recommender system using sas - spears business · collaborative recommender system 8 • create...

Post on 20-May-2020

10 Views

Category:

Documents

11 Downloads

Preview:

Click to see full reader

TRANSCRIPT

2017

SAS Analytics Day

Recommender System Using SAS

Shanmugavel Gnanasekar

Ravishankar Subramanian

2

Business Goal• Provide personalized suggestions to the

users based on their preferences. They aid in the decision-making process for the users and make their experience enjoyable.

Cons• These system suffers from inaccuracy.

• To build recommendation system using only ratings.

• Perform text mining on user reviews and combine it with original model to improve its accuracy

3

Content-Based Filtering Method

Collaborative-Filtering Method

Data Preparation Create User ProfileCreate Business

ProfileCreate IDF Attributes

Provide Recommendations

Evaluate

#SCSUG2016 4

Data SnapshotUser Review dataset.

Business Information Dataset

The data collected over 263,000 ratings provided by 21,000 unique users for over 4,000 different restaurants.

name

Content Based Filtering• It works by learning user preference or profile which is inferred from user ratings

and reviews.

• Then restaurants matching user’s tastes are recommended

#SCSUG2016 5

Definitions• Business Profile: Provided in the dataset.

• IDF(Inverse Document Frequency): Created based on number of times an attribute appears in restaurants.

𝐼𝐷𝐹=1

(𝑚𝑎𝑥(1, 𝑛 𝑡𝑖𝑚𝑒𝑠 𝑖𝑡 𝑎𝑝𝑝𝑒𝑎𝑟𝑠 𝑖𝑛 𝑜𝑡h𝑒𝑟 𝑟𝑒𝑠𝑡𝑎𝑢𝑟𝑎𝑛𝑡𝑠))

• User Profile: Build it based on the ratings provided by the user to a restaurant.

Profiles for Content-Based Filtering

#SCSUG2016 6

User Profile is created by aggregating all the individual ratings given by a user to various restaurants.

Business Profile

#SCSUG2016 7

IDF values for various features

IDF Table

Collaborative Recommender System

8

• Create Business and User profile.

• Identify n Neighbors for current user (For our study we used 20 neighbors)

• Recommend top restaurants rated by neighbor weighted by their similarity measure to the given user

Create User ProfileCreate Business

ProfileFind Neighbors

Create Recommendations

Evaluate

Flow for collaborative-based filtering method

Top Five Suggestions Based On Rating (Collaborative)

#SCSUG2016 9

Cluster User Review

10

DJCrowd

Music

Club

LoudDance Rock

Concept Link

Review Clusters

11

Cluster Name Descriptive Terms Frequency

Pizza Loversalways + beer + cheese + Crust + good + order + pepperoni + pizza + place + salad + sauce + slice + taste + thin 8,192

Night Life Appetizer + happy hour + beer + bar + great + half + night + roll + price + special 6,055

French Foodback + bread + cheese + chicken + delicious + French + line + long + lunch + minute + night + order 19,853

Chinese Food beef + chicken + Chinese + dis + egg + food + fry + good + lunch + noodle + pork + portion 24,855

Method Content-based filtering

Collaborative filtering

Root Mean Square Error 0.447 0.316

Mean Absolute Error 0.2 0.1

Fit Statistics

2017

SAS Analytics Day

Shanmugavel Gnanasekarshanmg@okstate.edu(813) 810 5630https://www.linkedin.com/in/shan-g/

Ravi Shankar Subramanianrsubram@ostatemail.okstate.edu(405) 762 3625www.linkedin.com/in/ravi-shankar-subramanian-b088a079

top related