The present invention is related to a user profile learning method and an associated profiling engine.

Knowing one's customer is an essential ingredient for success in any business, but, in the telecommunications market, it has become more important than ever. Users are increasingly demanding personalized services that enhance their quality of life and deliver personalized information, anywhere, anytime, and on any device. For instance, users can now access hundreds of television (TV) channels with thousands of programs, and they need efficient support to automatically locate a program corresponding to their interests. Mastering knowledge of user profiles is therefore becoming one of the technical cornerstones in new business areas such as content placement.

Many personalized applications have already been successfully commercialized in the domain of Web technologies and e-commerce by pioneers such as Google™ and Amazon™. Delivering this level of customization in converging telecommunications and content services means going beyond a single personalized application. This can be done by implementing network intelligence that leverages the user profile and supports multiple applications such as social networking, personalized content push, and targeted advertising. Indeed, with converged triple screen offers for TV, mobile devices, and personal computers (PCs), telcos and service providers hold large amounts of data on end user service consumption and they are well positioned to accurately infer user preferences, interest domains, and behaviors. However, without an efficient profiling tool making benefit, the provider cannot leverage this information and deliver value-added services.
Accordingly, it is a general object of the present invention to provide a profiling engine that automatically learns the profile of each telecommunication (or telco) customer. In the following approach, it is proposed a generic profiling engine that:

The object of the present invention, according to an embodiment, is a user profile learning method comprising learning and updating the profile data defined in a user model, using two mechanisms which are an explicit profiling process and an implicit profiling process, in which, in the explicit profiling process, the user declares his topics of interest, e.g. via a Web portal, given that, at the initialization stage, the absence of active indication of interest in a topic implies a lack of user interest, and the implicit profiling process comprising learning and updating the profile data from all kinds of usage traces through observation and tracking of actual service consumption by the user such as the user's context or his presence status when consuming the services.

According to an embodiment of the invention, the implicit profiling process is based on the analysis of heterogeneous usage traces in a mobile environment.

According to an embodiment of the invention, three measurable quantities are used to describe contents, consumption events, and user profiles, the three measurable quantities being :

According to an embodiment of the invention, the Quantity of Consumption values are computed by modulating the Quantity of Affiliation values of the consumed content by numerical parameters held in an homogeneous raw data generation module destined to transform all platform-specific data into a common, pre-defined format, said Quantity of Consumption function being modeled as: $${\mathit{QoC}}_n^i=\mathit{QoA\_Modulation}\left({\mathit{QoA}}_n^i{\mathit{HRD}}_n\right)$$

where the index i refers to the relevant semantic concept and n represents consumption event ordering.

According to an embodiment of the invention, the consumption event-based Quantity of interest (QoI) learning function refreshes the Quantity of Interest data on user interest by combining their previously known values with a newly-observed interest manifestation defined by the Quantity of Consumption, this Quantity of Interest (QoI) learning function being modeled as: $${\mathit{QoI}}_{n+1}^i=\mathit{QoI\_Learning}\left({\mathit{QoI}}_n^i{\mathit{QoC}}_n^i\right)$$

According to an embodiment of the invention, it is considered a particular family of functions where the new Quantity of Interest is obtained by a weighted addition of the newly observed Quantity of Consumption with the old Quantity of Interest: $${\mathit{QoI}}_{n+1}^i={\mathit{QoI}}_n^i+W\left({\mathit{QoI}}_n^i\right)*{\mathit{QoC}}_n^i$$

the weight $W\left({\mathit{QoI}}_n^i\right)$

being a function of the current Quantity of Interest and defining how much the new observation is influencing the profile evolution.

According to an embodiment of the invention, a time-based decay function is introduced to account for the aging of profile data: $${\mathit{QoI}}_{k+1}^i=\mathit{QoI\_Decay}\left({\mathit{QoI}}_k^i{P}_k^i\right)$$

this time-based decay function being called with a given periodicity indexed by k.

According to an embodiment of the invention, said periodicity is significantly larger than the average time interval between consumption events.

According to an embodiment of the invention, in order to decide how to decrease a given Quantity of Interest, this time-based decay function take into account parameters $\left(P_k^i\right)$

like frequency or recency of consumption events on each considered semantic concept.

The object of the present invention, according to an embodiment, also concerns a profiling engine for Converged Service Delivery Platforms, comprising learning and updating means for learning and updating the profile data defined in a user model, using two mechanisms coming from explicit profiling means and implicit profiling means, wherein the explicit profiling means has inputs for receiving users' declarations about their topics of interest, e.g. via a Web portal, given that, at the initialization stage, the absence of active indication of interest in a topic implies a lack of user interest, and the implicit profiling means has inputs for receiving users profile data from all kinds of usage traces through observation and tracking of actual service consumption by the user, the learning/updating means receiving for learning/updating the profile data as a function of users' profile data provided by the explicit profiling means and the implicit profiling means.

According to an embodiment of the invention, the profiling engine offers intelligent query interfaces providing applications with generic access to all profile data thus allowing personalization of various applications.

According to an embodiment of the invention, these query interfaces include different distances to compute the similarity between a content and a unique user, e.g. to recommend specific content, or to compute the similarity between users, e.g. to define communities of interest.

According to an embodiment of the invention, the profile engine comprises a homogeneous raw data generation module to transform all platform-specific data into a common, pre-defined format processable by the profile engine.

Other objects and further features of the present invention will be apparent from the following detailed description when read in conjunction with the accompanying drawings :

A description will hereinafter be given of embodiments of the present invention, by referring to the drawings. It is first provided a functional description of the profiling engine, next the profiling technology in terms of the data model, the algorithmic approach, and the implementation. Finally, it is described the approach to privacy management which is as well an important element in profiling and personalization systems.

In this section, it is first presented the overall multi-platform and application-agnostic approach proposed for the profiling engine, as illustrated in Figure 1. Then, the different functional components of the engine are introduced.

The main objective of the profiling engine 1 is to ensure the automatic learning of each user's profile to develop an accurate estimation of his interest domains, service/content consumption habits, and purchasing behavior. For this purpose, large operators have the opportunity to collect and consolidate service usage and content consumption traces over their service delivery platforms 2. In addition to Web-based usage traces, they can largely exploit television 3 and mobile video 4 consumption data as well as the usage of IP communication services. The profiling engine 1 aggregates the usage traces coming from all these different platforms 2 and builds an end user profile based on a common model.

At the same time, this consolidated profile can support personalization for numerous applications 5 : content recommendation, personalized content search 6, or social networking applications 7. In addition, the engine can be put to immediate use for targeted ads 8 -a new revenue stream the service providers intend to tap into-in compensation for declining voice revenues.

Figure 1 illustrates this multi-platform and multi-application paradigm. Having such a profiling engine is critical for service providers 1) to derive benefit from all these sources of information, and 2) by valorizing the so- obtained profile to offer the end user a unified and consistent experience across personalized services.

The user profile is obtained by aggregating the usage traces available on different SDPs and by combining them with content descriptions (metadata) available in content management systems (CMS). The principle of the profiling engine according to an embodiment is depicted on Figure 2. When a user consumes a service, for example, views video-on-demand (VoD) on a mobile phone 9, some consumption traces are generated on the SDP or on the terminal. In the proposed approach, all these traces are injected into the profiling engine which dynamically updates the user profile in order to closely follow its real-life evolution.

First of all, the profiling engine relies on a user model 10 which is shaped according to the structure and the semantics of the content metadata available in the CMS 11. Nevertheless, the profiling engine 1 and its underlying update algorithms are not designed specifically for a given user profile and content metadata structure; they can be automatically applicable to a new structure and semantics.

Then the profiling engine 1 uses profiling algorithms running module 11 to learn and update the profile data defined in the user model 10. It is distinguished two mechanisms: explicit profiling and implicit profiling. In the explicit profiling process, the user declares his topics of interest, e.g., via a Web portal. At the initialization stage, the absence of active indication of interest in a topic implies a lack of user interest. The implicit profiling comprising learning and updating the profile from all the usage traces. It relies on content metadata describing the semantics of consumed contents and services. The combination of explicit and implicit profiling procedures allows the operator to build and maintain the most relevant and up-to-date profile. An intelligent query interface 12 is another important component of the profiling engine 1. It provides the personalized applications with generic access to all profile data comprised in user profile database 17. Various request types are enabled starting from a simple database (DB) request, e.g., "find all users with a high interest in tennis, and a heavy video consumption profile" to more sophisticated query types like "evaluate the interest of a given user for a given content," or "find users with a profile 'similar' to a given one".

Finally, the profiling engine addresses privacy issues, via a privacy protection module 13, in order to conform with legal requirements as well as to ensure the user's consent for being profiled. Our approach relies on privacy policies which allow, on the one hand, controls around access to personal profile data, and on the other hand, mechanisms to configure the level of intrusiveness of the profiling method itself.

This section details the technology realized in the profiling engine according to one embodiment. It covers the data model 10 (to describe the semantic concepts 14 and the measurable quantities 15), the profiling method (to incrementally build the profile), the profile data access (to personalize a wide range of applications), and finally provides an insight on privacy which is an important element of such profiling technology.

It is thus proposed a comprehensive approach comprising:

At a starting point, it is defined semantics concepts 14. Semantic concepts constitute a core element of the profiling engine's data model. They are used to represent the semantic characteristics of user profiles and contents. For example, the "action" film genre is a semantic concept, it can characterize a given content, and at the same time, a given user's interest in this kind of content in general. Of course, both content and user profile are generally characterized by several semantic concepts. Semantic concepts usually have relationships between each other, expressing notions such as composition, inheritance, or dependency. For example, the semantic concepts "action," "adventure," and "comedy" which are usually used to describe a movie, are related to the respective semantic concept "movie" by an inheritance relation, or a so-called "is-a" relation. Semantic concepts can therefore be organized in a structure reflecting their relationships, e.g., a taxonomy or an ontology. In some cases, they can simply form a vocabulary of flat keywords without any structure.

In order to represent a user profile, it is associated one numerical value in the interval of [0,1] to each semantic concept; higher values indicate higher user interest in the respective semantic concepts. Figure 3 illustrates how the same principle can be used to characterize the contents and the consumption events. The diagram uses the following three measurable quantities which allow to describe the contents, the consumption events, and the user profiles, respectively:

Note that for a given user profile (or a content or consumption event), one needs to store only non-zero values so that the problem of memory explosion is avoided when the potential number of semantic concepts is large.

The global user profiling process is described in Figure 4 . At first, it uses platform-specific usage traces 18 (SDP raw data) like those stored in log files, call data records (CDRs), or cookies, and contains the description of consumption data 19 like watched/ranked videos, time watched, price paid, accepted offers, or search strings. In order to facilitate profiling in a multi-SDP environment, this platform-specific data must first be transformed, thanks to a HRD generation module 20, into a common format called homogeneous raw data (HRD). It is called homogenization of raw data the process of transforming all the heterogeneous gathered data into a pre-defined format, comprehensible by the profiling engine.

The next stage consists of characterizing each consumption event in terms of values associated with semantic concepts. Independently from the source of that observation, each QoC (Quantity of Consumption) value provides a normalized measure, QoC∈[0,1], of the observed user interest for a given semantic concept. This measure is based on the following assumptions:

The QoC values are computed by modulating the QoA (Quantity of Affiliation) values of the consumed content by the numerical parameters held in the HRD. This function can be modeled as: $${\mathit{QoC}}_n^i=\mathit{QoA\_Modulation}\left({\mathit{QoA}}_n^i{\mathit{HRD}}_n\right)$$

where the index i refers to the relevant semantic concept and n represents consumption event ordering. In order to achieve a normalization, the modulation function may be domain-dependent. Indeed, each SDP offers a specific service and the usage of different services cannot be put on the same scale. For example, listening to two hours of music should reflect a higher interest in that type of music than watching a 2 hour movie for the type of movie watched, because an average music album is less than 60 minutes long, whereas an average movie is 1 hour and 40 minutes long. In a more general case, within a given SDP, the normalization function can also vary from one application to another. Some simple examples of QoC computation for VoD services are given below.

This stage describes how the user profile is incrementally updated. In the present approach, two complementary update functions co-exist: 1) consumption event-based QoI learning, and 2) time-based QoI decay.

The consumption event-based QoI learning function refreshes the QoI data on user interest by combining their previously known values with a newly-observed interest manifestation (defined by the QoC).

This function can be modeled as: $${\mathit{QoI}}_{n+1}^i=\mathit{QoI\_Learning}\left({\mathit{QoI}}_n^i{\mathit{QoC}}_n^i\right)$$

It is considered a particular family of functions where the new QoI is obtained by a weighted addition of the newly observed consumption, QoC, with the old QoI: $${\mathit{QoI}}_{n+1}^i={\mathit{QoI}}_n^i+W\left({\mathit{QoI}}_n^i\right)*{\mathit{QoC}}_n^i$$

The weight $W\left({\mathit{QoI}}_n^i\right)$

is a function of the current QoI; it defines how much the new observation is influencing the profile evolution. Such a variable weight allows us to obtain a "learning curve" behavior where secondary interests build relatively slowly (because of a lower weighted ranking) and primary interests are saturated by the upper limit of one (again, because of a diminishing variable weight). Here, the term "learning curve" makes reference to a relationship between the duration of a student's learning period and the knowledge or experience gained. In fact, the QoI evolution for a stable consumption pattern represents a sigmoid form, as shown in Figure 7 .

In addition to the previous phases describing profiling as a consumption event-driven iterative process, a time-based decay function is introduced to account for the aging of profile data: $${\mathit{QoI}}_{k+1}^i=\mathit{QoI\_Decay}\left({\mathit{QoI}}_k^i{P}_k^i\right)$$

This function is called with a given periodicity indexed by k. This periodicity should be significantly larger than the average time interval between consumption events, for example it can be on a monthly or quarterly basis. In order to decide how to decrease a given QoI, this function can take into account parameters $\left(P_k^i\right)$

like frequency or recency of consumption events on that semantic concept. For example, depending on the consumption frequency, the QoI can be diminished linearly, exponentially, or without decay for a fixed period of non-consumption followed by some decay curve.

When the raw data contains some information on the context in which the consumption event occurred, it can be used to obtain a context-aware profile. As illustrated in Figure 5 , this means that some QoI values will be contextualized. For example, a user's interest in "news" may be higher in the morning than in the evening, when a higher interest in "movies" can be observed. Here, the term of context can be applied in its broadest sense including geographical localization and presence (at home, at work, or on vacation), social (with friends, with family, or alone), or temporal (morning or evening) interpretations.

The intelligent profile query interface 12 enables personalized applications by providing Web service access to user profiles, and by providing some profile exploitation tools. In order to deal with many different applications such as targeted ads, content recommendation, or community-based applications, this interface should provide generic reusable features. In fact, three main types of requests are possible:

This query interface is critical as it will ensure the independence of the profiling module with respect to different personalized applications. Indeed, it is decoupled the similarity measuring from the personalization techniques deployed in the application logic. So, it is enabled both well known types of personalization techniques: the content-based (CB) approach and the collaborative filtering (CF) approach. The CB algorithms look at the "similarity" between the user (profile) and the item (metadata) to recommend, while the CF algorithms recommend the item if it has been appreciated by "similar" users.

The privacy of the end user is an important element to take into account when designing a profiling system. First, the service provider must ensure the compliance with the legal privacy rules in each country where the solution is deployed. In addition, user acceptance is a major issue, as user profiling can easily be perceived by end users as a threat and an intrusion into their private life. Furthermore, the user should be provided with a comprehensive interface for setting his privacy options.

The present approach encompasses two aspects of user privacy protection: access control of personal profile data and configuration of the intrusiveness level in the profiling process. This is achieved by using high-level privacy policy rules. Some of these rules are introduced by the service provider in order to define its global profiling and personalization policy in conformance with the existing legislation. For example, among the requirements put forward by the European Union, there are three core principles:

The second aspect of the privacy policy is introduced by each user in order to tune his personal privacy preferences. For example, a user can specify the types of services (IPTV, mobile video, or Web browsing) and traces (watching, interactivity, or zapping) that can be used for his profiling. With respect to access to his profile data, a user can define variable restrictions depending on the type of personalized application. For example, if a user does not want to receive targeted ads based on his video interests (action, sport, or adult), this part of the profile must be hidden to the targeted ad selector. The restrictions can also impact the granularity of the information made available in a given domain.

As a part of the user interface for privacy management, it is also included an explicit profiling feature, the user's read/write access to his own profile data. It not only allows a user to initialize the system, but also to update and rectify the learned profile. Once updated, the profiling process continues with the new current profile modified explicitly by the user in the same way as in the initialization phase.

A prototype of this process has been implemented in the scope of a video-on-demand service for converging fixed (IPTV) and mobile (mobile video) content delivery platforms where the customers can use a diversity of terminals: TV/set-top box, mobile phone, and laptop. In the prototype, the profiling has been done with two key personalized applications, namely targeted advertisement and content recommender. The implementation combines ontology Web language (OWL) for the description of the user/metadata model and Web services technology, notably for intelligent query interface. The use of open technologies allows rapid integration of new evolutions of the profile model as well as support for a wide range of personalized applications.

The prototype was also used to carry out different simulations and performance tests. For instance, Figure 7 illustrates a QoI evolution in a user profile showing stable consumption of content with a given semantic concept, and an average of three weekly consumption events over approximately 140 days. In this case, the decay function is calculated on the basis of the number of consumptions in a monthly time window (frequency). The higher the frequency, the lower the decay applied to QoI.