JP2015521769A - Method and apparatus for obfuscating user demographics - Google Patents

Abstract

A method of obfuscating accurate detection of new user demographic information that provides ratings to a digital content service having a recommendation system includes training an inference engine that detects five demographic information. The training set includes movie ratings and demographic information from multiple other users. The new user inputs a rating, such as a movie rating, and the inference engine determines demographic information for the new user. The obfuscation engine then adds movie ratings to the recommendation system such that the recommendation system's inference engine fails to accurately detect the new user's demographic information.

Description

(Cross-reference of related applications)
This application is based on US Provisional Patent Application No. 61 / 664,618, “Method and Apparatus for Obfuscating User Demographics Based on Ratings, filed June 21, 2012, and a method of obfuscating user demographics based on ratings and "Device" priority is claimed and is incorporated herein by reference in its entirety.

  The present invention generally relates to user profiling and user privacy in a recommendation system. The present invention relates more particularly to the inference of demographic information.

  User demographic reasoning has been studied on various types of user-generated data in different contexts. In the context of interactive networks, graph structures have proved useful for inferring demographics using linkbase information from social network data from blogs and Facebook. Other tasks rely on text characteristics obtained from user writing to infer demographics.

  The main disadvantage of text-based reasoning is that these methods are not applicable because most users do not write reviews. Similarly, the recommendation system may not be able to obtain the social network of the user who wants to infer details.

  It can be seen that a method for inferring user demographics based on as little information as possible is desired. The present invention relates to such an inference method.

  This summary is a brief introduction to some of the concepts described further below in the detailed description of the invention. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to limit the scope of the claimed subject matter.

  The present invention includes a method and apparatus for obfuscating demographic information that can be determined from user ratings for digital content. In one embodiment, gender information may be determined from the user's movie rating. In order to address privacy concerns, an obfuscation method and an obfuscation device are presented. The obfuscation method includes training an inference engine that communicates with the obfuscation engine. The inference engine determines demographic information using a training data set that includes movie ratings and demographic information from multiple other users. Thereafter, movie ratings from new users are received. Movie ratings from specific users are received without demographic information. New user demographic information is determined using a trained inference engine. The additional movie rating is then added to the user generated rating. Additional ratings are generated against the results of the user's demographic information when performed by an external inference engine. The external reasoning engine may be part of a recommendation system that recommends movies for users to watch.

  Additional features and advantages of the present invention will become apparent from the following detailed description of exemplary embodiments, which proceeds with reference to the accompanying drawings.

  The foregoing summary, as well as the following detailed description of exemplary embodiments, can be better understood when read with reference to the appended drawings, which are given by way of illustration only and limit the claimed invention. Not what you want.

2 illustrates an exemplary environment embodiment for an inference engine according to aspects of the present invention. Fig. 4 shows receiver operating characteristic (ROC) plots of different classifiers for the Flixster training data set. Figure 7 shows receiver operating characteristic (ROC) plots of different classifiers for the Movielens training data set. The accuracy improvement with the size of the Flixster training data set is shown. The cumulative distribution function (CDF) of the confidence value of the Flexster is shown. FIG. 6 illustrates an example flow diagram for use of an inference engine according to an aspect of the present invention. 2 illustrates an example inference engine according to an aspect of the present invention. 2 illustrates an example of a first embodiment of an obfuscation engine environment. 2 illustrates an example of a second embodiment of an obfuscation engine environment. 2 illustrates an example of an obfuscation engine according to an aspect of the present invention. FIG. 4 illustrates an example flow diagram of an obfuscation engine according to an aspect of the present invention.

  In the following detailed description of various exemplary embodiments, reference is made to the accompanying drawings that form a part hereof. The accompanying drawings show various possible embodiments of the present invention by way of example. It should be understood that other embodiments may be utilized and structural and functional changes may be made without departing from the scope of the invention.

  Profiling users via demographic information such as gender, age, income, race, etc. is very important for targeted advertising and content distribution tailored to individuals. The recommendation system can also benefit from such information to make personalized recommendations. However, users of recommendation systems often do not provide demographic information voluntarily. This may be intentional in order to protect privacy, and may be unintentional such as bothersome and indifferent. Therefore, the conventional collaborative filtering method of collecting user ratings from a large number of users and extracting meaningful information from the appearing patterns does not use such information depending only on the ratings provided by the users.

  At first glance, disclosing ratings to the recommendation system may seem less harmful. Users will certainly benefit from this disclosure. That is, the ability to find appropriate content / items. Nevertheless, inferring user demographics from user activity by correlating user demographics with user activity on social networks, blogs, microblogs, etc. requires a significant amount of work Met. Therefore, it is natural to ask whether demographic information, including not only age, gender, and race but also political direction, can be inferred from information disclosed in the collaborative filtering system. In fact, the mere fact that the user interacted with the item (eg, watched a particular movie, heard a particular song, or bought a product), regardless of the rating value, correlated with demographic information You may let me.

  The possibility of such inferences being successful has several important implications. On the one hand, from the point of view of the recommender, profiling a user with demographic information leads to several applications. That is, such profiling can generate additional revenue from advertisements as well as recommendations. Advertisers are inherently interested in targeting specific demographic groups. The present invention relates to such an inference technique. The information that the user wants to infer is assumed to be the user's gender, but the method of the present invention is also applied when inferring other demographic characteristics (age, race, political orientation, etc.). Also, although specific embodiments are directed to movie ratings, this is only an example. Any type of rating may be used, including but not limited to ratings for songs, digital games, products, restaurants, and the like. For the sake of easy and clear understanding, an example in which demographic information is determined using movie ratings is mainly used, but the present invention can also be applied to other types of ratings.

  FIG. 1 illustrates an environment for an exemplary system 100, ie, an inference engine described herein. Other environments are possible. The system 100 of FIG. 1 shows a recommendation system 130 that recommends content to users on the network 120. Common examples of recommendation systems include content recommendation systems operated by content providers such as Netflix “registered trademark”, Hulu “registered trademark”, and Amazon “registered trademark”. Typically, the recommendation system 100 provides candidate digital content to subscriber users. Such content can include streaming video, DVD mailing, books, articles, and merchandise. As an example, in the case of streaming video, a candidate movie can be recommended to the user based on the user's past movie selection or selection of user profile characteristics. As an example embodiment, consider the case of streaming video.

  In the context of the present invention, inference engine 135 may be a data processing device that can infer demographic information from non-demographic information provided by user 125 sending movie ratings to recommendation system 130. The inference engine 135 serves to infer demographic information by processing movie ratings provided by the user 125. In the example case, the demographic information considered is gender. However, those skilled in the art will recognize that other demographic information may be inferred in accordance with aspects of the present invention. Such demographic information includes, but is not limited to, age, race, political direction, and the like.

  In accordance with aspects of the present invention, inference engine 135 operates using training data (105, 110-115, respectively) acquired via users 1, 2-n, as described below. These users provide movie rating data and demographic information to the inference engine 135 via the recommendation system 130. The training data set may be acquired over time by the users 105-115 using the recommendation system. Alternatively, the inference engine can import the training data set by directly importing one or more data loads via the input port 136. Port 136 may be used to input a training data set from a network, training disk, or other data source that contains training data.

  The inference engine 135 processes the training data set using an algorithm. The inference engine 135 then uses the input of the user 125 (user X) including movie ratings. Movie ratings include one or more that can identify a movie, such as a movie title, movie index or reference number, and a rating value for inferring demographic information about the user 125. A “movie title” or, more generally, a “movie identifier” as used in this description is the name or title of a movie, show, documentary, series program, digital game, or other digital content viewed by the user 125. Or an identifier such as a database index. The rating value is a subjective measure of the viewed digital content determined by the user 125. Usually, the rating value is a quality assessment performed by the user 125, and is rated according to the criteria of 1 to 5. 1 is a low subjective score and 5 is a high subjective score. Other evaluation methods are used in the same way, such as 1-10 numerical evaluation, alphabetic evaluation, 5-star evaluation, 10-level evaluation of half stars, or evaluation with words from “not good” to “great” Those skilled in the art will recognize that this may be the case. Note that in accordance with aspects of the present invention, the information provided by user 125 does not include demographic information, and inference engine 135 determines user 125 demographic information from user 125 movie ratings only.

In accordance with aspects of the present invention, the training data set is used to educate the inference engine 135. The training data set may be available in both the recommendation system 130 and the inference engine 135. Here, the characteristics of the training data set are described. The training data set is N = {1,. . . , N}, each user giving a rating to a subset of movies M in the catalog. S _i ⊆M represents a set of movies in which the rating of user iεN is in the data set, and r _ij , jεS _i represents a rating given to movie jεM by user iεN. In addition, for each iεN, the training data set also includes a binary variable y _i ε {0,1} that indicates the gender of the user (bit 0 is mapped to a male user). The training data set is assumed to be pure, i.e. neither the rating label nor the gender label is touched and obfuscated.

The recommendation mechanism via space is commonly used in commercial systems and is assumed to be matrix factorization. Although matrix factorization is used as an example, any recommendation mechanism may be used. Alternative recommendation mechanisms include neighborhood methods (user clustering), item context similarity, or other mechanisms known to those skilled in the art. The rating of the set M \ S ₀ is generated by factoring the provided rating with the training set's rating matrix. More specifically, each user iεN∪ {0} is associated with a latent feature vector u _i εR ^d . Associated with each movie jεM is a latent feature vector v _j εR ^d . The normalized mean square error is defined by the following equation.

Here, μ is an average rating of the entire data set. The vectors u _i and v _j are constructed by minimizing the MSE by slope descent. The values d = 20 and λ = 0.3 are used. Profiling both the user and the movie in this way, the rating of user 0 for movie jεM \ S ₀ ′ is predicted through <u ₀ , v _j > + μ.

  Consider an example of two training data sets, Flixster and Movielens. Flixster is a publicly available online social network for movie ratings and reviews. Using Flixster, users can enter demographic information into their profiles and share their movie ratings and reviews with friends and the public. The data set has 1 million users, of which 34,200 users share their age and gender. This subset of 34,200 users has rated 17,000 movies and provided 5.8 million ratings. 12,800 men have provided 2.4 million ratings and 21,400 women have provided 3.4 million ratings. Using Flixster, the user can provide half-star ratings, but the ratings are rounded up to an integer of 1-5 to match the evaluation data sets. Another data set is Movielens. This second data set is publicly available from the Grouplens ™ research team. This data set consists of 3700 movies and 1 million ratings by 6000 users. 4331 men have provided 750,000 ratings and 1709 women have provided 250,000 ratings.

  In order to determine the demographic information, the inference engine uses a classifier. As noted above, demographic information can include many characteristics. As an example of a demographic, gender determination is described as an embodiment of the present invention. However, determination of another or more demographic characteristics of the user is also within the scope of the present invention.

To train the classifier, for _{j∈S i, x ij = r ij} , otherwise, so that x _ij = 0, the eigenvector each user i∈N the training data set x _i ∈R ^M Associate with. Recalling that the binary variable y _i indicates the gender of user i, gender acts as a dependent variable in the classification. XεR ^NXM represents a matrix of eigenvectors, and Yε {0,1} ^N represents a gender vector.

It is.

  Next, class prior probability classification will be described. Class prior probability classification serves as a reference method for evaluating the performance of other classifiers. Given a data set in which the gender classes of the population are unevenly distributed, this basic classification strategy classifies all users as having a majority gender. this is,

Is equivalent to using equation (1) based on the generated model P (y | x) = P (y) estimated from the training set set as

  Next, the mixed naive Bayes according to an embodiment of the present invention will be described. This is an alternative to the above-mentioned multiple naive Bayes, and the inventor has called mixed naive Bayes. This model is based on the assumption that the user usually has a normal distribution rating. More specifically,

Next, the use of logistic regression in the present invention will be described. A significant drawback of all the above Bayesian methods is that it assumes that movie ratings are independent. To address this shortcoming, the inventor used logistic regression. Recalling that linear regression produces a set of coefficients β = {β ₀ , β _1... , Β _M }, the classification of user iεN with eigenvector x _i is

Calculate and do. The user is classified as female if P _i <0.5, and otherwise classified as male. The value P _i also serves as a confidence value for the classification of user i. One major advantage of using logistic regression is that the coefficient β captures the degree of correlation between each movie and class. In this case, a large positive β _j indicates that movie j is correlated with class men, and a small negative β _j indicates that movie j is correlated with class women. The normalization parameters are selected to have at least 1000 movies correlated with each gender with non-zero coefficients.

  In machine learning, a support vector machine (SVM) is a supervised learning model that has an associated learning algorithm that analyzes data to recognize patterns and is used for classification and regression analysis. As is well known in the art, SVM intuitively finds hyperplanes that separate users belonging to different genders that minimize the distance from the hyperplane of misclassified users. SVM retains many of the advantages of logistic regression. That is, the coefficient is generated without assuming independence in the feature space. Since the feature space (number of movies) is already quite large, a linear SVM is used for classifier evaluation. Performing a logarithmic search over the parameter space (C), the inventor found that C = 1 gave the best results.

  All algorithms were evaluated for both the Flixster and the Movielens data sets. Ten-fold cross validation was performed, average accuracy and recall were calculated for both data sets, and average receiver operating characteristic (ROC) curves were calculated through the ten-segment data. For ROC, the true positive rate is calculated as the proportion of men correctly classified from men in the data set, and the false positive rate is calculated as incorrectly classified men from women in the data set. Table 1 gives an overview of the classification results for three metrics, AUC, accuracy, and recall. Table 2 shows the same results divided by gender. ROC curves are shown in FIGS. 2a and 2b. Table 1 summarizes the classification results for the three metrics, AUC, precision, and recall. Table 2 shows the same results divided by gender.

  As can be seen from the ROC curves, SVM and logistic regression are superior to any Bayesian model because SVM and logistic regression curves are superior to other curves for both data sets. Specifically, logistic regression was the best for Flixster, and SVM was the best for Movielens. The performances of the Bernoulli model, the mixed model, and the multinomial model are not significantly different from each other. This result is further confirmed by the AUC values in Table 1. This table also shows the weaknesses of simple class prior probabilistic models that are inferior in performance by all other methods.

  In general, the accuracy in a classification task depends on the number of true positives (ie, the number of items correctly labeled as belonging to the positive class) to the total number of elements labeled as belonging to the positive class (ie, true Divided by the sum of the positive and the false positive, which is an item incorrectly labeled as belonging to the positive class. Recall in this context is true positives, the total number of elements that actually belong to the positive class (i.e., false positives, items that should have been labeled as true positives and belong to positive classes). Defined as the number divided by the sum of negative).

  Table 2 shows that logistic regression outperforms all other models for both genders in terms of accuracy and recall. For the users of Movielens, SVM outperforms all other algorithms, and logistic regression is second best. In general, inferences perform well with respect to the majority gender (Fixster is female and Movielens is male) in each data set. This is particularly evident with SVM. SVM shows very high recall for the majority class and low recall for the minority class. The mixed model is significantly improved with the Bernoulli model and the results are similar to the multinomial model. This indicates that using a Gaussian distribution may not be a sufficiently accurate estimate for the rating distribution.

  The effect of training set size was evaluated. Since 10-fold cross validation is used, the training set is large relative to the evaluation set. Using Flixster data, we evaluate the impact of the number of users in the training set size on the inference accuracy. In addition to 10-fold cross validation with an evaluation set having 3000 users, 100-fold cross validation was performed using an evaluation set of 300 users. Furthermore, for the first time from 100 users, the training set was gradually increased by adding 100 users each time it was repeated.

  FIG. 2c represents the accuracy of the logistic regression inference for Flixster for the size of the two evaluation sets. The figure shows that for both sizes, an algorithm with an accuracy of over 70% requires about 300 users in the training set, and an accuracy of over 74% for 5000 users in the training set. Show that you reach. This indicates that a relatively small number of users is sufficient for training.

  Having detailed the characteristics of the SVM classifier and the linear regression classifier with respect to two obtainable data sets and obtained desirable results, a novel method and apparatus for implementing an inference engine has been invented. FIG. 3 illustrates a method according to an aspect of the present invention that generates demographic information from a user's rating without demographic information and uses the results for useful purposes. The ultimate goal of using such generated demographic information includes targeted advertisements for the user 125 and / or enhanced recommendations via the recommendation system 130.

  The method 300 of FIG. 3 begins at step 305 by inputting a training data set having ratings representing demo users and demographic information into an inference engine. FIG. 1 shows an inference engine 135 that is part of the recommendation system 130. This step may be accomplished using a recommendation system connection 137 to the network 120 or may be accomplished by direct input to the inference engine 135 via port 136. When input is made via the recommendation system network connection 137, the training data set may be one that stores demographic information and rating information (movie ratings, or any other digital content rating) one by one. It is also possible to load one or more training data sets having demographic information and rating information of at least one user. When inputting directly into the inference engine 135 via the input port 136, the data may be one or more downloaded training data sets of at least one user. In step 210, the recommendation system 135 trains the inference engine using information from the training data set. If the inference engine 135 receives a download directly via port 136, step 210 can be omitted. In either case, steps 205 and 210 represent training the inference engine 135 using a training data set having both user demographic information and user rating information.

  In step 315, a new user, such as user 125, not included in the training data set interacts with the recommendation system 130 to provide only ratings. As described above, these ratings may be, for example, movie ratings having movie identifier information and objective rating value information. The rating provided by the user 125 does not have demographic information that the inference engine retrieves. After the new user 125 inputs the rating into the recommendation system, in step 320, the inference engine 135 uses the classification algorithm to determine demographic information for the new user based on the new user's rating. The classification algorithm is preferably one of the aforementioned support vector machines (SVM) or logistic regression.

  Once demographic information for a new user is determined, the demographic information determined, such as gender, may be used for many useful purposes. Two examples are shown in FIG. In one example, the demographic information determined in step 320 is used by the recommendation system 130 in step 325 to make enhanced recommendations to new users. For example, if the recommendation system 130 is a movie recommendation system operated by Netflix (trademark) or Hulu (trademark), demographic information such as gender is specific to the user's gender for viewing by the new user. It can be used to select a more exacted movie. Alternatively, the recommendation system 130 can use the demographic information determined in step 320 for targeted advertising to the new user in step 330. For example, when the gender of a new user is determined, an advertisement specialized for that gender may be made targeting the new user. Such advertisements may include perfume purchase discount suggestions for women and shaving device purchase discounts for men. The recommendation system may have access to potential advertisements from an internal database, an external database, or a network server (not shown).

  As a useful action to take advantage of the demographic information extracted from the ratings provided by the new user 125, either or both of steps 325 and 330 may be performed. Steps 315 to 330 may be repeated for each new user who uses the service of the recommendation system 130. A user receiving an enhanced recommendation or advertisement from a recommendation system receives the enhanced recommendation or advertisement on a display device associated with the user, such as user 125. Such user display devices are well known and are associated with home television equipment, stand-alone televisions, personal computers and handheld devices such as personal digital assistants, laptops, tablets, mobile phones, web notebooks, etc. including.

  FIG. 4 is a block diagram of the inference engine 135. The inference engine 135 interfaces with the recommendation system 130 as shown in FIG. The inference engine interface 410 serves to connect the communication component of the inference engine 135 to the communication component of the recommendation system 130. The link from the inference engine interface 410 at 405 to the recommendation system may be a serial link or a parallel link, as known to those skilled in the art, and may be embedded or external. Thus, the inference engine may be coupled with the recommendation system or may be separated from the recommendation system. The interface port 405 allows the recommendation system 130 to provide training data to the inference engine 135 and to provide inference results to the recommendation system. An alternative training data set interface is an input port 136 through which training data can be input in an easy-to-use manner from a network or other digital data source such as a storage media source.

  The processor 420 provides a calculation function to the inference engine 135. The processor may be any form of CPU or controller that utilizes the communication between the elements of the inference engine to control the inference engine communication and computation processes. Those skilled in the art recognize that bus 415 provides a communication path between the various elements of inference engine 135 and that other point-to-point interconnections can also be implemented.

  Program memory 430 may provide a repository of memory associated with method 300 of FIG. The data memory 440 can provide a repository for storing information such as training data sets, downloaded ones, uploaded ones, scratchpad calculations, and the like. Those skilled in the art will recognize that the memories 430 and 440 may be coupled, may be separate, and may be incorporated in whole or in part into the processor 420. The processor 420 performs the steps of the method 300 by executing instructions, such as computer instructions, using program memory storage and search properties to generate demographic information for use by the recommendation system 130.

  The estimator 450, which may be separate or part of the processor 420, serves to provide computational resources for determining demographic information from new user ratings. As such, the estimator 450 can provide computational resources for a classifier, preferably SVM or logistic regression. The estimator may provide the data memory 440 or the processor 420 with provisional calculations in determining new user demographic information. This provisional calculation includes the probability of demographic information related to the new user who gave only his rating information. The estimator 450 may be hardware, but is preferably a combination of hardware and firmware or software.

  Given a relatively small training set, the inference algorithm accurately predicts the user's gender with 70% to 80% accuracy. However, the above techniques for determining demographic information from user ratings may raise concerns about user privacy. Some users may want to obfuscate their demographic information so that it is not reliably determined. An obfuscation mechanism that protects detectable demographic information from reliable detection is described below.

  FIG. 5a shows an exemplary environment 500 where an obfuscation mechanism for the inference engine 135 of the recommendation system may exist. An obfuscation mechanism can exist in multiple locations. The obfuscation mechanism may be in a cloud connected to the network 120 or in the user 125 device. When present in the cloud (not shown), the obfuscation mechanism is a network service provided to many users. When present in the user device, the obfuscation mechanism basically includes an inference engine with additional computational elements. For example, as shown in FIG. 5, the obfuscation engine 126 monitors recommendations from the user 125 and adds additional ratings to the user's ratings to reduce the accuracy of the inference engine present in the recommendation system 130. be able to.

  In another embodiment, a content aggregator that delivers content to the user may also serve to protect the user's demographic information by providing an obfuscation engine along with a content aggregation service. FIG. 5b shows such a content aggregator service. In the configuration of FIG. 5b, content aggregator 560 can connect to network 120 via link 555 to gain access to digital content that user 125 may be interested in. User 125 may access the content aggregator directly via link 582 or may access it via network 120 via link 581. In either case, the content aggregator acts as a provider of digital content for the user 125 and collects a fee to provide the content to the user. The content provider may be a recommendation system 130. Thus, the content aggregator 560 acts as a conduit for digital content that can be rated by the user 125. As a privacy service, a content aggregator can provide an obfuscation service to a user via an obfuscation engine 570 that operates with an inference engine 575. The obfuscation engine 570 allows the user 125 to add an additional obfuscation rating to the rating sent to the recommendation system 130 when the user 125 rates the digital content acquired from the recommendation system 130 that is a content provider. It works to obfuscate 125 demographic information. The added rating goes against the exact determination of demographic information. Accordingly, the inference engine 135 associated with the recommendation system cannot accurately determine the demographic information of the user 125 via the user 125 rating.

  FIG. 5 c shows an exemplary block diagram 590 of the obfuscation engine 599. The obfuscation engine 599 interfaces with a network such as 120 in FIG. 5b via a network interface 591. The network interface 591 can access user data such as user ratings and a training data set via a network such as the Internet. Therefore, the receiving unit of the network interface can receive the training data and the rating provided by the user such as movie rating. Furthermore, the additional rating generated by the rating generation unit 595 can be transmitted to the network by the transmission unit in the network interface 591. In one embodiment, the additional ratings and user-provided ratings are sent to the recommendation system 130 where an inference engine such as 135 in FIG. 5b accurately determines the user's demographic information. Disturb.

  The processor 592 provides a calculation function to the obfuscation engine 599. The processor may be any form of CPU or controller that utilizes communication between elements of the obfuscation engine to control the obfuscation engine communication and computation processes. Those skilled in the art recognize that bus 597 provides a communication path between the various elements of obfuscation engine 599 and that point-to-point connections can also be implemented instead of a bus architecture.

  Program memory 593 may provide a repository to memory associated with method 600 of FIG. Data memory 594 may provide a repository for storing information such as training data sets, downloaded ones, uploaded ones, or scratch pad calculations. One of ordinary skill in the art will recognize that the memories 593 and 594 may be coupled, separate, or incorporated in whole or in part into the processor 591. The processor 591 generates obfuscated data that violates the accurate determination of the user's demographic information by executing a method such as method 600 using instructions in the program memory. The obfuscated data is transmitted to the network-based recommendation system via the network interface 591.

  Inference engine 596 may be separate from or part of processor 592 and serves to provide computational resources for determining demographic information from new user ratings. As such, the inference engine may be similar to the inference engine of FIG. 4 and may utilize the computational resources shown in FIG. 5c. The rating generator 595 operates to generate ratings used by the obfuscation techniques described below. Specifically, the rating generator imitates user ratings, but generates additional ratings that are contrary to the accurate determination of demographic information by the inference engine in the recommendation system. As described above, the rating generation unit creates a rating to be transmitted to an external inference engine such as an inference engine of the recommendation system (see FIG. 1). The additional rating sent to the external reasoning system acts to interfere with the rating from the user so that it is not easy to accurately determine the demographic information of the new user. The obfuscation engine 599 may be hardware based, but is preferably a combination of hardware and firmware or software.

The characteristics of the obfuscation engine are described below. A user such as user 125 with an index of 0 views and rates a digital content item such as a movie. Assuming that the universe of movies that the user can rate includes a catalog of M movies, the user will have a catalog M = {1, 2,. . . , M}, a subset S ₀ is rated. r _{0j ∈R} represents the rating of movie j∈S ₀ , and the user's rating profile is defined as a set of (movie, ranking) pairs Η ₀ ≡ {(j, r _0j ): j∈S ₀ }. . Referring to FIG. 5, the user, Eta ₀ (i.e., 139) to submit to the obfuscation mechanism, obfuscation mechanisms of _S'0 ≠ S _0, the rating profile after change _Η'0 = {(j, r _0j ′): jεS ₀ ′} (ie, 138) is output. Briefly, the obfuscation, to the following (a) _Η'0 is deduced user demographic information gender, etc. from which it (b) _Η'0 used for proper recommended to the user The goal is to successfully balance the two conflicting objectives.

More specifically, the obfuscated rating profile “ _0” is considered to be presented to the recommendation system 130 having a module that implements the gender inference engine 135. Recommendation system 135, using a _Η'0, to predict the user rating on M\S' _0, in some cases, a user would recommend a likely movie interested. Gender inference engine 135, and profiling the user by using the same _Η'0, is a classification mechanism to label male or female.

  Although the implementation of the recommendation system 130 may be known, the obfuscation engine 126 and the gender reasoning engine 135 are not known. As a first step in this problem, both the recommendation system 130 and the inference engine 135 take a simple approach that they do not realize that any kind of obfuscation is taking place. Both of these mechanisms take a profile Η ′ by “value on the surface” and do not retrograde the “true” profile Η.

  As described above, the recommendation system 130 and the inference engine 135 have access to the training data set. It is assumed that the training data set is pure, that is, the rating and demographic information such as gender labels are not touched or obfuscated. The obfuscation engine 126 may see part of the training set. In one embodiment, the training data set is public and the obfuscation engine 126 has full access to the training data set.

  In general, the confidence value of the classifier used in the inference engine 135 is an obstacle that the obfuscation engine needs to overcome when trying to hide demographic information such as gender from the classifier. The obfuscation engine attempts to reduce this confidence value of the inference engine 135 classifier. Therefore, it is evaluated whether the classifier has different confidence values when the classifier outputs an accurate or inaccurate classification. Regarding the evaluation of the classifier used in the inference engine, FIG. 2d shows the cumulative distribution function (CDF) of confidence values for correct and incorrect classification. From FIG. 2d, the confidence value is higher when the classification is accurate, and the median of the confidence values is 0.65 for the incorrect classification and 0.85 for the accurate classification. Furthermore, close to 20% of the correct classification has a confidence value of 1.0, which is less than 1% for an incorrect classification.

The obfuscation engine receives as input the rating profile ユ ー ザ_{i of} user i, the parameter k representing the number of allowed changes, and the information from the training set, and minimizes the impact on the quality of received recommendations. On the other hand, it has a mechanism for outputting a rating profile Η _i that has been changed so that it is difficult to infer gender of the user. In general, such a mechanism can change Η _i by adding, deleting, or changing movie ratings. Here we focus on a setting where the obfuscation engine is only allowed to add k movie ratings. This is because movie deletion is impractical for most services, and rating changes are not as effective as adding ratings when viewing events are a powerful predictor of user demographic attributes. Since the number of movies that a user has rated with their profile varies from user to user (some users are small), do not use a fixed number k, but correspond to a given percentage of movies in the user's rating profile Use additional numbers. In order to add a movie to the user's profile, the obfuscation engine needs to make two important decisions: the movie to add and the rating assigned to each movie.

  These added movie ratings are referred to as additive ratings. Additional ratings are contrary to the accurate determination of the user's demographic information. The rating value of an additional rating rating pair (title, rating value) is not assigned as “noise” but has some useful value. For example, if this rating corresponds to an average rating for all users or a predictive rating for a particular user (using matrix factorization), the rating value will be what if that user watches the movie. It is a reasonable estimate of whether the rating was made.

To build an obfuscation scheme, the obfuscation mechanism has full access to the training data set and can use the training data set to extract information to select movies and ratings to add This is the premise. Given the choice of movies for the obfuscation engine, the inventor has chosen three strategies for choosing movies. Each strategy is a set S _i of the movie that the user i was rating, and the number k of the movie to be added, and a list L _F ordering of movie correlated with the correlated list L _M and the woman ordered the movie to men as input The set S ′ _i after receiving and changing the movie is output. Here, S _i ⊆S ′ _i . The lists L _M and L _F are stored in descending order L _M ∪L _F → R in the value of the scoring function w. Here, w (j) indicates how strongly the movie jεL _M ∪L _F correlates with the related gender. A specific example of a scoring function is to set w (j) = β _j , where β _j is a coefficient of movie j obtained by learning a logistic regression model from a training data set It is. This instantiation of the scoring function is used for evaluation. Further, it is assumed that k <min (| L _M |, | L _F |) − | S _i | and L _M ∩L _F =.

The movie selection process is as follows. Initialize S ′ _i = S _i for a given female (or male) user i. Each strategy is repeated retrieves the movie j from L _M (or L _F), if the J∈S' _i, until the k movie is added, to add j to _S'i. The set S ′ _i is a desirable output. The three strategies differ in how movies are taken from the ordered list of movies.

Given a random strategy for a given female (male) user i, the movie j is randomly and randomly extracted from the list corresponding to the opposite gender L _M (L _F ), regardless of the score of the movie. Given a sampling strategy, movies can be sampled based on the distribution of scores associated with movies in the list corresponding to the opposite gender. For example, if there are _three movies j ₁ , j ₂ , and j ₃ with scores 0.5, 0.3, and 0.2, respectively, j ₁ is extracted with a probability of 0.5. Given the greedy strategy, you can pick the highest scoring movie from the list corresponding to the opposite gender.

  Given the rating assignment of rating value pairs (title, rating value), binary events on whether a profile is included or not in a profile (indicating whether or not it was viewed) are almost as strong as gender inference I noticed that it was a signal for This has important implications for obfuscation mechanisms that need to do two things: determine the movie to add to the user profile and determine the rating value to give the movie. This finding suggests that the choice of movies to add can have a significant impact on preventing gender inference. However, if the actual rating does not significantly affect gender inference, a rating value can be selected that helps maintain the quality of the recommendation. Assuming that the binary event itself, whether the profile includes or does not include a movie, is a signal for gender inference, then rating the additional movie that has less impact on recommendations provided to the user via the recommendation system 130 Can be assigned. Two rating assignments are proposed: average movie rating and predictive rating.

In the average movie rating, the obfuscation mechanism calculates the average rating of all movies jεS ′ _i −S _i using the available training data, and the calculated average rating is the rating profile after the change of user i. Add to Η´ _i . In predictive rating, the obfuscation mechanism calculates the latent factor of the movie by performing matrix factorization on the training data set, and predicts the user's rating using the latent factor. The predicted rating of all of the movie j∈S' _i -S _i, is added to the _Η'i.

  In the above, it is assumed that the obfuscation engine 126 is not restricted from accessing the training set. However, the above mechanism does the following quantities: (a) an ordered list of movies correlated with men and movies correlated with women for movie selection; and (b) average movie ratings for rating assignments. Only the movie's latent factors for predicting the user's movie rating are needed. Note that this information can be found from publicly available data sets, such as the Netflix Prize ™ data set. Assuming that users of such public data sets are statistically similar to users of a particular recommendation system as a whole, the complete training data set used by the inference engine 135 is fully There is no need to assume secure access.

All permutations of the proposed movie selection and rating assignment strategies were evaluated. Evaluate the value of k corresponding to 1%, 5% and 10% | Si | for each user i. Score movie list L _M and L _F is set to the corresponding logistic regression coefficients.

  Obfuscation increases privacy by reducing the performance of gender inference. Table 4 shows the inference accuracy for all three movie selection strategies (ie, random, sampling, greedy) when the assigned rating is the average movie rating. The accuracy is calculated using 10-fold cross validation. That is, the model is trained on pure data and tested against obfuscated data. Since the reasoning accuracy is highest for a logistic regression classifier, the logistic regression classifier would be a natural choice as the inference mechanism of the recommendation system. Adding only 1% additional rating using a greedy strategy, the accuracy drops to 15% (ie, decreases by 80%) for the Fixster data set compared to 76.5% for pure data. ) Adding 10% additional rating brings the accuracy closer to zero. The trade-off between privacy and utility for different obfuscation mechanisms shows that adding only 1% rating to the user's profile reduces inference accuracy by 80%.

  Thus, if the obfuscation mechanism selects a movie according to a greedy strategy, adding a small number of movies is sufficient for gender obfuscation. Even when selecting a movie using a random strategy (ignoring the movie's score, and thus the logistic regression coefficient), adding only 10% of the opposite gender-correlated movies can improve the accuracy of gender inference (76 Sufficient to reduce by 63% (to an accuracy of .5% to 28.5%). A similar trend is seen for the Movielens data set.

The obfuscation mechanism uses an ordered list that better corresponds to the inference mechanism's idea that the movie is correlated with men or women. However, in general, obfuscation mechanism does not know what inference algorithm is used, the list of such L _M and L _F may seldom match such idea within the inference algorithm. An obfuscation mechanism is evaluated in this scenario, along with a multinomial naive Bayes classifier and an SVM classifier. As can be seen in Table 4, the obfuscation is still performing well, and the inference accuracy of the multinomial classifier is 71% for the Flixster data set (using 10% additional rating and greedy strategy). From 42.1% to 762.1% for the Movielens data set.

  We considered the effect on the quality of recommendations seen by users when they obfuscate their gender. This effect is measured by calculating the root mean square error (RMSE) of the factorization of a matrix of submitted test sets of 10 ratings for each user. Again, 10-fold cross validation was performed. Here, 9/10 is the user's data is pure and 1/10 is the rating with additional noise. That is, Η ′ is used for 1/10 of the user and Η is used for the rest. This is equivalent to assessing the change in RMSE for 10% of users of the system that obfuscated their gender. Overall, the inventor has discovered that obfuscation has only a negligible impact on RMSE. With respect to Flixster, RMSE increased with the addition of ratings compared to the case without additional ratings, but was negligible. For the Movielens training data set, RMSE is slightly reduced due to the additional rating. This can occur because adding additional ratings increases the density of the original rating matrix, which may improve the performance of the matrix factorization solution. The additional rating is not arbitrary, but is another explanation because it is meaningful to some extent (ie, the average of all users). For both datasets, the change in RMSE is not significant, with a maximum of 0.015 for Flixster (10% additional rating for random strategy) and 10% additional for Movielens (for sample strategy). The main result is that the maximum rating is 0.058. Therefore, the obfuscation engine maintains the quality of recommendation for users of the recommendation system.

  Consider the proposed trade-off between obfuscation privacy and utility. Here, the desired high level of privacy reduces the accuracy of gender inference, and the higher the utility, the lower the RMSE that is often used as a proxy for high quality recommendations. Upon evaluation, the inventor has found that the utility decreases with increasing privacy with respect to the Flixster training data set. As noted above, the use of the Movielens training data set increases utility, but only slightly, with increased privacy. An obfuscation mechanism can significantly reduce the accuracy of gender inference, and it causes only a small change in the quality of the recommendation.

  Maintaining the quality of the recommendation is an attractive feature for the obfuscation engine. In one evaluation, consider the trade-off when rating assignment corresponds to a “predictive rating” approach. The motivation behind this rating assignment is that, in principle, this obfuscation does not change the RMSE compared to the RMSE for unmodified data. In other words, with this selection of rating assignments, there is no trade-off with respect to utility front. Table 5 shows the accuracy of gender inference when using this rating assignment. This result is similar to the result in Table 4 where the rating assignment is average movie rating. For the Movielens training data set, the accuracy of gender reasoning is slightly lower than the predictive rating. For example, for a 1% additive rating greedy strategy, the accuracy of the logistic regression classifier drops from 57.7% to 48.4%. This advantage comes without sacrificing the quality of the recommendation. In conclusion, experimental evaluation shows that with a small amount of additional ratings, it is possible to protect the user's gender by obfuscation without significantly changing the quality of recommendations received by the user.

  FIG. 6 illustrates an exemplary method 600 for creating a set of user ratings (title, rating value) that can hide the demographic information of the user from accurate detection. Also, the advantage of this method is that using the inference engine 135 of the recommendation system 130 does not adversely affect the recommendations that will be received. The method begins at step 605 by introducing a rating training set from another user. The training data set includes both other users' ratings (title, rating value) and demographic information. In step 610, the training data set is used to train an inference engine such as 575 in FIG. 5b or 596 in FIG. 5c. The trained inference engine can determine demographic information for the user 125. As such, the trained inference engine somewhat emulates the functionality of an inference engine in a recommendation system, such as 130 in FIG.

  After training the inference engine, the obfuscation engine is ready for use by new users. In step 615, a new user who is not a user in the training data set provides a rating to the obfuscation engine. As a result, the obfuscation engine receives ratings such as movie ratings. The received movie ratings are only (title, rating value) rating pairs, and do not include demographic information of new users.

  In step 620, an inference engine, such as 575 or 596, uses a classification algorithm to determine demographic information for the new user based on the user's rating. In step 625, the obfuscation engine generates a rating that violates the correct determination of the demographic information by another inference engine. That is, the generated rating can be added to the user's rating and is an additional rating that helps obfuscate the user's detectable demographic information. To give a simple example, if the inference engine infers the gender of the user 125 as female, the additional rating generated by the obfuscation engine will provide data that infers the user's gender incorrectly. Accordingly, an external inference engine such as the recommendation system inference engine cannot accurately determine the gender demographic information of the new user 125. Thus, additional ratings are contrary to accurate detection of demographic information for new users.

  The additional rating is sent to the recommendation system (RS) by the obfuscation engine at step 630. This has the effect of obfuscating the demographic information of the user 125 detected by the inference engine of the recommendation system 130. This obfuscation includes not only the user's normal rating generated by an external inference engine such as 135 in FIG. 5b, but also an additional rating with a rating pair (title, rating value) that violates the determination of accurate demographic information. Occurs because it receives. That is, the additional rating serves to prevent the inference engine from accurately determining the user's demographic information. In accordance with aspects of the present invention, the recommendation system 130 with the inference engine 135 is prevented from making an accurate determination of the user's demographic information using additional ratings. However, the quality of recommendation received by the user 125 from the recommendation system 130 is not significantly degraded by the addition of additional ratings. Basically, the quality of recommendations received by the user 125 from the recommendation system 130 is maintained in the same way when additional ratings are added compared to when no additional ratings are included. Steps 615 to 630 may be repeated for new users. Thus, a large number of new users can obfuscate demographic information by the method 600.

  Although a specific architecture has been shown for the obfuscation engine implementation of FIGS. 5a, 5b, 5c, implementations such as component function distribution, component integration, location in the server as a service to the user regarding user privacy, etc. Those skilled in the art will recognize that there are alternatives. Such an option is equivalent to the function and structure of the configuration shown and described.

Claims (14)

In a method of obfuscating demographic information of a specific user, performed by an obfuscation engine,
Training an inference engine communicatively coupled to the obfuscation engine to determine demographic information using a training data set including ratings from other users and demographic information;
Receiving a rating from the specific user that includes only rating information and sent to a recommendation system for evaluating the rating from the specific user;
Determining the demographic information of the particular user from the rating provided by the particular user;
Generating a rating against the determined demographic information of the particular user by the obfuscation engine;
Transmitting the generated rating to the recommendation system;
The generated rating of the specific user obfuscates the determination of the demographic information of the specific user by the recommendation system;
Method.   The method of claim 1, wherein the rating from the particular user is a movie rating that includes a movie title and movie rating value information and no demographic information.   The method of claim 1, wherein the demographic information of the particular user is gender information.   The method of claim 1, wherein transmitting the generated rating to the recommendation system includes transmitting an additional rating that is against the determined demographic information of the particular user.   The method of claim 4, wherein the additional rating is about 10% of a movie rating provided by the particular user.   The method of claim 1, wherein the determining step includes determining the demographic information of the particular user using a classifier.   7. The method of claim 6, wherein the classifier is one of a support vector machine, a logistic regression algorithm, and a Bayesian approach such as a naive Bayes model, a polynomial model, a mixed model.   The method of claim 1, wherein the generated rating maintains a recommendation system quality. An obfuscation device that obfuscates an accurate determination of demographic information of a particular user providing movie ratings to a recommendation system via the device,
A receiver in the network interface for inputting a training data set including movie ratings and demographic information from a plurality of other users;
A processor having access to memory for determining demographic information based on the movie rating by executing a program using an inference engine;
A rating generation unit for generating an additional rating against the determined demographic information;
A transmitter in the network interface for transmitting both the movie rating provided by the user and the additional rating to the recommendation system;
A combination of the user-provided rating and the additional rating prevents the recommendation system from determining the demographic information;
Obfuscation device.   The apparatus of claim 9, wherein the apparatus is part of a user device.   The apparatus of claim 9, wherein the movie rating includes movie title information and a movie rating value.   The apparatus of claim 1, wherein the determined demographic information of the particular user is gender information.   The apparatus of claim 1, further comprising a classifier that assists the processor in determining the demographic information for the particular user.   The apparatus of claim 1, wherein the classifier is one of a support vector machine and a logistic regression algorithm.

Images