify inflection points in plateauing process
2. Use regression methods to fit a function
3. Find expected number of days to reach threshold
4. Rate function has linear, quadratic, seasonal (weekly, bi-annual, annual) components

Geographical “clusters”

Identifying Clusters

Topic Modeling

• Trained LDA and NMF models on reviews
• Analyzed top words in each topic
• Generated probability distributions among all the topics
• Compared topics generated from topic modeling with the categories generated on Yelp

Topic Top Words

Other Characteristics of Restaurants

Types of Restaurants

Identifying User Clusters

Steps:

1. Review Vectors�- word2vec → Reviews
2. User Vectors
3. Dimension Reduction
4. Fit Gaussian Mixture Models�→ Classify users

​

​

Result

​

​

​

​

• Latitude and Longitude has very high importance
• Certain categories are much more influential in making threshold
• More interaction terms need to be added
• ​

The End

Thank you!