WO2015006942A1

WO2015006942A1 - A method and apparatus for learning user preference with preservation of privacy

Info

Publication number: WO2015006942A1
Application number: PCT/CN2013/079546
Authority: WO
Inventors: Jilei Tian; Alvin CHIN; Yang Cao; Guangfu SUN
Original assignee: Nokia Corporation; Nokia (China) Investment Co. Ltd.
Priority date: 2013-07-17
Filing date: 2013-07-17
Publication date: 2015-01-22

Abstract

A method for learning user preference with preservation of privacy may comprise: collecting, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user; learning group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and estimating item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group.

Description

A METHOD AND APPARATUS FOR LEARNING USER PREFERENCE WITH

PRESERVATION OF PRIVACY

FIELD OF THE INVENTION

The present invention generally relates to communication networks. More specifically, the invention relates to a method and apparatus for learning user preference with preservation of privacy.

BACKGROUND

The modern communications era has brought about a tremendous expansion of communication networks. Communication service providers and device manufacturers are continually challenged to deliver value and convenience to consumers by, for example, providing compelling network services, applications, and contents. The development of communication technologies has contributed to an insatiable desire for new functionality. One area of interest is the development of services and technologies for bringing intelligent and personalized network (such as personalized Internet services and recommendations) to a user' s pocket. Web content is becoming explosive, and offering relevant contents by recommendation would greatly improve personalized user experience, because individual users have obviously different needs on the huge content space that the user cannot exhaustively explore with. Learning user preference plays a very important role for offering a personalized service to a user to discover the contents relevant to them, but often intrudes the user's privacy. Thus, it is desirable to design a privacy-preserving approach to learn user preference.

SUMMARY

The present description introduces a novel approach to learn a user's preference according to the user's feedback information over groups of items. By applying this approach, it is easy to collect data about the user's preference while keeping the user's privacy.

According to a first aspect of the present invention, there is provided a method comprising: collecting, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user; learning group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and estimating item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group.

According to a second aspect of the present invention, there is provided an apparatus comprising: at least one processor; and at least one memory comprising computer program code, the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus at least to: collect, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user; learn group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and estimate item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group.

According to a third aspect of the present invention, there is provided a computer program product comprising a computer-readable medium bearing computer program code embodied therein for use with a computer, the computer program code comprising: code for collecting, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user; code for learning group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and code for estimating item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group.

According to a fourth aspect of the present invention, there is provided an apparatus comprising: collecting means for collecting, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user; learning means for learning group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and estimating means for estimating item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group.

According to a fifth aspect of the present invention, there is provided a method comprising: facilitating access to at least one interface configured to allow access to at least one service, the at least one service configured to at least perform the method in the first aspect of the present invention.

According to exemplary embodiments, the feedback information may indicate: one or more user actions, one or more user comments, staying time, or a combination thereof. The group rating data of the one or more users may be used as user-group rating parameters for a latent factor approach. For example, the latent factor approach may be based at least in part on a matrix factorization model to reflect relationships among users, items and groups. The users, the items and the groups in the matrix factorization model may be represented by corresponding latent factors.

In accordance with exemplary embodiments, said estimating the item rating data of the one or more users may comprise: applying a data mining algorithm to the group rating data of the one or more users; and mining reversely the item rating data of the one or more users, based at least in part on the group rating data of the one or more users. For example, the data mining algorithm may comprise a matrix factorization algorithm.

According to exemplary embodiments, a personalized service may be recommended to at least one of the one or more users, based at least in part on the estimated item rating data of the at least one of the one or more users. For example, said recommending the personalized service to the at least one of the one or more users may comprise: offering a group of items ranked based at least in part on preferences of the at least one of the one or more users on individual items. Optionally, feedback information related to the personalized service may be collected from the at least one of the one or more users, for adaptively updating the estimated item rating data of the one or more users.

In exemplary embodiments of the present invention, the provided methods, apparatus, and computer program products can enable user data to be collected easily with less bothering and better user experience. Particularly, the proposed solution collects user data with a high-level granularity or a large granularity, instead of having user data with a refined low-level granularity or a small granularity, thus user privacy can be preserved. Meanwhile, good quality data can be got due to the high-level granularity, especially considering that the detailed data usually have much noise.

BRIEF DESCRIPTION OF THE DRAWINGS

The invention itself, the preferable mode of use and further objectives are best understood by reference to the following detailed description of the embodiments when read in conjunction with the accompanying drawings, in which:

Fig.l illustrates an exemplary system architecture in which a solution proposed in accordance with exemplary embodiments of the present invention can be implemented;

Fig.2 shows an exemplary user interface for providing personalized browsing discovery user experience, in accordance with an embodiment of the present invention;

Fig.3 is a flowchart illustrating a method for learning user preference with preservation of privacy, in accordance with embodiments of the present invention;

Fig.4 illustrates an exemplary matrix factorization model applicable to privacy friendly user preference learning, in accordance with an embodiment of the present invention;

Fig.5 exemplarily shows experimental results in accordance with an embodiment of the present invention;

Fig.6 exemplarily shows experimental results in accordance with another embodiment of the present invention; Fig.7 exemplarily shows experimental results in accordance with still another embodiment of the present invention; and

Fig.8 is a simplified block diagram of various apparatuses which are suitable for use in practicing exemplary embodiments of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

The embodiments of the present invention are described in detail with reference to the accompanying drawings. Reference throughout this specification to features, advantages, or similar language does not imply that all of the features and advantages that may be realized with the present invention should be or are in any single embodiment of the invention. Rather, language referring to the features and advantages is understood to mean that a specific feature, advantage, or characteristic described in connection with an embodiment is included in at least one embodiment of the present invention. Furthermore, the described features, advantages, and characteristics of the invention may be combined in any suitable manner in one or more embodiments. One skilled in the relevant art will recognize that the invention may be practiced without one or more of the specific features or advantages of a particular embodiment. In other instances, additional features and advantages may be recognized in certain embodiments that may not be present in all embodiments of the invention.

Personalized services may be offered or recommended to individual users by learning user preferences. The recommendation services are applicable to various communication environments such as Internet, online mobile network, or any other systems suitable for offering personalized services. Many recommendation methods may be applied in the commercial services. For example, with the content-based approach, a user is recommended items similar to those the user preferred in the past, while the collaborative filtering approach recommends items which people with similar tastes and preferences liked in the past. The two approaches can be combined for recommending items to a user. When applying these approaches, user preference or rating for a given item is required, and the content-based approach even requires the detailed information for the given item. Some recommendation systems collect a user's feedback information indirectly using machine learning inference, such as how long does the user stay, but they still monitor the user' s log or footprint information to make user annoying. Due to the user's concern on the privacy, he/she does not allow to give such information, which makes the recommendation hardly perform well. Some of recommendation systems ask a user to answer a series of questions or obtain the user' s opinion over each of contents in order to learn the user' s preference for offering the personalized service, due to knowing more detailed, better profile learning. However, it brings a challenge of improving user experience because it is not natural for a user to follow, and more important, it is often intrusive for the user's privacy. Thus, it is benefit to propose a novel solution to collect user data at a large granularity instead of a refined granularity, so as to learn user preferences on the provided contents with preservation of privacy.

Fig.l illustrates an exemplary system architecture in which a solution proposed in accordance with exemplary embodiments of the present invention can be implemented. As shown in Fig.l, the system architecture 100 may comprise one or more User Equipments (UEs) such as UE 101a and UE 101b which may communicate with a server 107 or other UEs through a communication network 105.

The server 107 can query or acquire specified information (such as user profile data) related to the UEs from a database 109. The UE can obtain content services offered by one or more service providers l l la-l l ln via the communication network 105. By way of example, the communication network 105 may comprise one or more networks such as a data network, a wireless network, a telephony network, or any combination thereof. It is contemplated that the data network may comprise any of

Local Area Network (LAN), Metropolitan Area Network (MAN), Wide Area Network (WAN), a public data network (such as the Internet), a self- organized mobile network, or any other suitable packet- switched network, such as a commercially owned, proprietary packet- switched network like a proprietary cable or fiber-optic network. In addition, the wireless network may comprise, for example, a cellular network, a peer-to-peer network or the like, and can employ various technologies comprising Enhanced Data rates for Global Evolution (EDGE), General Packet Radio Service (GPRS), Global System for Mobile communications (GSM), Internet protocol Multimedia Subsystem (IMS), Universal Mobile Telecommunications System (UMTS), etc., as well as any other suitable wireless medium, such as Worldwide interoperability for Microwave Access (WiMAX), Wireless Local Area Network (WLAN), Long Term Evolution (LTE) networks, Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Wireless Fidelity (WiFi), satellite, Mobile Ad-hoc Network (MANET), and the like.

As shown in Fig.l, the UEs 101a- 101b may be utilized to perform one or more applications such as a browser application, a music application, a reading application, among other application typically used within a mobile device or computing device.

For example, the application 103a- 103b can communicate to the server 107 for accessing contents (such as web pages, songs, movies and/or the like) from one or more service providers l l la-l l ln, and subsequently rendering the accessed contents to a user via a user interface. In an exemplary embodiment, the contents from the service provider 111 can be provided to the UE 101 by the server 107 through the communication network 105. For example, the information format and layout of web pages in the service provider 111 are generally designed for a computer device which has a big-size display and a strong processing capability. Thus, these web pages are not suitable to be rendered at a mobile phone which has a small-size display. Then, the server 107 can filter out some unnecessary information in the web pages, adjust the layout of the web pages according to the condition of the display of the UE 101 and then provide the adjusted web pages to the UE 101. In an exemplary embodiment, the server 107 can recommend and/or push a group of contents/items to the UE 101, such as hot news, articles, images, books, advertisements and/or the like. Optionally, the server 107 can obtain user information from the database 109, for example user profile data, user settings, user browsing histories and/or other information relevant to user actions or comments on the recommendation. Thus, personalized services can be provided to a user according to relevant and available information.

Fig.2 shows an exemplary user interface for providing personalized browsing discovery user experience, in accordance with an embodiment of the present invention. Internet browsing user experience has gone through from search (for example, user intent driven), to portal (for example, browsing with organized information by a service provider), and then to discovery now (for example, personalized, local and timely contents). The recommendation presented through the exemplary user interface in Fig.2 can provide personalized browsing discovery user experience. To enable personalized discovery user experience, it may be needed to automatically learn a user's browsing behavior and preference from the browsing history and recommend the relevant group of contents to the user, as shown in Fig.2, which may comprise for example web domain, story page, image, video, download, advertisement, and/or the like.

The most challenge in the personalized services is that the service provider needs to learn user preferences from user data, while users are not willing to share their own data due to the privacy issue. On the other hand, it may be hard to collect detailed and accurate user information since it is a tedious task. For example, it may be difficult to collect user information on every detailed item, especially when the user may also feel hard to tell the exact preference on every item. According to exemplary embodiments, a novel solution is proposed to learn a user's preference based at least in part on the user's feedback over groups of items, instead of the user's feedback over individual items, so it is easy to collect user data while keeping user privacy. More details of the proposed solution will be illustrated hereinafter by way of example with reference to the accompanying drawings.

Fig.3 is a flowchart illustrating a method for learning user preference with preservation of privacy, in accordance with embodiments of the present invention. It is contemplated that the method described herein may be used with any apparatus (for example, UE 101a or UE 101b shown in Fig.l) connected to a service network such as Internet, online mobile network or any other systems suitable for providing or supporting personalized services and recommendations. The UE may be any type of mobile terminal, fixed terminal, or portable terminal comprising a mobile handset, station, unit, device, multimedia computer, multimedia tablet, Internet node, communicator, desktop computer, laptop computer, notebook computer, netbook computer, tablet computer, personal communication system (PCS) device, personal navigation device, personal digital assistants (PDAs), audio/video player, digital camera/camcorder, positioning device, television receiver, radio broadcast receiver, electronic book device, game device, or any combination thereof, comprising the accessories and peripherals of these devices, or any combination thereof.

Additionally or alternatively, it is also contemplated that the method described herein may be used with any apparatus (for example, server 107 or service provider 111 shown in Fig.l) providing or supporting personalized services and recommendations, such as a network node operated by services providers or network operators. The network node may be any type of network device comprising Base Station (BS),

Access Point (AP), server, control center, service platform, or any combination thereof. In an exemplary embodiment, the method may be implemented by processes executing on various apparatuses which communicate using an interactive model

(such as a client- server model) of network communications. For example, the proposed solution may be performed at a UE, a network node, or both of them through communication interactions for personalized services and/or recommendations.

According to exemplary embodiments, feedback information of one or more users can be collected with a granularity at a group level, as shown in block 302 of Fig.3, and the feedback information of an associated user among the one or more users may be related to at least one group of items offered to the associated user. The at least one group of items may comprise service contents provided to the user, such as web page, browsing widget, music playlist, movie, book, shopping and/or the like. For example, as used herein, the term "item" refers to any data that can be represented in machine-readable form, including data that can be used to present content for observation by a human, such as photo, text, audio, music, image, video, thumbnail representation of a larger image, game, graph, table, map, diagram, identifier such as Uniform Resource Locator (URL), document, publication and/or spreadsheet, among others. According to exemplary embodiments, the same group of items may be offered to different users, or the one or more users may be provided with distinct groups of items. For example, groups 1&2&3 may be offered to user A, while user B may be provided with groups (such as groups 1&2&3 or groups 1&3 or groups 2&5&7) which have an overlap with groups 1&2&3, or with groups (such as group 4 or groups 6&8) which have nothing in common with groups 1&2&3. The group size which represents the number of items in a group may be fixed or distinct for different groups. The same item may be assigned into different groups optionally. Thus, different users may observe the same item even if they are offered different groups.

In accordance with exemplary embodiments, the feedback information of the associated user may indicate: one or more user actions (such as open, share, forward, filter, favorite and/or the like), one or more user comments (such as like or not, recommend or not, subscribe or not, and so on), staying time (for example, less than five minutes taken for browsing a web page, two hours taken for looping a music playlist, and so on), or a combination thereof. For example, the feedback information can be collected through a user interface, a tracker, a proxy, an information platform, or any other units suitable for collecting user data at a large granularity. The method illustrated with respect to Fig.3 enables user data (such as preferences or opinions on a group of items) to be collected easily without revealing personal privacy of a user, since the feedback information of the user is collected with the granularity at the group level, instead of a granularity at an item level. As such, the collected feedback information of the associated user can reflect user preferences or ratings over the respective offered groups, while keeping preservation of the detailed user preferences or ratings over individual items. Similar metaphor from a location-aware service is that a user is getting nervous if the service provider or system records the user's Global Positioning System (GPS) data all the time. However, user experience may be much improved if the service provider or system only records the user's location at a level of city or state.

In block 304 of Fig.3, group rating data of the one or more users can be learned from the feedback information of the one or more users, and the group rating data of the associated user indicates a rating of the associated user on the at least one group. For example, for a given music playlist, a user may perform some actions on the music playlist, such as repeatedly listen to some of tracks, and skip for some other tracks. From the feedback information (which may indicate these user actions, the user's comment on the music playlist, the user's overall duration for the music playlist and/or the like), the group rating data of the user for the music playlist can be learned, which may reflect a preference or rating of the user on this music playlist. In another example, tens of stories are recommended as top trending contents for a given user. Then the user's preference or rating over the entire recommended contents can be learned, for example, by monitoring the overall duration for which the user stays with the recommended contents, the user's implicit/explicit opinion on the entire recommended contents, and/or the like. Based at least in part on the group rating data of the one or more users, item rating data of the one or more users can be estimated, as shown in block 306 of Fig.3, and the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group. The estimation of user preferences on the individual items may involve a process of reverse data mining, which can be illustrated by a simple example where the user has a high rating for group 1, and a low rating for group 2. Assuming that there is only one item different between group 1 and group 2, it is natural to believe that the low rating is highly associated with this item. In case that there is a plurality of groups shared among many users, overlaps or gaps among the groups can offer a clue for learning user preferences of these users on individual items in the groups. Thus, preference/rating data of a user over one or more groups of items can be used to reveal his/her interests over respective items along the time. In an exemplary embodiment, the individual items on which the preferences of the associated user is indicated by the corresponding item rating data may further comprise items in one or more groups offered to other users among the one or more users. In this case, even if the user does not observe those items in the one or more groups offered to other users, feedback information and associated group rating data of the other users may be utilized together with those of the user to estimate the item rating data of the user, for indicating the preferences of the user on those items not in the at least one group offered to this user. For example, if user A and user B have similar preferences considering their history records, it is reasonable to believe that user A would make the same or similar rating on a specified group or item with user B, even though the specified group or item is only offered to user B but not to user A.

In accordance with an exemplary embodiment, the item rating data of the one or more users can be estimated by: applying a data mining algorithm to the group rating data of the one or more users; and mining reversely the item rating data of the one or more users, based at least in part on the group rating data of the one or more users. For example, a latent factor approach can be used in data mining. Particularly, the data mining algorithm may comprise a matrix factorization algorithm. It is noted that the latent factor approach and the matrix factorization algorithm mentioned here are merely exemplary and not intended to suggest any limitation as to the scope of use or functionality of embodiments of the invention described herein. It could be understood that the item rating data of the one or more users may be estimated by using any other data mining algorithm suitable for performing reverse data mining on the group rating data of the one or more users.

In an exemplary embodiment where the latent factor approach is used, some specified elements (such as user, item, group of items, and even category or tag associated with an item and a group) can be factored into a latent space. For simplicity, it is feasible to take only part of the elements (such as user, item and group) into account. The group rating data of the one or more users learned in block 304 of Fig.3 may be used as user-group rating parameters for the latent factor approach. Then the associated user-item rating parameters can be predicted or estimated with the latent factor approach by reversely data mining. As an example, the latent factor approach may be based at least in part on a matrix factorization model to reflect relationships among users, items and groups. The users, the items and the groups in the matrix factorization model can be represented by corresponding latent factors. For example, the latent factor can be a type or a topic which may illustrate the users' preferences, the items' features and/or the groups' semantic meanings. An exemplary matrix factorization model applied in the latent factor approach will be described hereafter with respect to Fig.4.

According to exemplary embodiments, the method described with respect to Fig.3 may further comprise: recommending a personalized service to at least one of the one or more users, based at least in part on the estimated item rating data of the at least one of the one or more users. For example, said recommending the personalized service to the at least one of the one or more users may comprise: offering a group of items ranked based at least in part on preferences of the at least one of the one or more users on individual items. Such personalized service recommendation can protect user privacy by offering a group of items according to the user's opinion on the group instead of a particular item, while providing better user experience with less bothering. In an exemplary embodiment, feedback information related to the recommended personalized service may be collected from the at least one of the one or more users, for adaptively updating the estimated item rating data of the one or more users. Actually, a reciprocity relationship exists between personalized service recommendation and user preference learning. The user preference learning can output preference estimation results to assist in recommending the personalized service which can in turn be observed by the user to provide feedback information for improving the performance of user preference learning. Optionally, the method described with respect to Fig.3 can be performed iteratively for the personalized service recommendation. In this case, a user's opinion on any new recommended group of contents or items can be collected for automatically learning group rating data of the user on the new recommended group, so as to adaptively update preference estimations on individual contents or items to improve the recommendation performance. In general, the more the number of groups about which the feedback information is collected, the better the preference estimation performance can be achieved statistically. Alternatively or additionally, the preference estimation performance is also affected by the number of items in a group, which can be observed from the experimental results shown in Figs.5-7.

The various blocks shown in Fig.3 may be viewed as method steps, and/or as operations that result from operation of computer program code, and/or as a plurality of coupled logic circuit elements constructed to carry out the associated function(s).

The schematic flow chart diagrams described above are generally set forth as logical flow chart diagrams. As such, the depicted order and labeled steps are indicative of specific embodiments of the presented methods. Other steps and methods may be conceived that are equivalent in function, logic, or effect to one or more steps, or portions thereof, of the illustrated methods. Additionally, the order in which a particular method occurs may or may not strictly adhere to the order of the corresponding steps shown.

Fig.4 illustrates an exemplary matrix factorization model applicable to privacy friendly user preference learning, in accordance with an embodiment of the present invention. To protect a user's privacy, the proposed solution can recommend a group of items (such as news, articles, URLs and/or the like) according to the user's comment and/or action (such as click, browsing behavior and/or the like) on the group instead of a particular item. For example, the group may contain many topics, which may be considered when the group is factored into a latent space to build relationships among users, groups and items in the matrix factorization model. Similarly, some features associated with the user and the item which are related to preference estimation or prediction also may be considered when the user and the item are factored into the latent space. In an exemplary embodiment, it is assumed that the statistic samples used for the matrix factorization model comprise N users, M groups and K items, and the number of latent factors of each user/item/group is D. Then, user factor matrix U e ¾^DxiV is a D N matrix, group factor matrix G e ¾^DxM is a D x M matrix, and item factor matrix V e 9l^DxK is a D x K matrix. The values in the factor matrices can be learned from the matrix factorization model.

In accordance with exemplary embodiments, group-items information and user ratings information at a group level (which can be used to form a user-group rating matrix) are collected and used for the matrix factorization model. The group-items information can explain what items are used to form a specific group, and a rating element in the user- group rating matrix can reflect a specific user's preference over a user privacy friendly group, rather than detailed preference at an item level. Thus consumers or users are happy on their personal data collection. With the information collected at a group level from a user, specific items can be recommended to the user, for example, by applying the matrix factorization model. As described previously, three low-rank matrices, user factor matrix U e ¾^DxiV _? group factor matrix G e ¾^ΰχΜ ₅ and item factor matrix V e 9l^DxK , are introduced here, where each row of the matrix represents a specific user/item/group's latent factor vector. For example, U_u represents a user-specific u latent factor vector, where u is a positive integer less than or equal to N; G_g represents a group- specific g latent factor vector, where g is a positive integer less than or equal to ; and V_v represents an item-specific v latent factor vector, where v is a positive integer less than or equal to K. In particular, a

N x M matrix R is also introduced here as the user-group rating matrix in which a value of an element can explain a user's preference on a specific group. For example, as shown in Fig.4, the observed rating of user u on group g can be denoted as matrix element R_u , which can explain a preference of user u on group g. Let N_g denote the item set in group g, and the size of item set N_g is represented by m = N as shown in Fig.4. For the specified group g, the item- specific (such as g\, g2 ... gm shown in Fig.4) latent factor vectors are denoted as V_gl , V_g2 ... V_gm , respectively.

Correspondingly, the weight of item k in group g is denoted as T _k , where k e [l,m]. The value of the weight T _k can be set according to the importance of

1

item k, or simply as ,— r for all items in group g.

N g

According to exemplary embodiments, the matrix factorization model can be trained to learn values in the factor matrices and get the trained matrices U , V and G . Based at least in part on the trained matrices U and V , a predicted rating R_{u v}

(a rating of user u on a specific item v) can be calculated as R_{u v} = U_U ^TV_V , where superscript T denotes vector transpose. According to the predicted ratings of the user on individual items, specific items can be recommended to the user. As such, the proposed solution can reconstruct user preferences at the item level for recommendation purpose with acceptable performance. In order to learn the values in the low-rank matrices, for example, a probability density function of the condition distribution over t s:

where w(R_»g g(t/jG_g ) ^ ) is a probability density function of the Gaussian distribution for R , g(t/ G_g )= l/(l + e^{~Uu Gg}

s a hyper-parameter of user-group rating matrix R , and is an indicator function which is equal to 1 if user u rated group g and equal to 0 otherwise. As an example, zero-mean spherical Gaussian priors are placed on the user factor vector U_u and the item factor vector V_v , which can be respectively expressed as:

where

of the Gaussian distribution for U_u with mean value 0 and variance σ² 1 , σ² is a hyper-parameter of user factor matrix U ,

i) is a probability density function of the Gaussian distribution for V_v with mean value 0 and variance I , is a hyper-parameter of item factor matrix V , and I is an identity matrix. For the group latent factor vectors, two features may be considered: (i) the zero-mean Gaussian prior to avoid over- fitting, as shown in equation (4); and (ii) the condition distribution of group latent factors given the latent features of their composed items, as shown in expression (5).

where w G_g |0,< ^ I is a probability density function of the Gaussian distribution for G_g with mean value 0 and variance σ I , σ is a hyper-parameter of group factor matrix G , N represents the item set in group g, and T is a weight of item v to

1

group g and can be equal to for all items in N or set according to the

N_r importance of item v. Then the probabilistic form of the above expression can be written as:

where symbol cc denotes proportional, W G Y T V ,σ_τ ² \ is a probability density function of the Gaussian distribution for G with mean value ∑ _v and veJV.

variance I, is a hyper-parameter of weight matrix T . Thus, the conditional distribution for the group latent factors can be written as:

The above distributions will keep the group latent factor vectors both small and close to the features of their composed items. For example, through a Bayesian inference, the following expression for the posterior probability of latent feature vectors can be derived given the group-items information and the ratings information at the group level. p(u,G,V Κ,Τ,σ²,σ²,σ²,σ_κ ²,σ²)

oc _P(R U,G,a_R ²)x p(cT,V,a_T ²)x p{ca_G ²)x p(u σ²)χ pfy

Then the logarithm of the above posterior distribution can be calculated as:

r(U_}G^F\ R, T, σ£_;> σ , σ£, < , o ) c

-

where C is a constant independent of other parameters. The posterior distribution can be maximized for getting the trained matrices U , V and G , which is equivalent to minimizing the lost function in equation (10).

where λ_υ , A_y , A_j, and A_G are the regularized hyper-parameters, and particularly, A_u =

. Then a local minimum of the above lost function as an objective function can be found by performing gradient decent respectively on U_u, G_g and V_v as follows. _8L_ M

(11)

(12)

(13) where g(i/ G 8) is the derivative of the logistic function g u_u ^TG_g which can be calculated as:

In an exemplary embodiment, the initial element values of the trained matrices

U , G and V may be samples from normal noise with zero mean. In each iteration for training these latent factor matrices, the element values of U , V and G can be updated based at least in part on the latent variables from the previous iteration. Then the user-item rating R_{u v} (a rating of user u on a specific item v) can be predicted or estimated by using the trained U and V , for example, through the following equation.

R = U^TV (15)

With the predicted user-item ratings, specific items can be recommended to the user. Optionally, the recommended items may be ordered by ranking the predicted user-item ratings.

Figs.5 -7 exemplarily show experimental results in accordance with various embodiments of the present invention. The experimental data set comprises 100,000 ratings (which may be set as for example any one of 1 to 5) from 943 users on 1682 movies. Each user gives ratings on some of the movies he/she knows well, and the user has rated at least 20 movies. According to an exemplary embodiment, the user factors and the item factors can be obtained by applying the latent factor approach (for example, through performing probabilistic matrix factorization) on training data from the experimental data set. The group information can be obtained by randomly assigning these movies as items into respective groups. For example, each item may at least belong to one group, and the group size may be fixed. It would be realized that the group size may not be fixed in practice, and it may be possible that a group is a subset of another group. As an exam le, user- roup ratings can be simulated as:

where ^N. is a number representing the size of the item set N_g in group g. Using the simulated user-group ratings as a training set, the corresponding predicted user-item ratings can be obtained from the proposed matrix factorization model. The Root Mean Square Error (RMSE) with respect to the predicted user-item ratings and the real user-item ratings in a test set (which may be from the experimental data set) can be calculated by:

where p_t denotes the ith predicted rating, r, denotes the ith real rating, and Q is the number of ratings in the test set. The variation of RMSE can reflect prediction performance on the user-item ratings. Fig.5 shows the impact of the number of groups in case that the group size is fixed as 5. It can be seen from Fig.5 that the greater number of groups yields the better experimental result, which means feedback information collected from more groups would result in the better performance of learning user preference on individual items. Fig.6 shows the impact of the group size in case that the number of groups is 1000. When the number of the groups is fixed, the smaller the group size is, the better the experimental result can be achieved. This is apprehensible since the smaller group size enables the overlaps or gaps among the groups to offer the more obvious clue for learning user preferences on individual items in the groups. Fig.7 shows the impact of parameter λ_τ in case that the number of groups is 1000 and the group size is fixed as 5. Parameter _j. controls the divergence of the groups' latent factors from their contained items. For example, a larger value of may indicate more reliance on the items' information. If is too large, it will lead to over fitting. Therefore, it is needed to select a suitable value for , as illustrated in Fig.7, especially from the view point of optimization.

The proposed solution can create a mechanism to predict or estimate a user's preference or opinion on individual items based at least in part on automatically learning of the user's browsing behaviors and preferences from the browsing histories and opinions over a plurality of groups of recommended contents/items. For example, for a recommended video or radio playlist, the user may repeatedly play some items therein, and skip for others. Based at least in part on the user's overall duration or rating on the given playlist, the user's preference or rating on the playlist can be found. Then personalized items may be provided to the user accordingly by applying the proposed solution to estimate the user' s preference or rating on individual videos

(which may comprise the videos in the playlist and even those not in the playlist). For example, the proposed solution can collect the user's opinion data on a group of individual items, learn the user's preference over the entire recommended items, and apply a suitable data mining algorithm to reversely mine the user's preference for each of individual items (some of them may be absent from the recommended group) by using the collected user's opinion data on the group of items. In an exemplary embodiment, a latent factor approach such as matrix factorization is used to model user-group ratings and build corresponding relationships in the common space among users, items and groups. It would be appreciated that the modeling also can be extended to content semantics of groups and items. As an example, parameters associated with the matrix factorization model can be estimated by using probability matrix factorization or matrix factorization, and user-item ratings can be predicted accordingly. Optionally, by knowing the predicted preferences or ratings on the individual items, the proposed solution can recommend the personalized service by offering a group of ranked items based at least in part on the individual ranking that matters to the user. Particularly, the proposed solution can be iteratively performed to adaptively update preference estimations on the individual items by getting the user' s opinion on the new recommended group of items, so as to improve the recommendation performance. Many advantages can be achieved by using the solution provided by the present invention. Since the solution can only collect the user's data with a large granularity (for example, at a group level), instead of having the user's data with a refined small granularity (for example, at an item level), the user's data can be gathered easily with less bothering, while the user's privacy is preserved, which brings better user experience. Meanwhile, good quality data can be obtained due to the large granularity, considering detailed data with the small granularity often have much noise.

Fig.8 is a simplified block diagram of various apparatuses which are suitable for use in practicing exemplary embodiments of the present invention. In Fig.8, a UE 810 (such as mobile phone, wireless terminal, portable device, PDA, multimedia tablet, desktop computer, laptop computer and etc.) may be adapted for communicating with a network node 820 (such as a server, an AP, a BS, a control center, a service platform and etc.). In an exemplary embodiment, the UE 810 may comprise at least one processor (such as a data processor (DP) 81 OA shown in Fig.8), and at least one memory (such as a memory (MEM) 810B shown in Fig.8) comprising computer program code (such as a program (PROG) 8 IOC shown in Fig.8). The at least one memory and the computer program code may be configured to, with the at least one processor, cause the UE 810 to perform operations and/or functions described in combination with Figs.1-4. In an exemplary embodiment, the UE 810 may optionally comprise a suitable transceiver 810D for communicating with an apparatus such as another UE, a network node (such as the network node 820) and so on. The network node 820 may comprise at least one processor (such as a data processor (DP) 820A shown in Fig.8), and at least one memory (such as a memory (MEM) 820B shown in Fig.8) comprising computer program code (such as a program (PROG) 820C shown in Fig.8). The at least one memory and the computer program code may be configured to, with the at least one processor, cause the network node 820 to perform operations and/or functions described in combination with Figs.1-4. In an exemplary embodiment, the network node 820 may optionally comprise a suitable transceiver 820D for communicating with an apparatus such as another network node, a UE (such as UE 810) or other network entity (not shown in Fig.8). For example, at least one of the transceivers 810D, 820D may be an integrated component for transmitting and/or receiving signals and messages. Alternatively, at least one of the transceivers 810D, 820D may comprise separate components to support transmitting and receiving signals/messages, respectively. The respective DPs 81 OA and 820A may be used for processing these signals and messages.

Alternatively or additionally, the UE 810 and the network node 820 may comprise various means and/or components for implementing functions of the foregoing method steps described with respect to Figs.1-4. According to exemplary embodiments, an apparatus (such as the UE 810 or the network node 820) may comprise: collecting means for collecting, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user; learning means for learning group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and estimating means for estimating item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group. Optionally, the apparatus may further comprise: recommending means for recommending a personalized service to at least one of the one or more users, based at least in part on the estimated item rating data of the at least one of the one or more users. Alternatively, the above mentioned collecting means, learning means, estimating means and recommending means may be implemented at either the UE 810 or the network node 820, or at both of them in a distributed manner. In an exemplary embodiment, a solution providing for the UE 810 and/or the network node 820 may comprise facilitating access to at least one interface configured to allow access to at least one service, and the at least one service may be configured to at least perform functions of the foregoing method steps as described with respect to Figs.1-4.

At least one of the PROGs 8 IOC and 820C is assumed to comprise program instructions that, when executed by the associated DP, enable an apparatus to operate in accordance with the exemplary embodiments, as discussed above. That is, the exemplary embodiments of the present invention may be implemented at least in part by computer software executable by the DP 81 OA of the UE 810 and by the DP 820A of the network node 820, or by hardware, or by a combination of software and hardware.

The MEMs 810B and 820B may be of any type suitable to the local technical environment and may be implemented using any suitable data storage technology, such as semiconductor based memory devices, flash memory, magnetic memory devices and systems, optical memory devices and systems, fixed memory and removable memory. The DPs 81 OA and 820A may be of any type suitable to the local technical environment, and may comprise one or more of general purpose computers, special purpose computers, microprocessors, digital signal processors (DSPs) and processors based on multi-core processor architectures, as non-limiting examples.

In general, the various exemplary embodiments may be implemented in hardware or special purpose circuits, software, logic or any combination thereof. For example, some aspects may be implemented in hardware, while other aspects may be implemented in firmware or software which may be executed by a controller, microprocessor or other computing device, although the invention is not limited thereto. While various aspects of the exemplary embodiments of this invention may be illustrated and described as block diagrams, flow charts, or using some other pictorial representation, it is well understood that these blocks, apparatus, systems, techniques or methods described herein may be implemented in, as non-limiting examples, hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controller or other computing devices, or some combination thereof.

It will be appreciated that at least some aspects of the exemplary embodiments of the inventions may be embodied in computer-executable instructions, such as in one or more program modules, executed by one or more computers or other devices. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types when executed by a processor in a computer or other device. The computer executable instructions may be stored on a computer readable medium such as a hard disk, optical disk, removable storage media, solid state memory, random access memory (RAM), and etc. As will be realized by one of skill in the art, the functionality of the program modules may be combined or distributed as desired in various embodiments. In addition, the functionality may be embodied in whole or in part in firmware or hardware equivalents such as integrated circuits, field programmable gate arrays (FPGA), and the like.

Although specific embodiments of the invention have been disclosed, those having ordinary skill in the art will understand that changes can be made to the specific embodiments without departing from the spirit and scope of the invention. The scope of the invention is not to be restricted therefore to the specific embodiments, and it is intended that the appended claims cover any and all such applications, modifications, and embodiments within the scope of the present invention.

Claims

CLAIMS What is claimed is:

1. A method comprising:

collecting, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user;

learning group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and

estimating item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group.

2. The method according to claim 1, wherein the feedback information indicates: one or more user actions, one or more user comments, staying time, or a combination thereof.

3. The method according to claim 1 or 2, wherein the group rating data of the one or more users is used as user-group rating parameters for a latent factor approach.

4. The method according to claim 3, wherein the latent factor approach is based at least in part on a matrix factorization model to reflect relationships among users, items and groups.

5. The method according to claim 4, wherein the users, the items and the groups in the matrix factorization model are represented by corresponding latent factors.

6. The method according to any one of claims 1 to 5, wherein said estimating the item rating data of the one or more users comprises:

applying a data mining algorithm to the group rating data of the one or more users; and

mining reversely the item rating data of the one or more users, based at least in part on the group rating data of the one or more users.

7. The method according to claim 6, wherein the data mining algorithm comprises a matrix factorization algorithm.

8. The method according to any one of claims 1 to 7, further comprising:

recommending a personalized service to at least one of the one or more users, based at least in part on the estimated item rating data of the at least one of the one or more users.

9. The method according to claim 8, wherein said recommending the personalized service to the at least one of the one or more users comprises: offering a group of items ranked based at least in part on preferences of the at least one of the one or more users on individual items.

10. The method according to claim 8 or 9, wherein feedback information related to the personalized service is collected from the at least one of the one or more users, for adaptively updating the estimated item rating data of the one or more users.

11. An apparatus, comprising:

at least one processor; and at least one memory comprising computer program code,

the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus at least to:

collect, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user;

learn group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and

estimate item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group.

12. The apparatus according to claim 11, wherein the feedback information indicates: one or more user actions, one or more user comments, staying time, or a combination thereof.

13. The apparatus according to claim 11 or 12, wherein the group rating data of the one or more users is used as user-group rating parameters for a latent factor approach.

14. The apparatus according to claim 13, wherein the latent factor approach is based at least in part on a matrix factorization model to reflect relationships among users, items and groups.

15. The apparatus according to claim 14, wherein the users, the items and the groups in the matrix factorization model are represented by corresponding latent factors.

16. The apparatus according to any one of claims 11 to 15, wherein the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus to estimate the item rating data of the one or more users comprises the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus at least to:

apply a data mining algorithm to the group rating data of the one or more users; and

mine reversely the item rating data of the one or more users, based at least in part on the group rating data of the one or more users.

17. The apparatus according to claim 16, wherein the data mining algorithm comprises a matrix factorization algorithm.

18. The apparatus according to any one of claims 11 to 17, wherein the at least one memory and the computer program code is further configured to, with the at least one processor, cause the apparatus to:

recommend a personalized service to at least one of the one or more users, based at least in part on the estimated item rating data of the at least one of the one or more users.

19. The apparatus according to claim 18, wherein the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus to recommend the personalized service to the at least one of the one or more users comprises: the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus to offer a group of items ranked based at least in part on preferences of the at least one of the one or more users on individual items.

20. The apparatus according to claim 18 or 19, wherein feedback information related to the personalized service is collected from the at least one of the one or more users, for adaptively updating the estimated item rating data of the one or more users.

21. A computer program product comprising a computer-readable medium bearing computer program code embodied therein for use with a computer, the computer program code comprising:

code for collecting, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user;

code for learning group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and

code for estimating item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group.

22. The computer program product according to claim 21, wherein the feedback information indicates: one or more user actions, one or more user comments, staying time, or a combination thereof.

23. The computer program product according to claim 21 or 22, wherein the group rating data of the one or more users is used as user-group rating parameters for a latent factor approach.

The computer program product according to claim 23, wherein the latent factor approach is based at least in part on a matrix factorization model to reflect relationships among users, items and groups.

25. The computer program product according to claim 24, wherein the users, the items and the groups in the matrix factorization model are represented by corresponding latent factors.

26. The computer program product according to any one of claims 21 to 25, wherein the code for estimating the item rating data of the one or more users comprises:

code for applying a data mining algorithm to the group rating data of the one or more users; and

code for mining reversely the item rating data of the one or more users, based at least in part on the group rating data of the one or more users.

27. The computer program product according to claim 26, wherein the data mining algorithm comprises a matrix factorization algorithm.

28. The computer program product according to any one of claims 21 to 27, wherein the computer program code further comprises:

code for recommending a personalized service to at least one of the one or more users, based at least in part on the estimated item rating data of the at least one of the one or more users.

29. The computer program product according to claim 28, wherein the code for recommending the personalized service to the at least one of the one or more users comprises: code for offering a group of items ranked based at least in part on preferences of the at least one of the one or more users on individual items.

30. The computer program product according to claim 28 or 29, wherein feedback information related to the personalized service is collected from the at least one of the one or more users, for adaptively updating the estimated item rating data of the one or more users.

31. An apparatus, comprising:

collecting means for collecting, with a granularity at a group level, feedback information of one or more users, wherein the feedback information of an associated user among the one or more users is related to at least one group of items offered to the associated user;

learning means for learning group rating data of the one or more users from the feedback information of the one or more users, wherein the group rating data of the associated user indicates a rating of the associated user on the at least one group; and estimating means for estimating item rating data of the one or more users based at least in part on the group rating data of the one or more users, wherein the item rating data of the associated user indicates preferences of the associated user on individual items comprising at least the items in the at least one group.

32. A method comprising facilitating access to at least one interface configured to allow access to at least one service, the at least one service configured to at least perform a method of at least one of claims 1 to 10.