WO2012013996A1 - Systèmes et procédés de recommandation - Google Patents

Systèmes et procédés de recommandation Download PDF

Info

Publication number
WO2012013996A1
WO2012013996A1 PCT/HU2011/000080 HU2011000080W WO2012013996A1 WO 2012013996 A1 WO2012013996 A1 WO 2012013996A1 HU 2011000080 W HU2011000080 W HU 2011000080W WO 2012013996 A1 WO2012013996 A1 WO 2012013996A1
Authority
WO
WIPO (PCT)
Prior art keywords
matrix
user
item
rating
feature matrix
Prior art date
Application number
PCT/HU2011/000080
Other languages
English (en)
Inventor
István PILÁSZY
Domonkos Tikk
Gábor TAKÁCS
András NÉMETH BOTTYÁN
Dávid ZIBRICZKY
Original Assignee
Gravity Research & Development Kft.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Gravity Research & Development Kft. filed Critical Gravity Research & Development Kft.
Publication of WO2012013996A1 publication Critical patent/WO2012013996A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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

Definitions

  • the present invention relates generally to the field of personalized content provision through the application of a recommender system.
  • the invention provides effective and robust algorithms to be implemented as engines of recommender systems in various peculiar fields.
  • media content such as audios, videos, or graphics
  • media content may include digital graphics, music, films, movies, broadcast television programs, cable or satellite television programs, home movies, user-created video clips, personal photographs, etc.
  • These contents are stored in many repositories and databases from which people can select and obtain the desired media content.
  • the amount of media content accessibly by the users grows rapidly day by day.
  • the Internet also broadens the horizon for many other types of businesses. By conducting transactions online, businesses are able to sell their products (or services) to customers all around the world. At the same time, businesses are able to offer much greater variety of products since they are no longer restricted by the spaces available in their physical stores.
  • personalized item recommendations collection of various types of information concerning the individual users or customers is needed in order to determine what a particular user may or may not like.
  • collaborative filtering is a process of filtering for information or patterns using techniques involving collaboration among multiple agents, viewpoints, data sources, etc.
  • collaborative filtering is a process of filtering for information or patterns using techniques involving collaboration among multiple agents, viewpoints, data sources, etc.
  • the collaboration is among the users or customers and the available products, service and media contents of a system.
  • known information is collected from the users and aggregated at a centralized location, such as a server. Patterns, such as similarities between two or more users or similarities between two or more items, among the collected data are located and used as the basis for making recommendations. Consequently, information that may be used to identify individual users and/or client devices associated with the users are also collected by the recommender system, since the system needs to be aware of each user's personal profile in order to personalize the recommendations for a particular user.
  • a recommender system becomes able to create the range/profile of interest of the user. Then, a provider can offer such novel content to the user which matches the user's interest. As a result of this, the user will subscribe more easily to (or at least require more information on) the content offered by the provider. In order to reach targeted users in a more personalized manner, hence, there is a need for enhancing the quality of recommendation.
  • recommender systems applied varies from field to field, since providers with interests in different fields have got highly different business models to be accomplished.
  • maximizing the time spent by the user with reading the articles presented in a given domain of the portal is the major object, since the provider's income might basically depend on the number of on-line ads actually read/visited by the user.
  • e-business selling typically e.g. electronic devices the most important aim is to generate as large business, and hence profit, as is possible.
  • matrix factorization is an efficient tool to make personalized recommendations of items, such as products, services or media content, while preserving complete user privacy.
  • US patent No. 7,685,232 discloses an item recommender system based on collaborative filtering, wherein a rating matrix is generated at a server from the rating vectors received from the users, more particularly from client devices operated by the users.
  • the rows correspond to a plurality of users
  • the columns correspond to a plurality of items, each of the cells representing a rating value associated with a corresponding item for a corresponding user.
  • the rating matrix is then approximated by the product of a user feature matrix and an item feature matrix (this operation is also referred to as factorization of the rating matrix) and the item feature matrix is then transmitted to selected ones of a plurality of client devices associated with the plurality of users.
  • a user feature vector is calculated, by means of approximation, from the item feature matrix and the rating vector available in the client device, wherein each of the cells of the rating vector represents a predicted rating value for a particular item to be recommended to the corresponding user.
  • Both of the above mentioned ALS and IALS method exhibits a computational complexity that is proportional to the cubic factor of the number of latent factors, which, in case of a huge number of users and items, might result in unreasonable computing time.
  • Such a high computational requirement provides a limitation on the frequency of updating the item feature matrix and user feature vectors according to the continuously changing user ratings.
  • the present invention relates to a method of providing personalized item recommendations in a communication system comprising a server and a plurality of client devices, the method comprising
  • A is a preset regularization factor, c legally, is a confidence value with respect to user u and item /, and /-, chorus is the rating value for an item evaluated by a user u
  • each scalar element of p u is computed from all other scalar elements of p u and the fixed matrix Q
  • each scalar element of q i? is computed from all other scalar elements of q t and the fixed matrix P
  • the present invention relates to a method of providing personalized item recommendations in a communication system comprising a server and a plurality of client devices, the method comprising
  • A is a preset regularization factor
  • c is a confidence value with respect to user u and item / '
  • r termi is the rating value for an item i evaluated by a user
  • each scalar element of p u is computed from all other scalar elements of p u and the fixed matrix Q, and similarly, in each of the Q-steps, each scalar element of q i ? is computed from all other scalar elements of q; and the fixed matrix P, ⁇ generating a predictive rating vector f u as the product of the associated user feature vector p u and the item feature matrix Q, and
  • the present invention relates to a server for providing personalized item recommendations in a communication system to a plurality of users associated with a plurality of client devices, the server comprising
  • ⁇ is a preset regularization factor
  • c is a confidence value with respect to user u and item / '
  • r ui is the rating value for an item i evaluated by a user u
  • each scalar element of p u is computed from all other scalar elements of p u and the fixed matrix Q
  • each scalar element of q is computed from all other scalar elements of q ; and the fixed matrix P, and • means for transmitting said item feature matrix Q to at least one client device.
  • the present invention relates to a system for providing personalized item recommendations to a plurality of users in a communication system, comprising:
  • each one of said client devices being adapted for
  • -generating a predictive rating vector f u for each selected user u associated with said selected client devices, said predictive rating vector f u being the product of the user feature vector p u and the item feature matrix Q, and - selecting, according to a predefined selection scheme, a plurality of items from the entire set of items included in the predictive rating vector f u for the recommendation thereof to said particular user u associated with the predictive rating vector f u , and
  • the present invention relates to a computer program product for providing personalized item recommendations to a plurality of users, the computer program product comprising a computer-readable medium having a plurality of computer program instructions stored therein, which are operable to cause at least one server and a plurality of client devices to carry out the respective steps of the first method according to the invention.
  • FIG. 1 is a schematic block diagram of the item recommender system according to a first embodiment of the invention
  • Figure 2 is a schematic block diagram of the item recommender system according to a second embodiment of the invention.
  • Figure 3a is a flow diagram illustrating the basic steps of a first item recommending method according to the invention
  • Figure 3b is a flow diagram illustrating the basic steps of a second item recommending method according to the invention.
  • Figure 4 is a graph depicting measurement results for the ALS and the ALS1 algorithms.
  • Figure 5 is a graph depicting measurement results for the IALS and the IALS1 algorithms.
  • Recommender systems suggest personalized recommendations on items to users based on various kinds of information on users and items.
  • Recommender systems intend to model user preferences on items and aim to recommend such items the user will probably like.
  • User preference information is twofold: explicit and implicit feedback.
  • the former mainly includes opinion expression via ratings of items on a predefined scale, while the latter consists of other user activities, such as purchasing, viewing, renting or searching of items.
  • the saved computational overhead obtained by replacing a slower method can be devoted for other purposes.
  • N denote the number of users and items, respectively.
  • u e ⁇ 1, ... , N] index for users
  • the rating of user it on item i is r ui
  • its prediction is f ui .
  • All r ui ratings are arranged in R; 31 denotes the set of (u, i) indexes of R where (a) a rating is provided (for explicit feedback), (b) a positive feedback is provided (for implicit feedback).
  • 3?* denotes each (u, i) pair of R, i.e. ⁇ l ⁇ ⁇ — N - M.
  • Matrix factorization approaches have been applied successfully for both rating-based and implicit feedback-based collaborative filtering (CF) problems.
  • MF methods perform a so-called low rank matrix approximation: the matrix R is approximated as a product of two lower rank matrices: R ⁇ PQ T , where P £ R NxK is the user feature matrix, Q £ R MxK is the item feature matrix, K is the number of latent factors that is a predefined constant, and the approximation is only performed at u, i) £ l positions.
  • the predefined regularization parameter ⁇ trades off between small training error and small model weights.
  • ALS alternating least squares
  • ALS alternates between two steps: the P-step fixes Q and recomputes P, the Q-step fixes P and recomputes Q.
  • the recomputation of P is performed by solving a separate RR problem for each user: for the u-th user it takes the feature vector (q;) of items rated by the user as input variables, and the value of the ratings (r ui ) as output variable, and finds the optimal p u by RR.
  • q feature vector
  • r ui the value of the ratings
  • p u (An u I + AJ- 1 ⁇ . (8)
  • a ridge regression is solved for each user, i.e. the user feature matrix is optimized on a vector basis by using all of the other user feature vectors currently available in the user feature matrix in order to optimize P.
  • the Q-step also requires a computational time of 0(K 2 ⁇ !R ⁇ + MK 3 ).
  • the number of recomputations of the user fuature vectors and the item feature vectors i.e. the number of P-steps and Q-steps, ranges between 10 and a "few tens" for the conventional ALS algorithm. .
  • each user rates each item either positively (user u viewed item t) or negatively (user u did not view item i). This is in contrast with the explicit case, where only a small subset of items are rated (but usually on a finer scale). Consequently, at implicit feedback the recommendation algorithm should handle a full rating matrix, in contrast to the explicit feedback case, where the rating matrix is typically a sparse matrix. Given that the sparsity of the explicit rating matrix is usually less than 1%, the difference in the size of the data is usually several orders of magnitude.
  • ALS A variant of ALS as proposed by Hu et al. and referred to as IALS below has been especially developed for implicit feedback datasets.
  • the key idea behind this approach is to assign a preference value r ui and a confidence level c ui to each element of H-.
  • Preference typically takes values 0 and 1, indicating whether a user watched a particular item or not.
  • the confidence level is related to how confident we are about user preferences. If a user watched the item for long, or visited it two times, we can be more confident about his preference. Then the optimal model is defined as:
  • the authors used ALS to solve the problem, but with the following two differences:
  • Hu et al. also proposed an efficient solution to the second issue: when recomputing P, they calculate an initial covariance matrix A 0 and a vector d 0 for a virtual user who watched no items:
  • a u A 0 + ⁇ i: (Uji)eS ⁇ (-c 0 + c ui ) ⁇
  • d-u d 0 + ⁇ i: (U> i ) g 3 ⁇ 4 ⁇ (— c 0 ⁇ r 0 + c ui ⁇ r ui ) ⁇ q ; .
  • the proposed SMF-variant is advantageous only when the average number of ratings per user and per item are both small.
  • the average number of ratings per user is 200 and the average number of ratings per item is 5000.
  • x a is the (i, /)-th element of X.
  • k the active feature.
  • the basic idea of the enhanced ridge regression method according to the invention is that only one scalar element of the vector w is optimized at one step, which results in a sequences of separate univariate optimization steps for the scalar elements of w. Due to the down-grading of the optimization task from a multivariate vector-based level to univariate scalar level in the above enhanced ridge regression method, a further speed-up of the ALS and IALS algorithms from K 3 to K may be achieved.
  • the weighted ridge regression method (referred to as algorithm RR1 below) may be described formally as follows:
  • n number of examples
  • K number of features
  • A regularization factor
  • a non-negative importance weight (confidence level), denoted here by C; may also be assigned to the the training examples when some training examples are regarded more important than other ones.
  • This enhanced weighted ridge regression method (referred to as RR1C below) may be described formally as follows:
  • n number of examples
  • K number of features
  • A regularization factor
  • Line 1 initializes the inner variables. The outermost for-cycle repeatedly optimizes each scalar element of w.
  • Lines 5— 6 set w k to 0 and adjusts the variables accordingly.
  • Lines 7— 8 compute the 1 X 1 variant of A and d respectively.
  • Line 9 actually executes a matrix inversion that is merely a division in a 1 X 1 case.
  • Line 10 adjusts the e t variables, since w k was changed.
  • RR1 and RR1C use only the diagonal of A £ R KxK that can be precomputed and does not change between two consecutive cycles.
  • the computational complexity of one cycle is 0(nK).
  • the total running time of optimiznig one vector w is O(LnK). If L is small, a significant speed-up can be achieved as compared to 0 (K 2 n).
  • the fourth optimization step results in an optimized value for w , being equal to 0.1 , and the target function value is reduced to 5.4.
  • the scalar elemets w ... , w 5 are repeatedly optimized several times by using all currently avalable scalar elements of the vector w (except the one to be computed). After a termination condition has met, the final set of the scalar elements of w are regarded as a near-optimal solution for w.
  • ALS1 A MODIFIED ALS ALGORITHM FOR EXPLICIT USER PREFERENCE FEEDBACK
  • each scalar element of the user feature vector p u is individually computed from all other scalar elements of p u and all scalar elements of the fixed item feature matrix Q .
  • the scalar elements of the user feature vector are rather "further optimized" than "recomputed".
  • a Q-step is carried out analogously, i.e.
  • A is a preset regularization factor
  • c ui is a confidence value with respect to user u and item t
  • r ui is the rating value for an item i evaluated by a user u
  • each scalar element of p u is computed from all other scalar elements of p u and the fixed matrix Q, and similarly, in each of the Q-steps, each scalar element of q i; is computed from all other scalar elements of q; and the fixed matrix P.
  • the running time of one epoch (loop) of ALSl is very favorable, as it is 0 ⁇ K ⁇ 3l ), which means a linear function of the number of latent factors K.
  • one epoch requires 0(K 2 ⁇ 3l ⁇ + (N + M)K 3 ) time.
  • the speed-up can be attributed to the approximate solution applied for the matrix inversion.
  • the ALSl algorithm according to the invention does not reproduce the results of ALS; with the same parameter setting, usually it achieves a slightly inferior prediction accuracy (as will be described below).
  • due to the speed-up of the ALSl algorithm trained with a somewhat larger number of latent factors K it can easily reach the accuracy of the original ALS algorithm under significantly smaller amount of time.
  • the ALSl algorithm may be formally described as follows.
  • K number of features
  • A regularization factor
  • the above described modified ALS algorithm may equally incorporate the weighted enhanced ridge regression algorithm RR1C instead of the basic RR1 algorithm, i.e. in the above formal description of the ALS l algorithm, the importance weights may be set to any non-negative value (instead of setting each of them to 1 ).
  • ALSl uses not only matrix Q, but also the values of previously computed elements of vectors p u in the computation of the new values of vector p u .
  • ALSl uses not only matrix P, but also the values of previously computed elements of vectors q ; in the computation of the new values of vector q;.
  • IALS1 A MODIFIED IALS ALGORITHM FOR IMPLICIT USER PREFERENCE FEEDBACK
  • the modified IALS method according to the invention will be described only through the optimization in a P-step; the Q-step can be treated analogously.
  • a full rating matrix should be provided.
  • a virtual user was introduced who rated all items negatively (i.e. he watched nothing).
  • the conventional ridge regression algorithm RR can be replaced with our proposed RRl algorithm to obtain a fast IALS algorithm for implicit user preference feedback, and how the favorable efficiency of the proposed RRl algorithm can be adapted to involve a virtual user to this case.
  • RR uses only the A 0 matrix and the d 0 vector.
  • RRl uses the diagonal of A 0 , the d 0 vector, and it also needs all the original examples (that generate A 0 and d 0 ), which makes inappropriate to combine the advantageous properties of RRl and IALS.
  • a 0 is a symmetric real matrix (and also positive semi-definite), its eigenvalue decomposition
  • G VAS t
  • G T SVX.
  • Regularization (usually) means that the diagonal of A is increased by a regularization factor A, which can be also user-dependent: A u .
  • regularization can be handled easily: assume that the regularization factor for user u is A u , i.e. we try to minimize the following cost function:
  • a 0 E R KXK a symmetric positive semidefinite matrix
  • a 0 e R K K a symmetric positive semidefinite matrix
  • the K + 1-th element of the (K + l)-th column is a large enough number (it should be greater to ensure that B is positive semidefinite
  • K number of features
  • A regularization factor
  • ⁇ i, ... , x n ⁇ is ⁇ gj, ... , g K ) U ⁇ q ; : (u, i) e 32 ⁇ ;
  • Ci c n is ⁇ 1 1 ⁇ U ⁇ (c ui - c 0 ): (u, t) e
  • ⁇ x 1 , ... , x n ⁇ is ⁇ g 1 , ... , g ⁇ U ⁇ p u : (u, 0 e ⁇ ;
  • ⁇ * n ⁇ is ⁇ r a r ⁇ U ⁇ c "' ri " ° r ° : (u, i) e 3? ⁇ ;
  • L is 1;
  • the invention relates to an item recommender system.
  • a system is typically based on a server-client architecture, wherein the server functions to control or manage a network of client devices interfacing the users of the item recommender system.
  • Figure 1 schematically illustrates an item reccomender system 10 comprising a server 1 1 and a plurality of client devices 12 connected to the server 1 1 through a communication network 13.
  • the client devices 12 are adapted for collecting user ratings and to forward the collected user ratings, preferably in the form of rating vectors, to the server periodically.
  • the server 1 1 is configured to receive the user ratings, preferably in the form of rating vectors r u , from the client devices 12 and to aggregate the user ratings into a rating matrix R.
  • the server 1 1 is typically adapted for factorizing said rating matrix R into a user feature matrix P and an item feature matrix Q using the above specified ALS1 or IALS1 factorization algorithms in accordance with the invention.
  • the server 1 1 may transmits the item feature matrix Q and selected user feature vectors p u to selected client devices 12, which can generate a predictive rating vector f u for each selected user u associated with the particular client device 12, said predictive rating vector f u being the product of the user feature vector p u and the item feature matrix Q.
  • Each one of the client devices 12 is further adapted for selecting, according to a predefined selection scheme, a plurality of items from the entire set of items included in the predictive rating vector f u for the recommendation thereof to said particular user u associated with the predictive rating vector f u .
  • the server 1 1 may be used only for collecting user ratings and to produce said rating matrix R, whereas the factorization itself may be performed in the client devices 12 which, for this purpose, receive the rating matrix R from the server 1 1 to generate a predictive rating vector r u .
  • a client device 12 may be, for example, a personal computer, a mobil phone, a set top box, or the like.
  • the communication network 13 may include, for example, the internet, a mobile telecommunication network, a video-on-demand network, a cable TV network, etc..
  • the client devices 12 are used by a plurality of users who evaluate any number of items among those that are presented them for evaluation.
  • the user ratings preferably in the form of rating vectors r u may be sent by the client devices 12 directly to the server 1 1, which, in turn, after factorization, may transmit the updated item feature matrix Q and the corresponding updated user feature vector p u to at least one of, but preferably all of the client devices 12.
  • the particular client device 12 may generate a predictive rating vector f u and make a recommendation to the associated user(s) according to the predictive rating vector f u .
  • This configuration is particularly suitable for an online web store, wherein the server 1 1 may reside at the merchant's site, and the client devices 12 may reside at the users who browse in the web store through the internet, for example.
  • the items correspond to movies offered by the broadcaster and the users correspond to the viewers.
  • the user ratings i.e. direct or indirect evaluations of the movies, are collected by the transceiver units coupled to the IPTV sets and forwarded thereby to the broadcasting centre, where the user ratings are periodically aggrageted into a user rating matrix, the factorization of which is then performed to produce an updated item feature matrix and an updated user feature matrix.
  • the item feature matrix which contains estimated latent ratings for each movie, and the user feature vectors, each of which contains estimated latent ratings of a particular user, are transmitted to the respective IPTV sets, in which a special program routine computes the estimated rating vector for the particular viewer.
  • the estimated rating vector contains an estimated rating for all of the movies on th basis of the prior (direct or indirect) ratings of the specific viewer.
  • the movies of likely the highest interest to the viewer will then be locally selected using a predefined selection scheme, and offered for viewing through a movie recommendation interface of the IPTV set.
  • At least one of the client devices 12' may be associated with a service provider (also referred to as an SP client device hereinafter), such as a media content provider, a DVD/video rental service, etc. which communicates with a plurality of user client devices 12 (the users of which are customers of the particular service provider), as well as with the server 1 1 through the communication network 13.
  • a service provider also referred to as an SP client device hereinafter
  • the user ratings are sent to the server 1 1 from the user's client device 12 indirectly, through the SP client devices 12' that continuously collect the ratings from the users and periodically forward the collected user ratings to the server 11.
  • the server 1 1 After a scheduled factorization of the available rating matrix, the server 1 1 sends the updated item feature matrix and selected ones of the updated user feature vectors to an SP client device 12' upon a corresponding request of said SP client device 12'. Based on the updated item feature matrix and the resquested user feature vectors, the SP client device 12' generates a predictive rating vector for selected users, and sends a list of items to the client devices 12 of said selected users so as to recommend items to them.
  • This configuration of the item recommendation system is particularly suitable for a video-on-demand network, wherein the server 11 may reside at an external computing center (performing the factorization for a huge number of users and items), whereas at least one SP client device 12' may be operated by a video content provider and the plurality of user client devices 12 may be connected to the SP client device 12' through a video broadcasting network.
  • the present invention relates to a server in an item recommender system, the server being adapted to perform factorization of a rating matrix, which is formed of user ratings, into an item feature matrix and a user feature matrix, wherein the server comprises
  • a server may be typically used, for example, for a movie recommender service in an IPTV network as described above.
  • the foregoing means of the server are preferably implemented by software tools, typically by computer program routines that can be run on a general-purpose processor or any application-specific processor.
  • the sofware implementiation of these means is a routin work for a skilled person on the basis of the above specified algorithms and pseudo codes, and therefore such implemenatation details are omitted here.
  • Preferred embodiments of the method are preferably implemented by software tools, typically by computer program routines that can be run on a general-purpose processor or any application-specific processor.
  • the present invention further relates to various methods of performing item recommendation in any one of the above specified item recommender systems.
  • the item recommendation methods according to the invention incorporate the above described ALS1 or IALS1 algorithms so as to reduce computational complexity, and thus the run time of the factorization of a rating matrix.
  • the factorization is performed at the server of an item recommendation system
  • the significant reduction of the computational complexity of the factorizing algorithm allows to perform the rating matrix factorization at a much higher frequency than it was achievable in the prior art item recommendation systems. Consequently, an updated item feature matrix and corresponding user feature vectors may be generated and sent to the client devices at a much higher frequency, therefore it is not necessary to generate an updated user feature vector in a client device any longer from a recent user rating vector and a relativly old item feature matrix by an additional factorization as applied in the prior art. Instead, both of the item feature matrix and the corresponding user feature vectors (i.e.
  • any client device can generate its predicitive rating vector by a simple multiplication of the item feature matrix and the respective user feature vector, which dramatically reducues the computational power requirements on the client devices.
  • the server may also be configured to only collect the user ratings, and then to forward the aggregated rating matrix (or at least a relevant portion thereof) to the client devices for local factorization.
  • the factorization within the client devices can also be preformed with much lower computational efforts to produce either the same quality of item recommendation with much less time, or much higher quality of item recommendation within the same time.
  • the basic steps of the first item recommendation method according to the invention comprises the following steps, also shown in the flow diagram of Figure 3a.
  • a plurality of user ratings preferably in the form of rating vectors r u , is received by a server from a plurality of client devices.
  • the user rating vectors r u may include explicit or implicit ratings.
  • step 1 10 the user rating vectors r u are aggregated by the server into a rating matrix R and in step S 120, the server performs the factorization of the rating matrix R into a user feature matrix P and an item feature matrix Q.
  • the factorization is executed periodically in the server, preferably at a predetermined frequency.
  • an approximate rating matrix R of said user feature matrix P and said item feature matrix Q is generated that approximates the rating matrix R.
  • the factorization is carried out by using the above mentioned ALS1 and IALS 1 factorization algorithms according to the invention, basically including the following steps.
  • the user feature matrix P and the item feature matrix Q are initialized with predefined initial values.
  • the matrices P and Q are initialized randomly.
  • the matrices P and Q are initialized according to the results of a previous factorization, assuming that only minor changes took place in the rating matrix R within the period between the previous factorization and the actual factorization. If this is the case, it has been experienced that such an initialization of P and Q results in a lower number of epochs of the factorization, thus further reducing the computation time of the factorization.
  • steps S122 and S123 the user feature matrix P and the item feature matrix Q, respectively, are alternately optimized until a termination condition is met in step SI 24.
  • the termination condition is predefined according to the desired accuracy of the prediction model that is calculated on the basis of certain criteria. It is preferred that for the ALS1 and IALS1 algorithms according to the invention, the well-known RMSE or the ARP measures may be used as performance measures. For a person skilled in the art, however, it is obvious that for defining a termination condition, any other performance measure can also be used depending on the particular field of application of the item recommendation method.
  • the above mentioned termination condition may include, inter alia, any of the following criteria: - the number of times that steps S122 and S123 were invoked, i.e. the number of factorization epochs, exceeds a preset limit;
  • the performance has reached a preset limit (i.e. the prediction model is regarded good enough);
  • the item feature matrix Q is sent from the server to every client device.
  • the corresponding user feature vector p u of the updated user feature matrix P may also be sent to the client devices.
  • step SI 50 a portion of the items are selected in the client devices for recommendation on the basis of the predictive rating vectors f u .
  • the factorization is performed in the client devices.
  • the aggregated rating matrix R is transmitted to at least one client device, but preferably every client device, to perform the factorization locally in step S230 to produce an updated item feature matrix Q and an updated user feature matrix R.
  • the factorization and the production of the predictive rating vector f u are performed in the same way as in the foregoing first embodiment of the method.
  • step S231 the item feature matrix and the user feature matrix are initialized, then in steps S232 and S233, the user feature vectors and the item feature vectors, respectively, are optimized using the ALS1 and IALS1 algorithms, and in step S234, it is checked whether or not the termination condition of the optimization loop is met.
  • the optimization routine proceeds with step S240 in which the a predictive rating vector f u is generated and finally, in step S250, a portion of the items are selected in the client devices for recommendation on the basis of the predictive rating vectors f u .
  • the client device may select, according to a predefined selection scheme, a plurality of items from the entire set of items included in the predictive rating vector f u for the recommendation thereof to said particular user u associated with the predictive rating vector f u .
  • a predefined selection scheme For the selection of the particular items to be recommended to a user, several specific selection schemes may be used. One of the best schemes is based on the Euclidean distances between the user feature vector and the item feature vectors. Such a selection scheme will be described in detail below.
  • the rating matrix R is factorized into a user feature matrix P and an item feature matrix Q.
  • p u denotes the u-th row of P
  • the user feature vector of user u the user feature vector of user u.
  • the factorization of the rating matrix R into a user feature matrix P and an item feature matrix Q may be performed at two stages, including a preliminary first approximation of the rating matrix, and then a final, more accurate approximation thereof, both at the server.
  • a first user feature matrix P x and a first item feature matrix Q t are generated by using a first number of latent factors K t for the preliminary factorization, thereby a first (preliminary) estimation of the rating matrix R is obtained.
  • a second user feature matrix P 2 and a second item feature matrix Q 2 are generated by using a second number of latent factors K 2 , wherein K 1 ⁇ K 2 .
  • the first item feature matrix Q x and selected rows of the first user feature matrix P j may be transmitted to the plurality of client devices.
  • a first item list including N 1 number of items are defined in the client device.
  • This list may then be transmitted from the client devices to the server, where predictions are generated for these N 1 items according to the second factorization (usingP 2 and Q 2 ), using the f ui — p u T q; formula, where p u now denotes the u-th row of P 2 and q; now denotes the t-th row of Q 2 .
  • the server may send back the result, and the client devices can select a reduced N 2 number of items from the first list of ⁇ items according to a predefined selection scheme for the recommendations thereof to a user.
  • the selection of N 2 number of items from the first list of N t items may also be performed by the server, in this case the selected N 2 number of items are returned by the server.
  • the present invention relates to a computer program product for providing personalized item recommendations to a plurality of users.
  • the computer program product comprises a computer-readable medium having a plurality of computer program instructions stored therein, which are operable to cause at least one server and a plurality of client devices to carry out the respective steps of the item recommendation methods according to the invention.
  • the NP dataset contains 100 480 507 ratings (i.e. (u, i, r ui ) triplets) of 480 189 users on 17 770 movies, on a 1 to 5 integer scale. Ratings are provided as a standard train-test split.
  • the test set contains 1 408 395 ratings and is called Probe and contains newer ratings for each user than the Train.
  • our ProbelO set (available at the applicant's homepage, http://www.gravityrd.com ) that contains only a 1/10 subset of the Probe ratings ( 140 840 ratings), and puts the rest of Probe into Train. ProbelO approximates very well the withheld Quiz set described in G. Takacs, I. Pilaszy, B. Nemeth, and D.
  • Figure 4 shows the results measured in each epoch. We depicted only the best RMSE results achieved with both ALS and ALS l under the same amount of time, independently from the number of epochs and the value of K. Both obtained curves are monotone decreasing.
  • the curve of ALS l is always under ALS which means that:
  • ⁇ ALSl can reach the same RMSE as ALS much faster.
  • ALSl offers a superior trade-off possibility between running time and performance compared to ALS.
  • the recommendation algorithm For a given user u in the test set, the recommendation algorithm has to order all the recommendable items. This ordering can be based on the f ui £ . prediction values: items with highest f ui are ranked first. For user , the set of items are divided into two parts: relevant and non-relevant ones for the user. Assume that the recommendable items are indexed by i ranging from 1 to M . Let r ui £ ⁇ 0,1 ⁇ denote whether an item is relevant to the user or not, e.g. user will view that item during the test period or not. Then the position for item i relevant to user u is defined as:
  • the relative position is defined as:
  • the average relative position is defined as:
  • Figure 5 compares the time— accuracy trade-off of the two algorithms. We used the same consideration to depict the curves as for Figure 1.
  • the curve of IALS1 is always located under IALS, so we can draw similar conclusions as before, i.e. IALS1 offers superior trade-off possibility between running time and performance compared to IALS.
  • the curves of IALS 1 and IALS have similar characteristics: steep start, flat in the middle, and gets steep again at the end.
  • both curves reach a point when their gradient decreases and then the two curves become closer.
  • the ARP is better for larger K values.
  • K 100— 300 (on the time scale between 2000 and 20000 seconds) this does not hold (the flat middle), and consequently the two curves get close to each other.
  • the enhanced factorization algorithms i.e. the ALSl and IALS1 algorithms
  • these factorization algorithms may equally be applied in various other information processing areas (e.g. for data mining, for text mining or even for image processing) and in different business domains (e.g. for content and application recommendation in interactive TV, for recommendations in e-commerce stores/auction portals/internet market places, for digital media & streaming, for online dating, for targeted advertisements and for streamlined direct marketing based on CRM databases) with the advantages as discussed above.
  • some typical application of the proposed factorization algorithms will now be introduced briefly as non-limiting examples.
  • each row corresponds to a word of the vocabulary
  • each column corresponds to a document
  • an element of the matrix tells how many times a given word occurs in a given document.
  • this matrix is modified (e.g. the rows or columns are normalized, etc.).
  • This matrix can then be analyzed by matrix factorization algorithms. For example, Singular Value Decomposition (SVD, a well known matrix factorization algorithm) was applied to the term-document matrix in 1990 (S. Deerwester: Indexing by Latent Semantic Analysis). Since the ALSl and IALS1 algorithms in accordance with the present invention are very fast, they can be applied to analyze huge term-document matrices. (10 epoch) IALS IALS1
  • Table 3 ARP of IALS and IALS1 with different number of factors on the RIF dataset, after 10 (upper) and 20
  • Each digitized image can be represented as a matrix.
  • grey shaded digitized images can be represented in the following way: rows of the matrix correspond to the rows of the image, columns of the matrix correspond to the columns of the image, and the numerically represented grey-level can be put in the appropriate cell of the matrix.
  • E-commerce stores differ from normal, everyday shopping; there are provided e.g. limited display capabilities, less "enjoyment factor” (e.g. bookstore effect), etc.
  • An efficient product recommendation in e-commerce stores / auction portals / internet market places offers faster product discovery, personal shopping experience to the end-user, increased revenue, optimized one-to-one merchandising options, inventory usage.
  • Products can be any consumer goods, food, media (music, mp3, dvd), etc. to mention only a few non-limiting examples.
  • Offerings can be places, events (collectively venue recommendations), and ticketing as well.
  • the present invention can be also exploited in the field of digital media & streaming.
  • Internet became the new interactive TV platform. More people consume content/programs on Internet than on TV (e.g. YouTube, Hulu, Last.fm).
  • Our solution helps similarly the content discovery such as in the case of Interactive TV.
  • the recommendable program set can be any kind of video/audio or mixed, can be live or recorded (e.g. YouTube, music sites, Ustream - live streaming, web cam sites, etc.).

Landscapes

  • Business, Economics & Management (AREA)
  • Engineering & Computer Science (AREA)
  • Accounting & Taxation (AREA)
  • Development Economics (AREA)
  • Strategic Management (AREA)
  • Finance (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Economics (AREA)
  • Marketing (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

L'invention porte sur un procédé de fourniture de recommandations d'objet personnalisées dans un système de communication comprenant un serveur et une pluralité de dispositifs clients. Le procédé comprend les étapes suivantes : au niveau du serveur, la réception (S100) d'une pluralité de vecteurs de classement d'utilisateur ru à partir d'une pluralité desdits dispositifs clients; au niveau du serveur, l'agrégation (S110) des vecteurs de classement d'utilisateur ru dans une matrice de classement R; au niveau du serveur, la factorisation (S120) de ladite matrice de classement R en une matrice de caractéristiques d'utilisateur P et en une matrice de caractéristiques d'objet Q, le produit R de ladite matrice de caractéristiques d'utilisateur P et de ladite matrice de caractéristiques d'objet Q étant proche de la matrice de classement d'utilisateur R, ladite factorisation comprenant les étapes de transmission (S130) de ladite matrice de caractéristiques d'objet Q du serveur à au moins un dispositif client; la génération (S140) d'un vecteur de classement ru prédictif en tant que produit du vecteur de caractéristiques d'utilisateur ru et de la matrice de caractéristiques d'objet Q associée; et par utilisation d'un schéma de sélection prédéfini, la sélection (S150) d'au moins un objet à recommander à l'utilisateur à partir des objets associés au vecteur de classement ru prédictif. L'invention porte également sur un serveur, un système et un produit programme d'ordinateur pour la mise en œuvre du procédé.
PCT/HU2011/000080 2010-07-30 2011-07-29 Systèmes et procédés de recommandation WO2012013996A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
HU1000408A HUP1000408A2 (en) 2010-07-30 2010-07-30 Recommender systems and methods
HUP1000408 2010-07-30

Publications (1)

Publication Number Publication Date
WO2012013996A1 true WO2012013996A1 (fr) 2012-02-02

Family

ID=89989863

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/HU2011/000080 WO2012013996A1 (fr) 2010-07-30 2011-07-29 Systèmes et procédés de recommandation

Country Status (2)

Country Link
HU (1) HUP1000408A2 (fr)
WO (1) WO2012013996A1 (fr)

Cited By (27)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102708525A (zh) * 2012-05-22 2012-10-03 华南理工大学 基于gpu加速的招聘职位智能推荐方法
EP2744219A1 (fr) * 2012-12-14 2014-06-18 Thomson Licensing Prédiction d'appréciation d'articles d'un utilisateur et procédé de recommandation correspondant
EP2746983A1 (fr) * 2012-12-21 2014-06-25 Alcatel-Lucent Calcul préservant la confidentialité de tendances populaires pour des services de recommandation fondés sur un groupe
EP2960849A1 (fr) * 2014-06-26 2015-12-30 Deutsche Telekom AG Procédé et système pour recommander un objet à un utilisateur
US9691035B1 (en) 2014-04-14 2017-06-27 Amazon Technologies, Inc. Real-time updates to item recommendation models based on matrix factorization
CN106971053A (zh) * 2016-01-08 2017-07-21 车海莺 一种基于混合协同过滤的推荐方法
WO2017178870A1 (fr) * 2016-04-15 2017-10-19 Telefonaktiebolaget Lm Ericsson (Publ) Système et procédé destinés à l'apprentissage de modèle côté client dans des systèmes de recommandation
CN109670141A (zh) * 2018-11-30 2019-04-23 网易传媒科技（北京）有限公司 预测方法、系统、介质和电子设备
CN110110214A (zh) * 2018-01-25 2019-08-09 重庆邮电大学 基于双向加权值及用户行为的动态推荐及加噪方法
WO2019169266A1 (fr) * 2018-03-02 2019-09-06 Alibaba Group Holding Limited Procédé et appareil de construction de système de recommandation
CN110321489A (zh) * 2019-07-11 2019-10-11 重庆邮电大学 一种基于改进推荐算法的智慧旅游推荐方法及系统
CN110674020A (zh) * 2019-09-02 2020-01-10 平安科技（深圳）有限公司 App智能推荐方法、装置及计算机可读存储介质
US20200043021A1 (en) * 2018-08-06 2020-02-06 Walmart Apollo, Llc Systems and methods for identifying and using micro-intents
FR3087921A1 (fr) * 2018-10-31 2020-05-01 Amadeus S.A.S. Systèmes et méthodes de recommandation utilisant des modèles d'apprentissage automatique en cascade
CN111340660A (zh) * 2019-07-01 2020-06-26 黑龙江省华熵助晟网络科技有限公司 一种在线学习辅助系统及方法
CN111492392A (zh) * 2017-12-22 2020-08-04 华为技术有限公司 用于更新客户端-项目矩阵的客户端、服务器和客户端-服务器系统
CN111492391A (zh) * 2017-12-22 2020-08-04 华为技术有限公司 用于生成个性化推荐的客户端、服务器和客户端-服务器系统
CN111581333A (zh) * 2020-05-07 2020-08-25 重庆大学 基于Text-CNN的影音播单推送方法及影音播单推送系统
CN112507216A (zh) * 2020-12-01 2021-03-16 北京奇艺世纪科技有限公司 一种数据对象推荐方法、装置、设备和存储介质
US20210142330A1 (en) * 2017-10-26 2021-05-13 Mastercard International Incorporated Systems and methods for detecting out-of-pattern transactions
WO2021113716A1 (fr) * 2019-12-04 2021-06-10 Canopy Crest Corporation Recommandation privée dans un environnement client-serveur
CN113254777A (zh) * 2021-06-07 2021-08-13 武汉卓尔数字传媒科技有限公司 信息推荐方法、装置、电子设备及存储介质
US11093846B2 (en) 2016-07-01 2021-08-17 International Business Machines Corporation Rating model generation
WO2021185428A1 (fr) * 2020-03-16 2021-09-23 Huawei Technologies Co., Ltd. Activation d'une génération de recommandations personnalisées
WO2021185427A1 (fr) * 2020-03-16 2021-09-23 Huawei Technologies Co., Ltd. Génération de recommandations personnalisées
CN113765959A (zh) * 2020-06-30 2021-12-07 北京沃东天骏信息技术有限公司 信息推送方法、装置、设备及计算机可读存储介质
US11884277B2 (en) 2014-06-27 2024-01-30 Bayerische Motoren Werke Aktiengesellschaft Method for producing a model of the surroundings of a vehicle

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117710020B (zh) * 2024-02-06 2024-05-17 湖南惟客科技集团有限公司 一种基于大数据的用户喜好的分析方法

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7685232B2 (en) 2008-06-04 2010-03-23 Samsung Electronics Co., Ltd. Method for anonymous collaborative filtering using matrix factorization

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7685232B2 (en) 2008-06-04 2010-03-23 Samsung Electronics Co., Ltd. Method for anonymous collaborative filtering using matrix factorization

Non-Patent Citations (6)

* Cited by examiner, † Cited by third party
Title
"STATEMENT IN ACCORDANCE WITH THE NOTICE FROM THE EUROPEAN PATENT OFFICE DATED 1 OCTOBER 2007 CONCERNING BUSINESS METHODS - PCT / ERKLAERUNG GEMAESS DER MITTEILUNG DES EUROPAEISCHEN PATENTAMTS VOM 1.OKTOBER 2007 UEBER GESCHAEFTSMETHODEN - PCT / DECLARATION CONFORMEMENT AU COMMUNIQUE DE L'OFFICE EUROP", 20071101, 1 November 2007 (2007-11-01), XP002456414 *
G. TAKACS, I. PILÁSZY, B. NÉMETH, D. TIKK: "Scalable collaborative filtering approaches for large recommender systems", JOURNAL OF MACHINE LEARNING RESEARCH, vol. 10, 2009, pages 623 - 656
J. BENNETT, S. LANNING: "The Netflix Prize", KDD CUP WORKSHOP AT KDD-07, 13TH ACM INT. CONF. ON KNOWLEDGE DISCOVERY AND DATA MINING, 2007, pages 3 - 6
R.M. BELL, Y.KOREN: "Scalable collaborative filtering with jointly derived neighborhood intelpolation weights", ICDM-07, 7`H IEEE INT. CONF. ON DATA MINING, 2007, pages 43 - 52, XP031238239
Y. HU, Y. KOREN, C. VOLINSKY: "Collaborative filtering for implicit feedback datasets", PROC. OF ICDM-08, 8TH IEEE INT. CONF. ON DATA MINING, PISA, ITALY, 2008, pages 263 - 272, XP031423705
Y.HU, Y.KOREN, C.VOLINSKY: "Collaborative filtering for implicit feedback datasets", ICDM-08, 8TH IEEE INT. CONF. ON DATA MINING, 2008, pages 263 - 272, XP031423705

Cited By (36)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102708525A (zh) * 2012-05-22 2012-10-03 华南理工大学 基于gpu加速的招聘职位智能推荐方法
EP2744219A1 (fr) * 2012-12-14 2014-06-18 Thomson Licensing Prédiction d'appréciation d'articles d'un utilisateur et procédé de recommandation correspondant
EP2746983A1 (fr) * 2012-12-21 2014-06-25 Alcatel-Lucent Calcul préservant la confidentialité de tendances populaires pour des services de recommandation fondés sur un groupe
US9691035B1 (en) 2014-04-14 2017-06-27 Amazon Technologies, Inc. Real-time updates to item recommendation models based on matrix factorization
EP2960849A1 (fr) * 2014-06-26 2015-12-30 Deutsche Telekom AG Procédé et système pour recommander un objet à un utilisateur
US11884277B2 (en) 2014-06-27 2024-01-30 Bayerische Motoren Werke Aktiengesellschaft Method for producing a model of the surroundings of a vehicle
CN106971053A (zh) * 2016-01-08 2017-07-21 车海莺 一种基于混合协同过滤的推荐方法
WO2017178870A1 (fr) * 2016-04-15 2017-10-19 Telefonaktiebolaget Lm Ericsson (Publ) Système et procédé destinés à l'apprentissage de modèle côté client dans des systèmes de recommandation
US11093846B2 (en) 2016-07-01 2021-08-17 International Business Machines Corporation Rating model generation
US11727407B2 (en) * 2017-10-26 2023-08-15 Mastercard International Incorporated Systems and methods for detecting out-of-pattern transactions
US20210142330A1 (en) * 2017-10-26 2021-05-13 Mastercard International Incorporated Systems and methods for detecting out-of-pattern transactions
CN111492391A (zh) * 2017-12-22 2020-08-04 华为技术有限公司 用于生成个性化推荐的客户端、服务器和客户端-服务器系统
CN111492392B (zh) * 2017-12-22 2023-11-17 华为技术有限公司 用于更新客户端-项目矩阵的客户端、服务器和客户端-服务器系统
CN111492392A (zh) * 2017-12-22 2020-08-04 华为技术有限公司 用于更新客户端-项目矩阵的客户端、服务器和客户端-服务器系统
CN110110214A (zh) * 2018-01-25 2019-08-09 重庆邮电大学 基于双向加权值及用户行为的动态推荐及加噪方法
US11551110B2 (en) 2018-03-02 2023-01-10 Advanced New Technologies Co., Ltd. Recommendation system construction method and apparatus
WO2019169266A1 (fr) * 2018-03-02 2019-09-06 Alibaba Group Holding Limited Procédé et appareil de construction de système de recommandation
US10902332B2 (en) 2018-03-02 2021-01-26 Advanced New Technologies Co., Ltd. Recommendation system construction method and apparatus
US20200043021A1 (en) * 2018-08-06 2020-02-06 Walmart Apollo, Llc Systems and methods for identifying and using micro-intents
EP3648041A1 (fr) * 2018-10-31 2020-05-06 Amadeus S.A.S. Systèmes et procédés de recommandation utilisant des modèles d'apprentissage machine en cascade
FR3087921A1 (fr) * 2018-10-31 2020-05-01 Amadeus S.A.S. Systèmes et méthodes de recommandation utilisant des modèles d'apprentissage automatique en cascade
US11538086B2 (en) 2018-10-31 2022-12-27 Amadeus S.A.S. Recommender systems and methods using cascaded machine learning models
CN109670141A (zh) * 2018-11-30 2019-04-23 网易传媒科技（北京）有限公司 预测方法、系统、介质和电子设备
CN111340660A (zh) * 2019-07-01 2020-06-26 黑龙江省华熵助晟网络科技有限公司 一种在线学习辅助系统及方法
CN111340660B (zh) * 2019-07-01 2023-09-01 黑龙江省华熵助晟网络科技有限公司 一种在线学习辅助系统及方法
CN110321489A (zh) * 2019-07-11 2019-10-11 重庆邮电大学 一种基于改进推荐算法的智慧旅游推荐方法及系统
CN110674020A (zh) * 2019-09-02 2020-01-10 平安科技（深圳）有限公司 App智能推荐方法、装置及计算机可读存储介质
WO2021113716A1 (fr) * 2019-12-04 2021-06-10 Canopy Crest Corporation Recommandation privée dans un environnement client-serveur
WO2021185428A1 (fr) * 2020-03-16 2021-09-23 Huawei Technologies Co., Ltd. Activation d'une génération de recommandations personnalisées
WO2021185427A1 (fr) * 2020-03-16 2021-09-23 Huawei Technologies Co., Ltd. Génération de recommandations personnalisées
CN111581333A (zh) * 2020-05-07 2020-08-25 重庆大学 基于Text-CNN的影音播单推送方法及影音播单推送系统
CN111581333B (zh) * 2020-05-07 2023-05-26 重庆大学 基于Text-CNN的影音播单推送方法及影音播单推送系统
CN113765959A (zh) * 2020-06-30 2021-12-07 北京沃东天骏信息技术有限公司 信息推送方法、装置、设备及计算机可读存储介质
CN112507216A (zh) * 2020-12-01 2021-03-16 北京奇艺世纪科技有限公司 一种数据对象推荐方法、装置、设备和存储介质
CN113254777B (zh) * 2021-06-07 2021-09-24 武汉卓尔数字传媒科技有限公司 信息推荐方法、装置、电子设备及存储介质
CN113254777A (zh) * 2021-06-07 2021-08-13 武汉卓尔数字传媒科技有限公司 信息推荐方法、装置、电子设备及存储介质

Also Published As

Publication number Publication date
HUP1000408A2 (en) 2012-03-28
HU1000408D0 (en) 2010-10-28

Similar Documents

Publication Publication Date Title
US8676736B2 (en) Recommender systems and methods using modified alternating least squares algorithm
WO2012013996A1 (fr) Systèmes et procédés de recommandation
Zarzour et al. A new collaborative filtering recommendation algorithm based on dimensionality reduction and clustering techniques
Guan et al. Matrix factorization with rating completion: An enhanced SVD model for collaborative filtering recommender systems
Mehta et al. A review on matrix factorization techniques in recommender systems
US20210027160A1 (en) End-to-end deep collaborative filtering
Yang et al. A survey of collaborative filtering-based recommender systems for mobile internet applications
Pilászy et al. Fast als-based matrix factorization for explicit and implicit feedback datasets
CN106940801B (zh) 一种用于广域网络的深度强化学习推荐系统及方法
US20110153663A1 (en) Recommendation engine using implicit feedback observations
KR20160064447A (ko) 협력적 필터링의 예측 선호도를 이용한 처음 사용자에 대한 추천 제공 방법
WO2014100321A2 (fr) Intégration d'attribut dans une factorisation matricielle
KR20160064448A (ko) 유사 집합의 예상 선호도 대비 기반 아이템 추천 제공 방법
CA3111094C (fr) Evaluation de contraste de bruit destinee au filtrage collaboratif
US20220197978A1 (en) Learning ordinal regression model via divide-and-conquer technique
Seo et al. Video on demand recommender system for internet protocol television service based on explicit information fusion
Salmani et al. Hybrid movie recommendation system using machine learning
US11653064B2 (en) Methods and systems for determining disliked content
CN108763515B (zh) 一种基于概率矩阵分解的时间敏感个性化推荐方法
Hidasi et al. Enhancing matrix factorization through initialization for implicit feedback databases
WO2017095371A1 (fr) Recommandations de produits en fonction de caractéristiques d'utilisateurs et de produits sélectionnées
Said et al. Benchmarking: A methodology for ensuring the relative quality of recommendation systems in software engineering
Alshammari et al. Improved movie recommendations based on a hybrid feature combination method
Mehta et al. An empirical analysis on SVD based recommendation techniques
Li et al. Matrix factorization model with dual preferences for rating prediction

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 11748983

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 11748983

Country of ref document: EP

Kind code of ref document: A1