DE202012013130U1 - Predictive behavioral targeting - Google Patents

Abstract

Apparatus configured to carry out the following method or comprising respective means for carrying out the method steps of the following method: obtaining information about an access behavior of a current user who is accessing content; Predictions, with a prediction module of the device and in response to receiving the information about the access behavior of the current user, of at least one characteristic of the current user based at least on a model for the characteristic and the obtained information about the access behavior of the current user, wherein the model is based on at least one set of parameters that is determined by a training module based at least on information from a plurality of training sets and is provided to the prediction module from time to time for update purposes, wherein each training set of the training sets is for a respective training user from several training users , Information about an access behavior of this respective training user and information about the characteristics of this respective training user, and - providing information about the predicted character istik of the current user, so that information can be aimed at the current user at least on the basis of the information about the predicted characteristic.

Description

AREA

Embodiments of the present invention relate to the field of predictive behavioral targeting.

BACKGROUND

Many owners or publishers of websites want to present information or advertisements to their users visiting the website, which information should match the characteristics of the user in order to increase the value of the advertisement. This process is commonly referred to as "targeting" users. To this end, web site and / or specialized service providers seek to obtain as much personal information as possible about users, and this information can be collected in a number of ways. For example, Behavioral Targeting analyzes, among other things, information about users' browsing or "clicking" behavior. When such information type is merged with demographic data and / or data obtained from inquiries and surveys and when information about users is provided with information about other users with similar characteristics, e.g. Through collaborative filtering, then it becomes possible to predict user characteristics based on their browsing behavior. This process is known as Predictive Behavioral Targeting.

The German patent specification DE 10 2004 047 815 B4 describes a method for generating and routing advertisements, combining a user's browsing behavior with data about the interests of the user and personal information about the user. Control data for selecting advertisements based on this information is stored in user profiles for each user, and these user profiles are then compared and their completeness and similarities are analyzed.

SUMMARY OF SOME EMBODIMENTS OF THE INVENTION

It is essential for predictive behavioral targeting that characteristics of a user to be targeted be predicted as quickly as possible, because based on the predicted characteristics, an ad server will usually provide a further selection of information about the adware User to target, so there is a further delay. For example, characteristics should be predictable within about a few seconds or even milliseconds.

One particularly straightforward approach to predicting a current user's characteristics is to compare his browsing behavior to a database that stores the browsing behavior and associated characteristics of a plurality of users by identifying the user with the best matching browsing behavior and its associated characteristics will be used as a prediction. Such an approach is commonly known as the k-nearest-neighbor approach. Since the entire database must be searched for each prediction, the prediction time is quite large. Furthermore, the entire database must be accessible if a prediction is to be made.

Furthermore, it is generally desirable to be able to provide a quick prediction already when there is little information about the browsing behavior of a current user, e.g. For example, in the case of a first-user, who is only aware that he has accessed the current website. In the in DE 10 2004 047 815 B4 described profiled system would first have to set up a profile and enriched with profile information from profiles of other users, who are considered similar to the current user, which requires a considerable amount of time.

It is therefore an object of the invention to provide an improved solution for predictive behavioral targeting.

• – Erhalten von Informationen über ein Zugriffsverhalten eines aktuellen Benutzers, der auf Inhalt zugreift;
• – Vorhersagen, mit einem Vorhersagemodul der Vorrichtung und als Reaktion auf das Erhalten der Informationen über das Zugriffsverhalten des aktuellen Benutzers, zumindest einer Charakteristik des aktuellen Benutzers zumindest auf der Basis eines Modells für die Charakteristik und auf der Basis der erhaltenen Informationen über das Zugriffsverhalten des aktuellen Benutzers, wobei das Modell zumindest auf einem Satz von Parametern basiert, der von einem Trainingsmodul zumindest auf Basis einer Vielzahl von Trainingssätzen ermittelt zumindest und dem Vorhersagemodul von Zeit zu Zeit für Aktualisierungszwecke bereitgestellt wird, wobei jeder Trainingssatz der Trainingssätze für einen jeweiligen Trainingsbenutzer von mehreren Trainingsbenutzern Informationen über ein Zugriffsverhalten dieses jeweiligen Trainingsbenutzers und Informationen über die Charakteristik dieses jeweiligen Trainingsbenutzers umfasst, und
• – Bereitstellen von Informationen über die vorhergesagte Charakteristik des aktuellen Benutzers, so dass Informationen auf den aktuellen Benutzer zumindest auf der Basis der Informationen über die vorhergesagte Charakteristik ausgerichtet werden können.

According to a first aspect of the invention, there is disclosed a method performed by at least one device, the method comprising:

• Obtaining information about an access behavior of a current user accessing content;
• Predicting, with a predictive module of the device and in response to obtaining the information about the access behavior of the current user, at least one characteristic of the current user based at least on a model for the characteristic and based on the obtained information about the access behavior of the current one A user, wherein the model is based at least on a set of parameters provided by a training module based at least on a plurality of training sets at least and the predictive module from time to time for update purposes, each training set of training sets for a respective training user of a plurality of training users Includes information about an access behavior of that particular training user and information about the characteristics of that particular training user, and
• Providing information about the predicted characteristic of the current user so that information can be aligned with the current user based at least on the predicted characteristic information.

According to a second aspect of the invention, there is disclosed a computer program comprising program code for carrying out the method according to the first aspect of the invention or any of its subsequently described embodiments when the computer program is executed on a processor. The computer program can, for example, via a network such. B. the Internet be distributed. For example, the computer program may at least partially represent software and / or firmware of a processor. The computer program can be stored and / or coded, for example, in a computer-readable medium. The computer-readable medium may be configured, for example, as an electrical, magnetic, electromagnetic, optical, or other storage medium and may be either a removable medium or a medium permanently installed in a device or terminal. Non-limiting examples of such a computer readable medium are Random Access Memory (RAM) or Read Only Memory (ROM). The computer-readable medium may, for example, a tangible medium such. B. be a tangible storage medium. A computer-readable medium is understood to be a medium that can be used by a computer such as a computer. B. a processor can be read.

According to a third aspect of the invention there is disclosed an apparatus configured to carry out the method according to the first aspect of the invention or any of its embodiments described below, or respective means for carrying out the method steps of the method according to the first aspect of the invention or any of its subsequently described Embodiments comprises.

According to a fourth aspect of the invention, there is disclosed a device comprising at least one processor and at least one computer program code memory, wherein the at least one memory and the computer program code are configured to, with the at least one processor, cause the device to at least Process according to the first aspect of the invention or to carry out any of its subsequent described embodiments.

In particular, the device according to the third and fourth aspects of the invention may be the device which carries out the method according to the first aspect of the invention and comprises the prediction module.

According to a fifth aspect of the invention, there is disclosed a system comprising a device according to the third or fourth aspect of the invention and the training module.

There will now be described exemplary features and exemplary embodiments that are equally applicable to the first, second, third, fourth, and fifth aspects of the invention. All exemplary features and exemplary embodiments are to be understood as disclosed in all possible combinations with each other.

Information about an access behavior of a current user accessing content is obtained. For example, the content may be content of any type of IT that a user, particularly a human user, may access, e.g. B. content of the Internet or any other network such. A social network. The content may, for example, electronic content such. B. Content that can be viewed or heard. In particular, a web page of the Internet is to be understood as an example of content that a user can access, e.g. B. via a web browser such. Internet Explorer, Mozilla Firefox or Apple's Safari browser. For example, a webpage may be part of a set of one or more related webpages that make up a web site. The term "current user" is used in this specification, a user who accesses content and for which targeting is to be performed, e.g. During and / or after accessing the content. Moreover, the term "current" is not to be construed in a limiting manner. For example, the access behavior of the current user may refer to accessing content by the current user. For example, if the content includes one or more web pages, then the access behavior may relate to, for example, accessing the web pages by the current user and may then be understood, for example, as the click behavior of the current user. The access behavior information may then include, for example, information about the types (eg, in terms of content categories) of Web pages accessed by the current user and / or frequencies that the current user accesses. Information about the access behavior of the current user can be collected, for example, by means of a cookie, in particular a third-party cookie. The information about the access behavior of the current user can already take into account the content that the current user is currently accessing or can not yet take into account. The information about the access behavior of the current user can be obtained, for example, from a cookie processing unit that can form part of the device or a unit that is separate with respect to the device.

In response to obtaining the information about the access behavior of the current user, at least one characteristic of the current user is predicted by a predictive module of the device at least (or only) based on a model of the characteristic and the obtained information about the access behavior of the current user. The characteristic may be, for example, a demographic characteristic of the user, such. Gender, age, age group, income or income group, to name just a few non-limiting examples. Likewise, the characteristic may relate to an interest of the user, such. An interest in a specific topic or product.

The model is based on at least one set of parameters that is determined by a training module and provided to the predictive module from time to time for update purposes. Therein, the training module may be part of the device or part of a unit that is separate from the device. The prediction module and the training module are to be understood as functionally separate and / or decoupled from each other. The prediction module applies the model to the obtained information about the access behavior of the current user to predict a characteristic of the current user. The training module provides, from time to time (eg, at regular or irregular intervals) the set of parameters upon which the model used by the prediction module is based. The prediction module can thus operate independently of the training module based on the set of parameters it currently has. If there is an update to this set of parameters, then that update is provided to the predictor module by the training module. In particular, however, it is possible to make the determination of the set of parameters by the training module and / or the provision of the update of the set of parameters to the prediction module not in response to receiving the access behavior of the current user. In contrast, determining and / or providing an update to the set of parameters may be triggered, for example, by other events, as discussed below. The prediction of the characteristic of the current user on the one hand and the determination and / or provision of updates for the parameter set on the other hand can thus be regarded as asynchronous with respect to each other. The connection between both processes may be considered to be formed by the updates provided to the prediction module by the training module. The functional separation of prediction and training allows a large reduction in the prediction time (ie, the time needed to predict a characteristic of the current user based on the obtained information about its access behavior). The prediction is therefore no longer slowed down by training (as would be the case with lazy learning). Further, the use of a parameter-based prediction model (rather than a vast database of respective profiles and / or sets of access behavior and associated characteristics of users that contain information for millions of users) also allows rapid prediction of characteristics of a current user even then if there is little information about the access behavior of the current user (for example, only information about what the current user is accessing). The functional separation of prediction and training also allows the prediction module to function as an online module (eg, as part of a web server) that has in-depth requirements for predicting the characteristics of the current user based on access behavior information handle the current user (eg together) with the request, and immediately (eg with a delay of only milliseconds, eg with an average delay of less than 10, 20, 50, 100) or 500 milliseconds) provides the predicted characteristic. The online module can thus work in real time. In contrast, the training module can be operated as an offline module, for example, no requirements in real time served. For example, a new set of parameters may be determined and / or provided to the prediction module by the training module when a sufficient number of new and / or updated training sets based on which the set of parameters is determined are obtained. Alternatively or additionally, the training module may perform the determination and / or provision of the set of parameters based on a regular schedule. In either case, a frequency of predicting the characteristic of a current user (as previously performed by the prediction module) may be greater than a frequency of determining and / or providing the set of parameters (as performed by the training module).

The training module determines the set of parameters based at least on information from a plurality of training sets, each training set of training sentences for a respective training user of multiple training users containing information about an access behavior of that particular training user and information about the characteristic (and possibly other characteristics) of that training user respective training user. Therein, the term "training user" defines a user for whom both information about their access behavior and information about the characteristic are available (and thus form a training set) and the training module for determining an update (or initial version) for the sentence of Parameters is used. This term is in no way to be considered as further limiting. A training user may possibly be different from a regular or normal user in that he or she gives information about the characteristic voluntarily (eg, by participating in a survey), or by providing information about the characteristic in a different way to that user available (for example, from a database, such as a Nielsen database). It should be noted that the current user may also be a training user, e.g. When the current user provides information about the characteristic. Because the training sets include information about both the access behavior and the characteristics for multiple training users, the training module may derive therefrom the set of parameters that it allows, including the characteristics of the current user based on information about the current user's access behavior predict the model based on this set of parameters. The set of parameters may not include training sets. Instead, this can be understood as a (mathematical) generalization of information about some or all training sets. In particular, at least one parameter of the set of parameters (or even some or all parameters of the set of parameters) depends on information from multiple training sets (or even all training sets) of the training sets. Thus, even if the prediction module and the training module both worked in response to receiving information about a current user's access behavior (ie, would not work asynchronously), the entire system would still be completely different from a lazy learning system because there is always an intermediate step of determining a set of parameters based on the training sets, and this set of parameters is then applied in a further step through the model to the information about the access behavior of the current user.

The predicted characteristic of the current user is then provided, for example, to an advertising server (either directly or indirectly) or other entity so that information can be aligned with the current user based at least on the predicted characteristic information. For example, if the characteristic is an age or sex or an interest of the current user, then this information may be valuable for an appropriate selection of information (such as advertisements or recommendations) to be provided to the current user. This targeting can, for. For example, while the current user is accessing the content, or thereafter.

According to an embodiment of all aspects of the invention, the characteristic of the current user is by classifying a vector representing the information about the access behavior of the current user into one of at least two classes, and the model is based on a support vector machine comprising the Classification allows. A support vector engine is a concept in statistics and computer science for determining a hyperplane or a set of hyperplanes in a multi-dimensional space to separate vectors contained in the multidimensional space associated with two or more classes. A good separation is achieved, for example, when the hyperplane has the greatest possible distance from the next vector of each class. An example of a representation of information about the current user's access behavior as a vector is, for example, the five-dimensional vector with components (0, 2, 3, 0, 6) indicating that the current user has 0 webpages about sports topics in the available observation period, 2 websites about business topics, 3 websites about art topics, 0 websites about vehicle topics and 6 websites about fashion topics. Further, if the characteristic of the current user is the gender of the current user associated with the class "male" or "female", such a vector may for example be classified as "female". The characteristic may of course alternatively be with a classification of a vector in more than two Classes be predictable. For example, if the characteristic is an age group of the current user, then it may be associated with a classification in the following age groups: 0-10, 11-20, 21-30, 31-45, 46-60 and 61-90 years ,

For example, in this embodiment, the set of parameters may define a support vector engine decision function that is applicable to perform the classification on the vector that represents the current user's access behavior information. For example, the decision function may refer to the one or more hyperplanes identified by the support vector engine. For example, if the characteristic of the current user is predictable by classifying a vector representing information about the access behavior of the current user into one of two classes, then the set of parameters may include, for example, components of a normal vector and a bias including a Defined hyperplane of the support vector machine. A prediction of the characteristics of the current user based on a decision function of a support vector machine can be computationally extremely cost-effective. For example, it may be sufficient to multiply each component (or just each non-zero component) of the vector with an associated parameter from the set of parameters, sum up the result, add one of the parameters from the set of parameters and the sign of the To determine the overall result. This low complexity of the decision function application is particularly advantageous as it must be applied in the prediction module where a real-time operation with the least possible delays is desired.

In this embodiment, the information about the access behavior of the respective training user contained in the training set for each training set of the training sets may represent a training vector associated with a class derivable from the information about the characteristic of the respective training user contained in the training set, and Set of parameters may be determined by the training module to reflect a separation of training vectors of different classes according to a predefined optimization criterion. For example, if the characteristic is a gender, then each training set may include, for example, the training vector (representing information about the access behavior of the training user) and information about the gender of the training user. The set of parameters is then determined to obtain a separation between the training vectors of male training users and the training vectors of female training users according to the predefined optimization criterion. The predefined optimization criterion may be, for example, maximizing the respective distance of the training vector closest to the hyperplane in both the male and the female class of training vectors (collectively called optimal hyperplane). Just as well, other optimization criteria can be applied. For example, an optimization criterion can be applied that leads to a soft-margin hyperplane.

According to one embodiment of all aspects of the invention, the current user's access behavior information for each content category of a set of one or more content categories includes information about how often accesses to content associated with the content category were observed by the current user in a time period , and the information about the access behavior of a respective training user included in the training sentences, for each content category of a set of one or more content categories, contains information about how often access to content belonging to the content category has been observed by the training user in a time period , For example, the access behavior information may take the form of a content category histogram for a user. Non-limiting examples of content categories are sports, business, economics, art, fashion, cars, etc. The time period may be, for example, a limited or unlimited period of time, e.g. For example, the time period since the beginning of the observation. For example, the set of one or more content categories used for the current user and the set of one or more content categories used for the training users may be the same, but may be at least partially different. The information about how many views of content that belongs to the content category has been observed can only belong to a limited number of content sources. For example, if the access to content is in the form of a visit to a web page and the web site visit is monitored based on a third party cookie, then only visits to web pages to which the cookie of the third party relates (eg the website contains an information request (eg a web bug) from the website of the third party who has set the third party's cookie).

In this embodiment, a decision as to whether content belongs to one of the content categories is made by a decision unit based on a received content classifier and / or on the basis of an analysis of at least part of the content. The decision unit may for example be part of the device or may be part of another device that is separate from the device. The decision unit may, for example, be part of a cookie processing unit, which in turn may be part of or separate from the device. The information about the access behavior of the current user can be obtained, for example, by the device or the prediction unit from the decision unit or the cookie processing unit. For example, if a current user / training user (hereinafter referred to as either the current user or the training user) accesses a web page (as an example of content) that has a request (eg, a web bug or a tracking Pixel or script) of data from a third party web site (or associated therewith, eg in the form of a script of a tracking pixel associated with some or all web pages of a web site), then the web server of the web site of the third party (called the third party's web server) - upon initial communication with the current user / training user's web browser - require the current user / training user's web browser to set a cookie (eg, a small text file) and content of the cookie on any request for content from the third party's website. The request for data (eg the web bug) from the third party web site may be provided with a content classifier or with an identifier (eg the uniform resource locator (URL)) of the web page in which the request from Data or with which the request for data is associated. Thus, when the request for data from the third party's web server is launched by the current user / training user's web browser, the third party's web server becomes aware of the content of the web page in which the request for data has been embedded (or associated with the request for data is) informed by the content classifier, or may obtain such information by analyzing the (eg retrieved) content of that web page based on the identifier (if no dedicated identifier is included in the request for data, then the in the request to the web server of the third party contained HTML referer are used). Thus, based on the knowledge of the content of the web page accessed by the current user / training user, a content category histogram may be started or updated or written back to the cookie of the current user / training user web browser. This content category histogram, along with the content classifier or identifier of the web page, is then provided to the web server of the third party each time a web page containing (or associated with) a request for data from the web site of the third party is owned by the current user / Training user is accessed.

In this embodiment, the prediction of the current user's characteristic may be based on a representation of the current user's access behavior information, in which the information about how many times the current user has accessed content matches the content categories of the set of one or more multiple sets of categories, has been normalized, and wherein the set of parameters may be determined based on training sets in which the respective information on how many times the respective training user has accessed content belongs to content categories of the set of one or more content categories , were normalized. For example, normalizing the information about how often the current user / training user has accessed content belonging to content categories of the set of one or more content categories may take into account that access to generally frequently visited content is less relevant than access to generally less haunted content. Normalizing the information about how often the current user accessed content belonging to content categories of the set of one or more content categories may be e.g. B. be part of the prediction. For example, the information about how many times the current user accessed content associated with content categories of the set of one or more content categories may be stored in a non-normalized form (eg, in a cookie as described below) and can only be normalized for the purpose of predicting the characteristics of the current user. Normalizing may include, for example, the number of clicks of a current user / training user per content category (eg, in a time period), a total likelihood of a content category being clicked, and a total number of clicks of a current user / training user (e.g. the period of time).

According to one embodiment of all aspects of the invention, at least part of the current user's access behavior information is stored in a cookie on a current user's web browser and provided by the web browser to, for example, a cookie processing system, part of or disconnected from the device is. For example, the portion of information about the access behavior of the current user may belong to content that the current user accessed prior to accessing the content (ie, the current content). For example, this portion of the current user's access behavior information may be augmented with information about the current content (eg, by a cookie processing unit or a decision unit that may, for example, form part of the cookie processing unit). This enriched information can then for example be stored back in the cookie on the web browser. The part of the information about that For example, access behavior of the current user stored in the cookie may be provided (eg, in an HTTP GET request) because the cookie is associated with a domain / path combination (a combination that specifies a domain and a path, where the path as well as just a "/" may be) that matches a domain / path combination of a request (from data such as an image tag such as an HTTP IMG or a script) that is in the Contains or is associated with content that the current user is currently accessing. For example, the cookie may be a third party cookie. For example, the domain / path combination may specify a web page (or an entire web site containing multiple web pages) associated with the device, e.g. Because the device is a web server (or part of it) for this website (or website). The initial setting of the cookie (and subsequent updating of the cookie) may, for example, be triggered by the device (eg, by a cookie processing unit thereof). The content of the cookie may be encrypted and / or compressed, for example.

In this embodiment, at least some of the information about the access behavior of a training user of the training user may be stored in a cookie on a training user's web browser, where at least one domain (or domain / path combination) for which the cookie was set, matches a domain (or domain / path combination) for which the cookie stored on the web browser of the current user has been set. Thus, the cookie set on the current user's web browser and the cookie set on the training user's web browser may both be cookies set by the same entity, for example, a web server associated with the domain (or domain / path combination) , For example, the cookie may be a third party cookie. The device may be, for example, the web server or a part thereof. The device may then, for example, direct the information about the access behavior of the training user stored in the cookie on the training user's web server to a database (which may be part of or separate from the device), the training module having access to the database. If the cookies set on the web browser of the current user or training user are set by the same entity, then that entity can control information collection about the access behavior of the current user and the training user, both of which are needed to predict the current user's characteristics.

According to one embodiment of all aspects of the invention, the information about the characteristic of a training user contained in a training set of training sets is obtained from a survey performed with the training user and associated with the training user's access behavior information to obtain the training set. For example, the survey can be an online survey. For example, the survey may be based on a set of questions to be answered by the training user. For example, the characteristic may be obtained in response to a question from the survey, or may be derivable from an analysis of respective responses to one or more questions of the survey. For example, participation in the survey may be offered to a training user when a training user accesses content, e.g. B. by an advertising server. For example, the survey may be handled by the ad server like other advertisements managed by the ad server. For example, the presentation of the survey may be controlled by the ad server in such a way as to ensure a minimum daily number of presentations of the survey. Optionally, the survey's presentation may be tied to a frequency limit to avoid having the survey too often presented to the same user. For example, a representation of responses given by the training user in response to the survey may be provided to the device (or other entity separate from the device) (eg, via an HTTP request) that provides this representation with information about the device Access behavior of the training user combined. For example, this combination may be based on an identifier (eg, an anonymous or anonymized identifier) of the training user and / or the web browser and / or the electronic device running the training user's web browser. For example, a training user may be identified by an identifier (eg, a unique user identifier, UUID) stored in a training user's cookie, which is read both when the training user participates in the survey and then when gathering information about the access behavior of the training user. This may be done by associating both the survey and the web pages accessed by the training user with a web bug or script to track the training user, for example, the web bug or script causing the same web server to participate of the training user to the survey and about accesses of the training user to the website. Alternatively, an electronic fingerprint of the device used by the training user to access web pages may be used as an identifier, for example. The identifier may include, for example, the representation of responses given by the training user in the survey and the access behavior information of the user Training user are associated. Additionally or alternatively, the information about the characteristic of a training user included in a training set of the training sentences may be derived from a data source such as a training sentence. A Nielsen database or a database that stores (eg demographic) information about the training users provided by the training users, e.g. For example, if they register for a service or conclude a contract.

According to one embodiment of all aspects of the invention, an actual address of the current user is anonymized by an anonymization unit, and the information about the access behavior of the current user is obtained with the anonymized address and not the real address. The address may be, for example, an IP (Internet Protocol) address of the user. For example, the anonymizer unit may be switched between a web browser of the current user and the device (which may be, for example, a web browser or a portion thereof). In this way, the actual address of the current user of the device is unknown, so that the privacy of the current user is protected. Likewise, the real address of the training user can be anonymized by the anonymization, z. B. when the information about the access behavior and / or the information about the characteristics of the training user are provided. The anonymization unit may be, for example, a proxy server such as a reverse proxy server.

In this embodiment, the information about the predicted characteristic of the current user may be provided, for example, with the anonymized address and not the real address.

According to one embodiment of all aspects of the invention, the current user's predicted characteristic information is formatted to conform to a format required by an advertising server, which is to direct the information to the current user based at least on the information performs the predicted characteristic. For example, different advertising servers may require different formats for the predicted characteristic. An example of a format is a key-value (such as gender - male, or age - 20). For example, different advertising servers may use different key-value pairs so that appropriate formatting may be unavoidable. The formatting can be done, for example, by a formatting unit, which may be part of the device or another unit that is separate from the device.

According to one embodiment of all aspects of the invention, the current user's predicted characteristic information is provided to an advertising server that performs the targeting of the information for the current user based at least on the predicted characteristic information. The information about the predicted characteristic can be provided to the advertising server, for example, directly or indirectly. For example, the predicted characteristic information may be provided to a current user's web browser and then provided by the web browser to the ad server, e.g. By means of a location redirect enriched with information about the predicted characteristic. For example, information about the current user's access behavior may first be provided to the device (eg, a web server) that predicts the characteristics of the current user in response to the obtained information about the access behavior of the current user. For example, the information about the access behavior of the current user may be at least partially stored in a cookie on the current user's web browser (another part of this information may be, for example, a content classifier provided with the content of the cookie, as explained above and is then provided (eg, via HTTP GET) when the current user accesses content associated with the cookie (ie, provided on a cookie-associated domain / path combination). The response (eg via an HTTP response) to the current user's web browser may then include, for example, an instruction to store updated information about the access behavior of the current user into the cookie on the web browser and a location redirect request (e.g. B. via HTTP LOCATION), which specifies the address of the advertising server and also contains information about the predicted characteristic. In this way, information about the access behavior of the current user may be collected and information about the predicted characteristic provided to the ad server, the only intervention being in or the only change of content (eg, the publisher's web page) to which the current user which is the addition of a request for data (from a device-associated domain) to the content.

In this embodiment, the ad server may trigger the storage of the predicted characteristic information in a cookie on a current user's web browser. The information about the predicted profile is then available to the ad server on each web page that causes the web browser of the current user to provide the content of the ad server cookie to the ad server, e.g. Because a request for data (eg, a web bug) associated with the ad server or a script for communicating with the ad server is included in or associated with the web page. The advertisement server may thus perform predictive behavioral targeting (based on a past prediction of the current user's characteristic provided by the device) even if a web page accessed by the current user does not contain a request for data associated with the device (which would trigger a new prediction), but only includes a request for data associated with the ad server (the predicted characteristic stored in the cookie cookie, of course, becomes obsolete after some time, unless it is refreshed by the device).

According to one embodiment of all aspects of the invention, an updated version of the set of parameters is determined by the training module and provided to the prediction module each time a number of new and / or at least partially updated training sets have reached a predefined value. For example, the training sets may be stored in one or more databases. Therein, for example, the respective information about the access behavior of the training users can be stored in a first part of the database (or in a first database) and the respective information about the characteristic of the training user can be stored in a second part of the database (or a second database) where information belonging to the same training user, e.g. B. can be linked via an identifier (eg., Anonymous or anonymous identifier) of the training user. Alternatively, the training sets may be stored together in the same database, e.g. B. after the information about the access behavior of a training user and the information about the characteristics of the training user were associated with each other. For example, information about the access behavior of a training user included in a training set may be updated each time new information about the access behavior of the training user becomes available (eg, when content of a cookie is provided on a web browser of the training user is set and associated with an entity (eg, the device) that passes this information to the database). Similarly, the information on the characteristics of a training user may be updated as new related information becomes available, such new information being obtainable, for example, from an online survey and / or from other data sources. For example, a training set may be considered new if at least one of the training user's access behavior information and the training user's characteristic information was previously unavailable. For example, a training set may be considered to be at least partially updated when at least one of the training user's access behavior information and the training user's information has been updated. For example, a single counter may be used to count both new incoming training sets and updated training sentences. If the counter exceeds the predefined value, then the training module may be encouraged to determine an updated version of the set of parameters (eg, based only on the new and updated training sets, or at least based on the new and updated training sets some training sets that were neither new nor updated (eg, all available training sets)), and provide them to the predictive engine. The predefined value may, for example, have a first value in a start phase and a second value in a phase following the start phase. For example, the first value may be less than the second value, such that the set of parameters for the model quickly becomes available for prediction. Higher quality of the set of parameters is then achieved in the post-launch phase.

According to one embodiment of all aspects of the invention, the device is a web server or a part thereof. For example, the web server may be associated with a domain for which third party cookies have been set in the current user's web browsers and training users.

According to one embodiment of all aspects of the invention, the device further comprises the training module. A processor of the device may then, for example, jointly implement the functionality of the prediction module and the functionality of the training module. However, these functionalities can be handled by different threads of the processor. Alternatively, the device may include separate processors for respectively performing the functionality of the prediction module and the functionality of the training module. Further alternatively, the training module may not be included in the device but may be included in an entity other than the device (eg, in a server (eg, a backend server) that communicates with the device).

According to one embodiment of all aspects of the invention, two or more characteristics of the current user are predicted, and each of the characteristics is predicted by a respective model. Two or more characteristics of the current user may then be taken into account, for example, to form a profile of the current user. For each model, a respective set of parameters may then be determined by the training module (or respective training modules) and provided to the prediction module (or respective prediction modules) from time to time for updating purposes.

In this embodiment, a plausibility check may be performed with respect to at least two characteristics predicted from their respective models. For example, if a characteristic is a marital status of the current user and another characteristic of the current user is an age group of the current user, then it can be checked if the family estates are plausible in terms of age group to avoid conflicting results. Such conflicting results may be reported back to the training module (or training modules) to enhance the process of determining the respective parameter sets.

According to one embodiment of all aspects of the invention, the characteristic relates to demographic information, in particular information about one of sex, race, age, marital status, disabilities, mobility, home ownership, employment status, place of residence and income, or also information about an interest, in particular an interest in one specific topic or product type.

Further advantageous embodiments of the invention will become apparent from the detailed description given below in conjunction with the accompanying drawings. It is to be understood, however, that the drawings are merely for the purpose of illustration and not as a definition of the limits of the invention, with reference to the appended claims. Further, it should be understood that the drawings are not to scale and that they are merely illustrative of the structures and procedures described therein. In particular, the presence of features in the drawings is not to be understood as requiring these features for the invention.

BRIEF DESCRIPTION OF THE FIGURES

In the figures show:

1 3 is a schematic block diagram of an exemplary embodiment of a predictive behavioral targeting system in accordance with the invention;

2a Fig. 3 is a flowchart of an exemplary embodiment of a method according to the invention;

2 B Fig. 3 is a flowchart of another exemplary embodiment of a method according to the invention;

3 FIG. 3 is a schematic block diagram of an exemplary embodiment of a device according to the invention; FIG.

4 10 is a schematic representation of a web page with an embedded web bug triggering predictive behavioral targeting, according to an example embodiment of the invention;

5a : A representation of multiple training sets according to an exemplary embodiment of the invention;

5b 3 is a schematic illustration of two sets of vectors separated by a hyperplane of a support vector machine, according to an example embodiment of the invention; and

6 3 is a flowchart of an exemplary embodiment of a method for predicting a characteristic of a user based on a set of parameters and on an observed click behavior based on an exemplary embodiment of the invention.

DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION

1 Figure 10 is a block diagram of an exemplary embodiment of a system according to the invention.

The system of 1 includes one from the user 6 operated terminal 1 , an advertising server 3 , an optional anonymization unit 4 and a forecasting system 5 which in turn comprises several components discussed in more detail below.

terminal 1 is represented by any computer by a user 6 to interact with a website 2 used on the internet. The user's computer may be a personal computer or other network-enabled device such as a digital assistant (PDA), a mobile phone such as a smartphone, a pager, a video game system, and so on. The website 2 is an Internet site owned by people who, for example, provide media pages, social networks and / or retailer pages.

The forecasting system 5 For example, the owner of the website 2 be operated on his website 2 Advertising information to a user 6 based on one or more predicted characteristics of the user 6 would like to align. However, the invention is also particularly advantageous if the prediction system 5 is operated by a service provider who performs the service of predicting one or more characteristics (eg, a profile) of the user 6 the website 2 offers to target advertising information through the ad server 3 on the user 6 on the basis of the predicted characteristics. Then the service of the forecasting system 5 Owners or publishers of websites offered to deliver appropriate advertisements to visitors to their websites.

It's easy to understand that 1 To simplify the presentation, the presence of only one user 6 pointing to the website 2 accesses. On the website 2 For example, multiple users may access through respective terminals, and for each of these users, the prediction system could be used 5 provide a predicted profile. Likewise, in 1 the presence of only one webpage 2 and an associated advertising server 3 simplify the presentation. The forecasting system 5 Of course, you can work with several different websites and associated advertising servers.

The forecasting system 5 includes a cookie processing unit 50 , a predictive module 51 , a training module 52 , a database 53 and an optional output formatting unit 54 , In principle, all components can be implemented as separate devices that communicate with each other. These devices can be positioned at remote locations, e.g. B. with distances of several hundred meters in between. As well, some of these components may be combined in one or more respective separate devices, or all components may be combined into one device. Parts of the forecasting system 5 can implement functionality of a web browser. Accordingly, at least parts of the prediction system 5 form a web browser with extended functionality.

Particularly relevant to the operation of the forecasting system 5 is a predictor module 51 , the information about the click behavior (as a special case of the access behavior) of a user 6 (of a current user) on one of the terminal 1 operated web browser on the website 2 accesses, receives (see step 101 ) and in response to and based on the information on the click behavior and on the basis of a predictive model, which in turn is based on a set of parameters, a characteristic of the user 6 predicts (see step 103 ). The prediction module 51 may equally well predict several characteristics based on respective predictive models with respective parameter sets. These one or more predicted characteristics form a predicted profile, that of the prediction module 51 to the output formatting unit 54 is issued (see step 103 ) and the basis for selecting an advertisement for inclusion in the website 2 that makes special on the user 6 is aligned. The output formatting unit 54 however, is optional, its functionality can be, for example, in the prediction module 51 be included or may not be necessary because, for example, the output of the prediction module 51 already the formatting request of the advertising server 3 Fulfills.

Another particularly relevant component of the forecasting system 5 is the training module 52 , which detects updates of the sets of parameters and updates these to the prediction module 51 provides (see step 110 ), allowing them from its predictive models to predict the characteristics of the user 6 can be used. These updates are from the training module 52 on the basis of Training sets determined in the database 53 stored and obtained from this (see step 109 ), wherein each training set for a respective training user associated with the training set includes information about its click behavior as well as information about its characteristics, which may be derived, for example, directly and indirectly from online surveys.

According to embodiments of the invention, the operation of the prediction module 51 and the training module 52 decoupled and asynchronous. The prediction module 51 is an online module that focuses on in-depth information about a user's click behavior 6 on the website 2 accesses, responds (see step 101 ) and in real time a predicted profile for the user 6 generated (see step 103 ). In contrast, the updates of the sets of parameters are not determined and provided in real time. A possible trigger event for the training module to update the sets of parameters 52 can be, for example, that in the database 53 a number of newly received and updated training sets has reached a predefined number. This functional separation of the prediction module 51 and the training module 52 is possible through the use of parametric predictive models that do not access a database 53 to require for any prediction that is to be made. It allows a much faster prediction of profiles by the prediction module 53 (on the order of only a few milliseconds) and has the further advantage that predictions can be made for users for whom no dedicated profile is available, as is generally required in state-of-the-art behavioral targeting systems. Therefore, profiles can be predicted for users who use the predictive system 5 before the first access to the website 2 were completely unknown ("first clickers"), and in principle with the same speed as with a user for whom a detailed click behavior is already available.

The cookie processing unit 50 of the prediction module 5 handles the setting of cookies (see step 102 ) and receiving information stored in cookies (see step 100 ) the web browser of Terminal 1 , This can be done in part with functionality implemented by a web browser. If information in the cookies is compressed and / or encrypted, then the compression / decompression and / or encryption / decryption can also be performed by the cookie processing unit. Cookies represent a way to obtain information about the user's click behavior 6 , For this it is necessary that a so-called web bug 20 (also referred to as tracking pixel or blind GIF) from the publisher of the website to the website 2 is installed. Alternatively, for example, a tracking script may be centrally deployed in the website comprising the website and may then be associated with some or all of the website's websites. For example, the web bug may be an HTML image tag that specifies a URL (Uniform Resource Locator) where image data is to be downloaded (alternatively, the web bug may also be implemented as a script, such as a JavaScript image object ). This image data is usually just a 1 × 1 pixel image that is the rest of the webpage 2 does not bother (it may be transparent, for example). However, the path bug causes the web browser to request the image data from the web server associated with the domain specified in the URL. In the forecasting system 5 is the cookie processing unit 50 responsible for setting cookies via web browser, the data from the with the prediction system 5 request associated domain and have not yet set a cookie for this domain (see step 102 ). The cookies may be limited in time, for example. For example, one year. The following is an example assuming that the domain of the prediction system 5 "Nugg.ad" is. If now the user 6 for the first time on the website 2 with the web bug 20 accesses it (or accesses the site again after the deletion or expiration of an already existing cookie) and if the web bug 20 Image data in the domain "nugg.ad" pretends to be the cookie processing unit 50 upon receiving the request for image data with instructions for the user's web browser 6 to save a cookie in the web browser (HTTP Set-Cookie) for the domain "nugg.ad" (the path parameter of the cookie may simply be "1", for example). If the user 6 with his web browser again on a the web bug 20 containing website 2 or accessing another web page that has a web bug that specifies image data in the domain "nugg.ad", then the web browser represents user 6 automatically the contents of the cookie of the cookie processing unit 50 ready (see step 100 ). In its response then the cookie processing unit 50 Write updated information to the cookie, which in turn should be stored in the web browser (see step 102 ). Because the cookie belongs to a domain that is different from the domain of webpage 2 This cookie is commonly referred to as a "third party cookie". In embodiments of the present invention, the cookie is used by the user in the web browser 6 (and equally in other web browsers of other users) for storing information about user click behavior 6 used in a decentralized way.

In embodiments of the invention, the information about the click behavior is represented by a content category histogram indicating how often user accesses to web pages that belong to a specific content category. Examples of content categories include culture, sports, art, entertainment, business, politics, etc. This content category histogram is stored in a cookie.

The decision on which content category the website 2 belongs to which the user 6 accessed by the decision unit 501 , in the 1 by way of example as a subcomponent of the cookie processing unit 50 Is accepted. Equally well, the decision-making unit could also have its own component of the forecasting system 5 be. The one of the decision-making unit 501 For example, the decision taken is based on a content classifier, for example, from a limited set of content categories by the publisher or owner of the website 2 is selected and used as a parameter in the web bug 20 from website 2 is included. Alternatively, the decision unit 501 a content category from a title or at least part of the content of the website 2 derive, for. By parsing. An identifier of website 2 It's a decision-making unit 5 allows to the website 2 For example, access can be obtained as an HTTP referer or may be explicitly by the publisher or owner of the website 2 as a parameter in the web bug 20 be included. If the user 6 on the website 2 accesses and the web browser from the user 6 the web bug 20 parse and accordingly image data from that with the forecasting system 5 requesting associated domain "nugg.ad", then determines the decision unit 501 thus the content category of website 2 , If previously a cookie for the domain "nugg.ad" on the user's web browser 6 has been set, then the content of this cookie is the cookie processing unit 50 provided upon requesting the image data. The decision-making unit 501 can then update the content category histogram obtained from the cookie and write the updated content category histogram to the cookie. If not previously a cookie for the domain "nugg.ad" on the user's web browser 6 has been set, then generates the decision unit 501 an empty content category histogram, adds a hit for the content category of webpage 2 to the content category histogram and writes this content category histogram to a cookie for the nugg.ad domain on the user's Web browser 6 , Consequently reflected when the web bug 20 is contained in several different web pages each with different content and if the user 6 accesses several of these web pages, the content diagram stored in the cookie on the user's web browser adequately the content categories in which the user is interested. It is also quite clear that the decision-making unit 501 the cookie processing unit 50 a content category diagram (e.g., the prediction module 51 as by arrow 101 indicated) already with a first click of the user 6 can deploy on a web page that has the web bug 20 contains, that is, if not yet a cookie for the domain "nugg.ad" on the user's web browser 6 was saved. It is also obvious that the cookie measuring unit 50 an updated version of the content category histogram for the user 6 the database 53 (as by arrow 108 indicated) every time the user can provide 6 on a website (like the website 2 ) accesses a web bug 20 (or a similar web bug associated with the domain "nugg.ad"). The updated content category histogram of a user (referred to in this context as a "training user") forms part of one in the database 53 stored training set, wherein the other part includes information about one or more characteristics of the training user, for example, were obtained from an online survey conducted with the training user.

The output formatting unit 54 of the forecasting system 5 receives predicted profiles from the prediction module 51 (see step 103 ) and converts these profiles into a format that the advertising server 3 as the final recipient of the predicted profile (see step 104 ). Formatting may involve logical and / or technical formatting. Logical formatting may include, for example, combining one or more predicted characteristics into a new characteristic or deriving a characteristic from one or more predicted characteristics. Technical formatting may include, for example, converting a data format and / or syntax of one or more predicted characteristics. The format requirements of ad server 3 For example, in the web bug 20 , z. B. defined by one or more parameters thereof.

The anonymizer unit 4 is an optional component for solving privacy-related issues that may arise when tracking user click behavior. This will be the communication between the user's web browser 6 and the forecasting system 5 via the anonymisation unit 4 passed the IP address of the user 6 anonymized and the forecasting system 5 only provide this anonymized IP address. The forecasting system 5 thus can the click behavior of user 6 not with the real IP address of the user 6 get in touch. Even so, the predicted profile may return to the user's web browser 6 can be sent, the forecasting system 5 the anonymous IP address of the user 6 then used by the anonymizer unit 4 to the original IP of the user 6 is shown. The anonymizer unit 4 is for example from Third parties who have agreed not to associate the recorded click behavior and / or the resulting predicted profiles in any way with the IP address of a user. The anonymity 4 serves as a processor of data for the forecasting system 5 and preferably has a contractual agreement that any data communication between the prediction system 5 and web browsers of users (like users 6 ) is sent by an anonymization server in a data center, which only from the anonymizer unit 4 is managed. The forecasting system 5 does not have access to this server. The anonymizer unit 4 generates an anonymized IP address (for example, by applying a hash function to the IP address). The anonymizer unit 4 preferably does not store data about the click behavior of a user. The anonymizer unit 4 only forwards this data. The anonymizer unit 4 For example, it may represent a proxy, in particular a reverse proxy. As another feature, the anonymizer unit 4 Information about the location of a user (so-called geo information), e.g. B. derive information about the state or even the state of the IP address of the user, for. Using an Apache server module such as Geo-IP. Accurate information about the actual location of the user can not be obtained with this module, this is merely a probable assumption about the state / state in which the user is currently located. For example, this geo-information can be used by the forecasting system 5 can be provided and, for example, as a characteristic to those in the database 53 be added to stored training sets. Alternatively, this information may be from the prediction module 51 predicted profile added as another characteristic and the advertising server 3 be provided for targeting (potentially after formatting in the output formatting unit 54 ). Targeting can then have components of predictive behavioral targeting and geo or location-based targeting.

The advertising server 3 is for selecting and delivering advertisements to the website 2 responsible. For example, the website 2 include an empty horizontal space in the upper part where to place a banner ad. For this purpose, the website 2 a script for communicating with the ad server 3 include. If the script is executed by a web browser, then a request for an advertisement to the advertising server 3 which selects the advertisement from a pool of available advertisements and sends parameters (and for example the selected advertisement or a link to the advertisement) back to the web browser. To implement predictive behavioral targeting, the ad server must 3 with information about that for the user 6 from the forecasting system 5 predicted profile are supplied. This can be done, for example, by using, as a reaction (see step 104 ) on the way-bug 20 initiated request (see step 102 ), a location redirect statement, where the redirect statement is the address of the advertising server 3 and information about the predicted profile. This will cause the web browser (from the terminal 1 operated) by user 6 the advertising server 3 contacted (see step 105 ) and the advertising server 3 provides the information about the predicted profile, then from the ad server 3 to select an advertisement for the website 2 that can be used to that for the user 6 predicted profile fits. Furthermore, the advertising server 3 in response to the request of step 105 the web browser of user 6 instruct you to set an ad server cookie (see step 107 ) that for the user 6 predicted profile stores. In this way, the advertising server has 3 always access the predicted profile when the user 6 accesses a web page that matches the domain of the advertising server 3 associated with the website (eg because the website requests data from the ad server domain, eg via a web bug). Predictive behavioral targeting is thus for the ad server 3 even if possible with the domain of the prediction system 5 associated web bug 20 not included in the website to which the user 6 accesses and for the advertising server 3 to deliver an advertisement.

So far 1 used to explain how a profile for a user (user 6 ), which is currently on a website (website 2 ) accesses. As explained above, the profile is determined by the prediction module 51 of the forecasting system 5 based on respective prediction models for each characteristic contained in the profile, wherein a set of parameters exist for each prediction model, and wherein the sets of parameters from time to time are derived from the training module 52 on the basis of several in the database 53 stored training sets are generated and updated.

1 can also be used for explanation, such as the training records in the database 53 to be obtained. This is done by users 6 from 1 no longer viewed as a user for whom a profile is to be predicted, but as a user pointing to the webpage 2 and a survey presented to him 21 participates. The user 6 will then be considered a training user as he becomes one of the database users 53 contributes to stored training sets. It is assumed that the mechanism of the website 2 included web bug 20 is maintained, ie that access to the website 2 by the user 6 causes the cookie processing unit 50 is activated, so that one Content category histogram for the user 6 initialized and by a hit for the content category of webpage 2 is enriched if no cookie for the domain "nugg.ad" on the user's web browser 6 or an existing content category histogram stored in a cookie on the user's web browser 6 is stored with the content category of website 2 is updated (see step 100 ). In both cases, the content category histogram will be from the cookie processing unit 50 the database 53 along with the anonymized IP address of user 6 provided for storage.

If now the user 6 , in addition to accessing the website 2 , at the poll 21 Participates by answering the questions of the survey, then information about the results of the survey also the database 53 along with the anonymized IP address of user 6 provided (eg via an HTTP request, eg also via the cookie processing unit 50 ). Using the anonymized IP address, the information from these responses can now be used with the content category histogram for users 6 linked and thus to form a training set for the training user 6 get saved. For example, the information about the survey results may correspond to at least some of the characteristics provided by the predictive module 51 can be predicted. For example, a gender and / or age group in the survey may be provided directly by the user 6 be requested. More generally, the survey may ask questions about the demographics, lifestyle and product interests of the user 6 included, just to name a few examples.

The survey 21 does not need every user 6 to be submitted to the website 2 but only a small random part of users of the website 2 , Thus, the website can 2 Ask users to join the survey, and only when a user agrees can the information be retrieved. For example, the survey may be submitted by an ad server running the website 2 is associated, such. B. the advertising server 3 ,

Users who receive an online questionnaire are preferably informed transparently and clearly that participation in the survey is voluntary. Users can in particular be informed about the purpose of the survey (improvement of advertising). For example, survey results do not need to be stored in a cookie. The proportion of users to whom the survey is submitted may be less than 1% of the total number of users (for example, all users who access the website 2 access). The percentage is preferably between 0.02% and 0.05%.

The survey results and click-through information can be used for targeted advertising and collected and evaluated to produce predictive models. The underlying idea is that users with similar click behavior regularly demonstrate a similar demographic profile and vice versa. These predictions allow publishers and website owners to submit profiles that are more relevant to advertising. Thus, the publisher / web site owner may provide each user appropriate advertisements based on the forecasting system 5 made predictions, z. In terms of demographics, product interests and / or lifestyle of that user. For example, the result may be that advertisements for specifically eco-friendly products are offered to a web site visitor to a grocery store that is believed to be financially well-off and has a "sustainable" lifestyle.

If the user 6 Visited several websites from different publishers, all of which have the profile forecasting service of the forecasting system 5 then use that on the web browser of user 6 set cookie data of all publishers concerned. As a result, the click behavior information (eg, the content category histogram) is stored separately in the cookie for the various publishers and is processed and used separately. Preferably, a publisher can not access data about the visits from websites of other publishers. Furthermore, profiles can only be predicted on the basis of cookie information, which relates directly to the visit of the websites of the specific publisher and not to visits to websites of other publishers. This also applies to the derivation of the sets of parameters from the training sets, ie sets of parameters to be used for predicting a profile for a publisher's web page are derived only from training sentences that contain information pertaining to this publisher's web pages ,

2a shows a flowchart 200 an exemplary embodiment of a method according to the invention. The steps 201 - 206 This flowchart reflects functionality provided by the prediction engine 51 is implemented (see 1 ). In a first step 201 It checks for an update of one or more sets of parameters from the training module 52 was obtained (see 1 ). Each of these sets of parameters belongs to a respective predictive model, which in turn predicts a Characteristic of a current user (like user 6 from 1 ), that is, a user currently accessing a web page for which predictive behavioral targeting is to be performed. If the test determines that an update has been received, then the updated set (s) of parameters will become / are in step 202 recorded, z. By replacing the exiting sentence (s) of parameters with the new set (s) of parameters. Then in one step 203 Checked if information about click behavior of the current user was received. If this is the case, then be in one step 204 selected one or more behavioral models, each needed for the prediction of one or more characteristics of the current user. For example, information about a choice of the characteristic (s) to be predicted is obtained along with the information about the click behavior of the current user (eg, as a parameter of a web bug). In one step 205 the characteristic (s) for the current user are predicted based on the selected predictive model (s). If the selected predictive model (s) is on (one) in step 202 based on updated respective set / sets of parameters, then the updated set (s) of parameters for the prediction in step (s) will become / are updated 205 used. Otherwise, non-updated set (s) of parameters will be used. step 205 Optionally, it may include normalization of the clickthrough behavior of the current user, as discussed with reference to equation (4) below. Finally, will / will be in step 206 the predicted characteristic (s) are output, e.g. To the output formatting unit 54 (please refer 1 ).

2 B shows a flowchart 300 Another exemplary embodiment of a method according to the invention. The steps 301 - 316 This flowchart describes the reception and storage of training set information in the database 53 (please refer 1 ), as by the steps 301 - 314 and determining or updating one or more sets of parameters based on the training sets by the training module 52 and the output of it, as through the steps 315 - 316 represents. The steps 301 - 314 For example, they can be executed by a unit that owns the database 53 controls, z. From the cookie processing unit 50 , The steps 315 - 316 for example, from the training module 52 be executed.

In a first step 301 a counter variable is initialized to zero. In one step 302 is then checked for information about a user's click behavior (such as user 6 in 1 ) were received. If this is not the case, then the flowchart jumps to step 306 , Otherwise, it is checked whether a training set (with information about the user's click behavior as well as information about one or more characteristics of the training user) is already in the database 53 is available (step 303 ). If so, then the information about the click behavior of the training user in the existing training set may be in step 310 are updated and the counter variable is incremented by one (step 311 ). Otherwise, in step 304 whether information about one or more characteristics of the training user is available (but not yet stored in a training set for that training user). If this is not the case, then the information about the click behavior of the training user in step 305 stored (eg in the database 53 or at another location, but not yet as part of a training set). Otherwise, a new training set that contains the training user's click behavior information and the training user's characteristic (s) can be found in step 312 generated and in the database 53 are stored and the counter variable is incremented by one (step 313 ).

After step 305 , or if the exam in step 302 A negative result is shown in step 306 Check to see if information about one or more characteristics of a training user has been received. If this is not the case, then the flowchart jumps to step 314 , Otherwise, it is checked whether a training set (with information about the user's click behavior as well as information about one or more characteristics of the training user) is already in the database 53 is available (step 307 ). If so, then the information about the trait user's characteristic (s) in the existing training set may be in step 310 are updated and the counter variable is incremented by one (step 311 ). Otherwise, in step 308 Verifies that information about a user's click behavior is available (but not yet stored in a training set for that training user). If this is not the case, then the information about the characteristics of the training user in step 309 stored (eg in the database 53 or at another location, but not yet as part of a training set). Otherwise, a new training set that contains the training user's click behavior information and the training user's characteristic (s) can be found in step 312 generated and in the database 53 are stored and the counter variable is incremented by one (step 313 ).

In step 314 it is then checked whether the counter variable is greater than a predefined trigger value. If so, then this is considered to have a sufficient number of training sets to make a determination (if no set / sets of parameters exist) or update (if already a set) / Sets of parameters) of one or more sets of prediction model parameters based on the one in the database 53 to allow stored training sets, and this determination or update is done in step 315 (through the training module 52 ), and the determined or updated set (s) of parameters will become (s) in step 316 (to the prediction module 51 ). The flowchart then returns to step 310 back where the counter variable is reset to zero. The flowchart is then repeated as described above.

step 315 Optionally, it may also include normalization of the training sets of the training users, as discussed below for the current user with reference to equation (4).

In the above description of the flowchart 300 from 2 B By way of example, it has been assumed that training sets for training users will only be generated and stored in the database 53 stored when both information about the click behavior of a training user and information about the characteristics of the training user are available. Of course, it is also possible to include information related to a training user (ie information about the click behavior of a training user or information about the characteristics of a training user), always directly in the database 53 regardless of whether the additional information needed to form a training set is already available or not. Nevertheless, it can still be assumed that a training set for a training user in the database 53 is present only if both the information about the user's click behavior and information about the trainer user's characteristics are present. Once this is the case for a training user (in response to receiving either the click behavior information or the training user's characteristic (s) that complete the training set), the counter for a new training set may be in the database 53 be incremented by one. Subsequent receipt of information about a click behavior and / or characteristic (s) of a training user for whom a training set is already assumed to be present leads to an update of this training set and thus also to an increment of the counter by one.

3 FIG. 10 is a schematic block diagram of an exemplary embodiment of a device. FIG 30 according to the invention. The device 30 may be, for example, a web server or at least part of it. The device 30 For example, it may be configured to follow the steps of the flowchart 200 from 2 B and / or the flowchart 300 from 3b performs. Accordingly, the device 30 for example, the functionality of the prediction module 51 (please refer 1 ) and / or the training module 52 and / or the database 53 to implement. Furthermore, the device 300 also functionality of the cookie processing unit 50 and / or the output formatting unit 54 to implement.

The device 30 includes a processor 31 which may be configured, for example, as a microprocessor, digital signal processor (DSP) or application specific integrated circuit (ASIC), to name just a few non-limiting examples. The processor 31 leads one in program memory 32 stored program code. The program memory 32 can also be in the processor 31 be included. The program memory 32 can be stuck with the processor 30 connected or from the processor 30 be removable, z. B. in the form of a memory card or a stick. The program memory 32 and the computer program code stored therein may be communicated with the processor 31 be configured to cause the device 30 at least executes the method according to the first aspect of the invention or its exemplary embodiments described above. The program memory 32 may constitute a tangible storage medium comprising a computer program with program code for carrying out the method according to the first aspect of the invention or one of its embodiments described above, when the computer program on the processor 31 is executed comprises.

The processor 31 is also with a main memory 33 which may, for example, be configured as random access memory (RAM) to support its operation.

The device 30 can optionally have a database 34 include, for example, the database 53 from 1 represents. The processor 31 can then go to the database 34 access (eg, read from and / or write to), which may be configured, for example as a mass storage, z. B. with capacities of several gigabytes or several terabytes. He can be stuck with the processor 31 connected or detachably connected to it.

The processor 31 also controls a network interface 36 that is configured to receive and / or output information. About this network interface 36 are, for example, communications of the device 30 with at least some of the device not 30 implemented components of the forecasting system 5 (please refer 1 ) possible. Via the network interface 36 can the device 30 for example, with the user's web browser 6 (if no anonymization unit 4 used) or with the anonymizer unit 6 (if an anonymization unit is used) communicate, for example, when the device 30 the cookie processing unit 50 and / or the output formatting unit 54 implemented.

The network interface 36 may, for example, allow communications with one or more partners according to the HTTP protocol.

The device 30 can optionally also have a user interface 35 configured to be a user / operator of the device 30 Information presents and / or information (eg, configuration parameters such as the trigger value of step 314 of flowchart 300 (please refer 2 B )) from such a user / operator.

It should be noted that due to the components of device 30 may be implemented only in hardware, partly in hardware and in software or only in software.

4 is a schematic representation of a website 40 with an embedded web bug 401 , which initiates predictive behavioral targeting according to an exemplary embodiment of the invention. The web bug 401 has here for example the form of an image tag. The general syntax of Web Bug 401 is explained below by way of example:
<img src = "http://subdomain.nuggad.net/bk?nuggn=12345&nuggsid=12345&nuggrid=http://site.com&nuggtg=contentClassifier&nuggl=location">

Therein, the actual URL (Uniform Resource Locator) of the image (eg, a 1x1 transparent pixel) "http://subdomain.nuggad.net/bk" is followed by a series of parameters defined as follows: nuggn Network ID, defined by the operator of the forecasting system 41 nuggsid Site ID, defined by the operator of the forecasting system 41 nugg.ad (optional) nuggrid Referrer URL (eg specified by the publisher / owner of the web page), if http referrer can not be used, URL encoded (optional, but recommended if, for example, nuggsid can not be used) nuggtg Content classifier (ie 'sport'), URL encoded (optional, for example if http referrer can not be used) nuggl Redirect location, URL encoded (http header LOCATION; ad server cookie writes API)

If on the website 40 by a user (eg user 6 in 1 ), then causes the web bug 401 a request for the image data defined by the URL of the web bug from that with the prediction system 41 (the forecasting system 5 from 1 corresponds) associated domain ("nugg.ad"). The parameters contained in the web bug then also become the predictive system 41 transfer and allow it to be a content of the website 40 classify (eg based on content classifier nuggtg). In response to this requirement gives the forecasting system 41 the instructions to be executed by the user's web browser 410 back. In particular, a cookie for the domain "nugg.ad" is set via HTTP SET-COOKIE (which stores the content category histogram based, at least in part, on the content category of Web page 40 based). Furthermore, a redirection via HTTP LOCATION is triggered, ie the user's web browser is instructed to use the advertising server 42 to contact based on the URL contained in the nuggl parameter of the web bug. Therein, the value "NUGGVARS" of parameter cvalue 1, which appends to the URL of the ad server, as in the parameter nuggl of Web-Bug 401 contained in response to the request to the prediction system 41 by the forecasting system 41 for the user who is on the website 40 This predicted profile "21999930001" follows the format required by the ad server and / or its cookie as discussed below. Thus, each time a user accesses the web bug 401 containing website 40 the content category of the website 40 and the contents of the cookie stored for the nugg.ad domain Forecast System 41 provided, and a profile for the user is predicted and a redirect to the web bug 401 given advertising server 42 returned.

The advertising server 42 then sets up an advertisement server cookie containing the predicted profile (via HTTP SET-COOKIE) for its domain (eg, "adxy.de") on the user's web browser, such as through the HTTP response 420 shown. After writing this ad serving cookie, it may be used to target advertisements on each web page accessed by the user's web browser and include a request for data from the ad serving server 42 in or associated with it, such as Eg a script 402 that the advertising server 42 contacted for an advertisement in the website 40 should be included. When contacting the advertising server 42 the predicted profile for the user becomes the ad server 42 provided through the ad server cookie, allowing the ad server 42 select a user-targeted advertisement based on the predicted profile.

It will be easy to understand that the HTML image tag described above 401 just an example of a web bug. As well, for example, a script code, z. For example, a JavaScript Image Object can be used for this task, and this script can refer, for example, to multiple Web pages on a Web site.

5a shows in table 60 a representation of several training sets according to an exemplary embodiment of the invention. Every row of table 60 can be understood as a training sentence. The first column 61 from table 60 includes respective identifiers of the training users to which the training sentences belong, respectively. For example, these identifiers can be anonymized identifiers, such as those from the anonymization unit 5 be provided (see

1 ). In 5a In order to simplify the presentation, these identifiers have been simplified to simple numbers such as "1", "2" and so on. The second and third columns, with reference number 62 , each contain characteristics of the user. By way of example only two characteristics are available here, "gender" and "age group". In the other columns, with reference number 63 , the content category histogram of each training user is included. Here, the content categories are exemplified as "vehicles", "cars", "computers", "consumer electronics", "finance", "economy" and "fashion and beauty". Empty table cells in the column area 63 express that they have not seen traffic to web pages belonging to the respective content category. The non-empty table cells in the column area 63 express - in (exemplary) normalized form (as explained with reference to equation (4) below) - how often visits to web pages belonging to the respective content category have been observed for the respective training user. For example, for a user with id = 1, it is known (eg, from a survey) that he has the following two characteristics: he is male (gender = 0) and has an age from the group = 3 on its click behavior the following content category histogram is known: vehicles 0.4733 cars 0.4993 computer 0.4897 Consumer electronics 0.4853 finances 0 economy 0.5065 Fashion and beauty 0

This content category histogram can be expressed as a seven-dimensional vector:
x ₁ = (0.4733, 0.4993, 0.4897, 0.4853, 0, 0.5065, 0)

Similar vectors can be set up to match the respective click behavior of the other training users of Table 60 represent. According to embodiments of the invention, characteristics of a user are predicted based on a predictive model that allows classification of a vector associated with the user into one of at least two classes. For example, if the characteristic is the gender of the user, then a vector representing the user's click behavior may be classified as belonging either to a class "male" or to a class "female".

Now the training sets of table 60 for determining a set of parameters for a prediction model for the characteristic "gender" and for determining a set of parameters for a predictive model for the characteristic "age group". For example, such a set of parameters may be derived from the training sets based on a support vector engine, as discussed in more detail below.

A description of support vector machines is in the Article "Support Vector Networks" by Corinna Cortes and Vladimir Vapnik in Machine Learning Vol. 20, No. 3 (1995), pp. 273-297 which is incorporated herein by reference in its entirety.

In particular, in the exemplary case of a linear support vector machine having an optimal hyperplane, a support vector engine is determined based on a set of N p-dimensional training vectors x _n (where n = 1..N) with, respectively associated class identifiers y _n ε {-1, 1} a hyperplane: wx + b = 0 (1) such that: w x _i + b ≥ 1 if y _i = 1, and w x _i + b ≤ 1 if y _i = -1. (2)

Therein, the hyperplane is defined according to its p-dimensional normal vector w and the so (scalar) bias b and determined such that the margin, ie the distance of the training vectors from each of the two classes of training vectors (indicated by the class identifiers y _n ) which are closest to the hyperplane is maximized.

5b is a schematic representation of two sets of training vectors x _n 72 . 73 passing through a hyperplane 71 (with the normal vector w 710 and the margin 711 ) of a support vector machine according to an exemplary embodiment of the invention.

In 5b becomes only a two-dimensional scenario (ie a two-dimensional space 70 ) and only two classes of training vectors x _n 72 . 73 are exemplified by the hyperplane 71 separated. The first class of training vectors x _n is denoted by black dots 72 represented and belongs to male training users (eg, y _n = 1). The second class of vectors x _n is represented by circles 73 represents and belongs to female training users (eg y _n = -1). For some of the vectors, the content categories, which are particularly pronounced for the respective training user, are also shown in brackets. For example, it can be seen that a training user who is interested in consumer electronics is closer to the hyperplane 71 lies as a training user who is interested in fashion, computer and consumer electronics.

The principle of separating vectors of different classes by a hyperplane can of course also be extended to the multidimensional case. For example, instead of two-dimensional vectors each containing the content category histograms of training users, seven-dimensional training vectors each having the content category histograms as in the column area 63 from table 60 from 5a can be used in a seven-dimensional space and could be assigned to either a male or a female class, according to the information in the second column of the table 60 , For these two classes of training vectors, a hyperplane may then be determined that yields a set of parameters that serves as a decision function for classifying a corresponding content category histogram vector of a current user. The application of this decision function to a vector representing the click behavior of a current user for predicting character characteristics of the current user will now be explained in more detail.

When a hyperplane defined by its normal vector w and bias b is derived from a set of training vectors (in the training module 52 from 1 ), then a particularly simple classification of a vector x _i into one of the classes is possible. The decision function for this classification can be written as follows: y _i = sgn (<w, x _i > + b). (3)

Sgn (·) returns the sign of its argument, ie 1 for a positive argument and -1 for a negative argument, and <·, ·> returns the inner vector product of its two arguments. This decision function can have very little computational complexity in the prediction module 51 be calculated (see 1 ), thus allowing real-time prediction of user characteristics, e.g. Within a few milliseconds.

6 FIG. 3 is a flowchart of an exemplary embodiment of a method for predicting a characteristic of a current user based on a set of parameters 80 and an observed click behavior according to an exemplary embodiment of the invention.

In the left part of 6 is an exemplary set of parameters 80 which is derived from several training sets for predicting the user characteristic "gender". In it were the parameters 80 determined according to the Support Vector Machine approach described above. The set of parameters 80 includes an "offset" value that represents the bias of the hyperplane. Further, the set of parameters includes 80 Weightings representing the components of the normal vector w of the hyperplane. Each of these weights is associated with a component of the current user's vector and thus with the content categories "vehicles", "cars", "fashion and beauty", "economy", "finance", "computer", "consumer electronics".

The right part of 6 shows the actual flowchart 90 for the prediction process (such as in the prediction module 51 from 1 is performed). If the current user (for example, user 6 from 1 ) accesses a web page containing the web bug for tracking click behavior, then the content category histogram for the current user is updated and may be, for example, the one in step 91 of flowchart 90 This means that there were 12 tracked clicks on web pages that belong to the content category "cars", 10 tracked clicks on web pages that belong to the content category "vehicles" and 5 tracked clicks on web pages with the content category "business". In an (optional) step 92 the content category histogram is normalized to that in step 93 shown normalized content category histogram.

The normalization in the optional step 92 For example, to account for the fact that a user's clicks on web pages that are generally accessed frequently have less of an impact than a user's clicks on web pages that are generally less frequently accessed (indicating special interests of the user). , For example, normalization can be based on the following formula:

Is p _n, denotes the normalized value α is a constant factor, z. In the range [0.01..1] (eg 0.4), n _a means the number of clicks of the current user on web pages of a specific content category (eg in a specific (eg predefined) time interval ), p means the total likelihood of clicks on web pages of this specific content category (eg, determined by statistical analysis of the number of accesses by multiple users to web pages of this specific content category), and N _{a represents} the total number of clicks of the current user on web pages ( eg in the specific time interval) (eg, determined by adding the user's values n _a across all content categories). For example, normalization of the content category histogram of the current user may be done by the prediction module 51 be carried out (see 1 ), for example as part of step 205 of the flowchart 200 from 2a , When normalization is applied to the current user's content category histogram, then it also applies to the content category histograms of the training users, e.g. B. by training module 52 (please refer 1 ), for example in step 315 of flowchart 300 from 2 B ,

Then in step 94 the on the set of parameters 80 based decision function as described by equation (3) above. This gives a value for the decision function of y _i = 1. In step 95 Then, since the class identifier of the vector representing the click behavior / content category histogram of the current user is y _i = 1, it is determined that the current user is male.

It should be understood that the invention is not limited to support vector machines as classification algorithms. As well as other classification algorithms can be used, which allow a decoupling of training and prediction and thus fast forecasts (in real time). If a support Vector machine is chosen, then not necessarily optimal hyperplanes to be determined. Instead, for example, soft-margin hyperplanes can also be determined. It should also be understood that support vector machines are not tied to a two-class classification. Similarly, a classification of vectors in more than two classes is possible, for. By applying so-called multiclass support vector machines. One approach here is to reduce the multiclass problem to multiple binary (two-class) classification problems.

With respect to the above-described embodiments of the invention, it is to be understood that a disclosure of an action or a step of a method according to the present embodiment should be understood as a disclosure of a corresponding (functional) configuration of a device (for example, a configuration of the computer program code and / or or the processor and / or other means of the device) according to this embodiment, a corresponding computer program code defined to effect such action or step in execution and / or a corresponding (functional) configuration a system (or parts thereof) according to the present embodiment.

The above-presented embodiments of the invention and their individual features should be understood as disclosed in all possible combinations. It should also be understood that the sequence of process steps in the flowcharts presented above is not mandatory, but alternative sequences are also possible.

The invention has been described above with reference to embodiments, which are to be understood as non-limiting examples. In particular, it should be noted that there are alternative manners and variations which are obvious to those skilled in the art and which may be implemented without departing from the scope and spirit of the appended claims.

QUOTES INCLUDE IN THE DESCRIPTION

This list of the documents listed by the applicant has been generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• DE 102004047815 B4 [0003, 0006]

Cited non-patent literature

• Article "Support Vector Networks" by Corinna Cortes and Vladimir Vapnik in Machine Learning Vol. 20, No. 3 (1995), pp. 273-297 [0097]

Claims (46)

Apparatus arranged to carry out the following method or comprising respective means for carrying out the method steps of the following method: Obtaining information about an access behavior of a current user accessing content; Predicting, with a predictive module of the device and in response to obtaining the information about the access behavior of the current user, at least one characteristic of the current user based at least on a model for the characteristic and the obtained information about the access behavior of the current user, wherein the model is based on at least one set of parameters determined by a training module based on information from multiple training sets and provided to the predictive module from time to time for update purposes, each training set of training sets for a respective training user of a plurality of training users , Information about an access behavior of this particular training user and information about the characteristics of this particular training user includes, and Providing information about the predicted characteristic of the current user so that information can be aligned with the current user based at least on the predicted characteristic information. The apparatus of claim 1, wherein the characteristic of the current user can be predicted by classifying a vector representing the information about the access behavior of the current user into one of at least two classes, the model being based on a support vector machine comprising the classification allows. The apparatus of claim 2, wherein the set of parameters defines a support vector engine decision function applicable to the vector representing the current user's access behavior information for performing the classification. The apparatus of any one of claims 2-3, wherein for each training set of the training sets, the information about the access behavior of the respective training user contained in the training set represents a training vector associated with a class comprised of the information about the characteristic of the one contained in the training set respective training user, and wherein the set of parameters is determined by the training module to reflect a separation of training vectors from different classes according to a predefined optimization criterion. The apparatus of any one of claims 1-4, wherein the current user access behavior information for each content category of a set of one or more content categories includes information about how often accesses to content associated with the content category by the current user in one Time period, and wherein the information about the access behavior of a respective training user included in the training sentences for each content category of a set of one or more content categories includes information about how often accesses to content belonging to the content category by the training user in a time period were observed. The apparatus of claim 5, wherein a decision as to whether content belongs to one of the content categories is made by a decision unit based on a received content classifier and / or on the basis of an analysis of at least part of the content. The apparatus of any one of claims 5-6, wherein predicting the current user's characteristic is based on presentation of the current user's access behavior information, the information on how often the current user has accessed content from the content category of the set is heard from one or more content categories, normalized, and wherein the set of parameters is determined based on training sets in which the respective information about how often the respective training user has accessed content, the content categories of the set of one or more content categories heard, were normalized. Device according to one of claims 1-7, wherein at least part of the information about the access behavior of the current user is stored in a cookie on a web browser of the current user and provided by the web browser. The apparatus of claim 8, wherein the information about the access behavior of a training user of the training user is stored in a cookie on a web browser of the training user where at least one domain for which the cookie was set matches a domain for which the cookie stored on the web browser of the current user has been set. The apparatus of any one of claims 1-9, wherein the information contained in a training set of the training sentences about the characteristic of a training user is obtained from a survey performed with the training user and associated with the information about the access behavior of the training user to the training set to obtain. Device according to one of claims 1-10, wherein an actual address of the current user is anonymized by an anonymization unit and wherein the information about the access behavior of the current user with the anonymized address and not with the real address are obtained. The apparatus of claim 11, wherein the information about the predicted characteristic of the current user is provided with the anonymized address and not the actual address. The apparatus of any one of claims 1-12, wherein the predicted characteristic information of the current user is formatted to conform to a format required by an advertising server that aligns the information to the current user, at least on the basis the information about the predicted characteristic executes at least. The apparatus of any one of claims 1-13, wherein the provided information about the predicted characteristic of the current user is provided to an advertisement server that at least executes the alignment of the information to the current user based at least on the information about the predicted characteristic The apparatus of claim 14, wherein the advertisement server initiates storage of the predicted characteristic information in a cookie on a current user's web browser. The apparatus of any one of claims 1-15, wherein an updated version of the set of parameters is determined by the training module and provided to the prediction module each time a number of new and / or at least partially updated training sets have reached a predefined value. The device of any of claims 1-16, wherein the device is a web server or a part thereof. The device of any one of claims 1-17, wherein the device further comprises the training module. Apparatus according to any one of claims 1-18, wherein two or more characteristics of the user are predicted, each of the characteristics being predicted by a respective model. The apparatus of claim 19, wherein a plausibility check is performed with respect to at least two characteristics predicted from their respective models. Device according to one of claims 1-20, wherein the characteristic relates to demographic information, in particular information about one of sex, race, age, marital status, disabilities, mobility, home ownership, employment status, place and income, or information about an interest, in particular an interest on a specific topic or product type. Device according to one of claims 1-21, wherein the device further comprises the training module. Apparatus according to any one of claims 1-22, wherein the apparatus is a web server or a part thereof. A system comprising: - A device according to any one of claims 1-21 or 23, and - the training module. A computer program comprising program code for performing the following procedure when the computer program is run on a processor: Obtaining information about an access behavior of a current user accessing content; Predicting, with a prediction module and in response to obtaining the information about the access behavior of the current user, at least one characteristic of the current user based at least on a model for the characteristic and the obtained information about the access behavior of the current user Model based at least on a set of parameters determined by a training module based at least on information from multiple training sets and provided to the predictive module from time to time for update purposes, each training set of training sets information for a respective training user of a plurality of training users includes an access behavior of that respective training user as well as information about the characteristics of that particular training user, and providing information about the predicted characteristic of the current training user len so that information can be targeted to the current user at least on the basis of the information about the predicted characteristic. The computer program of claim 25, wherein the characteristic of the current user can be predicted by classifying a vector representing the information about the access behavior of the current user into one of at least two classes, the model being based on a support vector machine comprising the classification allows. The computer program of claim 26, wherein the set of parameters defines a support vector engine decision function applicable to the vector representing the current user's access behavior information for performing the classification. A computer program according to any one of claims 26-27, wherein for each training set of the training sets the information about the access behavior of the respective training user contained in the training set represents a training vector associated with a class comprised of the information about the characteristic of the training sentence contained in the training set respective training user, and wherein the set of parameters is determined by the training module to reflect a separation of training vectors from different classes according to a predefined optimization criterion. The computer program of claim 25, wherein the current user access behavior information for each content category of a set of one or more content categories includes information about how often accesses to content associated with the content category by the current user in one Time period, and wherein the information about the access behavior of a respective training user included in the training sentences for each content category of a set of one or more content categories includes information about how often accesses to content belonging to the content category by the training user in a time period were observed. The computer program of claim 29, wherein a decision as to whether content belongs to one of the content categories is made by a decision unit based on a received content classifier and / or based on an analysis of at least part of the content. The computer program of claim 29, wherein predicting the current user's characteristic is based on a presentation of the current user's access behavior information, wherein the information about how often the current user has accessed content belongs to content categories of the sentence is heard from one or more content categories, normalized, and wherein the set of parameters is determined based on training sets in which the respective information about how often the respective training user has accessed content, the content categories of the set of one or more content categories heard, were normalized. Computer program according to one of claims 25-31, wherein at least part of the information about the access behavior of the current user is stored in a cookie on a web browser of the current user and provided by the web browser. The computer program of claim 32, wherein the information about the access behavior of a training user of the training user is stored in a cookie on a training user web browser, wherein at least one domain for which the cookie was set matches a domain for which the one on the Web browser's current user's stored cookie has been set. A computer program according to any one of claims 25-33, wherein the information contained in a training set of training sets about the characteristic of a training user is obtained from a survey performed with the training user and associated with the information about the access behavior of the training user to the training set to obtain. A computer program according to any one of claims 25-34, wherein an actual address of the current user is anonymized by an anonymization unit and wherein the information about the access behavior of the current user is obtained with the anonymized address and not the real address. The computer program of claim 35, wherein the information about the predicted characteristic of the current user is provided with the anonymized address and not the actual address. The computer program of any one of claims 25-36, wherein the predicted characteristic information of the current user is formatted to match a format required by an advertising server that aligns the information to the current user, at least on the basis the information about the predicted characteristic executes at least. The computer program of any one of claims 25-37, wherein the provided information about the predicted characteristic of the current user is provided to an advertisement server that at least executes the alignment of the information to the current user based at least on the information about the predicted characteristic The computer program of claim 38, wherein the ad server triggers storage of the predicted characteristic information in a cookie on a current user's web browser. The computer program of any one of claims 25-39, wherein an updated version of the set of parameters is determined by the training module and provided to the prediction module each time a number of new and / or at least partially updated training sets have reached a predefined value. The computer program of any one of claims 1-17, wherein the prediction module and the training module are comprised of the same device. The computer program of claim 41, wherein the device is a web server or a portion thereof. The computer program of any of claims 25-42, wherein two or more characteristics of the user are predicted, each of the characteristics being predicted by a respective model. The computer program of claim 43, wherein a plausibility check is performed with respect to at least two characteristics predicted by their respective models. A computer program according to any one of claims 25-44, wherein the characteristic relates to demographic information, in particular information about one of sex, race, age, marital status, disabilities, mobility, home ownership, employment status, location and income, or information about an interest, in particular an interest on a specific topic or product type. A computer readable medium having a computer program stored thereon according to any of claims 25-45.

Images