Download presentation
Presentation is loading. Please wait.
Published byShannon Snow Modified over 6 years ago
1
Advisor: Prof. Shou-de Lin (林守德) Student: Eric L. Lee (李揚)
Collaborative Filtering Based Model for Privacy-Preserving Course Recommendation Advisor: Prof. Shou-de Lin (林守德) Student: Eric L. Lee (李揚)
2
Outline Motivation and Introduction Baselines Our Methods
Experiment and Results Conclusion
3
Motivation There are courses for students to choose in 2012, it costs lots of time for students searching courses they want to take!! We aims to develop a course recommendation system using only course records to protect the privacy of users in this paper.
4
Example My Model Recommended list Machine Learning Statistics Stu Cou
… Most recommend Stu Cou Time Calculus 2009-1 Statistics 2010-2 1 Probability 2010-1 2 Physics 2008-2 .. … My Model Least recommend Former and current Students’ Course Record Output recommended list for each current students
5
Related Work Collaborative Filtering [1] [2] [3] (our baselines)
Privacy-Preserving Course Recommendation Our Work!!! [1] Deshpande, Mukund, and George Karypis. "Item-based top-n recommendation algorithms." ACM Transactions on Information Systems (TOIS) 22.1 (2004): [2] Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer, (8), [3] Rendle, S., Freudenthaler, C., Gantner, Z., & Schmidt-Thieme, L. (2009, June). BPR: Bayesian personalized ranking from implicit feedback. In Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence (pp ). AUAI Press.
6
One Class Collaborative Filtering Problem
1 ? 5 ? 3 2 4 1 Vs. OCCF problem Traditional Recommendation Problem
7
Lack of positive training data in the next grade
1st ~3rd grade 4th grade Former students Current Students Current students we want to predict don’t have training data in the 4th grade.
8
Courses’ Order Traditional CF assume each course actions are i.i.d
However, we think the course records in the consecutive year have a closer relationship.
9
Our Framework P1 P2 P3 MF Matrix Factorization (MF) Memory-Based CF P2
Bayesian Personal Ranking (BPR) Reduce Search Space P2 Two Stage Method P1: OCCF P2: Lack of training data in next grade P3: courses’ order P3 Course Network Regularization
10
Modified Memory Based CF
Origin CF Modified CF Search space for similar students All other students Only Former students Course Candidates All the courses taken by the compared students Only 4th grade courses taken by the compared students
11
Modified Memory-Based Collaborative Filtering
1st ~ 3rd grade 4th grade Former Students Calculate the similarities Current Students Student u’s course record 𝑠∈𝑓𝑜𝑟𝑚𝑒𝑟 𝑠𝑡𝑢𝑑𝑒𝑛𝑡𝑠 𝑠𝑖𝑚(𝑢,𝑠) Score(u, 𝑐) ≔ ∗𝑝𝑟𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒(𝑠,𝑐)
12
Our Framework P1 P2 P3 Matrix Factorization (MF)
Bayesian Personal Ranking (BPR) P1: OCCF P2: Lack of training data in next grade P3 : courses’ order P2 Two Stage Method P3 Course Network Regularization
13
MF Minimizing Square Error
2 ? 1 3 4 ≈ X Q (courses) R P (student) min P,𝑄 𝑢,𝑖 ∈𝑟𝑎𝑡𝑖𝑛𝑔𝑠 [ ( 𝑅 𝑢,𝑖 − 𝑃 𝑢 ∙ 𝑄 𝑖 ) 2 ]+ 𝐶( 𝑃 2 + |𝑄| 2 ) Koren, Yehuda, Robert Bell, and Chris Volinsky. "Matrix factorization techniques for recommender systems." Computer 8 (2009):
14
Our Framework P1 P2 P3 Matrix Factorization (MF)
Bayesian Personal Ranking (BPR) P2 Two Stage Method P1: OCCF P2: Lack of training data in next grade P3: Courses’ order P3 Course Network Regularization
15
One class collaborative Filtering (OCCF) Problem
Course 1 Course 2 Course 3 Course 4 Student 1 ? 1 Student 2 Student 3 Student 4 Characteristics: The rating matrix only has one class(take / or not take the course).
16
Bayesian Personal Ranking : Intuition
student u 𝑗∈𝐼 𝑢 − 𝑖∈𝐼 𝑢 + … … Maximize the likelihood!!! 𝐼 𝑢 + :𝑇ℎ𝑒 𝑐𝑜𝑢𝑟𝑠𝑒𝑠 𝑡ℎ𝑎𝑡 𝑡ℎ𝑒 𝑠𝑡𝑢𝑑𝑒𝑛𝑡𝑠 ℎ𝑎𝑣𝑒 𝑡𝑎𝑘𝑒𝑛 𝐼 𝑢 − :𝑇ℎ𝑒 𝑐𝑜𝑢𝑟𝑠𝑒𝑠 𝑡ℎ𝑎𝑡 𝑡ℎ𝑒 𝑠𝑡𝑢𝑑𝑒𝑛𝑡𝑠 ℎ𝑎𝑣𝑒 𝑛 ′ 𝑡 𝑡𝑎𝑘𝑒𝑛
17
Bayesian Personal Ranking: Intuition
We want to maximize the AUC that + rank higher than ? Let 𝐷 𝑆 ≔ 𝑢,𝑖,𝑗 |(𝑖,𝑗)∈ 𝐼 𝑢 + × 𝐼 𝑢 − max 𝜃 (𝑢,𝑖,𝑗)∈ 𝐷 𝑠 𝑃𝑟(𝜃| 𝑅 𝑢𝑖 > 𝑅 𝑢𝑗 ) m𝑖𝑛 𝜃 (𝑢,𝑖,𝑗)∈ 𝐷 𝑠 ln(1+ 𝑒 −(𝑦 𝑢𝑖 𝜃 − 𝑦 𝑢𝑗 (𝜃)) ) + ? Rendle, Steffen, et al. "BPR: Bayesian personalized ranking from implicit feedback." Proceedings of the Twenty-Fifth Conference on Uncertainty in Artificial Intelligence. AUAI Press, 2009.
18
Matrix Factorization using Bayesian Personal Ranking (BPRMF)
≈ X Rating Matrix P Q 𝑢,𝑖,𝑗 ln 1+ exp − 𝑃 𝑢 𝑄 𝑖 − 𝑄 𝑗 C 1 P + C 2 Q
19
Our Framework P1 P2 P3 Matrix Factorization (MF)
Bayesian Personal Ranking (BPR) P1: OCCF P2: Lack of training data in next grade P3: Model the course relation P2 Two Stage Method P3 Course Network Regularization
20
Recall of BPRMF Bayesian Personal Ranking … …
student u … … 𝐼 𝑢 + 𝐼 𝑢 − What courses has been taken by student u Who take the course c ? Course c latent features Student u latent features
21
Problems Lack of training data in the next grade.
1st ~3rd grade 4th grade Former students Current Students Lack of training data in the next grade. Therefore, these courses are all in the set of 𝐼 𝑢 − for current students
22
Problems Courses in 1st ~3rd grade Student u Courses in 4th grade Because the students u doesn’t have the positive training data in 4th grade, the final model tends to have an unreasonable bias to rank courses in 4th grade lower.
23
Two Stage Method - Motivation
Students The latent features of course A The latent features of course B The latent features of two courses would be similar if they are taken by the same group of students
24
Two Stage Method - Motivation
Although BPRMF have bias on 4th grade courses, it still learn the type of courses. If a student often take certain type of courses in the 1st ~ 3rd grade, he may take the same type of courses in the 4th grade.
25
Two Stage Method - Intuition
Learn the type for courses Rank the type of courses the current students like Notation: P => The latent matrix for students Q => The latent matrix for courses
26
First Stage P Q min 𝑃,𝑄 𝑢,𝑖,𝑗 ln 1+ exp − 𝑃 𝑢 𝑄 𝑖 − 𝑄 𝑗 +𝐶|𝑃|
Former students Get the Courses’ latent features Current Students P Q min 𝑃,𝑄 𝑢,𝑖,𝑗 ln 1+ exp − 𝑃 𝑢 𝑄 𝑖 − 𝑄 𝑗 𝐶|𝑃|
27
Second Stage Q P min 𝑃 𝑢,𝑖,𝑗 ln 1+ exp − 𝑃 𝑢 𝑄 𝑖 − 𝑄 𝑗 +𝐶|𝑃|
Former students Get the current students’ latent feature Current students Q P min 𝑃 𝑢,𝑖,𝑗 ln 1+ exp − 𝑃 𝑢 𝑄 𝑖 − 𝑄 𝑗 𝐶|𝑃|
28
Review of Two Stage Methods
X Use all the training course records Q P Use the course records of current students only X P Q
29
Our Framework P1 P2 P3 Matrix Factorization (MF)
Bayesian Personal Ranking (BPR) P1: OCCF P2: Lack of training data in next grade P3: Courses’ Order P2 Two Stage Method P3 Course Network Regularization
30
Review of Two Stage Methods
First Stage Learn the type of courses Second Stage Rank what the type of courses the students like to take
31
Idea Observation: Some courses will be very likely to be taken after certain courses in a consecutive year. And the Courses are often belongs to the same type Idea: Let these courses have similar course latent features
32
Recommender with social regularization
Idea : The behaviors of the neighbors might be similar. More likely to happen Less likely to happen
33
Use the regularization term to let the latent features similar
Individual based regularization 𝛽 2 𝑖 ∈𝑛𝑜𝑑𝑒𝑠 𝑓∈𝑁(𝑖) 𝑙𝑖𝑛𝑘(𝑖,𝑓)|| 𝑄 𝑖 − 𝑄 𝑓 || 2 Where N(i) is the set of neighbors of node i link(i,f) is the weight between node i and node f Ma, Hao, et al. "Recommender systems with social regularization." Proceedings of the fourth ACM international conference on Web search and data mining. ACM, 2011.
34
Objective Function 𝑢,𝑖,𝑗 ln 1+ exp − 𝑃 𝑢 𝑄 𝑖 − 𝑄 𝑗 C 1 P + C 2 Q + β 2 𝑖 𝑁 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (𝑖) | 𝑄 𝑖 − 𝑄 𝑗 | 2 ∗F𝑜𝑙𝑙𝑜𝑤(𝑖,𝑗) 𝑾𝒉𝒆𝒓𝒆 𝒇𝒐𝒍𝒍𝒐𝒘 𝒊,𝒋 𝒊𝒔 𝒕𝒉𝒆 𝒘𝒆𝒊𝒈𝒉𝒕 𝒐𝒇 𝒕𝒉𝒆 𝒍𝒊𝒏𝒌 𝒃𝒆𝒕𝒘𝒆𝒆𝒏 𝒊 𝒂𝒏𝒅 𝒋 𝑾𝒉𝒆𝒓𝒆 𝑵 𝒕𝒉𝒓𝒆𝒔𝒉𝒐𝒍𝒅 𝒊 𝒊𝒔 𝒂 𝒔𝒆𝒕 𝒕𝒉𝒂𝒕 𝒇𝒐𝒍𝒍𝒐𝒘 𝒊,𝒋 >𝒕𝒉𝒓𝒆𝒔𝒉𝒐𝒍𝒅 Non Prunning Prunning
35
Example Name 1st grade 2nd grade 3rd grade Alice AB C D Bob A CD B
Christine AD Follow (A,B) = # B courses taken after courses A in one year # courses taken after courses A in one year = 1 4 =0.25
36
Final Course Network of the simple small university
Name 1st grade 2nd grade 3rd grade Alice AB C D Bob A CD B Christine AD
37
Combining Social Network to MF model
Pruning ≈ X
38
Stochastic Gradient Descent
Target Function: 𝑢,𝑖,𝑗 ∈ 𝐷 𝑠 ln 1+ exp − 𝑃 𝑢 𝑄 𝑖 − 𝑄 𝑗 C 1 P 2 + C 2 Q β 2 𝑖 𝑁 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (𝑖) | 𝑄 𝑖 − 𝑄 𝑗 | 2 ∗𝑓𝑜𝑙𝑙𝑜𝑤(𝑖,𝑗) For each update: 𝑠𝑖𝑔= − 𝑒 − 𝑘 𝑃 𝑢𝑘 ( 𝑄 𝑖𝑘 − 𝑄 𝑗𝑘 ) 1+ 𝑒 − 𝑘 𝑃 𝑢𝑘 ( 𝑄 𝑖𝑘 − 𝑄 𝑗𝑘 ) 𝑃 𝑢 ≔ 𝑃 𝑢 −𝑎𝑙𝑝ℎ𝑎 ∗(𝑠𝑖𝑔∗ 𝑄 𝑖𝑘 − 𝑄 𝑗𝑘 + 𝐶 1 𝑃 𝑢 ) 𝑄 𝑖 ≔ 𝑄 𝑖 −𝑎𝑙𝑝ℎ𝑎∗(𝑠𝑖𝑔 ∗ 𝑃 𝑢 + 𝐶 2 𝑄 𝑖 + 𝛽 𝑛𝜖 𝑁 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (𝑖) 𝑄 𝑖 − 𝑄 𝑛 ∗𝑓𝑜𝑙𝑙𝑜𝑤(𝑖,𝑛) ) 𝑄 𝑗 ≔ 𝑄 𝑗 −𝑎𝑙𝑝ℎ𝑎∗(𝑠𝑖𝑔 ∗( −𝑃 𝑢 )+ 𝐶 2 𝑄 𝑖 + 𝛽 𝑛𝜖 𝑁 𝑡ℎ𝑟𝑒𝑠ℎ𝑜𝑙𝑑 (𝑖) 𝑄 𝑗 − 𝑄 𝑛 ∗𝑓𝑜𝑙𝑙𝑜𝑤(𝑗,𝑛) )
39
Personalized page rank
Parameters: 𝛼 , T(the set of courses the student has taken): Assuming the student has taken C, D and E courses previously Choose a course from the neighbors with probability proportional to the weight of the link A B 𝑃(𝛼) E C D 𝑃(1−𝛼) Randomly choose a course from set T with a uniform distribution
40
Overview of our Methods
CF- based model Stu Cou Time 2 2011-1 3 2012-2 1 4 2011-2 .. … Apply BPR Two Stage Method Social Regularization Perform personalized page rank Construct the course networks Graph-based model
41
Ensemble CF-Based Model Linear RankSVM Graph-Based Model
42
Data Set Description Use the triplets
<student_id, course_id, Time(Year-Semester)> Year # Students # Course Actions 民國97 4736 311283 民國98 4686 299772 民國99 4555 285561
43
Training, Validation and Testing set
1~3 grade 4th grade Use the students in the fourth grade in 民國98 as validation set to choose the parameters 民國97 民國98 Test our model use the course records of fourth grade students enter NTU in 民國99 民國99
44
Evaluation The average AUC (Area under Reciever Opterating Characteristics) of students True positive Rate Ranking of the courses XOXXOOOXXOOOXXO X: the student doesn’t take the course O: the student does take the course False Positve Rate
45
Memory Based Collaborative Filtering
Let 𝐴 𝑠 be the set of courses of student a take in s semesters Let 𝐵 𝑠 be the set of courses of student b take in s semesters Number of Intersection 𝑠𝑖𝑚(𝑎,𝑏)= 𝑠∈𝑠𝑒𝑚𝑒𝑠𝑡𝑒𝑟𝑠 𝐴 𝑠 ∩ 𝐵 𝑠 Jaccard Similarity sim(a,b)= 𝑠∈𝑠𝑒𝑚𝑒𝑠𝑡𝑒𝑟𝑠 𝐴 𝑠 ∩ 𝐵 𝑠 / 𝐴 𝑠 ∪ 𝐵 𝑠 TFIDF similarity 𝑠𝑖𝑚(𝑎,𝑏)= 𝑠∈𝑠𝑒𝑚𝑒𝑠𝑡𝑒𝑟𝑠 𝑐∈ 𝐴 𝑠 ∩ 𝐵 𝑠 𝑇𝐹𝐼𝐷𝐹(𝑐)
46
Memory-based CF Methods AUC Most Popular 0.817208 Normal CF 0.744253
Number of Intersection Jaccard TFIDF TFIDF_Whole (X以偷看到testing) Jaccard_Whole
47
Result of Model-Based Model
Methods AUC One Stage BPRMF Two Stage BPRMF Two Stage BPRMF with Course Network P-value Target Baseline P-value Two Stage BPRMF One Stage BPRMF 𝟑.𝟐𝟒 ∗ 𝟏𝟎 −𝟏𝟐 Two Stage BPRMF with Course Network 𝟎.𝟎𝟏𝟑𝟒𝟓𝟔
48
Personalized Pagerank and Ensemble
Methods AUC Personalized Page Rank Ensemble
49
Result – AUC with 95% confident interval
Methods AUC with 95% confident interval Most Popular ± Normal CF ± Number of Intersection Similarity ± Jaccard Similarity ± TFIDF Similarity ± Personalized Page Rank ± BPRMF (one stage) ± BPRMF (two stage) ± BPRMF with course network prior (one stage) ± BPRMF with course network prior (two stage) ± Ensemble ±𝟎.𝟎𝟎𝟏𝟎𝟏𝟑
50
Conclusion and Future work
We propose an accurate privacy-preserving course recommendation system. We can think the course record as a recommendation problem with seasonal product. Our method has the potential to apply on these kinds of data sets.
51
Thank you for listening
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.