Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation...

20
Amazon.com Recommendations Item-to-Item Collaborative Filtering CIS 601: Graduate Seminar Prof. S. S. Chung Presented By:- Amol Chaudhari CSU ID 2682329

Transcript of Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation...

Page 1: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Amazon.com RecommendationsItem-to-Item Collaborative Filtering

CIS 601: Graduate Seminar

Prof. S. S. Chung

Presented By:-

Amol Chaudhari

CSU ID 2682329

Page 2: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Outline

� Introduction

� Problems

� RecommendationAlgorithms

� Comparison

� Conclusion

Page 3: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Introduction

Page 4: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Introduction

� Best for use in e-Commerce websites.

� Input – Customers Interest.

� Output-List of recommendation.

� Attributes like items viewed, demographic data, subject interest etc.

Page 5: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Challenges for algorithm :

� Retailer with huge amount of data, catalogue andmillions of customers

� Real-time output should be in no more than halfsecond with high quality results.

� New customers have limited info about product

� Older customer with huge Information.

� New data gets added rapidly. Customers data isVolatile

Page 6: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Three common approaches to solving the problem� Traditional collaborative filtering

� Cluster models

� Search-based methods

Page 7: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Traditional Collaborative Filtering

� Represent a customer as an N-dimentional vector of items

� Typically multiplies the vector component by inverse frequency.

� Generates recommendations based on few customers who are most similar to the user.

Page 8: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Traditional Collaborative Filtering

Disadvantage:� examines only a small customer sample...

� If most popular or unpopular items are discarded they will never appear as recommendation.

� Very expensive and time complexity is O(MN) for worst case and for best case O(M+N)

Page 9: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Cluster Models

� Divide the customer base into manysegments and assign the user to the segment containing the most similar customers

� Segments created using clustering.

� Groups similar customers

� Any clustering or learning algorithm can be used.

Page 10: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Cluster Models

Advantages:-

� better onlinescalability and performance than Collaborative

Disadvantages:-

� The recommendation quality is low when it runs offline.

� Equal expensive to collaborative

Page 11: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Search Based Methods

� Content based method

� Constructs a search query , find other popular items by similarity.

� Impractical to base query on all items.

� If the user has few purchases or ratings,search- based

recommendation algorithms scale and perform well

Page 12: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Disadvantages:� too general

� too narrow

Page 13: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Item-to-Item Collaborative Filtering

� Amazon uses recommendations as a targeted marketing tool in many email campaign and web site pages.

� Eg. “your recommendations”

Page 14: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized
Page 15: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Item-to-Item Collaborative Filtering

How it Works

� Instead of matching the user to similar customer , it matches each of the users purchased items.

� Combine items to make list.

� Similar Item table which are purchased together.

� Build product to product matrix by iterating through all item pairs and similarity metric.

Page 16: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Item-to-Item Collaborative Filtering

� Algorithm

� Increase the scalability and performance

Page 17: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Scalability: A Comparison

Traditional Collaborative filtering:� Online computation , � Impractical on large dataset, unless dimensionality

reduction , sampling , and partitioning is one.

Cluster models:� Offline computation� Quality is poor� Increase number of segments to improve quality but cost

increases.

Search based Models:� Build keyword, category author indexes offline , but no

recommendation as per interest titles.� Scale poorly for customers with numerouspurchases and

ratings

Page 18: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Scalability: A Comparison

Item-to-item collaborative filtering:

� fast for extremely large data set

� creates the similar-items table offline

� performs well with limited userdata

� quality is excellent

Page 19: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

Conclusion:

� Recommendation algorithm provides an effective form of targeted marketing by creating personalized experience for each customer.

� Large retailers like Amazon it is scalable over larger dataset.

� It takes only few sub seconds processing time to generate online recommendation.

� In future it is expected that retail industry will use this algorithm for targeted marketing , both offline and online.

Page 20: Amazon.com Recommendationscis.csuohio.edu/~sschung/CIS601/AmolAmazonRecommendation.pdfRecommendation algorithm provides an effective form of targeted marketing by creating personalized

THANK YOU