recsplorer: recommendation algorithms based on precedence mining acm sigmod conference 2010 1

25
Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

Upload: natalie-elliott

Post on 14-Dec-2015

217 views

Category:

Documents


2 download

TRANSCRIPT

Page 1: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

1

Recsplorer: Recommendation Algorithms Based on Precedence Mining

ACM SIGMOD Conference 2010

Page 2: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

2

Outline

Introduction

Approach

Algorithms

Popularity Algorithm

Single Item Max-Confidence Algorithm

Joint Probabilities Algorithm Approximation

Joint Probabilities Support Variant

Joint Probabilities Hybrid Variant

Joint Probabilities Hybrid Reranked Variant

Evaluation

Conclusions

Page 3: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

3

Introduction

Recommender systems provide advice on products, movies…,and so on.

collaborative filtering (CF)

without regard to order

few items are rated by few users

precedence mining

based on temporal

does not suffer from the sparsity of ratings problem

Page 4: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

4

Approach

Page 5: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

5

Approach_Collaborative Filtering

Page 6: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

6

Approach_Precedence relationships

Page 7: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

7

definition

f(x):

the number of transcripts in T that contain x.

g(x; y):

the number of transcripts in T in which x precedes course y.

f(a)=2,f(b)=2

g(a,d)=2,g(e,f)=2,g(g,h)=2

Page 8: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

8

definition

# of transcripts containing a and x / # of transcripts containing a

But our user's transcript does not have x before a.(ignore #5)

Page 9: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

9

definition

Page 10: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

10

Top-k Recommendation Problem

Given a set T of transcripts over D for n users, the extra transcript T of a target user, and a desired number of recommendations k,our goal is to

Assign a score score(x) to every course ,

Using the score function, select the top k courses to recommend to the target user.

Page 11: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

11

RECOMMENDATION ALGORITHMS

Popularity Algorithm

Page 12: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

12

example

D = {a , b , c , d}

n = 50 students=6/50=0.12

=4/50=0.08

Page 13: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

13

RECOMMENDATION ALGORITHMS

Single Item Max-Confidence Algorithm

Page 14: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

14

example

D = {a , b , c , d} , T={a , b}

n = 50 students

Page 15: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

15

RECOMMENDATION ALGORITHMS

Joint Probabilities Algorithm Approximation

Page 16: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

16

example

D = {a , b , c , d} , T={a , b}

n = 50 students

Page 17: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

17

RECOMMENDATION ALGORITHMS

Consider a course x that has appeared in 1000 transcripts,

while y appeared in 10 transcripts.

Assume the student has not taken neither x nor y.

If there are 20 course (small) (big)

Page 18: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

18

RECOMMENDATION ALGORITHMS

Joint Probabilities Support Variant

where for any not-taken course x

if f(x)<θ(for some threshold)

then assign score(x) = 0

Page 19: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

19

RECOMMENDATION ALGORITHMS

Assume a set of courses all appearing in the transcript T .

Consider course x that we wish to recommend to a user.

It may be the case (especially when the data is sparse) that x is strongly suggested by courses

but we may still not recommend x because of .

Page 20: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

20

RECOMMENDATION ALGORITHMS

Joint Probabilities Hybrid Variant

Step 1:

assign a score of 0 to not-taken courses where

Step 2:

assign a score to a remaining course x we proceed as follows.

We set top-I(T) to be the top I courses from T

ranked by

Step 3:

Page 21: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

21

RECOMMENDATION ALGORITHMS

Reranked Hybrid Variant

Step 1:

Take the set O of courses recommended by the Joint Probabilities Hybrid Variant

Step 2:

pick the best m courses recommended

Step 3:

for each courses x in the remaining courses,

Assign

Page 22: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

22

RECOMMENDATION ALGORITHMS

Step 4:

Order this top-m set in inverse order of f(x).

i.e. if x and y are in the set, x is ranked higher

than y if f(x) < f(y)

Step 5:

for the first coure in top-m set, assign score= for the second coure in top-m set, assign score=

for the final coure in top-m set, assign score= ,

and so on

Page 23: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

23

EVALUATION

Page 24: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

24

EVALUATION

Page 25: Recsplorer: Recommendation Algorithms Based on Precedence Mining ACM SIGMOD Conference 2010 1

25

CONCLUSIONS

The Single Item Max Confidence approach has the highest precision when we have little information about the student.

Joint Prob. Hybrid works best with more information at hand.

we found that algorithms beat popularity-based recommendations and collaborative filtering.