WO2015188885A1

WO2015188885A1 - Method and system for determining a recommendation for content

Info

Publication number: WO2015188885A1
Application number: PCT/EP2014/062450
Authority: WO
Inventors: Bin Cheng; Benjamin HEBGEN
Original assignee: Nec Europe Ltd.
Priority date: 2014-06-13
Filing date: 2014-06-13
Publication date: 2015-12-17

Abstract

The present invention relates to a method for determining a recommendation for a content on a mobile device, preferably a companion device of a user, comprising the steps of a) Generating an event stream based on events of captured user interactions when interacting with a provided service, wherein said event stream is a sequence of events, wherein each event is associated at least with information of one or more performed actions by the user on an object and context information related to said performed action, b) Constructing one or more composite user activities based on matching of one or more predefined user activity patterns with an actual event stream, c) Generating an engagement score for said constructed composite user activities, wherein an engagement score represents a level of engagement of a user when performing the corresponding composite user activity, and d) Determining one or more content recommendations based on the engagement scores of the composite user activities. The present invention further relates to a system for determining a recommendation for a content on a mobile device, preferably a companion device of a user.

Description

METHOD AND SYSTEM FOR DETERMINING A

RECOMMENDATION FOR CONTENT

The present invention relates to a method for deternnining a recommendation for content on a mobile device, preferably a companion device of a user.

The present invention further relates to a system for determining a recommendation for content on a mobile device, preferably a companion device of a user, preferably operable to perform a method according to one of the claims 1 - 17.

Although applicable to any device in general the present invention will be described with regard to a mobile device, preferably a companion device of a user. Mobile devices have become more and more popular nowadays and people are connected with to the Internet at any time from any place with their mobile devices. Mobile devices have entered the daily live of users in a plurality forms like Wi-Fi scales, fitness trackers, smart refrigerators or even coffee machines having a Wi- Fi connection. Further with smartphones people are able to buy a product when shopping in a mall, check information when seeing an art in a museum, or vote for some actor while watching a television show at home. Fitness trackers enable to follow progress in training and the coffee machine can be remotely activated that it is ready when standing up in the morning or the refrigerator sends a notification message when the milk is empty so that further milk can be bought in advance.

Mobile devices are in such cases used as companion devices together with main things, trigger users to perform an action and interact with the companion devices, for example a digital signage in a shopping mall, an art associated with an iBeacon in a museum, the fitness tracker linked to a user's calendar or a coffee machine reacting to an alarm of a smartphone. By triggering the interactions from each individual user, companion devices enable service providers to recommend personalized content for example in form of offers, products, relevant programs, descriptions or services, etc., to each user according to their behaviour and context on the companion devices of the respective users. The personalized content recommendation on the companion devices and evaluation has to be performed based on how users are engaged with the main things via companion their devices, preferably in real-time. Conventional recommendation systems rely on users to explicitly give their inputs for rating content: For example, online shops provide a rating possibility after a user has bought a product, so that user preference between user and content can be calculated based on their rating by their recommendation systems. However, most of users do not want to make the effort to give explicit inputs that are only useful for the corresponding shop operator. Instead of asking feedback from users explicitly, other conventional methods for providing recommendations collect user inputs and/or feedback from social networks. For example in US 2013/0262191 A1 a system and method for measuring and scoring engagement in organizations is shown. Key performance indicators are selected and received so as to reflect an organization's goal with regard to an individual or organization's level of engagement with another individual, another group or organization. The key performs indicators are then weighted against each other in accordance with their importance relative to other selected key performance indicators with more important key performance indicators receiving a greater weight. Then criteria for translating performance data associated which each key performance indicator to a normalized point value are determined and received. A normalized point value associated with a particular key performance indicator performance is typically representative of the importance the organization has placed on that level of performance. In the last step a composite engagement score is calculated by locating, mining or otherwise receiving performance data specific to the selected key performance indicators translating the performance data to a normalized point value for each selected key performance indicator in accordance with the previous defined criteria and adding the normalized point values together.

However, one of the problems is that feedback from social networks cannot be collected in real-time. However, for content recommendation in companion devices real-time feedback is important to achieve better end optimized content recommendations. Personalized content recommendation need to be made immediately according to the latest context and user behaviours. Otherwise business chances decreases

In the non-patent literature of Paper: Georgios Meditskos, Stamatia Dasiopoulou, Vasiliki Efstathiou, loannis Kompatsiaris , "Ontology Patterns for Complex Activity Modelling", in Proceedings of the 7th International Symposium, RuleML 2013, Seattle, WA, USA, July 1 1 -13, 2013" ontologies are used for modelling and reasoning about complex activities and situations in context-aware applications. An activity pattern ontology is shown serving as meta model over domain activity classes capturing structural notions of atomic and compound activities and interoperable activity models. The activity pattern ontology represents relationships and is used for derivation of complex activities in terms of activity types and temporal relations. In the non-patent literature of Paper: Menaka Gandhi J., Gayathri K. S., "Activity Modeling in Smart Home using High Utility Pattern Mining over Data Streams", International Journal of Computer Science and Network, Vol. 2, Issue 3, June 2013" frequent patent mining and high utility patent mining is shown and used to mine activity patterns of residents of a home from collected sensor data for a smart home environment.

It is therefore an objective of the present invention to provide a method and a system for determining a recommendation for content on a companion device of a user enabling an accurate and precise estimation of a user engagement on a recommended content and therefore provide a precise content recommendation.

It is a further objective of the present invention to provide a method and a system for determining a recommendation for content on a mobile device of a user which enable a real-time recommendation for content on a mobile or companion device of the user.

It is an even further objective of the present invention to provide a method and a system for determining a recommendation for content on a mobile device of a user. It is an even further objective of the present invention to provide a method and a system for determining a recommendation for content on a mobile device of a user which are easy to implement and cost-effective.

The aforementioned objectives are accomplished by a method of claim 1 and a system of claim 18.

In claim 1 a method for determining a recommendation for content on a mobile device, preferably a companion device of a user, is defined.

According to claim 1 the method is characterized by the steps of

a) Generating an event stream based on events of captured user interactions when interacting with a provided service, wherein said event stream is a sequence of events, wherein each event is associated at least with information of one or more performed actions by the user on an object and context information related to said performed action, b) Constructing one or more composite user activities based on matching of one or more predefined and preferably weighted user activity patterns with an actual event stream,

c) Generating an engagement score for said constructed composite user activities, wherein an engagement score represents a level of engagement of a user when performing the corresponding composite user activity, and

d) Determining one or more content recommendations based on the engagement scores of the composite user activities.

In claim 18 a system for determining a recommendation for content on a mobile device, preferably a companion device of a user, preferably operable to perform a method according to one of the claims 17 is defined.

According to claim 18 the system is characterized by

means operable to perform the following steps: a) Generating an event stream based on events of captured user interactions when interacting with a provided service, wherein said event stream is a sequence of events, wherein each event is associated at least with information of one or more performed actions by the user on an object and context information related to said performed action, b) Constructing one or more composite user activities based on matching of one or more predefined and preferably weighted user activity patterns with an actual event stream,

In other words the present invention provides a method and a system for a content recommendation on companion devices driven by user engagement being evaluated based on preferably high level user activities recognized from captured event streams using pattern recognition. A set of composite activities is derived from captured user interactions and then user engagement is determined based on the composite activities recognized from the collected user interactions in realtime. The estimated user engagement is used to serve personalized and context aware content to users on the companion devices.

A composite user activity may comprise only one interaction specified by a certain context condition. In a more complex case the composite activity can be defined by a set of interactions following a more complex pattern derived from multiple contextual dimensions such as patterns with both time and location constraints or pattern with both content and action constraints. In general activity patterns can be shared for a similar service in the same business domain. For example the activity pattern that has been dug out from a traditional online shopping website can be generally applied into other online shopping websites for measuring user engagement. Further activity pattern knowledge base can be built and shared for many service providers in the same business domain. According to the invention it has been recognized that no explicit user inputs are required for a recommendation. The user engagement score can be automatically calculated based on the constructed composite user activities recognized by the predefined user activity patterns from the collected event streams in real-time without having the users to explicitly rate preferences or expressing feedbacks to service providers.

According to the invention it has been further recognized that a real-time contextual recommendation on companion devices is enabled: Content recommendation on companion devices can be provided in real-time based on the latest engagement score that can be adapted in a fast manner to changing context and user engagement. According to the invention it has been even further recognized that a more accurate and precise estimation of user engagement on recommended content is enabled: The activity-based engagement score provides more accurate information of user preferences on content and/or services when users interact with content and/or services on the companion devices. The user engagement on recommended content can be measured to justify how much positive effect the previous recommendation has been taken for users and can be utilized as feedback to improve content recommendation.

Further features, advantages and preferred embodiments are described in the following sub claims.

According to a preferred embodiment one or more user activity pattern candidates are generated by performing data mining, preferably frequent data mining on one or more previous event streams and the predefined user activity patterns are provided by interpreting said one or more user activity pattern candidates. By using pattern mining to determine user activity pattern candidates and interpreting the same a very precise content recommendation in real-time is enabled. For example interpreting may be performed by explicitly rating each pattern candidate according to some predefined variables. According to a further preferred embodiment the user interactions are captured on the mobile device, preferably in form of a companion device of a user and sent to a sever for performing at least step b). This enables to use for example big processing power of computers or elastic cloud computing platforms for performing the pattern mining process ensuring a fast content recommendation in real-time without having to use the resources of the mobile device for data mining, but only capturing the user interactions. According to a further preferred embodiment for capturing a user interaction an action of a user is recorded together with a context related to the action. This provides that - when generating the event stream - actions and context related to the actions are available together ensuring a fast generation of the event stream while providing a close association between context of the action and the action itself.

According to a further preferred embodiment an identification of a user is associated with an event and provided by a user ID associated with a user profile. This allows for example to save resources, since only a user ID has to be included in the event stream and the user profile can be - if necessary - every time it is needed downloaded without having to include it into the transmission from the companion device to the server.

According to a further preferred embodiment an object is represented by a content path and/or context information is represented by a set of attributes. This ensures that for example additional attributes can be easily included. Thus flexibility and processing of the events are enhanced.

According to a further preferred embodiment one or more predefined constraints are respected when performing the pattern mining, preferably wherein said constraints are predefined by a service operator. For example with regard to the scale of users, length of the sequence of events of a user and the multidimensional property of a data of the events the search space for the patterns exponentially increasing. Therefore said constraints can be used to optimize the pattern nnining procedure. The constraints can be given by a service operator or provider according to any requirement or can be specified by prior engagement analytics in advance. For example some constraints can be defined to specify attributes of a group of users in order to limit a user sample space or a pattern constraint can be given to define which type of patterns is to be searched.

According to a further preferred embodiment for performing pattern mining users are clustered. This also optimizes the pattern mining process limiting the user sample space in particular.

According to a further preferred embodiment the generated user pattern candidates are interpreted based on one or more requirements of an operator, preferably wherein said user pattern candidates are presented to a human operator for interpretation. This even further enhances the precision of the later content recommendation taking into account both the computer capability and the human capability for determining and interpreting the activity patterns. Usually when a human operator, for example a service operator, is involved he may enrich the above mentioned constraints for pattern mining defining user clustering rules and interpreting pattern candidates since he has usually the knowledge of his provided service. Further such an interpretation does not use too much calculation or processing time.

According to a further preferred embodiment the number of generated user pattern candidates is controlled by the total number of occurrences of a pattern and/or a probability of occurrence of a pattern. This enables in an easy way to control how many candidates pattern will be output by the pattern mining. Thus efficiency is enhanced.

According to a further preferred embodiment the generated user pattern candidates are location-, action-, device- and/or time-based for indicating user engagement. This enables to indicate a user engagement beyond a single interaction level. For example location-based patterns show how a user is engaged with different places in an environment. For instance in a shopping mall the number of customer revisits in some area of the shop may indicate that this customer is considering buying some product but is still a bit concerned about the price. If a customer is just looking around the customer may just walk through the environment without coming back or staying longer at any location. Action-based patterns may show a series of user interactions that happened together a certain frequency and device-based pattern show cross-device user interactions which can be used as an indication that a user likes certain services and wants to make efforts to finish something across devices. Time-based patterns may be used describing user interactions that are repeated in a daily or weekly pattern or the patterns show some interactions happening very often at a specific time period of the day. Such time-based or temporal patterns of user interactions may represent great user royalty and indicate high user engagement.

According to a further preferred embodiment the user activity patterns are translated into one or more condition rules for said matching. This enables for example an activity pattern that can be understood by an event processing engine to generate a composite user activity in real-time.

According to a further preferred embodiment for matching a plurality of event processing engines is used, wherein events of a user are provided to the same processing engine. This allows in an efficient way to detect activity patterns while enabling the work load to be spread across different event processing engines for load balancing. According to a further preferred embodiment actual event streams received during predefined time periods are saved and used as previous event streams for the next time period, preferably together with incremental frequent pattern mining. This further enhances the efficiency and precision since for example incremental pattern mining procedures can be utilized to reuse some previous performed pattern mining while using the save event streams for further enhancing pattern candidates and/or activity patterns.

According to a further preferred embodiment the engagement score is generated based on weighted user activity patterns, preferably wherein the weight associated with a user activity pattern is determined by the service provider. This enables to convert interaction events to weighted composite activities from activity patterns. Thus content recommendation is further enhanced.

According to a further preferred embodiment the engagement scores of all composite activities are accumulated for each user using an aggregation function. This enables to present the user the most relevant content based on the generated user engagement score providing a further optimized content recommendation.

According to a further preferred embodiment collaborative filtering and/or content- based filtering is used to rank the content recommendations based on the engagement score. This provides a reliable ranking of the content recommendation by using well-known collaborative filtering or content-based filtering.

There are several ways how to design and further develop the teaching of the present invention in an advantageous way. To this end it is to be referred to the patent claims subordinate to patent claim 1 on the one hand and to the following explanation of preferred embodiments of the invention by way of example, illustrated by the figure on the other hand. In connection with the explanation of the preferred embodiments of the invention by the aid of the figure, generally preferred embodiments and further developments of the teaching will be explained. the drawings

Fig. 1 shows a system and steps of a method according to a first embodiment of the present invention; Fig. 2 shows part of a method according to a second embodiment of the present invention;

Fig. 3 shows part of a method according to a third embodiment of the present invention; Fig. 4 shows part of a method according to a fourth embodiment of the present invention and Fig. 5 shows part of a method according to a fifth embodiment of the present invention.

Fig. 1 shows a system and steps of a method according to a first embodiment of the present invention.

In Fig. 1 an engagement driven recommendation on companion devices is shown. In more detail Fig. 1 illustrates how recommendation works for companion devices CD in a digitalized environment at the high level. Within the digitalized environment DE one or more main devices MD are deployed. Some main content MC is associated with each main device MD. When a user U with a mobile device CD, for example in form of a smartphone, a tablet or the like, is nearby, the main device MD can trigger the user U to check some relevant content of the main content MC of his mobile device CD which is also called companion device here. The relevant content RC on the companion device CD is personalized and context-aware to the user U.

On the user device CD a content sensing component CS is running used to capture contextual events when the user U interacts with the relevant content RC and also provides context information CI when the companion device CD is requesting relevant content RC. To capture user interactions of a user U with a companion device CD a series of events e with detailed context information CI will be recorded and further reported to a backend server. Each event e means that a user U performs an action a with an object o under a certain context c. Here an object o can be a predefined service or specific content like a webpage or product or the like.

An event e is therefore used to capture which user u does what to which object o under what context c. The content sensing component CS transmits - as already mentioned above - the sensed context to a receiver R. A receiver R receives a user interaction and transmits them to a contextual event entity CEE creating an event stream ES of the contextual events CE.

On the event stream ES pattern mining PM is performed. Since the collected events e are multiple dimensional data and the event stream ES, i.e. the sequence of events e, reflects how users U are interacting with the provided service PS then some pattern are dug out of the event stream ES and the discovered patterns are used to improve the way of measuring a user engagement more accurately. Based on the mined patterns, i.e. candidates for activity patterns, activity patterns AP are provided/chosen and used for recognizing an activity by an activity recognition entity AR. For example when a new event stream ES is transmitted to a receiver R the receiver R forwards the new event stream ES also to the activity recognition entity AR. The activity recognition AR matches the provided new event stream ES with the former generated/selected activity patterns AP. Based on a matching of the activity patterns AP with the corresponding event stream ES an engagement calculation EGMC is performed and the engagement is qualified based on the recognized activities based on some weighted recognized activity patterns AP.

The activity patterns AP may be selected and weighted by the service provider SP, or any other user U. The content recommendation CR is made based on the user ID and current context in the request and the estimated engagement qualifying how users U are engaged with content under different context in the previous time. After content recommendation CR the relevant content RC is served by the provided service PS on request according to the recommendation giving by the content recommendation CR.

The pattern mining PM explores the frequent sequential event patterns from the contextual event data CE under different constraints, generates a list of pattern candidates PC and may visualize the generated pattern candidates PC for the service provider to group pattern selection and perform weighing.

Fig. 2 shows part of a method according to a second embodiment of the present invention. ln Fig. 2 a presentation of a captured event e is shown. An event e comprises four properties u, a, o, and c, where u means the profile of the user, such as the user ID, age, gender, or profession; a means the action that the user does, for example, click, review, select, scan, register, or pay; o means the object associated with the action; c means the context which has been collected when the user does the action, for example, location, time, device, the associated main content MC, network type, or social relationship with other uses in the system. Since the user profile u is relatively static and may be represented by a unique user id, the reported event e only needs to contain the user ID in the event message.

The object o can be presented by a content path, for example, service->category- >product. The context information CI can be represented by a set of attributes. These properties need to be extracted by the user device CD in the first place when the user U interacts with it.

Fig. 3 shows part of a method according to a third embodiment of the present invention.

In Fig. 3 the procedure to dig out activity patterns is shown. For digging out patterns user interaction data has been collected beforehand. The user interaction data comprises the sequence of events e reflecting how users U are interacting with the provided service PS. In Fig. 3 the collected data comprising events e is clustered according to users U so that events e of each of the users U are separated.

After the user clustering UC a frequent pattern mining FPM procedure is performed to find out the most frequent pattern candidates PC. With regard to the scale of users U, the length of event sequence per user U and the multiple dimension property of the event data e, the pattern searching space is expanded exponentially and therefore constraints may be applied provided by a service provider SP according to his requirements or which can be specified by a system in advance. For example one or more rules can be defined to specify attributes of a user group in order to limit or narrow down the user sample space or one or more frequent pattern constraints can be given to define which type of frequent patterns can be searched in the data.

After the pattern candidates PC are generated under the defined constraints those pattern candidates PC are preferably visualized (pattern visualization PV) for a service provider SP for performing pattern interpretation PI. When searching for a pattern candidate PC frequent pattern mining FPM procedures focus on a current frequency of a sequence of events e across users U but may not consider an actual mining of event sequences. Visualizing PV the pattern and its associated event information enable to consider an actual meaning: A service provider SP may interpret the pattern candidates PC and check its actual meaning with his understanding of his business.

If a pattern candidate PC does not show an indication of a high level user behavior this pattern candidate PC will be dropped out. If a pattern candidate PC shows a high level user behavior, preferably from his service providers SP point of view, this pattern candidate PC will be regarded as an activity pattern AP and the service provider SP will assign a score to this activity pattern AP in terms of how influential this type of user activity means to his service business.

During the mining and interpretation process the service provider SP is preferably involved: He is asked to provide one or more constraints (mining assistance Ml) for the frequent pattern mining FPM, to define user clustering rules for the user clustering UC and/or to provide pattern interpretation PI of visualized PV pattern candidates PC. Usually the service provider SP has the knowledge of the provided service PS but the mining assistance Ml and the pattern interpretation PI do not require much calculation or processing time. Since frequent pattern mining FPM requires compared the mining assistance Ml and the pattern interpretation PI a much greater amount of processing the assistance from the service provider SP provides an even higher efficiency and scalable big data for assessing.

By using frequent pattern mining FPM and user clustering UC different types of pattern candidates PC can be identified from the event stream ES of the interactions of a user U according to predefined clustering rules and if applicable constraints for the frequent pattern mining FPM. To control how many pattern candidates PC are provided the following parameters may be preferably used: "Support" representing the total number of occurrences of a pattern and "confidence" meaning a probability of the pattern occurrence in the corresponding data set for the pattern mining.

In general patterns may be location-based, action-based, device-based and/or time-based as already mentioned above. In general activity patterns AP can be shared for similar services in the same business domain. For example the activity patterns AP that have been dug out from a traditional online shopping website can be generally applied into other online shopping websites for measuring user engagement. Further an activity pattern AP knowledge base can be built up and shared for many service providers in the same business domain. Fig. 4 shows part of a method according to a fourth embodiment of the present invention.

In Fig. 4 a pattern-based activity recognition via complex event processing CEP is shown.

When the service provider SP picks up some meaning patterns from identified behavioural pattern candidates PC, the selected activity patterns AP are transformed into a set of condition rules which can be understood by complex event processing engines CEP. As Figure 4 shows, the rule sets are broadcasted to all complex event processing CEP instances by a scheduler S. The front-end scheduler receives all incoming interaction events and forwards them to back-end complex event processing engine CEP instances for further activity detection in a way that events e from the same user U are given to the same complex event processing engine CEP instance and workload can be spread across different complex event processing engine CEP instances for load balancing. To be scalable, the back-end complex event processing engine CEP instance cluster is scaled up and down according to the current incoming workload. When an engine CEP instance detects a pattern from its received event stream ES based on the pattern rule set, it outputs a composite activity CA to other components in real time for quantifying user engagement and reasoning about the results statically.

On the other hand, all incoming events e will be saved and a batch processing BP to do pattern discovery are triggered again in a large time period, for example once per day during the time when the system has lowest workload, or once per week, or the like. Therefore, the rule set to detect composite activities CA can be updated to reflect a change of user behaviour patterns. To be more efficient, some incremental frequent pattern mining FPM procedures can be utilized to reuse some previous computation.

Fig. 5 shows part of a method according to a fifth embodiment of the present invention. In Fig. 5 a procedure to calculate the user engagement is shown.

In detail Fig. 5 illustrates how a user engagement score UES can be derived from a captured event stream ES: In a first step the user interaction events e-i , e_m from each user U are queued up into an event buffer in a certain time window. The time window could be an hour, per day, per session or per visit. It may be predefined by the service provider SP. For example it depends on the time period the service provider SP wants to look back for checking user behavior patterns. The buffered event stream ES comprising the events e-i , e_m is updated as more events e-i , e_m are collected.

In a second step all selected pattern candidates PC are applied to match the event stream ES in the defined time window. A pattern candidate PC can be formulated and expressed by complex event rules. Based on the complex event processing engine CEP once a pattern P is matched with the event stream ES a composite activity CA will be constructed together its associated pattern P and its associated score w. When a pattern P is selected to recognize composite activities CA it will be assigned the score w to indicate the importance of its associated composite activities CA on measuring user engagement. For example the pattern p1 has a score w1 . When p1 is used to identify the composite activity CA1 for a user U, CA1 will have the score w1 . In a third step for each user U a series of composite activities CA can be recognized from the incoming event stream ES together with the scores W. Compared with the corresponding lower level events, composite activities CA are more explanatory for service providers SP to see how influential they are for indicating user engagement. For a given user U the sequence of his activities can be regarded as a short summary of what he did until now.

In a fourth step an aggregation function AF is used to accumulate the scores W of all recognized composite activities CA for each user. For example, the aggregation functions can be a sum function. In this case, the engagement calculation for a given user U can be formulated as follows. Assume that there are N composite activity patterns {p-i , p2, P3, PN}, and for a given user U, m composite activities CA have been recognized from the event stream ES of this user U, the aggregated engagement score UES can be denoted by the following equation: E

represents how many times the pattern p, has been recognized from the event stream ES of this user, and 0 < n, < m.

Given the context provided in the request from the companion device CD and the user information, the content recommendation CR component can select the most relevant content based on the measured/calculated engagement score UES. The provided context include location, time, the main content MC on the main device MD and the user information can indicate the user profile u and the user's previous behaviour. These two types of inputs will be used by the content recommendation CR component to figure out what could be the content candidates. Then some recommendation models, for example, collaborative filtering or content-based filtering, can be used to rank the content candidates based on the engagement score UES. Finally, the content recommendation CR component returns the top n content candidates to the companion device CD. ln summary the present invention enables activity pattern selected and weighted by service providers to convert interaction events to weighted composite activities and further to aggregated engagement scores. The present invention further enables engagement scores to be applied between user and content to drive better and real-time content recommendation on companion devices.

Further no explicit user inputs are required for recommendation. An automatic calculation of the user engagement score based on weighted composite activities being recognized by the weighted activity patterns from collected interaction event streams in real-time is enabled without users having to explicitly rate their preferences or expressing their feedback to service providers. The present invention further provides real-time content recommendation on companion devices: Content recommendation on companion devices can be given in real-time based on the latest engagement score and can be fast adapted to the changing context and user engagement. Even further the present invention enables a more accurate and precise estimation of a user engagement on recommended content: An activity based engagement metric is enabled providing more accurate information of user preference on content and/or services on companion devices. The user engagement on recommended content can be measured to justify how much positive effect of a previous recommendation has been taken by the user and can be further utilized as feedback to improve content recommendations.

Even further the present invention enables weighted composite activities being understandable for service providers and which are recognizable by a computational system for example. Behavior context can be taken into account to measure user engagement and to provide personalized and context aware content recommendations on companion devices in real-time. ln other words the present invention provides preferably a method for service and content recommendation on user's companion devices in a digitalized environment where some devices or things in general are deployed and trigger users to further check relevant information or receive personalized services on their personal companion devices like mobile phones or other interactive connected things, for example in digitalized shops, smart homes, or smart museums and cities, with preferably the following steps:

- report contextual user interaction events captured on user companion devices to a remote server,

- recognize composite activities from contextual event stream according to the weighted activity patterns mined from previous user interaction data,

- calculate user engagement score based on the recognized composite activities and their weights,

- recommend content for companion devices on request based on the measured engagement scores between users and content.

The concept of the present invention in particular of companion devices can be presented to any content in general which can receive recommended content and/or services and/or engage users through their interactions with the recommended content.

Many modifications and other embodiments of the invention set forth herein will come to mind the one skilled in the art to which the invention pertains having the benefit of the teachings presented in the foregoing description and the associated drawings. Therefore, it is to be understood that the invention is not to be limited to the specific embodiments disclosed and that modifications and other embodiments are intended to be included within the scope of the appended claims. Although specific terms are employed herein, they are used in a generic and descriptive sense only and not for purposes of limitation.

Claims

C l a i m s

A method for determining a recommendation for a content (RC) on a mobile device, preferably a companion device (CD) of a user (U),

characterized by the steps of

a) Generating an event stream (ES) based on events (e) of captured user interactions when interacting with a provided service (PS), wherein said event stream (ES) is a sequence of events (e), wherein each event (e) is associated at least with information of one or more performed actions (a) by the user (U) on an object (o) and context information (CI) related to said performed action (a),

b) Constructing one or more composite user activities (CA) based on matching of one or more predefined and preferably weighted user activity patterns (AP) with an actual event stream (ES),

c) Generating an engagement score (UES) for said constructed composite user activities (CA), wherein an engagement score (UES) represents a level of engagement of a user (U) when performing the corresponding composite user activity (CA), and

d) Determining one or more content recommendations (CR) based on the engagement scores (UES) of the composite user activities (CA).

The method according to claim 1 , characterized in that one or more user activity pattern candidates (PC) are generated by performing data mining (FPM), preferably frequent data mining, on one or more previous event streams (ES) and the predefined user activity patterns (AP) are provided by interpreting said one or more user activity pattern candidates (PC).

The method according to one of the claims 1 -2, characterized in that the user interactions are captured on the mobile device (CD), preferably in form of a companion device (CD) of a user (U) and sent to a server (1 ) for performing at least step b).

4. The method according to claim 3, characterized in that for capturing a user interaction an action (a) of a user (U) is recorded together with a context (c) related to the action (a).

5. The method according to one of the claims 1 -4, characterized in that an identification of a user (U) is associated with an event (e) and provided by a user id associated with a user profile (u).

6. The method according to one of the claims 1 -5, characterized in that an object (o) is represented by a content path and/or the context information (CI) is represented by a set of attributes.

7. The method according to one of the claims 2-6, characterized in that one or more predefined constraints are respected when performing the pattern mining (PM, FPM), preferably wherein said constraints are predefined by a service operator (SP).

8. The method according to one of the claims 2-7 characterized in that for performing pattern mining (PM, FPM) users (U) are clustered.

9. The method according to one of the claims 2-8, characterized in that the generated user pattern candidates (PC) are interpreted (PI) based on one or more requirements of an operator (SP), preferably wherein said user pattern candidates (PC) are presented (PV) to human operator (SP) for interpretation.

10. The method according to one of the claims 2-9, characterized in that the number of generated user pattern candidates (PC) is controlled by the total number of occurrences of a pattern and/or a probability of occurrence of a pattern.

1 1. The method according to one of the claims 2-10, characterized in that the generated user pattern candidates (PC) are location-, action-, device- and/or time-based for indicating user engagement.

12. The method according to one of the claims 1 -1 1 , characterized in that the user activity patterns (AP) are translated into one or more condition rules for said matching.

13. The method according to one of the claims 1 -12, characterized in that for matching a plurality of event processing engines (CEP) is used, wherein events (e) of a user (u) are provided to the same processing engine (CEP).

14. The method according to one of the claims 1 -13, characterized in that actual event streams received during predefined time periods are saved and used as previous event streams (ES) for the next time period, preferably together with incremental frequent pattern mining (FPM).

15. The method according to one of the claims 1 -14, characterized in that the engagement score (UES) is generated based on weighted user activity patterns (AP), preferably wherein the weight associated with a user activity pattern (AP) is determined by the service provider (SP).

16. The method according to one of the claims 1 -15, characterized in that the engagement scores (UES) of all composite activities (CA) are accumulated for each user (u) using an aggregation function.

17. The method according to one of the claims 1 -16, characterized in that collaborative filtering and/or content-based filtering is used to rank the content recommendations (CR) based on the engagement score (UES).

18. A system for determining a recommendation for a content on a mobile device, preferably a companion device (CD) of a user (U), preferably operable to perform a method according to one of the claims 1 -17,

characterized by means operable to perform the following steps: