Classifications

• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
  • H04N21/20—Servers specifically adapted for the distribution of content, e.g. VOD servers; Operations thereof
  • H04N21/25—Management operations performed by the server for facilitating the content distribution or administrating data related to end-users or client devices, e.g. end-user or client device authentication, learning user preferences for recommending movies
  • H04N21/251—Learning process for intelligent management, e.g. learning user preferences for recommending movies
  • H04N21/252—Processing of multiple end-users' preferences to derive collaborative data
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N5/00—Computing arrangements using knowledge-based models
  • G06N5/02—Knowledge representation; Symbolic representation
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N5/00—Computing arrangements using knowledge-based models
  • G06N5/04—Inference or reasoning models
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
  • G06Q30/00—Commerce
  • G06Q30/02—Marketing; Price estimation or determination; Fundraising
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
  • H04N21/40—Client devices specifically adapted for the reception of or interaction with content, e.g. set-top-box [STB]; Operations thereof
  • H04N21/47—End-user applications
  • H04N21/475—End-user interface for inputting end-user data, e.g. personal identification number [PIN], preference data
  • H04N21/4756—End-user interface for inputting end-user data, e.g. personal identification number [PIN], preference data for rating content, e.g. scoring a recommended movie
• • H—ELECTRICITY
  • H04—ELECTRIC COMMUNICATION TECHNIQUE
  • H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
  • H04N21/00—Selective content distribution, e.g. interactive television or video on demand [VOD]
  • H04N21/60—Network structure or processes for video distribution between server and client or between remote clients; Control signalling between clients, server and network components; Transmission of management data between server and client, e.g. sending from server to client commands for recording incoming content stream; Communication details between server and client 
  • H04N21/65—Transmission of management data between client and server
  • H04N21/658—Transmission by the client directed to the server
  • H04N21/6582—Data stored in the client, e.g. viewing habits, hardware capabilities, credit card number
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N20/00—Machine learning

Definitions

• Recommender systems are fast becoming one of the cornerstones of the Internet. In a world with ever increasing choices, they are one of the most effective ways of matching users with items. Today, many websites use some form of such systems. Recommendation engines ask users to rate certain items, e.g. books, movies or music, infer the ratings of other items from this data, and use these inferred ratings to recommend new items to users.
• items e.g. books, movies or music
• differential privacy In order to allow for improved privacy of users providing ratings a new approach also referred to as differential privacy has been developed, recognising that, although perfect privacy is impossible to be had, a user may be put in control of the information provided through deliberately corrupting sensitive data before submitting ratings, and declaring to do so.
• US201 1/0064221 discloses a differential privacy recommendation method following the centralised model in which noise is added centrally in the correlation engine only after the correlation has been done. Further information on the design of recommender systems with differential privacy under the centralised model can be found in F. McSherry and I.
• WO 2008/124285 suggests a recommendation system using a sample group of users that is a sub-group of a larger group, for rating items. The rated items are suggested to other users if the rating from the sub-group exceeds a preset value.
• US 201 1/0078775 discloses adjusting a trust value associated with content through a trust server based on locally collected credibility information.
• the present invention provides improved privacy under the local model, in which users store their data locally, and differential privacy is ensured through randomization under the control of the user before submitting data to the recommender system.
• the invention provides a user-adjustable degree of privacy while still allowing for generating recommendations with a decent accuracy.
• user ratings generated under privacy at the user's location, of items are requested by a content provider or a recommendation engine for making recommendations to other users.
• the request is done as bulk request, i.e. a list of items, without knowledge which or indeed whether any of the items in the list have been "consumed” or rated by the user.
• "consumed” includes any activity involving an item that allows for rating the item.
• the user's device provides an aggregate rating in response.
• the user can add a value indicating the likelihood that the returned rating is exact and reliable, i.e. a "trust value”.
• the trust value may be provided for individual items in the rated list, or as a global value for the entire list.
• the content provider receives the rated item lists and the associated trust value from a plurality of users and clusters items from the received lists into clusters of items similarly "liked" by a number of users in according with their ratings and the "trust value".
• the content of each cluster is likely to be liked by the set of users who also liked other contents within this cluster.
• the content provider or recommendation engine returns the results of the clustering to the users.
• a user's device locally stores its response provided to the content provider or to the recommendation engine and uses this reliable information for extracting items contained in the cluster such that recommendations can be made to each individual user.
• Aspects of the invention pertain to improving the accuracy of content recommendation depending on whether the basic data used for clustering originates from an information-scarce or an information-rich environment, while maintaining improved privacy.
• Information-rich in this context relates to a situation in which the rated item is known to have a rating assigned to it, while the rating itself is privatized.
• Information-scarce as used in the present specification relates to an environment, in which a user provides ratings only for a small fraction of content items of a set of content items, and possibly even a varying fraction. Also, no information may be provided whether a user has consumed an item at all.
• [N] is used as a reference to a set ⁇ 1 , 2,..., N ⁇ of N ordered items
• [U] is used as a reference to a set of U users.
• A denotes the matrix of user/item ratings, each row corresponding to a user and each column corresponding to an item. For simplicity it is assumed that Aij ⁇ ⁇ 0, 1 ⁇ . This could, for example, correspond to 'like/dislike' ratings.
• a randomized function ⁇ X ⁇ Ythat maps data X ⁇ Xto Y ⁇ Y is said to be e-differential private, or e-DP, if for all values y ⁇ Y in the range space of ⁇ , and for all 'neighbouring' data x, x' the following (in-)equation holds true:
• 'neighbouring' data is chosen according to the situation and determines the properties of the data that remain private.
• Two databases are said to be neighbours if the larger database is constructed by adding a single row to the smaller database; if the rows correspond to the data of a single individual, then differential privacy can be thought as guaranteeing that the output does not reveal the presence or absence of any single individual in the database.
• two matrices can be neighbours if they differ in a single row, corresponding to per-user privacy, or if they differ in a single rating, corresponding to per-rating privacy.
• Recommender systems generate predictions of user preference by
• One way to do this is by clustering the items into a set of classes, e.g. by finding and using correlations between the rankings given to each item by a multiplicity of users, and then releasing this classification.
• the items considered by the present recommender system are thus assumed to have a cluster structure, i.e. items may be clustered in accordance with specific item properties.
• users can determine their potential ratings for new content, taking into account algorithms which are used by the recommender system for determining the classification.
• the recommendation method in accordance with the invention comprises two main phases: a learning phase and a recommendation phase.
• the learning phase is performed collaboratively with privacy guaranteed, while the actual recommendation is performed by the user's devices, based upon a general recommendation that is a result of the learning phase, without additional interaction with the system.
• the clusters are populated.
• the recommendation phase the populated clusters are revealed to the users.
• Each user can then derive recommendations based upon the populated clusters and locally stored data about the user's ratings of items and respective trust values associated with rated items. ln a hypothetical information-rich environment a minimum number ULB of users required for learning a concept class C under e-differential privacy can be calculated as
• a user is provided with a list of items to be rated. These items could be, for example, content items as found in audiovisual content. However, the items could as well relate to anything else that can be electronically rated.
• the user rates, in privacy, e.g. by using a device to which the user has exclusive access, at least some of the items.
• the rating of an item list through the users is a form of "bulk" rating without knowledge which, or indeed whether any of the items in the list have been rated, or consumed, by the user.
• the user selects a degree of trust that is assigned to individual ratings or assigned globally to all items in the list of items, thereby "privatizing" the ratings.
• the degree of trust is a value indicating the likelihood that the returned rating is exact and reliable. This adds noise to the ratings, which enhances the user privacy.
• the rated and "privatized" list is submitted to a recommender device.
• the returned rating can be considered an aggregate rating.
• the recommender device may be operated by the provider of the items to be rated, or operated by a third party providing a recommendation service.
• the recommender device collects rated lists of items from a predetermined minimum number of users and performs a clustering on the lists, clustering items of similar content and/or similar rating.
• the clustering takes the degree of trust into account that has been assigned to the items or to the lists.
• the content of each cluster is likely to be liked by the set of users who also liked other contents within this cluster.
• the clustered items are used to generate a list of recommended items, which is returned to the users that had provided their rating on one or more items.
• Storing true ratings and related data only locally in the user's device, and providing "bulk” rating is one aspect that provides the desired degree of privacy.
• Assigning a trust value at the user's side and providing the ratings only after assigning the trust value is a further aspect that provides the desired degree of privacy. It can be compared to adding noise that "hides” data a user wishes to keep private, and at the same time reduces the confidence that a user must have in the provider of the recommendation service.
• the "consumption" habits of the users is not considered private, while their ratings are, i.e. the information whether a user "consumed” an item from the list of items is public and true, while information how a user actually rated the item is private. This is applicable in many scenarios, as in order to "consume” the item, e.g. view a movie, the user most likely will leave a trail of information.
• each user is first presented two movies, both of which the user has watched. These two movies may be picked at random.
• the user converts the rating for these two movies to normalized rating that lies in ⁇ 0,1 ⁇ . This may be done by a scheme which maps each possible rating to ⁇ 0,1 ⁇ according to a randomized function, which can be user-preset in the user's device, or user-controlled for each rating.
• a private 'sketch' is determined, which may consist of a single bit.
• the sketch bit may also be referred to as similitude indicator. More generally, the similitude indicator may indicate if the rating value of each item of the pair of items (A, B) is above a predetermined value.
• the sketch bit is privatized according to a given procedure, e.g. a trust value is assigned in accordance with which the true sketch bit is provided to the recommendation device with a certain specified probability, or else the complement of the sketch bit is provided. This step ensures differential privacy.
• the recommendation engine then stores these 'pairwise preference' values in a matrix, and performs spectral clustering on this matrix to determine the movie clusters, which are then released.
• the spectral clustering may include adding the sketch values provided by the users for each identical pair.
• N items can be rated by U users.
• Each user has a set of w ratings ⁇ W u , R u ), W u ⁇ ( ⁇ ), Ru ⁇ ⁇ 0, 1 ⁇ W .
• Each item / ' is associated with a cluster C/v(/) from a set of L clusters, ⁇ 1 , 2,..., /. ⁇ .
• a user's device determines a private sketch that is 1 if the ratings of the two items are identical, or, if the normalized ratings thereof are identical, and otherwise is 0. If an item has no rating, the rating is assigned a value of 0.
• This private sketch also referred to as S° , is then privatized as 6
• Each row (node) is projected into the L-dimensional profile space of the eigenvectors, and k- means clustering is performed in the profile space for obtaining the item clusters.
• both the "consumption" habits i.e. whether or not an item has been consumed, as well as the ratings of the users are considered private information.
• this information-scarce environment requires collecting information for a much larger set of items from each user than only two, as being sufficient in the preceding example for the information-rich environment.
• a user's device is presented with a sensing vector of binary values.
• Each vector component corresponds to an item from the set of items.
• each item corresponds to a movie.
• the sensing vector determines whether a movie from the set of movies is probed. A movie may be probed for example if the corresponding
• the user's device converts all ratings to normalized ratings for the probed movies, and sets the normalized rating for unwatched movies to 0.
• the user's device calculates the maximum normalized value among the probed movies, i.e., movies for whom the corresponding component in the sensing vector was ⁇ ', and releases a privatized version of this vector component.
• the privatization mechanism is the same as before, i.e. the -I l logical value is flipped or not in accordance with the probability, or trust value assigned.
• a sketch bit indicates if a movie has the same normalized rating as the maximum normalized value determined before.
• a trust value indicates the probability of the sketch bit being true.
• the recommendation engine then combines these vector components in the following way: for each movie, the engine calculates the sum of the privatized user ratings over all users for whom the corresponding movie was probed, i.e., for whom the corresponding element in the sensing vector was set to ⁇ '. Next, the engine uses these 'movie sums' to perform k-means clustering in 1 -dimension, and returns the so-obtained movie classifications.
• the sensing vectors may also be randomly generated by the users. In this case the vector must be provided to the recommendation engine in addition to their privatized rating. In either case differential privacy is maintained as long as the sensing vector is independent of the set of items actually rated by the user.
• N items can be rated by U users.
• Each user has a set of w ratings (W u , R u ), W u ⁇ ( ⁇ ), Ru ⁇ ⁇ 0, 1 ⁇ W .
• Each item / ' is associated with a cluster C/v(/) from a set of
• a user's device determines the maximum rating within the ratings for items corresponding to the sensing vector in accordance with
• normalization may not be needed, depending on the items' features and/or properties considered when rating.
• the sensing vector H ui essentially asks for ratings of those items whose indices are 1 in this vector.
• the user device constructs a sketch S°, which is 1 if at least one of these items is rated 1 , and 0 otherwise.
• the user's device is presented a multiplicity of sensing vectors, each sensing vector being determined to cover a randomly chosen subset of the entire set of items. Ratings and privatization are processed as described before. It is obvious that this variant also allows for sensing vectors randomly generated at the user's device, the sensing vectors not being private information and being transmitted to the recommendation engine.
• Each item / ' is associated with a cluster C/v(/) from a set of L clusters, ⁇ 1 , 2,..., L ⁇ .
• the user's device returns a privatized response S( U , q ) to the recommendation engine.
• the recommendation engine determines a count B / from the privatized responses in accordance with
• Figure 1 shows a schematic block diagram of a system in accordance with the invention
• Figure 2 shows a schematic block diagram of a user device in accordance with the invention
• Figure 3 schematically shows the distribution of information flow and processing operations for a first embodiment of the invention
• Figure 4 schematically shows the distribution of information flow and
• Figure 5 schematically shows the distribution of information flow
• Figure 6 represents an exemplary and schematic view of vector operations performed in accordance with the second embodiment of the invention.
• Figure 1 shows a schematic block diagram of a recommender system in accordance with the invention.
• User device 100 is located in a private zone, which is does not reveal any data considered private to a public zone. The border between the private zone and the public zone is indicted by the dashed line.
• Recommending engine 200 is located in the public zone, i.e. data that is processed in this zone may be visible to an arbitrary set of users, entities, etc.
• User device 100 and recommending engine 200 are connected via network 300, e.g. an IP network, a cable network, a telephone network, or the like. Multiple (not shown) user devices 100 are connected to recommending engine 200, making up the recommender system.
• network 300 e.g. an IP network, a cable network, a telephone network, or the like.
• Multiple (not shown) user devices 100 are connected to recommending engine 200, making up the recommender system.
• FIG. 2 shows a schematic block diagram of a user device 100 in accordance with the invention.
• microprocessor 101 user interface 102, network interface 103, memory 104 and storage 106 are connected via a general bus system 107.
• the actual interconnection may vary depending on the system architecture, and it is conceivable and within the scope of the invention to use several connection buses in parallel, not all buses connecting all of the components with. Direct connections between individual components are likewise possible and considered within the scope of the invention.
• Memory 104 and storage 106 may be separate and of different type, or may be combined.
• User interface 102 may be provided in the user device, or may be provided as an interface to external components such as a display, a keyboard, a remote control, and the like.
• Figure 3 schematically shows the distribution of information flow and processing operations for a first embodiment of the invention.
• the request crosses the border between the public and the private zones, indicated by a dashed line.
• the request may be public but, as will be shown in the following, the response has been privatized to some extent while being processed in the private zone.
• the privacy in the private zone is established through measures under the control of the user, including physical privacy, i.e. restricted access to the user's device or the user's premises.
• the user's device rates the pair of items A, B, in accordance with current or previous user input and/ or by evaluating user behaviour monitored and stored while the user consumed the items to be rated.
• the rating value is privatized by assigning a trust value indicating the probability of the rating value actually being true, i.e. the probability that the binary value is flipped or not.
• the rating value and the trust value are returned as a response to the recommending engine.
• recommending engine collects a multiplicity of ratings for a multiplicity of pairs from a plurality of users.
• the pairs may be identical for various users, but may as well be made from different items, including combining one item already rated before with another item that had not been rated before, so as to create a larger database for analysis and clustering. Sending only a probable rating provides for a certain degree of privacy.
• the trust value is user adjustable, i.e. a user can select different trust values for individual ratings.
• Figure 4 schematically shows the distribution of information flow and processing operations for a second embodiment of the invention.
• recommending engine located in the public zone issues a request to a user device, asking to provide a rating for items indicated 'to rate' in a 'sensing vector'. For example, a logical value '1 ' indicates 'rate this item', while a logical value '0' indicates 'do not rate'.
• the user device rates the items indicated by the sensing vector and assigns a logical value ⁇ ' to all items having the same, maximum rating.
• the result is privatized by assigning a trust value in the same manner as before.
• the response has the form of a rating vector, and a trust value valid for all rated items is assigned.
• the so-privatized rating vector is then returned to the recommending engine, along with the trust value or trust values.
• the recommending engine collects a multiplicity of ratings for a multiplicity of sensing vectors from a plurality of users.
• the sensing vectors may be identical for various users, but may as well be made from different items, including combining one item already rated before with another item that had not been rated before, so as to create a larger database for analysis and clustering. Also, the sensing vectors may only cover a subset of all rateable items, multiple subsets covering the entire set of rateable items in a non-overlapping or overlapping manner.
• FIG. 5 schematically shows the distribution of information flow
• the process differs from the one shown in figure 4 in that the sensing vector is generated by the user's device either following a timed schedule or in response to a request to rate (not shown).
• the user's device returns the privatized rating as before, including a representation of the locally generated sensing vector, or by adding corresponding information to the rating vector.
• the user's device may provide multiple different rating vectors to the
• FIG. 6 represents an exemplary and schematic view of vector operations performed in accordance with the second embodiment of the invention.
• the individual items to be rated are shown at the left side of the figure.
• each item to be rated is assigned a logical value ⁇
• all other items are assigned a logical value ⁇ '.
• Each component of the sensing vector is multiplied with the corresponding individual rating in a scalar multiplication.
• the sensing vector is used as a selection mask for a rating vector.
• the result is a rating vector having ratings only for components that had a logical value ⁇ in the sensing vector.
• the resulting vector is normalized, indicated by indices assigned to the ratings. Any rated item that has the same rating as the maximum rating value, optionally: has a rating that lies within a predetermined range of the maximum rating, is assigned a logical value '1 ' in a similitude vector. All items having a logical value '1 ' in the similitude vector are considered being rated identical, or optionally: rated sufficiently identical.
• a sketch for the similitude vector is generated and then privatized in the same manner as described before.

Landscapes

• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• Computing Systems (AREA)
• Data Mining & Analysis (AREA)
• Evolutionary Computation (AREA)
• General Engineering & Computer Science (AREA)
• Multimedia (AREA)
• Mathematical Physics (AREA)
• Software Systems (AREA)
• Computational Linguistics (AREA)
• Business, Economics & Management (AREA)
• Signal Processing (AREA)
• Artificial Intelligence (AREA)
• Development Economics (AREA)
• Accounting & Taxation (AREA)
• Databases & Information Systems (AREA)
• Strategic Management (AREA)
• Finance (AREA)
• Economics (AREA)
• General Business, Economics & Management (AREA)
• Human Computer Interaction (AREA)
• Entrepreneurship & Innovation (AREA)
• Marketing (AREA)
• Game Theory and Decision Science (AREA)
• Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
• Management, Administration, Business Operations System, And Electronic Commerce (AREA)
• Storage Device Security (AREA)

Priority Applications (1)

Applications Claiming Priority (3)

Publications (1)

Family

ID=47681901

Family Applications (2)

Family Applications Before (1)

Country Status (2)

Families Citing this family (8)

Family Cites Families (3)

• 2012
  • 2012-02-15 EP EP12305170.8A patent/EP2629248A1/de not_active Withdrawn
• 2013
  • 2013-02-08 EP EP13703400.5A patent/EP2815355A2/de not_active Withdrawn
  • 2013-02-08 WO PCT/EP2013/052533 patent/WO2013120780A2/en not_active Ceased

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events