WO2018176083A1

WO2018176083A1 - "recommending media items"

Info

Publication number: WO2018176083A1
Application number: PCT/AU2018/050266
Authority: WO
Inventors: Jonathon Osland Satterley; Jason William O'DONNELL
Original assignee: Village Roadshow IP Pty Ltd
Priority date: 2017-03-31
Filing date: 2018-03-23
Publication date: 2018-10-04

Abstract

This disclosure relates to a computer system for recommending media items to a viewer. Multiple reviewer devices display a review form to reviewers and receive from the reviewers input data indicative of a multi-parameter characterisation of multiple media items. A processing server associates each of the multiple media items to one media item cluster based on the multi-parameter characterisation. Multiple viewer devices display a questionnaire to each of multiple viewers. The questionnaire is disjunct to the multi -parameter characterisation of the multiple media items. The processing server receives viewer answers from the questionnaire, maps the viewer answers to a tribe, determines an association between the tribe and one of the multiple media item clusters, and recommends to the viewer one or more media items from the media item clusters associated with the tribe of that viewer.

Description

"Recommending media items"

Cross-Reference to Related Applications

[0001] The present application claims priority from Australian Provisional Patent Application No 2017901176 filed on 31 March 2017, the content of which is incorporated herein by reference.

Technical Field

[0002] The present invention relates to the field of recommender systems. In particular to a method, system and software for recommending media items to viewers.

Background

[0003] Current online video portals provide recommendations for media items to users. However, it is difficult for the video portal server to provide recommendations that are useful for the viewer. Further, with growing numbers of viewers and media items processing time increases which makes it difficult to implement more

sophisticated recommendation mechanisms. In particular, it is desirable that the processing time is reduced such that recommendations can be generated in real-time while improving recommendation accuracy.

[0004] Any discussion of documents, acts, materials, devices, articles or the like which has been included in the present specification is not to be taken as an admission that any or all of these matters form part of the prior art base or were common general knowledge in the field relevant to the present disclosure as it existed before the priority date of each claim of this application.

[0005] Throughout this specification the word "comprise", or variations such as "comprises" or "comprising", will be understood to imply the inclusion of a stated element, integer or step, or group of elements, integers or steps, but not the exclusion of any other element, integer or step, or group of elements, integers or steps. Summary

[0006] This disclosure provides a recommender system where viewers are grouped into tribes and media items are clustered into media item clusters. There are also associations between the tribes and the clusters, which means that recommendations can be generated for one tribe by recommending media items from the associated cluster. This can greatly reduce processing time and enable real-time generation of recommendations while using a multi-parameter characterisation of the media items.

[0007] A computer system for recommending media items to a viewer comprises: multiple reviewer devices to display a review form to reviewers and to receive from the reviewers input data indicative of a multi-parameter characterisation of each of multiple media items;

a processing server to associate each of the multiple media items to one of multiple media item clusters based on the multi-parameter characterisation;

multiple viewer devices to display a questionnaire to each of multiple viewers, the questionnaire being disjunct to the multi-parameter characterisation of the multiple media items;

wherein the processing server is configured to

receive viewer answers from the questionnaire;

map the viewer answers to a tribe;

determine an association between the tribe and one of the multiple media item clusters; and

recommend to the viewer one or more media items from the media item clusters associated with the tribe of that viewer.

[0008] Preferably, recommending to the viewer one or more media items is based on a personalised recommendation for that viewer.

[0009] Preferably, the personalised recommendation includes a subjective assessment of the viewer's current emotional state. [0010] Preferably, the multiple viewer devices display an interactive user interface for the viewer to determine the subjective assessment of the viewer's current emotional state.

[0011] The processing server may be configured to determine one or more clusters of media items based on the input data indicative of a multi-parameter characterisation.

[0012] The processing server may determine one or more tribes based on the input data indicative of a multi-parameter characterisation.

[0013] The processing server may determine the one or more tribes based on similarities in the multi-parameter characterisation as calculated by a similarity index.

[0014] The processing server may determines the one or more tribes based on similarities in the multi-parameter characterisation as calculated by a distance measure.

[0015] The similarity index may be a Jaccard similarity coefficient.

[0016] The processing server may utilise a shared nearest-neighbour algorithm to determine one or more clusters of media items.

[0017] The processing server may utilise a shared nearest-neighbour algorithm to determine one or more tribes of viewers.

[0018] A computer implemented method of operating a computer system for recommending media items to a viewer comprises:

displaying a review form to reviewers on a reviewer device;

receiving from the reviewers input data indicative of a multi-parameter characterisation of each of multiple media items;

associating each of the multiple media items to one of multiple media item clusters based on the multi-parameter characterisation;

displaying on a viewer device a questionnaire to a viewer, the questionnaire being disjunct to the multi -parameter characterisation of the multiple media items; receiving viewer answers from the questionnaire;

mapping the viewer answers to a tribe;

determining an association between the tribe and one of the multiple media item clusters; and

recommending to the viewer one or more media items from the media item clusters associated with the tribe of that viewer.

[0019] There is also provided software, being machine readable instructions, that when performed by a computer system causes the computer system to perform the method described above.

Brief Description of Drawings

[0020] Examples of the present disclosure will be described with reference to:

[0021] Fig. 1 illustrates an example system for recommending media items to viewers.

[0022] Fig. 2 illustrates example review forms.

[0023] Fig. 3 illustrates example viewer questionnaires.

[0024] Fig. 4a and Fig. 4b illustrate an example clustering of media items.

[0025] Fig. 5a and Fig. 5b illustrate an example determination of tribes.

[0026] Fig. 6 illustrates an example user interface displaying a tribe and personality assessment to the viewer.

[0027] Fig. 7 illustrates an example user interface displaying multiple tribe matches. [0028] Fig. 8 illustrates an example recommendation engine. [0029] Fig. 9 illustrates an example set of recommendations for the viewer.

[0030] Fig. 10 illustrates an example user interface on a viewer device for a subjective assessment of the viewer's current emotional state.

[0031] Fig. 11 illustrates an example method for recommending media items to viewers.

[0032] Fig 12 illustrates an example processing server. Description of Embodiments

[0033] The current disclosure relates to a method and system for the automatic recommendation of a media item for a viewer. This automatic recommendation is based on a series of questions posed to a reviewer from which the system collates the answers. The system may pose a series of questions to the viewer to ascertain viewer characteristics such that the viewers can be mapped to a tribe, which are groups of viewers that have a number of viewer characteristics in common.

[0034] Tribes are associated with clusters of media items such that the

recommendation engine can then take into account the tribe associated with the viewer such that the system can recommend a media item. The recommendation engine may additionally take into account personal preferences, previous likes or dislikes based on the recommendation, interaction with forums and other features which make the recommendation more personalised over a static (fixed) recommendation.

[0035] Fig. 1 illustrates a computer system 100 for recommending media items 160, 162 to a viewer 130,132. The computer system comprises multiple reviewer devices 111, 113 to display a review form to reviewers 110,112 and to receive from the reviewers input data indicative of a multi-parameter characterisation of each of multiple media items 114, 116. Computer system 100 further comprises a processing server 120 to associate each of the multiple media items 114, 116 to one of multiple media item clusters 180, 182 based on the multi -parameter characterisation. The computer system 100 also comprises multiple viewer devices 131, 133 to display a questionnaire 150,152 to each of multiple viewers 130, 132. The questionnaire is disjunct to the multi-parameter characterisation of the multiple media items 114,116.

[0036] The processing server 120 is configured to receive viewer answers from the questionnaire 150, 152 and to map the viewer answers to a tribe 140, 142. The processing server 120 is further configured to determine an association 170, 172, 174, 176 between the tribe 140, 142 and one of the multiple media item clusters 180, 182 and to recommend to the viewer 130,132 media items 160, 162 from the media item clusters 180,182 associated with the tribe 140,142 of that viewer 130,132.

Reviewer Devices

[0037] Each of the reviewer devices 111, 113 can be any device that has a capacity to receive questions and communicate the answers from the reviewer. Preferably a reviewer device 111,113 is a smartphone with a cellular or Wi-Fi connection. If it is using cellular, preferably the cellular connection is 4G, but the system may also work with other data communication technologies such as 2G and 3G. Where available, the reviewer device may also be able to utilise a Wi-Fi or other wireless data connection. In other embodiments, the reviewer device may not necessarily be a mobile device such as a personal computer or laptop.

[0038] In the example in Fig. 1, a reviewer 110 watches a media item 114 such as a film, television series, mini-series, music or even YouTube videos. Although not essential, this media item may be presented on the reviewer's mobile device 111. Alternatively, it may be screened on a device such as a television, a screen connected to the internet, or may be screened off-line such as at a cinema. In some cases, the reviewer 110,112 may simply be reviewing the media item 114,116 from memory as they may have seen it recently, multiple times, or just have a good recollection of the media item. [0039] The reviewer device 111,113 preferably has a user interface that allows interaction with a reviewer 110,112. Preferably the user interface would be a touch screen and the reviewer may therefore interact with the mobile device 111 simply by touching the screen.

[0040] In one example, the reviewer 110 is presented a review form on the reviewer device 111 for the media item 114. The reviewer 110 then inputs their review into the review form based on their assessment of the media item 114. The reviewer device 111 generates a display on the user interface that indicates the series of questions. This display is typically radio boxes or text boxes on the user interface which the user can enter to answer the questions posed by the system. The reviewer devices may store the answer data on the device or the reviewer devices may communicate the answer data to the processing server 120.

Viewer devices

[0041] Each of the viewer devices 131, 133 can be any device that has a capacity to receive questionnaires and communicate the answers from the viewer 130, 132 to the processing server 120. Similarly to the above in relation to the reviewer devices, preferably a reviewer device 131, 133 is a smartphone with a cellular or Wi-Fi connection. If it is using cellular, preferably the cellular connection is 4G, but the system may also work with other data communication technologies such as 2G and 3G. Where available, the reviewer device may also be able to utilise a Wi-Fi or other wireless data connection. In other embodiments, the viewer device may not necessarily be a mobile device such as a personal computer or laptop.

[0042] The viewer device 131, 113 preferably has a user interface to allow interaction with a viewer 130, 132. Preferably the user interface would be a touch screen and the reviewer may therefore interact with the mobile device 131 by touching the screen.

[0043] In one example, the viewer 130 is presented a questionnaire on the viewer device 131. The viewer 130 then inputs their answers based on their response to the questions. The viewer device 131 generates a display on the user interface that indicates the questions in the questionnaire and one or more areas within the user interface that the viewer 130 can use to enter his or her answer the questions posed by the processing server 120. This display is typically radio boxes or text boxes on the user interface which the viewer 130 can enter to answer. The viewer devices 131, 133 may store the answer data on the viewer device itself or the viewer devices may communicate the answer data to the processing server 120.

Processing server

[0044] The system as illustrated in Fig. 1 also comprises a processing server 120. The processing server 120 is preferably a web server that runs a server package such as Apache server. The processing server 120 listens for connections from the reviewer devices 111,113 and the viewer devices 131,133. Once a connection is established with the reviewer devices 111,113 or the viewer devices 131,132 the processing server 120 will be able to provide network services, or Web Services which are built and distributed using web development technologies such as PHP HyperText Preproccessor (PHP). The processing server 120 will typically handle the communications between the data store 140, the reviewer devices 111,113, the viewer devices 131,133 and the recommendation engine 190, which would typically be a process being executed on the processing server 120.

[0045] It should be noted that although the processing server 120 is shown as an independent network element in Fig. 1, the processing server 120 may also be part of another network element. Further, functions performed by the processing server 120 may be distributed between multiple network elements in Fig. 1.

[0046] The processing server 120 may send, from the output port of the processing server 120, a series of questions to the reviewer devices. The processing server 120 may receive data, such as reviewers' answers, from data memory as well as from the communications port. In one example, the processing server 120 receives and processes the reviewer answer data after all the questions are asked. This means that in this example the processing server 120 does not receive the data from the reviewer device 111, 113 before the reviewer is asked the next question. In this case, the data is sent from the reviewer device 111, 113 once the reviewer has completed their review.

[0047] When a new media item 114, 116 is to be reviewed, a new set of questions is sent to reviewer devices 111, 113. In this example, the new media items 114, 116 are to be reviewed by reviewers 110, 112. Although illustrated as two reviewers, in practice there is no limit to the number of reviewers. It is expected there would be potentially thousands or even millions of reviewers.

[0048] Once the multi-parameter characterisation of the media items 114,116 has been determined, the processing server 120 associates each of the media items 114,116 to one of the multiple media item clusters 180, 182 based on the multi-parameter characterisation.

[0049] Clusters 180, 182 are illustrated in the example in Fig. 1 as circles which represent that a number of media items are similar in one or more of the parameters of the multi-parameter characterisation. This is a simplification for the purposes of representation graphically. In practice, there are more sophisticated means to determine media item clusters as discussed below. It is worth noting that each media item may belong to multiple clusters. In the example in Fig. 1 the media item 160 is associated with the cluster 180 and the cluster 182. The media item 162 is only associated with the cluster 182 and not cluster 180.

[0050] The processing server 120 may cause multiple viewer devices 131, 133 to display a questionnaire 150,152 to each of multiple viewers 130, 132. The

questionnaire being disjunct to the multi-parameter characterisation of the multiple media items. Typically this would be triggered when a new viewer seeks a recommendation for a media item. Disjunct means that the questions in the questionnaire do not overlap with the multi-parameter characterisation of the media items. For example, the questionnaire is about human personalities while the media characterisation is about media characteristics. For example, the questionnaire may include questions related to whether the viewer is an introvert or extrovert. In contrast, the media items may be characterised in relation to whether they are action or romance movies.

[0051] In one example, the new viewers 130, 132 are being asked to answer on the questionnaire 150 questions according to the 16 personalities test by Myers Briggs. In another example, the questions relate to how the viewer responds to particular films. For example, a new viewer may be asked a series of pointed questions about the viewer's reactions to films they like and common tropes or themes they like, love or hate. In that sense ,the questions ask for subjective answers that differ from user to user, while the multi-parameter characterisation of movies is objective about the movie.

[0052] In one example, the questionnaire comprises questions about the viewer's personality using films as the subject of the question. For example, the question may be "which of the following films do you like?". This is a subjective question that is disjunct from the multi-parameter characterisation of the film itself.

[0053] Once the processing server 120 receives viewer answers from the

questionnaire, the processing server 120 is configured to map the viewer answers to a tribe. Each tribe represented a particular phenotype of film and the viewers in each tribe are people who responded to particular films with a passion. Each Tribe may be represented by an archetypal leader that was an expression of the tribal personality.

[0054] In one example, the clustering involves a machine learning process. Each response from the viewer serves as one learning sample in a supervised learning framework. In particular, the multiple parameters of the characterisation of the media items serve as features. The parameter values that are entered by the reviewers serve as feature values. The answers on the questionnaire by the viewer serve as labels. It can be appreciated again in this scenario that the labels are disjunct from the features. The processing server 120 can now perform machine learning to extract associations between the viewer's responses and the reviewer's characterisation. This may also involve unsupervised machine learning. For example, a self-organising map may be calculated to determine the tribes, such that each member of one tribe has provided similar answers to the questions of the questionnaire.

[0055] In one particular example, based on the answers to the questionnaire, the processing server 120 has determined that the viewer 130 is a 'Road Warrior' tribe 140 including viewers that have predominantly selected films characterised as 'Hollywood', 'special effects' and 'clear narrative' parameters. Similarly, the processing server 120 determined that the viewer 132 is a 'Gritter' tribe 142 including viewers that have predominantly selected films characterised as y 'Hollywood', 'critical acclaim' and 'believable.' This multi-parameter characterisation is simplified for illustrative purposes and in practice the number of parameters could be much higher.

[0056] The processing server 120 will determine an association 170, 172, 174, 176 between the tribe 140, 142 and one of the multiple media item clusters 180, 162. The associations are determined based on a match between the media item clusters and the multi-parameter characterisation of the tribes.

[0057] The associations 170,172,174,176 may be direct, but typically the associations may be weighted according to the degree of similarity. Therefore a match between media item clusters 180,182 and the multi-parameter characterisation of the tribes 140,142 would be a measure of their similarity. For example, the Road Warrior tribe has a high degree of similarity with a cluster of media items such as 'Fury Road', 'X- Men Apocalypse' and ' Star Wars: The Force Awakens.' The Road Warrior tribe has a low degree of similarity with a cluster of media items such as 'The English Patient', 'The Piano' and 'Cold Mountain.'

[0058] Each cluster of media items 180, 182 is measured against each tribe to associate one or more clusters of media items to the tribes. To save computation, the clusters of media items associated with a tribe may be stored in the data store 140. The clusters, tribes and associations may be updated when a sufficient amount of new information has been entered into the system such that the recommendations may change. This means that the calculations do not have to occur each time a new media item or viewer is added.

[0059] In the above example of features and labels of learning samples, each tribe may be associated with weights for some or all of the multiple parameters of the multiparameter characterisation of the media items. For example, the 'Road Warrior' tribe has relatively large positive weights for 'Hollywood', 'special effects' and 'clear narrative' . This means, processing server 120 can multiply the weights with the corresponding scores for each media item to calculate an association score for each media item with this tribe. In this sense, a high association score is indicative of a strong association between the tribe and the media item. Media items with a score higher than a predefined threshold, can be associated with the same cluster. In this sense, the tribe is associated with a cluster including all media items with a score above the threshold.

[0060] Once the association has been determined the processing server 120 is configured to recommend to the viewer 130 one or more media items (160, 162) from the media item clusters associated with the tribe of that viewer. In this example, the processing server 120 has determined that the viewer 130 matches the 'Road Warrior' tribe and that the media item cluster 'Fury Road', 'X-Men Apocalypse' and ' Star Wars: The Force Awakens' most closely matches the 'Road Warrior' tribe. On this basis, the processing server 120 may recommend 'Fury Road' to the viewer.

[0061] In the example in Fig. 1, the media items in the clusters 180, 182 are different sizes to represent the degree to which that media item is a good recommendation. For example, media item 160 in cluster 180 is a good recommendation for viewer 130, but a poor recommendation for viewer 132. Media item 162 is a better recommendation for viewer 132. This may be represented by the result of the weighted association score as above.

[0062] In one example, the tribes are allies or enemies. This may be encoded in terms of similar weights of the multi-parameter characterisation learned based on the viewer answers. Again, a self-organising map is well suited to extract this structure and can provide a distance or similarity score between tribes. If the distance score is sufficiently close, the processing server 120 can recommend media items associated with allied tribes. In some cases, there may not be any remaining media item in the tribe of the viewer or any allied tribes. In this case, processing server 120 may recommend a random selection. However, the random selection may exclude media items associated with enemy tribes.

[0063] In examples where the number of viewers, number of media items and number of parameters in the multi-parameter characterisation is large, a direct application of machine learning algorithms may be computationally difficult. In particular, it is desirable that the processing server 120 can generate recommendations in real time. That is, processing server 120 generates recommendations in less than 500 ms. The disclosed architecture using tribes and clusters and associations between the two, reduces the number of variables in the machine learning algorithm. As a result, computational complexity is reduced which leads to significantly faster execution time. This enables real-time execution and real-time generation of recommendations. As such, the disclosed split into tribes and media item clusters based on disjunct parameter sets is a technical solution to the technical problem of excessive computational complexity.

[0064] It is to be understood that throughout this disclosure unless stated otherwise, questions, answers, characteristics, attributes, variables, parameters and the like refer to data structures, which are physically stored on data memory or processed by a processor (described below). Further, for the sake of brevity when reference is made to particular characteristics, such as "character driven", "clear narrative", "believable", "structure " or "tropes," this is to be understood to refer to values of variables stored as physical data in the processing server 120 or in the data store 140. These values typically would be stored as part of the metadata associated with a media item and/or the viewers.

Reviewers and Reviewer Devices [0065] Fig. 2 illustrates simple example review forms 210, 212, 214, 216 on the reviewer devices 111,113 for the media items 114, 116. In the first example, the reviewers 110, 112 have finished watching the media item 'Fury Road' 114. The reviewer 110 has reviewed 'Fury Road' 114 on the review form 210 and considered that the 'style' was the most prominent characteristic of the media item 114. On the other hand, the reviewer 112 reviewed 'Fury Road' 114 on the review form 212 and, as can be seen, considered that story was the most prominent characteristic of 'Fury Road.'

[0066] Similarly, in the second example, the reviewers 110,112 have finished watching the media item 'Goodfellas' 116. The reviewer 110 has reviewed

'Goodfellas' 116 on the review form 214 and considered that the 'complexity' was the characteristic of the media item 116 that distinguished it from other films. The reviewer 112 reviewed 'Goodfellas' 116 on the review form 212 and, in agreement with the reviewer 110, considered that 'complexity' was the most distinguishing characteristic.

Viewers and viewer devices

[0067] Fig. 3 illustrates further examples of questionnaires 310, 312, 314, 316 on the viewer devices 131,133. In the first example, the processing server 120 has posed a question, 'what do you like most about movies?' along with three multiple choice answers: 'characters', 'special effects' and 'story' . Viewer 110 has responded with 'special effects' and viewer 112 has responded with 'characters'.

[0068] In the second example, the processing server 120 has posed a question 'which of the following most affects what you watch?' along with three multiple choice answers: 'filmography', 'style' and 'plot' . Viewer 110 has responded with filmography and viewer 112 has responded with plot. It is worth noting that these examples are simplified for illustrative purposes, and in practice the questionnaires may be longer significantly more complex, or require more complex interaction from the viewers. [0069] When a viewer has finished answering the questions in the questionnaire, the processing server 120 may receive viewer 110,112 answers from the questionnaire, map the viewer answers to a tribe, determine an association between the tribe and one of the multiple media item clusters; and recommend to the viewer one or more media items from the media item clusters associated with the tribe of that viewer.

Multi-parameter characterisation

[0070] As can be seen in the above examples, the system 100 utilises an analytical framework by which media items and viewers can be analysed by providing for a multi-parameter characterisation of media items 114,116 and viewers 130,132. This multi-parameter characterisation provides the basis of the clustering of media items, and the determination of tribes (which is discussed further below).

[0071] There are a number of parameters that are dimensions of assessable criteria of a media item. The assessable criteria include the following categories:

Semantics;

Syntactics;

Aesthetics; and

Pragmatics.

[0072] Semantics is the stylistic and symbolic building blocks that make up the media item's fictional world. Elements relating to semantics are the tropes and stereotypes, stars and actors, director, composer, iconography, authenticity and believability.

[0073] Syntactics are the underlying structural qualities and rhythmic beats that make up the media item's narrative. Elements relating to syntactics include the story, plot, narrative, structure, character, resolution, complexity and serial form. [0074] Aesthetics are the audiovisual and artistic qualities of the media item.

Elements of aesthetics include the look and feel, style, music, editing, filmography, three dimensions, special effects, animated, computer generated imagery (CGI).

[0075] Pragmatics are the practical elements external to the media item's narrative world that may influence the viewer's decision about what to watch. Pragmatics include two distinct sub-groups: function and effect. Elements of function include Hollywood production, indie / alternative, foreign, avant garde, genre, true story, source material, awards, budget, three dimensional, sexism, self-awareness, acclaim, tomatoe meter, animation, classification rating, notoriety and controversy. Elements of effect include credulity, emotion, feeling, senses, gut / heart, visceral, memory, challenge, wonder, Zeitgeist, friend factor, high stakes, peril, suspense, romance, fun, irony, comedy, self-reflexivity and surprise.

[0076] It is noted that each of the parameters above can have a numerical value, such as a value from 0 to 10. This provides a extremely granular characterisation of the media items, which may also be referred to as the DNA of a media item. As described above, this numerical characterisation leads to increased computational complexity that may be difficult to perform in a real-time manner using previously existing methods. In contrast, the proposed use of tribes and media item clusters allows this complex use of numerical multi-parameter characterisation with real-time recommendations.

[0077] Returning to Fig. 1, the review forms 150, 152 provide the basis of the multiparameter characterisation of media items. That is, a reviewer 110, 112 will be presented with a series of questions on the review form that correspond to the above dimensions of assessable criteria. The processing server 100 then receives from the reviewers 110, 112 input data indicative of a multi-parameter characterisation of each of multiple media items.

[0078] The degree to which each question affects the assessable criteria is variable and depends largely on the content of the question and the media item itself. For example, a review form question may be "On a scale of 0 to 10 are the characters mythic, archetypal characters and characters that fit an expected mould (10), or are they quirky and fit no mould? (0)" This example question is directed to partly semantics and partly syntactics. Multiple questions can be presented to the reviewer 110, 112 so that multiple assessable criteria can be assessed and the media item can be characterised by the assessment criteria.

[0079] A further aspect is that the dimensions of the assessable criteria characterise a media item well when they are covered adequately by the series of questions on the review form. That is, the questions would typically need to cover many different aspects of the media item.

[0080] However, not all the questions need to be answered by the same reviewer. For example, the system 120 may determine that a subset of the total group of reviewers will be asked a question about the characters of the movie. Another subset of the group of reviewers may be asked questions about the plot. This approach enables a large number of questions that are representable of the assessable criteria to be asked without overburdening each individual reviewer. Further, the answers to identical questions can be aggregated to form an 'average' characterisation. In another example, each answer from the reviewer may serve as a learning sample for the machine learning process described above. The aggregation or consolidation of answers from different reviewers is then an integral part of the machine learning process. In other words, the aggregation is a side product of the machine learning process and there is no separate process to aggregate the answers to identical questions on identical media items from different reviewers.

Clusters

[0081] Each media item 114, 116 is classified according to the above dimensions of assessable criteria. After a sufficiently large number of media items have been classified by the reviewers 110, 112 the system 120 may determine clusters 180, 182 in the media items. Clusters are groups of media items that are similar in one or more dimensions of assessable criteria as outlined above. [0082] The clusters may be predetermined and fixed, or they may be dynamic and flexible. Where the clusters are dynamic and flexible, the clusters may change where more media items are reviewed and added to the data store 140. Where the clusters are fixed, the clusters will not change when new media items are added to the data store 140. In order to save computation cost and complexity, not all media items are necessarily required to determine the clusters, in some cases a representative subset of media items may be used to identify the clusters 180, 182. In some examples, the clusters are determined in real-time and/or on demand, in the sense that media items that score highly in characteristics typical of a particular tribe are in one cluster.

[0083] Different clustering techniques can be used. One type of example are unsupervised learning techniques which can be used with the set of inputs from the classification of the media items. If used in this case, the unsupervised learning may be used to find the structure or relationships between different classification of media items. Several approaches to clustering using unsupervised learning techniques can be used, such as subspace clustering, projected clustering, hybrid approaches (that use a combination of both and heuristics to arrive at satisfactory answers), pattern based clustering and correlation clustering. This will create different clusters of media items and will be enable the system to put any newly classified media item in an appropriate cluster. If some clusters are known beforehand and some of the media items are classified accordingly, then a form of semi-supervised learning may be used.

[0084] Fig. 4a and Fig. 4b illustrate an example of clustering using unsupervised learning techniques in two dimensional space 410. While Figs. 4a and 4b illustrate only two dimensions, it is noted that the clustering may occur in (many) more dimensions. As can be seen the media items 412-424 are distributed over the two dimensional space but are not evenly distributed. For example, media items 412, 416 and 424 are relatively isolated from other media items whereas the media items 414, 418, and 420 are nearby. A distance measure, such as vector length or norm, can be used to determine how different a pair of two media items are in the two or more dimensions. [0085] Fig. 4b is the two dimensional space 430 but with clusters 440, 450 and 460 identified. In this case, the clusters cover all but one of the media items 416. That is, the clusters as represented by circles 440, 450 and 460 enclose all the media items except for media item 416. A more programmatic approach may be to determine clusters based on minimising the average distance but maximising the coverage of the dimensional space. In this example, not all the media items fit cleanly into a cluster. For instance the media item 424 is in both clusters 440 and 460. The media item 416 is not determined to be in any cluster. In this case, the system 120 may make an approximate estimation on the closest match, which by simple two dimensional distance in this example, is the cluster 440. The system 120 may take into account that the media item 416 does not quite fit into the cluster.

[0086] One approach of determining similarity of clusters is to the use the Jaccard index, also known as the Jaccard similarity coefficient, as this is a statistic used for comparing the similarity and diversity of sample sets. In simple cases two sets are compared the index is a simple decimal statistic between 0 and 1.0. The Jaccard index of two identical sets will always be 1, while the Jaccard index of two sets with no common elements will always yield 0.

[0087] The Jaccard coefficient measures similarity between finite sample sets, and is defined as the size of the intersection divided by the size of the union of the sample sets:

J {A, B) _^ \A n B\ \A n B

\A U B\ ^"" \A\ i- \B ^■■■■ A n B

And the Jaccard distance:

A u S\ ~ n Bi

A u B\ [0088] In more complex cases, multiple characteristics are being compared. But it is worth noting that for a broad range of data distributions and distance measures, the relative contrast diminishes as the dimensionality increases (known as the curse of dimensionality). In types of data distributions expected from this disclosure, the dimensionality of the data may be quite high given that the total number of

characteristics that are being measured is also high. As a result, the distance measure may have reduced utility as the number of dimensions increases.

[0089] In order to reduce the dimensionality of the data, it may be appropriate to use feature selection to try to find a subset of the characteristics. Generally three strategies may be employed: filter (selecting characteristics such as by information gain including Kullback-Leibler divergence), wrapper (that evaluate subsets of variables which allows, unlike filter approaches, to detect the possible interactions between variables) and embedded (features are selected to be added or removed while building the model based on the prediction error).

[0090] Using a filter approach, for example, could determine that only a specific five characteristics (such as tropes, structure, story, look and feel, and genre) are required in order to sufficiently distinguish the viewers. In practice, the dimensionality of the data is likely to be higher, but it is worth noting that it is potentially variable. In another example the system 120 may use heuristics to more efficiently determine sets of characteristics that would be more likely to be distinct such as complexity and

Hollywood . In another example sets of characteristics can be arbitrarily chosen such that a specified number of sets, such as 10 or 12, are representative of a majority of differences in user characteristics.

[0091] Given that the dimension of the data is controlled by the criteria by which the media items are assessed, the criteria that are used to assess media items can be limited to the criteria that ensure that the data distribution of the media items does not produce irrelevant or redundant characteristics. In some cases, it may be useful to consider a subset of assessable criteria (a filter method) so that the distance measure remains effective. In this case, the subset of assessable criteria may be chosen such that the distance measure discriminates effectively based on the data distribution.

[0092] If the dimensionality cannot be reduced (whether because of the data distribution or otherwise), it may be more appropriate to use a shared nearest-neighbour similarity measure, which has been shown to be more effective in with high

dimensional data.

[0093] Once the clusters 180,182 have been determined, the processing server 120 will then associate each of the multiple media items to one of multiple media item clusters 180,182 based on the multi-parameter characterisation.

Tribes

[0094] In the terms of this disclosure, a tribe represents a particular phenotype of a media item. Specifically, a tribe is a group of people with its own distinctive tastes, interests, and behaviour when it comes to the consumption of media items. That is, the set of observable characteristics or traits of a group of people. In particular the observable characteristics are in measured in relation to the media items.

[0095] Associated with the tribes are the media items people in the tribe would watch, interact with or otherwise consume. Media items are not associated with a single tribe, but rather could be associated with many or even all the tribes. In some embodiments, the media items may have degrees of association that differ between some or all tribes.

[0096] Fig. 5a and 5b, illustrate an example of determining tribes in two dimensional space 510. There are many approaches one may take to determining tribes but in practice it would be expected that a similar approach as the method described above for clustering can be used. Tribes are intended to be inherently distinguishable from each other by their observable viewer characteristics when measured in terms of the media items. For example, observable characteristics of a viewer include liking a strong stories with complex plots, or character based drama that focusses less on plot. [0097] In this case, the clustering can be performed on the entire set of viewer characteristics. However it is important to note that as the questionnaire is disjunct to the multi-parameter characterisation of the multiple media items, the underlying data is different and the tribes are determined on data that is not the same as the media item clusters. Although the tribes may be defined in relation to the multi-parameter characteristics, this is not necessary. In some cases, initially the processing server 120 may compare two arbitrary subsets of viewer characteristics. This may be used where there is not enough user data to determine tribes accurately or usefully.

[0098] Similarly to the above clustering, as can be seen the viewers 512-524 are distributed over the two dimensional space but are not evenly distributed. For example, media items 512, 516 and 524 are relatively isolated from other media items whereas the media items 514, 518, and 520 are nearby. As is typical in two dimensional space, a distance measure can be used to determine how different a pair of two media items are in the two dimensions. In higher dimensions, a measure that is analogous to a distance measure can still be used.

[0099] Again similarly as with the above approach for the clustering of media items, unsupervised learning techniques can be used to organise the set of viewer

characteristics into clusters such that each cluster represents a tribe. In order to for unsupervised learning to make this determination, there needs to be a metric of similarity between sets of viewer characteristics. Typically, one would use a metric of similarity (index) and a distance, which measures the dissimilarity between sets. These metrics can be used to measure similarities between clusters. A tribe may be defined as either a set of actual users or a set of characteristics. In the former, a set of existing users that have characteristics that are highly similar to each other can therefore be defined as a tribe. In the latter, each set of characteristics may be defined to be sufficient different (that is, in terms of the distance metric) from other sets of characteristics.

[0100] To make this tribe more easily discernible from other tribes, tribes can be allocated a name and an avatar that represents key characteristics of the tribe. In the case of the viewer 130, the tribe is called 'Road Warrior' represented by Max

Rockatanski, the lead character from the Mad Max films. These names and avatars are not intended to have a substantive bearing on the outcome from the recommendation system. This is just to reduce the cognitive burden of the viewer 130 by using names and images that the viewer 130 may more easily identify and engage with.

[0101] It is to be appreciated that there are other ways of determining similarity between sets and therefore tribes. Distance in one example may be simply a measure of the number of changes necessary to convert one cluster to another. Distance in other examples could be a best-match or k-centre measurement.

Allocating viewers to a tribe

[0102] As described above, in order for a viewer to be allocated to a tribe, the system 120 presents to the viewer 130,132 a questionnaire 150,152, which would typically contain a series of questions. These questions enable the processing server 120 to determine the multi-parameter characteristics of the viewer 130, 132 and to match these multi-parameter characteristics to the characteristics of the tribes. In the example in Fig. 6, the processing server 120 has mapped the viewer 110 to the 'Road Warrior' tribe. This is indicated on the user interface by the picture of the avatar for the 'Road Warrior' tribe: Max Rockatanski. There is also a description associated with the tribe 630. The indicator 620 illustrates the level of the viewer, which is in this case 'Rookie' because the viewer has just joined. As described below, this can be increased with achievements, goals, rewards and other incentives to encourage the viewer to continue to engage with the system 100.

[0103] In order to map the user to a tribe, again the Jaccard similarity coefficient can be used. In this case it is used to determine how well the characteristics of the viewer match the characteristics of a tribe, for each tribe definition that is currently in the data store 122. Typically, if there is a sufficient degree of match in the characteristics, the viewer can be allocated to a tribe. If there is an insufficient degree of match, then the processing server 120 may continue to ask more questions. If there are no further questions to ask, or if the processing server 120 has determined that there are no sufficient prospects of allocating the viewer to a tribe, then the viewer may be allocated one or more tribes based on the degree of similarity.

[0104] Although the tribes are not necessarily mutually exclusive, it is preferable that they are not so similar that it makes it difficult for the processing server 120 to allocate a viewer to a tribe. Further, the tribes should cover as much as possible the range of responses. This means that any viewer should be able to be allocated to a tribe regardless of how unusual the responses to the questions or combinations of responses to the questions may be.

[0105] In order to deal with a wider range of responses, the processing server 120 may allocate a viewer to more than one tribe and may also associate a numerical value that represents a closeness of fit. That is, a viewer may be a 67% match for a Road Warrior tribe and a 53% match for a Space Traveller tribe. Another viewer may be a 72% match for a Road Warrior tribe and a 62% match for an Explorer tribe.

[0106] Fig. 7 illustrates an example user interface for displaying the tribe of a viewer. In this case, the viewer 132 has not been mapped to a single tribe (indicated under 'Your Tribes' 710). Instead the viewer has been mapped to multiple tribes 712, 714, 716. The viewer 132 has also been recommended other tribe (indicated under

'Recommended Tribes' 722-729 which may also be close matches and therefore have interesting or relevant recommendations for that viewer 132.

[0107] In one example, server 120 determines if the film dimensions are sufficient data points in which to create discernible patterns of films into "tribes", by analysing the data collected as above to confirm it does in fact create heat maps of films that can become those tribes.

[0108] Further, artificial intelligence (AI) may be applied to provide the easy, carefully crafted questions a user would need to answer to assign them to a tribe. There may also be potential for those questions to be crafted by AI on the fly as it

progressively profiles the user.

[0109] It is noted that any information no matter how real world and unorganized can be meaningfully stored and analyzed. Machine learning can be used to make otherwise impossible new inferences from a mass of unstructured data. In that sense, tribe leaders can learn what the viewers like from a variety of unstructured places like web habits or a 1: 1 conversations.

[0110] Server 120 may personalize its responses and outputs based on its

understanding of the viewer, and how the viewer appears at a particular moment or day (e.g. sad, happy, stressed). By creating and using human-like Chabot's the Tribe Leaders can have personalities, and take & share information with viewers as if they were alive.

Recommendations

[0111] Fig. 8 is an example illustration of the components of the recommendation engine. Typically the recommendation as provided by the processing server 120 would take into account at least the tribes 830 and the media item clusters 840. That is, the tribe that the viewer is mapped to clusters of media items and their associations are the components of a typical recommendation. The system 100 will combine and weight the elements 830, 840 along with optional elements 820, 830 of the recommendation engine 810 to determine a personalised recommendation for each viewer 130,132.

Tribes

[0112] As Fig. 8 illustrates, the tribes 830 is made up of a number of individuals 832. This is a reflection of the fact that the recommendation engine 810 takes into account how other individuals affect the tribal recommendation. For example, if the

recommendation engine 810 determines that 'Goodfellas' is a recommendation for the 'Road Warrior' tribe, but then other individuals of the tribe do not watch it, or do not like it, then the recommendation engine 810 may adjust future recommendations for other members of the 'Road Warrior' tribe such as the viewer 130. The

recommendation engine 810 may also adjust the associations between the media item clusters that 'Good Fellas' belongs to and the 'Road Warrior' tribe.

Clusters

[0113] The clusters 840 are clusters of media item and this component reflects the fact that different tribes will be recommended different clusters of media items. The recommendation engine 810 takes into account the associations 170,172,174,176 in order to determine the appropriate recommended media item 160,162 for the viewer 130,132.

Social

[0114] Preferably the recommendation engine 810 may additionally take into account social 820 aspects of the viewer 110, 112. For example, the processing server 120 may keep track of the viewer's 130, 132 interactions in various known internet forums or forums that are set-up within the system. In this case, the viewer may login to forums such as reddit. During their interactions on the forums, they may make statements that reflect their opinion on a media item (160,162), whether they enjoyed it, and provide indications whether they would watch something like it in the future.

[0115] The processing server 120 may maintain an association between the viewer's 130 account and the login details of the viewer 130 for the relevant online forums. The processing server 120 can scan interactions for positive reactions based on the language used, word choice, grammar, punctuation, emoji, tags or other indicia of the engagement, emotion, intellectual stimulation or satisfaction of a viewer in viewing the media item. In the example, "Men in Black was so much fun!" the viewer has used a positive reaction word 'fun' along with an exclamation point to describe the media item the film Men in Black. [0116] The processing server 120 may take this into account as a positive reaction to allocate the viewer to a tribe. Context may be implied where the language is not apparent. In the example, "this movie was boring" it is clear that the viewer has had a negative reaction to the media item, but it is not clear which media item that the viewer is referring to. In this case though, the forum was specifically about the movie 'Under the skin.' As a result, the context of the comment from the viewer can be determined. It is worth pointing out that the forum is used as an example illustration only and it is to be understood that there are many different social elements that can be used such as social media (facebook, twitter etc), apps, or even comments on websites.

Personal

[0117] In addition to viewer characteristics, the recommendation engine 710 may take into account a set of positive evaluations (likes), and a set of negative evaluations (dislikes) of media items. This may be provided more directly than the social method described above. For example, the system 100 may simple ask a set of viewers whether they enjoyed a movie, or whether they enjoy big budget Hollywood films, special effects and 3D.

[0118] Preferably the system 100 may additionally take into account how the viewer's emotional state at which they are seeking a recommendation for a media item. Fig. 10 is an example illustration of a user interface 1010 on a viewer device 131 for the viewer 130 to determine the viewer's emotional state.

[0119] The user interface 1010 presents the user with a question, such as 'how are you feeling?' 1020. This may be essentially any question or series of question, but it is intended for these questions to evoke an answer on an emotional level. In this example, the system 100 presents two potential answers 'sad' and 'happy' 1020, 1022 on a sliding scale 1030, whereby the arrow 1040 represents the current emotional state of the viewer 130. Once the viewer 130 has made their subjective assessment of their emotional state, the viewer can press the finish button 1050. [0120] This embodiment enables the system 100 to take into account that a viewer 130,132 is more likely to want to the recommendations to be adaptable to the viewer's emotional state. In this case, the viewer 130 wants to watch 'sad' media items when they are themselves feeling sad, or 'happy' media items when they are feeling happy. The viewer 132 wants to watch 'happy' media items when feeling sad, presumably to cheer him or her up.

[0121] Fig. 9 illustrates an example set of recommended media items to the viewer. In this example, these are 'Must watch films' 910 which indicates a very close match therefore represent strong recommendations for the viewer 130. The films 911-918 are displayed in a list. In this case the list is ordered according to the degree of the closeness of the match but in other example the list could be unordered. In this example, the film 'Inception' is the highest recommended film and the encourages the viewer to watch the film if the viewer has not done so already.

Tribe Rewards

[0122] As referred to above in the discussion of tribes and Fig. 6, the system 100 may provide rewards for viewers 130,132. For example the system 100 may reward viewers 130,132 additional levels by interacting with the system 100 and the system may provide further incentives for doing so. For example, there may be a score allocated per media item which is then tallied and published on a regular basis to ascertain the viewers from each tribe that are most active or engaged. Other achievements might be to correctly answer trivia questions about the media items associated with the tribe or to provide the system with the media items they liked or disliked, or to provide written commentary or reviews. As the viewer increases in level, there may be additional features that are unlocked, or made available to him or her. This enables the tribe to feel like a community by which viewers would be encouraged to continue to engage with the system.

Monitoring Viewer Behaviour Visually [0123] In another embodiment, the processing server 120 may additionally analyse the reactions by the viewer directly (such as analysing the movement or other reactions of a viewer via a camera). In this direct method, certain actions or behaviour may be seen as a positive endorsement of a media item such as laughing, crying or prolonged eye focus. These actions indicate that the viewer is actively engaging with the content of the media item. In this method, the system may require additional image capture device such as one or more cameras. Many mobile devices have a front facing camera which may be utilised for this purpose. In this case, additional hardware would not be required.

[0124] The processing server 120 will need to be able to process the images received by the image capture device. In one example, the images are to be processed in realtime, but in other examples images can be processed to determine an overall level of engagement by the viewer with the media item. One issue with this method is that it delivers a media item while measuring the response, so will necessarily take more time than simply asking a question as to whether they enjoyed the media item or not.

Example method for recommending media items to viewers

[0125] Fig. 11 illustrates an example method for recommending media items to a viewer. As described above, the first step is displaying 1110 a review form to reviewers on a reviewer device. This allows the reviewer to analyse and review one or more media items.

[0126] Once the reviewer has completed the review, next step is receiving 1120 from the reviewers input data indicative of a multi-parameter characterisation of each of multiple media items.

[0127] Then the processing server 120 can associate 1130 each of the multiple media items to one of multiple media item clusters based on the multi-parameter

characterisation. [0128] The next step is to display 1140 on a viewer device a questionnaire to a viewer, the questionnaire being disjunct to the multi-parameter characterisation of the multiple media items.

[0129] Once the viewers have completed the questionnaire, the processing server 120 receives 1150 viewer answers from the questionnaire.

[0130] The processing server 120 can then map 1160 the viewer answers to a tribe.

[0131] Then the processing server 120 determines an association between the tribe and one of the multiple media item clusters.

[0132] Finally the processing server 120 recommends to the viewer one or more media items from the media item clusters associated with the tribe of that viewer.

Example processing server

[0133] The server 120 shown in Fig. 12 includes a processor 1202, a memory 1210, a reviewer device interface 1206, a viewer device interface 1207 and a network interface device 1208 that communicate with each other via a bus 1204. Although depicted as separate entities the device interfaces 1206, 1207 and 1208 may be combined as a single device interface which interfaces with a communication network. The memory stores instructions 1212, 1214, and 1216 and data for the processes described with reference to Figs. 1 to 11, and the processor performs the instructions from the memory to implement the processes.

The processor 1202 performs the instructions stored on memory 3110. Processor 1202 receives input from reviewer devices 111,113 and viewer devices . Processor 1202 determines an instruction according to the recommendation engine 1212. The instruction may be a function to execute according to the method to determine a recommendation for the viewer 130. The processor 1202 may execute instructions stored in the interface module 1216 to communicate with any of the reviewer devices 111, 113 or viewer devices 131, 133. The processor 1202 may execute instructions stored in the display module 1214 to displaying a review form 210, 212, 214, 216 to reviewers 110, 112 on a reviewer device 111, 113, to display on a viewer device 131, 133 a questionnaire 150, 152 to a viewer 130, 132, the questionnaire being disjunct to the multi-parameter characterisation of the multiple media items and to display a recommended one or more media items 160, 162 to the viewer 130,132 from the media item clusters 180,182 associated 170,172,174,176 with the tribe 140,142 of that viewer 130,132.

[0134] It is to be understood that any kind of communications data port may be used to receive data, such as a network connection, a memory interface, a pin of the chip package of processing server 120, or logical ports, such as IP sockets or parameters of functions stored on program memory in processing server 120 and executed by a processor. The multi -parameter characterisation (as describe above) may be stored on data memory and may be handled by- value or by-reference, that is, as a pointer, in the source code.

[0135] The processing server 120 may receive data through all these interfaces 1206,1207,1208, and this includes memory access of volatile memory, such as cache or RAM, or non-volatile memory, such as an optical disk drive, hard disk drive, storage server or cloud storage. The processing server 120 may further be implemented within a cloud computing environment, such as a managed group of interconnected servers hosting a dynamic number of virtual machines.

[0136] It is to be understood that any receiving step may be preceded by the processing server 120 determining or computing the data that is later received. For example, the processing server 120 may determine answer data and store the answer data in data memory, such as RAM or a processor register. The processing server 120 then requests the data from the data memory, such as by providing a read signal together with a memory address. The data memory provides the data as a voltage signal on a physical bit line and the processor receives the answer data via a memory interface. [0137] It will be appreciated by persons skilled in the art that numerous variations and/or modifications may be made to the above-described embodiments, without departing from the broad general scope of the present disclosure. The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive.

Claims

CLAIMS:

1. A computer system for recommending media items to a viewer, the computer system comprising:

multiple reviewer devices to display a review form to reviewers and to receive from the reviewers input data indicative of a multi-parameter characterisation of each of multiple media items;

wherein the processing server is configured to

receive viewer answers from the questionnaire;

map the viewer answers to a tribe;

2. A computer system according to claim 1, wherein recommending to the viewer one or more media items is based on a personalised recommendation for each viewer.

3. A computer system according to claim 2, wherein the personalised

recommendation includes a subjective assessment of the viewer's current emotional state.

4. A computer system according to claim 3, wherein the multiple viewer devices display an interactive user interface for the viewer to determine the subjective assessment of the viewer's current emotional state.

5. A computer system according to any of the preceding claims, wherein the processing server is configured to determine one or more clusters of media items based on the input data indicative of a multi-parameter characterisation.

6. A computer system according to any of the preceding claims, wherein the processing server determines one or more tribes based on the input data indicative of a multi-parameter characterisation.

7. A computer system according to any of the preceding claims, wherein the processing server determines the one or more tribes based on similarities in the multiparameter characterisation as calculated by a similarity index.

8. A computer system according to any of the preceding claims, wherein the processing server determines the one or more tribes based on similarities in the multiparameter characterisation as calculated by a distance measure.

9. A computer system according to claim 7 wherein the similarity index is a Jaccard similarity coefficient.

10. A computer system according to claim 5 wherein the processing server utilises a shared nearest-neighbour algorithm to determine one or more clusters of media items.

12. A computer system according to claim 7 wherein the processing server utilises a shared nearest-neighbour algorithm to determine one or more tribes of viewers.

13. A computer implemented method of operating a computer system for recommending media items to a viewer, the method comprising:

displaying a review form to reviewers on a reviewer device;

receiving from the reviewers input data indicative of a multi-parameter characterisation of each of multiple media items; associating each of the multiple media items to one of multiple media item clusters based on the multi-parameter characterisation;

displaying on a viewer device a questionnaire to a viewer, the questionnaire being disjunct to the multi-parameter characterisation of the multiple media items; receiving viewer answers from the questionnaire;

mapping the viewer answers to a tribe;

14. Software, being machine readable instructions, that when performed by a computer system causes the computer system to perform the method claim 13.