Model-based Approaches for Independence-Enhanced Recommendation
2 likes1,929 views
The document discusses model-based approaches for enhancing recommendation independence in recommender systems, focusing on separating recommendations from sensitive features such as user demographics or content provider information. It outlines two primary methods: a regularization approach using probabilistic matrix factorization and a model-based approach that incorporates sensitive features into a latent class model. Experimental results demonstrate improvements in recommendation independence, although some accuracy trade-offs were observed.
![Recommendation Independence
7
No information about a sensitive feature influences the result
The status of the sensitive feature is explicitly excluded from the
inference of the recommendation result
Recommendation Independence
the statistical independence
between a recommendation result, R, and a sensitive feature, S
Ratings of items are predicted
under this constraint of recommendation independence
[Kamishima 12, Kamishima 13]
Pr[R | S] = Pr[R]
R ⫫ S
≡](https://image.slidesharecdn.com/2016-ws-icdm-pr-161211210252/85/Model-based-Approaches-for-Independence-Enhanced-Recommendation-7-320.jpg)
![Application
Adherence to Laws and Regulations
10
A recommendation service must be managed
while adhering to laws and regulations
suspicious placement keyword-matching advertisement
Advertisements indicating arrest records were more frequently
displayed for names that are more popular among individuals of
African descent than those of European descent
sensitive feature = users’ demographic information
Legally or socially sensitive information
can be excluded from the inference process of recommendation
Socially discriminative treatments must be avoided
[Sweeney 13]](https://image.slidesharecdn.com/2016-ws-icdm-pr-161211210252/85/Model-based-Approaches-for-Independence-Enhanced-Recommendation-10-320.jpg)
![Application
Fair Treatment of Content Providers
11
System managers should fairly treat their content providers
The US FTC has investigated Google to determine whether the search
engine ranks its own services higher than those of competitors
Fair treatment in search engines
Fair treatment in recommendation
Marketplace sites should not abuse their position to recommend their
own items more frequently than tenants' items
sensitive feature = a content provider of a candidate item
Information about who provides a candidate item can be ignored,
and providers are treated fairly
[Bloomberg]](https://image.slidesharecdn.com/2016-ws-icdm-pr-161211210252/85/Model-based-Approaches-for-Independence-Enhanced-Recommendation-11-320.jpg)
![Application
Exclusion of Unwanted Information
12
Filter Bubble: To fit for Pariser’s preference, conservative people are
eliminated from his friend recommendation list in FaceBook
sensitive feature = a political conviction of a friend candidate
Information about whether a candidate is conservative or progressive
can be ignored in a recommendation process
Information unwanted by a user is excluded from recommendation
[TED Talk by Eli Pariser, http://www.filterbubble.com/]](https://image.slidesharecdn.com/2016-ws-icdm-pr-161211210252/85/Model-based-Approaches-for-Independence-Enhanced-Recommendation-12-320.jpg)
![Çr(x, y) = 𝜇 + bx + cy + pxqÒ
y
Probabilistic Matrix Factorization
15
Probabilistic Matrix Factorization Model
predict a preference rating of an item y rated by a user x
well-performed and widely used
[Salakhutdinov 08, Koren 08]
For a given training dataset, model parameters are learned by
minimizing the squared loss function with an L2 regularizer.
cross effect of
users and itemsglobal bias
user-dependent bias item-dependent bias
≥
D (ri * Çr(xi, yi))2
+ 𝜆 Ò⇥Ò2
Prediction Function
Objective Function
L2 regularizer
regularization parameter
squared loss function](https://image.slidesharecdn.com/2016-ws-icdm-pr-161211210252/85/Model-based-Approaches-for-Independence-Enhanced-Recommendation-15-320.jpg)
![Prediction:
Latent Class Model
17
[Hofmann 99]
z
y
x
r
Latent Class Model: A probabilistic model for collaborative filtering
A basic topic model, pLSA
extended so as to be able to deal with
ratings r given by users x to items y
Çr(x, y) = EPr[rx,y][level(r)]
=
≥
r Pr[rx, y] level(r)
the r-th rating value
A rating value can be predicted by the expectation of ratings
Model parameters can be learned by an EM algorithm
latent topic variable](https://image.slidesharecdn.com/2016-ws-icdm-pr-161211210252/85/Model-based-Approaches-for-Independence-Enhanced-Recommendation-17-320.jpg)
![Recommendation Diversity
23
[Ziegler+ 05, Zhang+ 08, Latha+ 09, Adomavicius+ 12]
Recommendation Diversity
Similar items are not recommended in a single list, to a single user,
to all users, or in a temporally successive lists
recommendation list
similar items
excluded
Diversity
Items that are similar in a
specified metric are excluded
from recommendation results
The mutual relations
among results
Independence
Information about a sensitive
feature is excluded from
recommendation results
The relations between
results and sensitive values](https://image.slidesharecdn.com/2016-ws-icdm-pr-161211210252/85/Model-based-Approaches-for-Independence-Enhanced-Recommendation-23-320.jpg)
Model-based Approaches for Independence-Enhanced Recommendation
- 1. Model-based Approaches for Independence-Enhanced Recommendation Toshihiro Kamishima*, Shotaro Akaho*, Hideki Asoh*, and Issei Sato** http://www.kamishima.net/ *National Institute of Advanced Industrial Science and Technology (AIST), Japan **The University of Tokyo, Japan The 1st Int’l Workshop on Privacy and Discrimination in Data Mining in conjunction with ICDM2016 @ Barcelona, Spain, Dec. 12, 2016 1START
- 2. Independence-Enhanced Recommender Systems 2 Providing independent information is useful in recommendation Adherence to laws and regulations Fair treatment of content providers Exclusion of unwanted information Independence-enhanced Recommender System The absolutely independent recommendation is intrinsically infeasible, because recommendation always depends on the preference of a specific user ↓ This system makes recommendation so as to enhance independence with respect to a specific sensitive feature
- 3. Contributions 3 Our Previous Work We advocated a concept of independence-enhanced recommendation We developed a regularization approach to enhance recommendation independence This approach is applied to a probabilistic matrix factorization (PMF) model This Talk We propose another approach to enhance recommendation independence, a model-based approach A sensitive feature is embedded into a graphical model for recommendation, while maintaining independence between recommendation and sensitive information
- 4. Outline 4 A concept of recommendation independence Applications of recommendation independence Two approaches to enhance recommendation Independence A regularization approach a regularizer to constrain independence is introduced to a probabilistic matrix factorization model A Model-based approach a sensitive feature is embedded into a latent class model, while maintaining independence between a recommendation result and a sensitive value Experiments Related work Conclusions
- 5. Outline 5 A concept of recommendation independence Applications of recommendation independence Two approaches to enhance recommendation Independence A regularization approach a regularizer to constrain independence is introduced to a probabilistic matrix factorization model A Model-based approach a sensitive feature is embedded into a latent class model, while maintaining independence between a recommendation result and a sensitive value Experiments Related work Conclusions
- 6. Sensitive Feature 6 S : sensitive feature It is specified by a user depending on his/her purpose A recommendation result is independent from this sensitive feature Sensitive value is determined depending on a user and an item As in a case of standard recommendation, we use random variables X: a user, Y: an item, and R: a rating A sensitive feature is restricted to a binary type We adopt an additional variable for recommendation independence Ex. Sensitive feature = movie’s popularity / user’s gender
- 7. Recommendation Independence 7 No information about a sensitive feature influences the result The status of the sensitive feature is explicitly excluded from the inference of the recommendation result Recommendation Independence the statistical independence between a recommendation result, R, and a sensitive feature, S Ratings of items are predicted under this constraint of recommendation independence [Kamishima 12, Kamishima 13] Pr[R | S] = Pr[R] R ⫫ S ≡
- 8. dislike like dislike like Effect of Independence Enhancement 8 Standard recommendation Independence-enhanced r. two distributions are largely diverged distributions become close by enhancing independence The bias that older movies were rated higher could be successfully canceled by enhancing independence ✽ each bin of histograms of predicted ratings for older and newer movies
- 9. Outline 9 A concept of recommendation independence Applications of recommendation independence Two approaches to enhance recommendation Independence A regularization approach a regularizer to constrain independence is introduced to a probabilistic matrix factorization model A Model-based approach a sensitive feature is embedded into a latent class model, while maintaining independence between a recommendation result and a sensitive value Experiments Related work Conclusions
- 10. Application Adherence to Laws and Regulations 10 A recommendation service must be managed while adhering to laws and regulations suspicious placement keyword-matching advertisement Advertisements indicating arrest records were more frequently displayed for names that are more popular among individuals of African descent than those of European descent sensitive feature = users’ demographic information Legally or socially sensitive information can be excluded from the inference process of recommendation Socially discriminative treatments must be avoided [Sweeney 13]
- 11. Application Fair Treatment of Content Providers 11 System managers should fairly treat their content providers The US FTC has investigated Google to determine whether the search engine ranks its own services higher than those of competitors Fair treatment in search engines Fair treatment in recommendation Marketplace sites should not abuse their position to recommend their own items more frequently than tenants' items sensitive feature = a content provider of a candidate item Information about who provides a candidate item can be ignored, and providers are treated fairly [Bloomberg]
- 12. Application Exclusion of Unwanted Information 12 Filter Bubble: To fit for Pariser’s preference, conservative people are eliminated from his friend recommendation list in FaceBook sensitive feature = a political conviction of a friend candidate Information about whether a candidate is conservative or progressive can be ignored in a recommendation process Information unwanted by a user is excluded from recommendation [TED Talk by Eli Pariser, http://www.filterbubble.com/]
- 13. Outline 13 A concept of recommendation independence Applications of recommendation independence Two approaches to enhance recommendation Independence A regularization approach a regularizer to constrain independence is introduced to a probabilistic matrix factorization model A Model-based approach a sensitive feature is embedded into a latent class model, while maintaining independence between a recommendation result and a sensitive value Experiments Related work Conclusions
- 14. Formalizing Task 14 Predicting Ratings: a task to predict a rating value that a user would provide to an item Dataset Random variables: user X, item Y, rating R, sensitive feature S Prediction Function Dataset Prediction Function Standard Recommendation Independence-Enhanced Rec. D = {(xi, yi, ri)} D = {(xi, yi, ri, si)} Çr(x, y) Çr(x, y, s)
- 15. Çr(x, y) = 𝜇 + bx + cy + pxqÒ y Probabilistic Matrix Factorization 15 Probabilistic Matrix Factorization Model predict a preference rating of an item y rated by a user x well-performed and widely used [Salakhutdinov 08, Koren 08] For a given training dataset, model parameters are learned by minimizing the squared loss function with an L2 regularizer. cross effect of users and itemsglobal bias user-dependent bias item-dependent bias ≥ D (ri * Çr(xi, yi))2 + 𝜆 Ò⇥Ò2 Prediction Function Objective Function L2 regularizer regularization parameter squared loss function
- 16. Independence-Enhaned PMF 16 a prediction function is selected according to a sensitive value sensitive feature Çr(x, y, s) = 𝜇(s) + b(s) x + c(s) y + p(s) x q(s) y Ò Prediction Function Objective Function ≥ D (ri * Çr(xi, yi))2 * ⌘ indep(R, S) + 𝜆 Ò⇥Ò2 independence parameter: control the balance between the independence and accuracy independence term: a regularizer to constrain independence The larger value indicates that ratings and sensitive values are more independent Matching means of predicted ratings for two sensitive values
- 17. Prediction: Latent Class Model 17 [Hofmann 99] z y x r Latent Class Model: A probabilistic model for collaborative filtering A basic topic model, pLSA extended so as to be able to deal with ratings r given by users x to items y Çr(x, y) = EPr[rx,y][level(r)] = ≥ r Pr[rx, y] level(r) the r-th rating value A rating value can be predicted by the expectation of ratings Model parameters can be learned by an EM algorithm latent topic variable
- 18. Independence-Enhanced LCM 18 z y x r s z y x r s Independence-Enhancement by a Model-based Approach A sensitive variable is embedded into the original LCM A rating and a sensitive variable are mutually independent A user, an item, and a rating are conditionally independent given Z A type 2 model can more strictly enhance recommendation independence, because in addition to X and Y, Z depends on a sensitive variable Type 1 model Type 2 model
- 19. Outline 19 A concept of recommendation independence Applications of recommendation independence Two approaches to enhance recommendation Independence A regularization approach a regularizer to constrain independence is introduced to a probabilistic matrix factorization model A Model-based approach a sensitive feature is embedded into a latent class model, while maintaining independence between a recommendation result and a sensitive value Experiments Related work Conclusions
- 20. Experimental Conditions 20 Data Sets ML1M-Year: Movie preference data. A sensitive feature is whether movies’ release year is old or new. ML1M-Gender: Movie preference data. A sensitive feature is a gender of a user. Flixster: Movie preference data. A sensitive feature is whether a movie is popular or not. Sushi (not presented here) Evaluation measures MAE (Mean Absolute Error) Precision measure. The smaller is the better. KS (Statistic of the two-sample Kolmogorov-Smirnov test) Independence measure. The smaller is the better. The area between two empirical cumulative distributions of predicted ratings for S = 0 and S = 1.
- 21. Experimental Results 21 Independence indexes were improved compared to their original method, but their precisions were slightly sacrificed Contrary to our expectation, no clear differences between Type 1 and Type 2 models Independence seemed to be less strictly enhanced by a model based approach ML1M-Year ML1M-Gender Flixster MAE KS MAE KS MAE KS PMF 0.685 0.1687 0.685 0.0389 0.655 0.1523 PMF-r 0.697 0.0271 0.694 0.0050 0.653 0.0165 LCM 0.729 0.1984 0.729 0.0487 0.671 0.1787 LCM-mb1 0.717 0.0752 0.719 0.0243 0.672 0.0656 LCM-mb2 0.720 0.1030 0.720 0.0364 0.672 0.0656 PMF: original PMF, PMF-r: independence-enhanced PMF, LCM: original LCM, LCM-mb1: Type1 LCM, LCM-mb2: Type 2 LCM
- 22. Outline 22 A concept of recommendation independence Applications of recommendation independence Two approaches to enhance recommendation Independence A regularization approach a regularizer to constrain independence is introduced to a probabilistic matrix factorization model A Model-based approach a sensitive feature is embedded into a latent class model, while maintaining independence between a recommendation result and a sensitive value Experiments Related work Conclusions
- 23. Recommendation Diversity 23 [Ziegler+ 05, Zhang+ 08, Latha+ 09, Adomavicius+ 12] Recommendation Diversity Similar items are not recommended in a single list, to a single user, to all users, or in a temporally successive lists recommendation list similar items excluded Diversity Items that are similar in a specified metric are excluded from recommendation results The mutual relations among results Independence Information about a sensitive feature is excluded from recommendation results The relations between results and sensitive values
- 24. Independence vs Diversity 24 short-head long-tail short-head long-tail standard recommendation diversified recommendation Because a set of recommendations are diversified by abandoning short-head items, predicted ratings are still biased Prediction ratings themselves are unbiased by enhancing recommendation independence
- 25. Privacy-preserving Data Mining 25 recommendation results, R, and sensitive features, S, are statistically independent In a context of privacy-preservation Even if the information about R is disclosed, the information about S will not exposed mutual information between a recommendation result, R, and a sensitive feature, S, is zero I(R; S) = 0 In particular, a notion of the t-closeness has strong connection
- 26. Conclusions 26 Contributions We proposed a new model-based approach to enhance recommendation independence This approach was implemented into a latent class model Experimental results showed the successful enhancement of recommendation independence by this approach Future work Developing a regularization approach for a latent class model, and comparing the performance with a model-based approach Bayesian extension Acknowledgment We would like to thank for providing datasets for the Grouplens research lab and Dr. Mohsen Jamali. This work is supported by MEXT/JSPS KAKENHI Grant Number JP24500194 and JP15K00327, and JP16H02864.