GB2442024A

GB2442024A - Context sensitive user preference prediction

Info

Publication number: GB2442024A
Application number: GB0618561A
Authority: GB
Inventors: Nicolas Lhuillier; Makram Bouzid; Sandra C Gadanho
Original assignee: Motorola Inc
Current assignee: Motorola Solutions Inc
Priority date: 2006-09-20
Filing date: 2006-09-20
Publication date: 2008-03-26
Also published as: GB0618561D0

Abstract

A method for determining a user preference prediction comprises receiving a plurality of user preference indications with each user preference indication comprising content data and at least some of the plurality of user preference indications comprising context data. Each of the plurality of user preference indications are placed into at least one group (movies, cartoons etc.) in response to the content data, but without consideration of the context data. A context profile/signature for each of the plurality of groups is determined. A user preference prediction is determined for a certain context (e.g. now) by selecting a group in response to a comparison of the (certain) context and the context profiles/signatures of the plurality of groups. The user preference prediction is then determined in response to the selected group. The context data may comprise time or location data.

Description

CONTEXT SENSITIVE USER PREFERENCE PREDICTION

Field of the invention

The invention relates to context sensitive user preference prediction and in particular, but not exclusively, to user preference prediction for content item recommendation.

Background of the Invention

Personalisation and user adaptation of applications is becoming of increasing importance. The personalization of an *,,,. 15 application typically consists in computing predictions about future user actions or desires based on user S...

preferences collected during a learning phase. Predictions * may be used to adapt the user interface, to propose recommendations, etc. * 20 S.....

* For example, in recent years, the availability and provision of multimedia and entertainment content has increased substantially. E.g., the number of available television and radio channels has grown considerably and the popularity of the Internet has provided new content distribution means.

Consequently, users are increasingly provided with a plethora of different types of content from different sources. In order to identify and select the desired content, the user must typically process large amounts of information which can be very cumbersome and impractical.

Accordingly, significant resources have been invested in research into techniques and algorithms that may provide an improved personalisation and e.g. assist a user in identifying and selecting content of specific interest to the user.

For example, Digital Video Recorders (DVRs) or Personal Video Recorders (PVRS) which comprise functionality for providing recommendations of television programs to the user based on user preferences are becoming increasingly popular.

More specifically, such devices can comprise functionality for monitoring the viewing/recording preferences of a user.

These preferences can be stored in a user preference profile which subsequently can be used to autonomously select arid 15 recommend suitable television programs for viewing or recording. E.g. a DVR may automatically record programs which are then recommended to the user, for example by * inclusion of the automatically recorded programs in a listing of all the programs recorded by the DVR. * 20

**.*** * * In order to enhance the user experience, it is advantageous to personalise the recommendations to the individual user as much as is possible. In this context, a recommendation consists in predicting how much a user may like a particular content item and recommending it if it is considered of sufficient interest. The process of generating recommendations requires that user preferences have been captured so that they can be used as input data by the prediction algorithm.

Advanced personalization actions do not only take user preferences into account but also considers the user context, e.g. the situation in which the user prediction is generated and/or used to predict the user's preference. This leads to the creation of contextual user profiles wherein the user's preference also depends on the context. However, context information is often difficult to capture and compare and conventional approaches tend to be inflexible, result in suboptimal predictions and/or tend to be impractical for obtaining, maintaining and customizing the context information.

Specifically, current context profile solutions almost always assume that the different user sub-profiles are linked to different context. Thus, the user profiles are divided into different sub-profiles for different contexts. S..

15 E.g., a user profile may include a first sub-profile which indicates the user's preferences in a first context, a second sub-profile which indicates the user's preferences in S.....

* a second context etc. Thus, current context profiles are divided on the basis of the different context and are generated by first defining the context and then the individual preferences are attached to it. In most typical applications, the user preference profile is thus made up be different and separate sub-profiles for different contexts.

However, such an approach tends to be suboptimal in many scenarios. For example, a change in the user context does not necessary imply that there is a change in the user preferences (e.g. a user may prefer the same type of music irregardless of where the user is).

Hence, an improved system for context sensitive user preference prediction would be advantageous and in particular a system allowing increased flexibility, improved predictions, facilitated maintenance and operation, reduced complexity and/or improved performance would be advantageous.

Summary of the Invention

Accordingly, the Invention seeks to preferably mitigate, alleviate or eliminate one or more of the above mentioned disadvantages singly or in any combination.

* According to a first aspect of the invention there is provided S...

1. An apparatus for determining a user preference prediction, the apparatus comprising: means for receiving a plurality of user preference *SS*S* * indications, each user preference indication comprising content data and at least some of the plurality of user preference indications comprising context data; profile means for generating a user preference profile by grouping the plurality of user preference indications into a plurality of sub-profiles in response to the content data and without consideration of the Context data; context profile means for determining a context profile for each of the plurality of sub-profiles; means for determining a context for the user preference prediction; selection means for selecting a first sub-profile of the plurality of sub-profiles in response to a comparison of the first context and the context profiles of the plurality of sub-profiles; and prediction means for determining the user preference prediction in response to the first sub-profile.

The invention may allow improved and/or facilitated context sensitive user preference prediction. In particular, improved and/or facilitated operation may be achieved by ignoring context data when generating sub-profiles of a user preference profile (e.g. during a learning phase) combined with selecting an appropriate sub-profile for a prediction based on the context of the prediction and context profiles of the sub-profiles (e.g. during the prediction generation phase). * I.

This invention may allow improved contextual personalization ***.

*,**. 15 actions. Contextual profiles may be defined progressively and only if there is a need for them. In particular, the invention allows a more natural grouping of contexts around * user preferences instead of context similarities. *. * * S * * S

The user preference prediction may be a prediction of a user rating of a content item. The user preference prediction may be a user recommendation rating of a content item.

According to an optional feature of the invention, 2. The apparatus of claim 1 wherein the profile means is arranged to allocate a new user preference indication to a sub-profile by determining a prediction performance improvement indication for each sub-profile and allocating the new user preference indication to a selected sub-profile having the highest prediction performance improvement indication.

This may allow improved performance and may in particular provide an efficient and practical means of maintaining a user preference profile without requiring a completely new grouping of all user preference indications when a new user preference indication is received.

According to an optional feature of the invention, 7. The apparatus of any previous claim wherein the context profile means is arranged to generate a context profile comprising a context probability distribution for each of a set of context parameters.

This may provide a practical and/or high performance means of determining user preference predictions. Specifically, a * 15 probabilistic approach for selecting the appropriate sub-profile allows a more flexible context comparison. * S *.**

S

S.....

* According to an optional feature of the invention, * 12. The apparatus of any previous claim wherein the profile means is arranged to group the user preference indications by performing a clustering algorithm optimising the user preference prediction based on content data only.

This may allow a particularly practical and/or high performance operation and may typically provide a grouping into sub-profiles which can provide improved context sensitive user preference predictions.

According to another aspect of the invention, there is provided 15. A method of determining a user preference prediction, the method comprising: receiving a plurality of user preference indications, each user preference indication comprising content data and at least some of the plurality of user preference indications comprising context data; generating a user preference profile by grouping the plurality of user preference indications into a plurality of sub-profiles in response to the content data and without consideration of the context data; determining a context profile for each of the plurality of sub-profiles; determining a context for the user preference prediction; selecting a first sub-profile of the plurality of sub- :*;. profiles in response to a comparison of the first context and the context profiles of the plurality of sub-profiles; and * . determining the user preference prediction in response s.es.s * * to the first sub-profile. ** S * S

* * 20 These and other aspects, features and advantages of the S.....

* invention will be apparent from and elucidated with reference to the embodiment(s) described hereinafter.

Brief Description of the Drawings

Embodiments of the invention will be described, by way of example only, with reference to the drawings, in which FIG. 1 illustrates an example of a content item recommender in accordance with some embodiments of the invention; FIG. 2 illustrates an example of a user preference indication; FIG. 3 illustrates an example of a generation of context profiles in accordance with some embodiments of the invention; FIG. 4 illustrates an example of context profiles in accordance with some embodiments of the invention; FIG. 5 illustrates an example of context profiles in accordance with some embodiments of the invention; FIG. 6 illustrates an example of a generation of context profiles in accordance with some embodiments of the invention; * * ** a'. a..

* * FIG. 7 illustrates a method of generating a user preference * profile in accordance with some embodiments of the * . . * S invention; d S FIG. 8 illustrates a method of generating a user prediction in accordance with some embodiments of the invention.

Detailed Description of Some Embodiments of the Invention The following description focuses on embodiments of the invention applicable to a prediction of user preferences for content items and in particular to a system for generating recommendations for content items such as music, television programmes, etc. However, it will be appreciated that the invention is not limited to this application but may be applied to many other applications and types of user preference predictions.

FIG. 1 illustrates an example of a content item recommender, such as e.g. used in a portable PVR. The content item recommender specifically comprises functionality for generating context sensitive recommendations. For example, the content item recommender may generate different preference predictions and thus recommendations depending on the time (and e.g. location) for which the recommendation is made. * S.

*.:. In contrast to conventional systems, the content item recommender is not based on a user preference profile organised or sub-divided based on context data. Rather, user S... preference profiles are generated in which sub-profiles are *5SS*S * generated in response to content data but ignoring context data. Thus, in the system a user preference profile is generated which contains both content and context data but is divided into sub-profiles (sets of preference inputs) based only on a content data preference similarity. The content item recommender only uses the context data when the system is making a prediction and specifically uses the context data to select which of the sub-profiles to use.

In more detail, the content item recommender comprises a preference indication receiver 101 which receives user preference indications. For example, when watching a television programme or listening to a piece of music, the user may enter a rating of the programme or music. In addition to the rating, the content item recommender may record context data such as the time and location at which the rating is provided. Thus, as illustrated in FIG. 2, the user preference indications comprise both content data including the rating and describing the content item (e.g. artist, title, content metadata etc) as well as context data which relates to the context in which the user preference indication was generated (e.g. time and location).

The preference indication receiver 101 is coupled to a user profile processor 103 which generates a user preference profile comprising a plurality of sub-profiles. The sub-profiles are generated by grouping the user preference indications depending on the preference. data. However, when :,:::. grouping the user preference indications into sub-profiles, *::::* 15 the context data is completely ignored.

Specifically, the user profile processor 103 can group the S.....

* user preference indications into sub-profiles using a :. * clustering algorithm such as a K-means cluster algorithm.

The clustering is performed based only on the content data and ignoring the context data. A clustering algorithm generally attempts to minimize a criterion such as a distance measure, the error rate of the prediction, etc. In the current invention, the clustering algorithm attempts to maximize the theoretical performance of sub-profiles in predicting preferences using the content part alone.

In the content item recommender of FIG. 1, the K-means clustering algorithm initially defines k clusters with given initial parameters. The user preference indications are then matched to the k clusters. The parameters for each cluster are then recalculated based on the user preference indications that have been assigned to each cluster. The algorithm then proceeds to reallocate the user preference indications to the k clusters in response to the updated parameters for the clusters. If these operations are iterated a sufficient number of times, the clustering converges resulting in k groups of user preference indications having similar properties.

The theoretical performance of a user profile can be obtained via some statistics and in the described embodiment, the performance of a sub-profile is defined by the difference between the actual and predicted value for preferences contained in the sub-profile using a cross validation. Specifically, the user preferences of a sub- 15 profile are split in n-folds with n-i folds being iteratively used for learning and the remaining being used for testing. Since the clustering algorithm only runs on the * content part of the preference, a decrease in the sub-profile performance indicates a divergence in user tastes * 20 but not necessary a change in context (and vice versa).

****** * S Therefore the discriminative contextual information is implicitly identified by grouping similar preferences together.

As an example of a grouping of user preference indications, a new user preference indication may be added to an existing user preference profile in the following way (e.g. after an initial clustering the following approach may be used for new user preference indication or in some embodiments the following approach may also be used to initially generate the sub-profiles).

When a new user preference indication is received, the user profile processor 103 includes the user preference indication in an existing sub-profile for which some predefined performance measurement improves. Specifically, the user profile processor 103 system temporarily includes the new preference in each of the existing sub-profiles and evaluates the resulting impact on each sub-profile performance (for example by generating predictions for other preferences in the sub- profile and comparing the resulting value to the actual recorded value). In the described embodiment, the preference is then included in the sub-profile which has the maximum increase of performance (or the minimum decrease if there is no increase). * ** * S * S...

* 15 Thus, in the example, the user profile processor 103 allocates a new user preference indication to the sub-profile having the highest prediction performance **.* * * improvement. *5 5 * 5 *

* . 20 In other embodiments the user profile processor 103 may *.**.

* include the new user preference indication in a plurality of sub-profiles and may specifically include the user preference indication in all sub-profiles which show an improvement above a given threshold (which may be zero corresponding to the user preference indication being included in all sub-profiles for which an improvement is found).

The content item recommender also comprises a group context processor 105 which is coupled to the user profile processor 103 and which is arranged to determine a context profile for each of the sub-profiles of the user preference profile.

As exemplified in FIG. 3, the group context processor 105 may evaluate the context data for all the user preference indications in a given sub-profile and may generate the context signature based thereon.

In a simple example, the context signature may simply comprise all the context indications of the sub-profile such as an indication of all the time instants and locations represented in the sub-profile. Such an example is presented in FIG. 4 corresponding to a situation where movies are predominantly watched during the evening and cartoons are preferable watched in the morning and afternoon. * ** * * * a...

15 In more complex embodiments, the context data may be processed to generate a context profile that indicates the distribution of one or more individual parameters of the *...* * 1 context data. For example, the time of the user preference indications of a sub-profile can be evaluated to determine a profile of the time context parameter which can indicate the *.*..

* probability of a user preference indication in the sub-profile being in a given time interval.

E.g., the time parameter may be evaluated by dividing all time instants of the sub-profile into a number of groups corresponding to different times of the day. For example, the number of user preference indications having a time indication from 8.00 AM to 9.OOAM, from 9.00 AM to 10.OOAM etc. may be calculated. By dividing the number in each interval by the total number of time indications an indication of the probability of a user preference indication of the sub-profile being in each time interval is generated.

The same approach may be used for other context parameters such as a location parameter. It will also be appreciated that averaging, smoothing or curve fitting can be applied depending on the characteristics and preferences of the individual embodiment.

FIG. 5 illustrates some examples of possible probability distributions derived using the above approach.

The context signature of each sub-profile is stored in the ::::. content item recommender together with the user preference * * 15 profile. **.* I...

In the following the exemplary operation of the content item recommender when generating rating predications and content * item recommendations will be described. * * * * * * 20

* The content item recommender comprises a recommendation processor 107 which generates recommendations for content items. For example, at a given time and location a user of the portable PVR may enter a request mode wherein one or more content items are recommended to the user. In response the recommendation processor 107 proceeds to control the content item recommender such that a recommendation is generated. Specifically, the recommendation processor 107 obtains user rating predictions for a plurality of content items stored on the PVR and recommends one or more of the content items based on these predictions.

-Iv"-'--The recommendation processor 107 is coupled to a prediction context processor 109 which determines a context for the predictions. For example, if a user of the content item recommender requests a content item recommendation, the prediction context processor 109 can proceed to determine the current time and location (the content item recommender may for example be a portable device with a built in GPS receiver capable of providing both a current time and location).

The prediction context processor 109 is coupled to a sub-profile selection processor 111 which is further coupled to the group context processor 105. The sub-profile selection processor 111 is arranged to select one of the sub-profiles **. 15 of the user preference profile based on a comparison of the context of the recommendation and the context profiles of S...

the plurality of sub-profiles. Thus, whereas the generation * * of the user preference profile and the division into sub-profiles is based on the content data while ignoring the context data, the selection of which sub-profile(s) to use *....

* for a prediction is based on the context data of the prediction and the sub-profiles.

Thus, during the recommendation phase, the content item recommender matches the current context information and the context profiles of the sub-profiles with the context profile being computed using the context information from each of the preference inputs that form the sub-profile.

This approach allows creation of contextual user profiles in an emergent way, with minimal usage of context information (no predefined, explicit or static contexts) while providing an efficient and accurate user preference prediction and content item recommendation.

In the specific embodiment, the context probability distributions built during the learning phase using the context information from the preference inputs for each context parameter (e.g. time, location, etc.) are used to select the user profile.

Specifically, the sub-profile selection processor ill calculates the match probability of the current context for each of the sub-profiles. For example, for the movie and cartoon sub-profiles of FIG. 4, a time context of the : .. prediction of early morning (i.e. a recommendation is S...

** 15 requested in the morning), will result in a high match S...

probability for the cartoon profile as there is a high *55 probability that a user preference indication for this profile falls in the early morning time slot. Accordingly, the sub-profile selection processor 111 will select the * 20 cartoon sub-profile to use for the prediction. * *

If the context parameters can be considered statistically independent, the match probability can be determined by multiplying the probabilities for each of the context parameters. Even if the context parameters are not statistically independent, this approach may provide a suitable approximate approach in many embodiments.

For example, and as illustrated in FIG. 5, the time and location context parameters may individually be evaluated to determine a sub-probability for each parameter matching each of the sub-profiles. The match profile for the entire sub- profile is then obtained by multiplying the sub-probabilities of each parameter. The sub-profile with the highest probability is then used for the prediction. E.g. in the example of FIG. 5, sub-profile 1 would be selected since it fits the current context data better than sub-profile 2.

It will be appreciated that if e.g. the current context conditions are close enough to more than one sub-profile, the content item recommender may e.g. use all sub-profiles to make the predictions or may e.g. ask the user to select a sub-profile. For example, the content item recommender may proceed to use all sub-profiles having a match probability above a threshold. * *. * * *

** 15 After a sub-profile (or set of sub-profiles) has been selected, the sub-profile is used to make a preference **** prediction taking into consideration the content data of the selected sub-profile as well as content data for the possible content items to be recommended. However, in the content item recommender of FIG. 1, the context data is not S.....

* used for anything else than the selection of the most appropriate sub-profile(s).

The sub-profile selection processor 111 is coupled to a prediction processor 113 which determines a predicted user preference for each of a number of content items based on the selected user profile. For example, the PVR may have stored a number of content items (e.g. video clips or music clips) with associated content information such as meta-data describing the content, the artist, the title etc. For each of the content items, the prediction processor 113 can compare the stored content data to the content data stored in the selected sub-profile and based on this a rating for each content item can be generated.

The predication processor 113 is coupled to the recommendation processor 107 which is fed the determined ratings for the different content items. In response, the recommendation processor 107 selects the content item having the highest rating (or a plurality of content items having the highest rating) and presents it to the user.

Thus, the content item recommender uses an approach wherein contextual user profiles are built by implicitly and automatically identifying distinct contexts based on : .. differences or conflicts between user preferences. S... * * ***.

In particular, rather than explicitly identifying the *..

different contexts and attaching preferences to them, a context change is defined by a change in user preferences but not necessarily a modification of context conditions.

*****.

* This approach may provide a number of advantages including: * It requires less training data. The filtering/grouping based on content reduces noise on the data and results in more focused sub-profiles.

* It allows more continuity of experience. Specifically, a preference input will always be taken into account whatever the context conditions were when this preference was provided by the user * It allows facilitated development and specifically does not require any effort in predetermining and identifying appropriate contextual data and categories.

In some embodiments, the user profile processor 103 also comprises functionality for dynamically managing and maintaining the user preference profile. Specifically, the user profile processor 103 can monitor the prediction performance when new user preference indications are added to the user profile.

For example, the user profile processor 103 can detect if the prediction performance actually degrades to an extent : .. where it falls below a given threshold. Specifically, if a **, 15 user preference indication does not fit any of the existing sub-profiles particularly well, it may result in performance *..... degradation for all subprofiles and may be allocated to the sub-profile for which this degradation is the lowest.

* However, if the resulting performance falls below a * 20 predetermined threshold, the user profile processor 103 can * ** *** * proceed to generate a new profile which includes the new user preference indication. It may then proceed to perform a new clustering process to result in a new grouping of the user preference indications into sub-profiles. FIG. 6 illustrates an example of such an approach.

In some embodiments, the user profile processor 103 may monitor how often the individual sub-profiles are selected by the sub-profile selection processor 111. If one of the sub-profiles has not been used in agiven interval, the user profile processor 103 may delete the sub-profile (e.g. following accept by a user). This provides a practical way of both deleting noisy information and tracking permanent preference changes. The deletion of a sub-profile may be associated with a re-clustering being performed.

FIG. 7 illustrates an example of a method of generating a user preference profile comprising a plurality of sub-profiles.

The method starts in step 701 wherein a plurality of user preference indications is received. Each user preference indication comprises content data and at least some of the plurality of user preference indications comprises context data. * *. * S * S...

***.** 15 Step 701 is followed by step 703 wherein a user preference profile is generated by grouping the plurality of user *...

preference indications into a plurality of sub-profiles in response to the content data and without consideration of ** * the context data.

S.....

* Step 703 is followed by step 705 wherein a context profile for each of the plurality of sub-profiles is generated.

FIG. 8 illustrates an example of generating a user prediction using the user preference profile generated by the method of FIG. 7.

The method starts in step 801 wherein a context for the user preference prediction is determined.

Step 801 is followed by step 803 wherein a first sub-profile of the plurality of sub-profiles is selected in response to a comparison of the first context and the context profiles of the plurality of sub-profiles.

Step 803 is followed by step 805 wherein the user preference prediction is determined in response to the first sub-profile.

It will be appreciated that the above description for clarity has described embodiments of the invention with reference to different functional units and processors.

However, it will be apparent that any suitable distribution of functionality between different functional units or processors may be used without detracting from the : **. invention. For example, functionality illustrated to be *.... 15 performed by separate processors or controllers may be I...

performed by the same processor or controllers. Hence, references to specific functional units are only to be seen as references to suitable means for providing the described functionality rather than indicative of a strict logical or * 20 physical structure or organization. * *

The invention can be implemented in any suitable form including hardware, software, firmware or any combination of these. The invention may optionally be implemented at least partly as computer software running on one or more data processors and/or digital signal processors. The elements and components of an embodiment of the invention may be physically, functionally and logically implemented in any suitable way. Indeed the functionality may be implemented in a single unit, in a plurality of units or as part of other functional units. As such, the invention may be implemented in a single unit or may be physically and functionally distributed between different units and processors.

Although the present invention has been described in connection with some embodiments, it is not intended to be limited to the specific form set forth herein. Rather, the scope of the present invention is limited only by the accompanying claims. Additionally, although a feature may appear to be described in connection with particular embodiments, one skilled in the art would recognize that various features of the described embodiments may be combined in accordance with the invention. In the claims, the term comprising does not exclude the presence of other : .. elements or steps. S... ** 15 S...

Furthermore, although individually listed, a plurality of S...

S..... means, elements or method steps may be implemented by e.g. a single unit or processor. Additionally, although individual features may be included in different claims, these may possibly be advantageously combined, and the inclusion in * S different claims does not imply that a combination of features is not feasible and/or advantageous. Also the inclusion of a feature in one category of claims does not imply a limitation to this category but rather indicates that the feature is equally applicable to other claim categories as appropriate. Furthermore, the order of features in the claims does not imply any specific order in which the features must be worked and in particular the order of individual steps in a method claim does not imply that the steps must be performed in this order. Rather, the steps may be performed in any suitable order.

Claims

1. An apparatus for determining a user preference prediction, the apparatus comprising: means for receiving a plurality of user preference indications, each user preference indication comprising content data and at least some of the plurality of user preference indications comprising context data; profile means for generating a user preference profile by grouping the plurality of user preference indications into a plurality of sub-profiles in response to the content data and without consideration of the context data; context profile means for determining a context profile : *. for each of the plurality of sub-profiles; means for determining a context for the user preference S...

prediction; S...

selection means for selecting a first sub-profile of the plurality of sub-profiles in response to a comparison of the first context and the context profiles of the plurality * 20 of sub-profiles; and *SSSS* * prediction means for determining the user preference prediction in response to the first sub-profile.

2. The apparatus of claim 1 wherein the profile means is arranged to allocate a new user preference indication to a sub-profile by determining a prediction performance improvement indication for each sub-profile and allocating the new user preference indication to a selected sub-profile having the highest prediction performance improvement indication.

3. The apparatus of claim 2 further comprising means for determining a prediction performance indication for the selected sub-profile after inclusion of the new user preference indication; means for comparing the prediction performance indication to a threshold; means for instigating a re-grouping of the user preference indications into sub-profiles if the prediction performance indication is below the threshold.

4. The apparatus of claim 2 or 3 comprising means for determining at least one of the prediction performance indication and the prediction performance improvement indication by determining a first user preference prediction : *** for a first user preference indication in response to a set of user preference indications belonging to a same sub-profile as the first user preference indication but excluding the first user preference indication and comparing the first user preference prediction to the first user preference prediction.

*0***S *

5. The apparatus of any of the previous claims 2 to 4 wherein the profile means is arranged to include the new user preference indication in all sub-profiles resulting in a prediction performance improvement indication above a threshold.

6. The apparatus of any previous claim wherein the profile means is arranged to group the user preference indications by performing a K-means clustering.

7. The apparatus of any previous claim wherein the context profile means is arranged to generate a context profile comprising a context probability distribution for each of a set of context parameters.

8. The apparatus of claim 7 wherein the selection means is arranged to generate a context match probability for each sub-profile by combining a sub-probability for each of the set of context parameters matching a corresponding context parameter for the user preference prediction and to select the first sub-profile as the sub-profile resulting in the highest context match probability.

9. The apparatus of claim 8 wherein the selection means is arranged to generate the context match probability for a sub-profile by multiplying the sub-probabilities for the set *... 15 of context parameters. 41.S

10. The apparatus of any of the previous claims 7 or 8 wherein the selection means is arranged to select a selected plurality of sub-profiles having a context match probability * * 20 above a threshold, and the prediction means is arranged to U.....

* determine the user preference prediction in response to the user preference indications of the selected plurality of sub-profiles.

11. The apparatus of any previous claim wherein the context data comprises at least data selected from one of: * time data; and * location data.

12. The apparatus of any previous claim wherein the profile means is arranged to group the user preference indications by performing a clustering algorithm optimising the user preference prediction based on content data only.

13. The apparatus of any previous claim wherein the apparatus further comprises means for deleting a sub-profile if the sub-profile has not been selected within a given time interval.

14. The apparatus of any previous claim wherein the user preference prediction comprises a predicted user rating for a content item, and the apparatus further comprises means for recommending content items in response to predicted user ratings for the content items. * U. * U a U.. )

15. A method of determining a user preference prediction, the method comprising: a...

receiving a plurality of user preference indications, each user preference indication comprising content data and ** at least some of the plurality of user preference : . 20 indications comprising context data; S. SU* * generating a user preference profile by grouping the plurality of user preference indications into a plurality of sub-profiles in response to the content data and without consideration of the context data; determining a context profile for each of the plurality of sub-profiles; determining a context for the user preference prediction; selecting a first sub-profile of the plurality of sub-profiles in response to a comparison of the first context and the context profiles of the plurality of sub-profiles; and determining the user preference prediction in response to the first sub-profile.

16. A computer program product enabling the carrying out of a method according to claim 15. * ** ** I * S.. **** * S S... * S *SS

S

S..... * I S. I * I

I

S.'...

S S