JP2009536412A - Method and apparatus for performing real-time audience estimation and commercial selection suitable for targeted advertising - Google Patents

Abstract

Input measurements from the measuring device are processed as Markov chains, where the chain transition depends on the signal. At this time, desired information related to the measuring device is obtained by estimating the state of the signal at the target time. A nonlinear filter system can be used to present an estimate of the signal based on the observation model. The nonlinear filter system includes a nonlinear filter model and an approximate filter that approximates an optimal nonlinear filter solution. The approximation filter may be a particle filter or a discrete state filter that allows a signal to be estimated substantially in real time based on an observation model. In one application, the click stream (208) input to the digital set top box (200) of the cable television network is analyzed and the digital set top box (such that the advertisement (204) can be targeted to the user (205). 206) for the user (205).

Description

  The present invention provides a non-linear filtering innovation in which the observation process is modeled as a Markov chain, and the user configuration of user equipment devices in a communication network, eg, the number and demographics of television viewers in a digital set-top box (DSTB) environment. Use of embodiments of the present invention to estimate. Furthermore, the present invention optimally determines which set of resources, eg, commercials, should be inserted into the available network bandwidth based on sampling of an optimal conditional estimate of current network usage (eg, audience rating). Provide a method.

  This application is filed on May 2, 2006, under US Patent Act 119, entitled “Method and Apparatus for Performing Real-Time Estimation and Commercial Selection Suitable for Targeted Advertisements” (METHOD AND APPARATUS TO PERFORM). US Provisional Application No. 60 / 746,244 entitled "REAL-TIME ESTIMATION AND COMMERCIAL SELECTION SUITABLE FOR TARGETED ADVERTISING"). In addition, this application is a US patent application No. 11/331, filed on Jan. 12, 2006, entitled “Impression Model for Targeting Broadcast Network Resource Distribution”. 835 claims priority. The contents of both of these applications are incorporated herein as if they were described in their entirety.

  Overall, the distribution of commercials to television viewers has changed relatively little over the last 50 years. Market research firms and advertisers use historical Nielson (TM) audience rating information to determine what the target audience is looking at. This data presents a demographic analysis (usually based on age, gender, income, and ethnicity) as well as an estimate of the number of households that saw a particular broadcast of a television program at a particular time. Such data (other audience rating data) currently uses “People Meter” data that automatically monitors what programs are being watched when the user informs them that they are watching TV. Collected. These samples are relatively small and currently only about 8000 households are used to estimate total ratings across the United States. As viewers transition from broadcast to cable TV and the number of available TV channels increases as the number of TV receivers in a household increases, actual viewing of TV programs based on these small samples It becomes increasingly difficult to accurately estimate the person. As a result, cable television stations with a relatively small share cannot correctly estimate the audience rating, and therefore, the advertiser cannot correctly grasp the target layer on which profits will increase.

  As DSTB continues to spread due to increased demand for digital cable supply, more accurate information about each household can theoretically be obtained. That is, the set top box has access to information about what channel is being watched, how many hours it has been watched, and so on. This wealth of information, when correctly processed, can provide insight into household behavior. However, this information cannot directly provide the type of information desired by the advertiser, i.e. what type of people are watching at a particular time. Advertisers want to best post their advertisements to target audiences in order to reduce marketing costs and increase effectiveness. In addition, advertisers want to avoid the negative advertising costs associated with broadcasting commercials to inappropriate audiences. The key to providing advertisers with the ability to maximize their investment is to change the way viewers are counted, which can “change” the comparative value of the entire genre and the overall demographic category. (Gertner, J; Our Ratings, Ourselves; New York Times; April 10, 2005).

  Various systems have been proposed or implemented to clarify current viewers or demographics of viewers. Some of these systems were cumbersome asking the user to explicitly input identification information and demographic information. Other systems have attempted to reveal viewer behavioral characteristics based on information from various sources. However, these systems generally have one or more of the following disadvantages. That is, 1) these systems are more focused on who is in the home than who is currently viewing them, and 2) they may only provide rough information about part of the family. ) That these systems require user involvement, which is undesirable for some users and may involve errors, 4) these systems determine when there are multiple viewers, or No framework for accurately determining demographics in the presence of multiple viewers 5) These systems are entirely static and do not properly handle changes in household composition and demographics And / or 6) These systems employ sub-optimal techniques, require extensive training, and do not require excessive resources or are put into practical use It is to be limited.

  The present invention relates to analyzing observations obtained from a measuring device in order to obtain information about a signal of interest. In one application, the present invention provides a user input for a user equipment device in a communication network that determines information about the user of the user equipment device (eg, the user's audience classification parameter) (User input click stream input with respect to (DSTB)). Some aspects of the invention process damaged data observations, distorted data observations, and / or partial data observations received from a measurement device to infer information about a signal, and The present invention relates to providing a filter system that estimates a signal state at a target time substantially in real time. In particular, such a filter system can provide a practical approximation of an optimal non-linear filter solution based on certain constraints on allowed states or combinations inferred from the observation environment.

  In accordance with one aspect of the present invention, a method and apparatus (“system”) for developing an observation model for data or measurements obtained from an apparatus to be analyzed is provided. In particular, the system models the input measurements as a Markov chain whose transition depends on the signal. The observation model may take into account extrinsic information, ie external information for the input measurement (not necessarily independent of the input measurement). In one embodiment, the input measurement represents a DSTB clickstream. The click stream may represent a channel selection event and / or other input related to volume control, for example. In this case, the observation model may include programming information associated with the selected channel (eg, downloaded from a network platform such as a headend). Therefore, the click stream information is processed as a Markov chain.

  Thereafter, desired information related to the device can be obtained by estimating the state of the signal at the target time. In the example of analyzing a DSTB click stream, the signal is sent to the click stream, such as user configuration (including one or more users and / or associated demographics), and channel changes as described in more detail below. It may represent additional factors that affect it. Once the signal is estimated, the state of the signal in the past, present, or future time can be determined, for example, to provide user configuration information used in connection with the resource target system.

  In accordance with yet another aspect of the invention, the system generates a substantially real-time estimate of the signal state based on the observation model. At this point, a non-linear filter system is used to provide an estimate of the signal based on the observation model. The non-linear filter system may include a non-linear filter model and an approximate filter that approximates an optimal non-linear filter solution. For example, the approximate filter may include a particle filter or a discrete state filter that allows a substantial real-time estimation of the signal based on the observation model. In the DSTB example, the non-linear filter system identifies a user configuration that includes multiple viewers, and changes in potential viewers, for example, the entry of previously unknown people or the exit of previous users with respect to potential viewers. Can be adapted.

  In accordance with yet another aspect of the invention, the system uses signals obtained by applying a filter to the observation model to obtain object information for the observation model. Specifically, information on the past, present, or future time is obtained based on the estimated state of the signal at the target time. When analyzing the use of the DSTB, the identity and / or demographics of the user of the DSTB at a particular time are determined from the signal status. This information may be used to select a resource from among the resource options to be delivered in the DSTB and / or to determine and / or report the suitability of the delivered resource for the user receiving the resource, or both. For example, it may be used to “vote” or identify resources suitable for the next commercial or programming spot.

The above aspects of the invention are provided in any suitable combination. Further, any or all of the above aspects are implemented in connection with a targeted resource delivery system.
In one embodiment of the present invention, a system is provided for use with a target resource of a user of a user equipment device in a communication network, eg, a cable television network. This system develops an observation model based on input by one or more users regarding a user equipment device, models the observation model as a signal reflecting at least the user configuration of one or more users of the user equipment device with respect to time, The determination of the user configuration at the target time as the signal state and the use of the user configuration determined for the resources related to the user equipment device. Thus, filtering theory is applied to the input of a user equipment device, such as a click stream, to generate a signal indicative of user configuration.

  The input is modeled as a Markov chain. Further, the signal model can represent the user configuration as including a plurality of users. Thus, multiple users that are used to target resources and / or to better assess viewer size and composition (eg, to improve evaluation and payment for resource delivery), or both The situation can become clear. Furthermore, it is preferable that the signal model can express a change in the user configuration, for example, an increase or decrease in user viewers.

  A non-linear filter is defined to obtain a signal based on the observation model. In this regard, the signal may represent a household's user composition over time, and audience classification parameters (eg, current demographics for one user or multiple users) may be determined as a function of signal state at the time of interest. In order to make a practical estimation of the optimal nonlinear filter solution, an approximation filter is provided that approximates the behavior of the nonlinear filter. For example, the approximate filter may include a particle filter or a discrete state filter. Furthermore, the approximate filter may implement at least one constraint on one or more signal components. In this regard, the constraint may work to treat one component of the signal as being invariant with time when the variation of the second component is allowed. Further, the constraint condition operates to treat at least one state of the first component as abnormal or to treat some combination of different signal component states as abnormal. May be. For example, in the case of a DSTB click stream, the occurrence of a click event indicates that there is at least one person. Therefore, only user configurations corresponding to the presence of at least one person are allowed when a click event occurs. Revenue may be related to location by other acceptable or unacceptable combinations. The constraint may be performed in connection with a finite space approximation filter. For example, a value that is likely to be in an abnormal cell may be moved to another location, for example, in proportion to a legitimate cell in the vicinity. In this way, the approximate filter quickly converges to a valid solution without requiring excessive processing resources. If the constraint serves to determine that at least one calculated latent state is abnormal, the approximate filter may redistribute one or more counts associated therewith.

  The approximate filter may function to prevent convergence to an abnormal state. Thus, the approximate filter is logically impossible or unlikely (click event when no user is present) or considered abnormal by convention (revenue range not allowed in a given location) Designed to avoid convergence to user configuration for DSTB. In one embodiment, this is accomplished by adding a seed count to the legitimate cells of the discrete spatial filter to prevent convergence associated with abnormal cells.

  Preferably, the user configuration information is determined at the digital set top box. That is, user information is calculated in a digital set top box and used for voting, resource selection, and / or reporting. Alternatively, the clickstream data may be directed to another platform where user configuration information may be determined, such as a headend, for example, with sufficient messaging bandwidth and limited DSTB processing resources. As yet another method, user configuration information (eg, as opposed to resource voting information) may be sent to the headend or other platform used in selecting the content to insert.

  The determined user configuration information may be used by the resource target system. For example, this information may be provided to a network platform such as a headend that can insert resources into the content stream of the network. In this regard, the platform may select resources to insert into the available network bandwidth using inputs from multiple DSTBs. Additional information, such as information representing the value for each resource delivery user, may be used in this context. The platform may process information from a plurality of user equipment devices as an observation model, and apply an appropriately configured filter to the observation model to estimate the overall configuration of the network viewer at the target time.

  In accordance with another aspect of the invention, probability control theory is applied to targeted resource distribution, eg, dynamic audience classification and / or television advertisement selection issues. Traditionally, stochastic control theory cannot directly calculate a signal or function, but can only estimate it based on observations that may contain noise or observations that may be incomplete Has been applied. In such a situation, the measured values from the measuring device can be processed by probability control theory to estimate a signal from which state information can be determined. For example, measurements from an input device such as a click stream from a remote control are considered as noisy observations, and probabilistic control is used to estimate the signal representing the household or viewer and the behavioral form of the input item to the input. Processed using. In the probability control, the signal can be traced so that the state of the signal at a specific time represents the configuration and form of the viewer at that time. This information can be used, for example, to select advertisements that are targeted for delivery by matching the audience classification parameters to the available advertisement target parameters.

In order that the invention and other advantages of the invention may be more fully understood, a detailed description thereof will now be given in conjunction with the drawings.
In the following description, the present invention will be described in the context of a targeted resource distribution (eg, targeted advertising) system for cable television networks, and the present invention will be embodied in the context as described herein. Offer a special advantage. However, it will be understood that the various aspects of the invention are not limited to this background. More precisely, the scope of the present invention is defined by the appended claims.

  Various advertising systems targeted for cable television networks have been proposed or implemented. These systems are generally based on understanding the current viewer configuration so that the commercial can be adapted to the viewer to maximize the value of the commercial. It will be appreciated that the structure and functionality of the present invention for identifying current viewer classification parameters (eg, demographics) may be useful in various such systems. Accordingly, a resource distribution system with a specific target will be referred to below for purposes of explanation, but it will be understood that the present invention is more broadly applicable.

  A targeted resource delivery system in connection with the fact that the present invention may be employed is described in the aforementioned US patent application Ser. No. 11 / 331,835 filed on Jan. 12, 2006. . For simplicity, all details of this system are omitted here. In general, in this system, multiple resource options are provided at a given time in a given programming channel. Although this may be targeted at various types of resources, as described in the description, targeted advertising (eg, targeted at commercials) is an exemplary application and is described herein. Used as a convenient and simple reference. Thus, a given programming channel may be supported by multiple resource (eg, advertising) channels that provide advertising options for one or more advertising spots at commercial times. The DSTB serves to switch (from the viewer) inconspicuously to the appropriate advertising channel during commercial hours to provide targeted advertising to the current viewer.

  The viewer identification structure and functionality of the present invention can be used in a variety of ways in the targeted resource delivery system described above. In the above system, an advertisement list including target parameters is sent to the DSTB prior to the commercial time. The DSTB determines the classification parameters for the current viewer, matches these classification parameters to the target parameters for each advertisement in the list, and sends a “vote” for one or more advertisements to the headend. The headend collects votes from multiple DSTBs and incorporates optimized advertisements into available bandwidth (which may include programming channels and multiple advertising channels). At commercial time, the DSTB selects the “route” of the ad group and delivers the appropriate ad. After this, the DSTB can report what advertisements have been delivered along with goodness-of-fit information indicating how much the actual viewer matches the target parameter.

  The present invention can be directly executed in the resource distribution system that defines the above-described objects. That is, using the techniques described herein, audience classification parameters for the current audience can be determined in the DSTB. This information can be used for voting, advertisement selection, and / or fitness determination as described in the above pending applications. Alternatively, the following description describes a headend advertisement selection system based on filter theory that is an alternative to the voting process. As yet another alternative, clickstream information can be provided to the headend, or another network platform, where audience classification parameters may be calculated. Thus, audience classification parameters, ad selection, and other functionality can be modified and distributed in various ways between the DSTB, headend, or other platforms.

The next section is divided into several parts. In the first part, the background of relevant nonlinear filter theory is explained. In part 2, the structure and model classes are described.
1.1 Non-Linear Filtering In order to properly solve targeted advertising audience (potential and current) problems, one may focus on the mathematically optimal field of filtering.

1.1.1 Overview of Traditional Non-Linear Filtering Non-linear filtering is a method of performing some non-linear random dynamic process (usually "" in real time based on damaged, distorted, or partial data observations of a signal. It deals with an optimal estimation of past, present and / or future states (referred to as signals). In general, X _i is a certain probability space

This is a solution to a martingale problem. Observations are usually made at discrete times t _m and sensor functions

Depends on the stochastic signal using. In practice, conventional theories and methods are built under this kind of observation, and the measured values are a distorted (due to nonlinear function h) sample, a damaged (due to noise V) sample, a partial (due to signal state) A sample) (depending on the possibility of dependence on only part h). Optimal filters provide conditional delivery of signal states, assuming that the observations are valid up to the present time.

The filter presents optimal estimates not only for the current state of the signal, but also for past and future states, as well as the entire path of the signal.

Here, 0 ≦ t _r ≦ t _s <∞.

  In certain linear environments, an effective optimal re-regression equation is available. Signal is Ito stochastic differential equation

Where A is linear and B is a constant. Furthermore, the observation function is

Where

Is an independent Gaussian random variable. This formula is known as the Kalman filter. The Kalman filter is very effective in making the estimation, but its application is inherently limited because of the exact description of the signal in the observation process. If the signal dynamics are non-linear, or if the observations have non-additive noise and / or correlation noise, the Kalman filter presents a sub-optimal estimate. As a result, other methods are sought that present optimal estimates in these more general scenarios.

  Equations for optimal nonlinear estimation have been available for decades, but until recently they proved to be almost useless. The optimal equation is infeasible on a computer because it requires the use of infinite memory and computational resources. However, in the past decade, an approximation of the optimal filtering equation has been generated that overcomes this problem. These approximations are usually asymptotically optimal. That is, the approximate expression converges to an optimal solution as the amount of resources used for calculation increases. The two most common methods of this kind are the particle method and the discrete space method.

1.1.2 Particle Filter The particle filtering method is

Where N _t is the number of particles used at time t, with the generation of an independent copy of the signal denoted by (called “particles”). These particles are generated over time according to the probability law of the signal. After this, each particle has a weight value

Are effectively incorporated to form a series of observations (Y ₁ ,..., Y _m ). This can be done so that the weight after the m observations is the weight after m-1 multiplied by a coefficient that depends on the _mth observation Ym. However, these weights are always very non-uniform, that is, many particles (particles with relatively small weights) are no longer important and do very little except to waste computer cycles. Rather, if we resample the particles by eliminating these particles and reducing the number of calculations to keep the number of particles down, both the particle positions and weights are meaningful for all conditional delivery calculations. , While no statistical bias is introduced in this adjustment. Early particle methods tended to do too much resampling, introducing excessive resampling noise into the particle system to degrade the estimate. After resampling, the particle weight after m observations is

It is assumed that At this time, the approximate expression of the particle filter for conditional delivery of the optimum filter is as follows.

For N _t → ∞, the particle filtering estimate is an optimal nonlinear filter estimate.

In US Patent Application No. 2002/0198681, entitled “Flexible Efficient Branching Particle Tracking Algorithms”, which is incorporated herein by reference, particles are stochastically replicated, destroyed, or invariant at each time step. To be kept. Based on the weights calculated for the current time step only (W _m (ξ ^j )), the particles are modified according to the following routine.

  1.

Then the particle ξ ^j

Make a copy and probability

Make another copy with

  2.

Then probability

The particles having

3. An unbiased control algorithm is run to return the particle count to the total amount that existed before resampling.
This “careful” approach is an improvement over previously known particle filters, and this application includes historical refinements along with a system that effectively implements the algorithm, where the current state is merely Rather, the route of X up to time t is to be estimated. Further improvement by reducing performance degradation and improving computational efficiency is introduced in US Pat. No. 7,058,550, entitled “Selective Resampling Particle Filter”, which is incorporated herein by reference. This method performs pair-wise resampling as follows.

  1. For maximum weighted particle j and minimum weighted particle i

While meeting
2. Probability of the state of particle i

Set j to have the probability of the state of particle j

Set to i with

  3. The weights of particles i and j

Reset to.

  In this method, the control parameter p is introduced in order to appropriately suppress the total amount of resampling performed. As described in application No. 2005/0049830, this value is dynamic over time to suit the current value and specific application of the filter. The application also includes an efficient system for storing and calculating the amount required for this algorithm in a computer.

1.1.3 Discrete Spatial Filter Discrete space and amplitude approximation are used when the signal state space is on some bounded finite dimensional space. Discrete spatial filters are described in detail in US Pat. No. 7,188,048, entitled “Refining Stochastic Grid Filter” (REST filter), which is incorporated herein by reference. In this form, the state space D is divided into discrete cells η _C. For example, this space may be a d-dimensional Euclidean space or some counting measure space. Each cell has a “number of particles” (

This is used to form a conditional distribution of discrete spatial filters.

The number of particles in each state cell is changed according to the signal operator and the observation data being processed. When the number of cells becomes infinite, the estimated value of the REST filter converges to the optimal filter. For clarity, this application considers discretizing the filtering equation directly rather than discretizing the signal and devising a feasible filtering equation for the discretized signal.

  In application 2005/0071123, the present invention describes a binary index with dynamic interleaving that organizes cells having a data structure for efficiently and recursively calculating a conditional estimate of a filter based on real-time processing of observations. A tree is used. Although this structure is suitable for certain applications when the state space dimensional complexity is low, the overhead of the data structure can reduce the usefulness of the method.

1.2 Stochastic control To properly solve targeted commercial selection problems, attention should be paid to the mathematically optimal field of probability control.

  Conceptually, a direct discretization particle method is devised to approximately solve the stochastic control problem on a computer. However, these have not yet been done, or at least not widely recognized. Instead, the implementation generally solves the discretization problem after discretizing the entire problem.

2.1 Targeted Advertising System Configuration FIG. 1 depicts the entire targeted advertising system. This system includes a head end 100 that controls one or more digital set-top boxes 200. The DSTB 200 is attempting to use the DSTB filter 202 to estimate the conditional probability of potential viewer status in the household 205, including the current members of the household watching the television. The DSTB filter 202 uses a pair of models 201 describing a signal (household) and an observed value (clickstream data 206). The DSTB filter 202 is initialized with the setting value 302 downloaded from the head end 100. Also, the DSTB filter 202 uses program information 207 obtained from the accumulated program information 208 (which may be current, proximity past, or future) to estimate household status. To do.

  The DSTB filter 202 passes the conditional distribution, or an estimate obtained from the conditional distribution, to the commercial selection algorithm 203, which then outputs the filter output, the downloaded commercial 301, and the viewer estimate. A commercial 204 is determined to be shown to the current viewer based on some rule 302 whose value defines a given acceptable commercial. Commercials that are shown to viewers are recorded and stored.

  DSTB filter 202 estimates and commercial delivery statistics and other information are randomly sampled 303 and collected 304 to provide information to headend 100. This information is used by the headend filter 102, which calculates conditional delivery on aggregate capabilities and the actual audience rating for the set of DSTBs with which it is associated (assuming resources are available). . The head end filter 101 presents the estimated value using the collected household and the DSTB feedback model 101. These estimates are used by the headend commercial selection system 103 to determine the commercials to be passed to the set of DSTBs controlled by the headend 100. The commercial selection system 103 also takes into account valid market information 105 regarding the current commercial contracts and the economics of these contracts. The commercial 301 resulting from the selection is later downloaded to the DSTB 100. The commercial selected for download affects the level setting 104, which imposes constraints on some commercials that are shown to certain individuals.

The following two sections describe certain details of the system.
2.2 Description of Household Signal and Observation Model In this section, a general signal and observation model is described along with examples of possible embodiments of this model.

2.2.1 Signal Model Description In general, household signals are modeled as various individual and household forms. In a preferred embodiment, this household represents people who are likely to watch a particular television using DSTB. Each individual at a given time t (denoted by X ⁱ ) has a state space sεS, where S represents the set of properties that we want to measure for each person in the household. For example, in one embodiment, suppose that one wishes to classify an individual's age, gender, income, and viewing status. Age and income may be considered actual values or discrete ranges. In this example, the state space is determined as follows.
S = {0-12, 12-18, 18-24, 24-38, 38+} × {male, female} × {0- $ 50,000, $ 50,000 +} × {yes, no}
The state space of the household members is

Where k represents the number of individuals and S ⁰ represents a single state with no individuals. Household members

Has a random number of members that change over time, where n _t is the number of members at time t. Since the order of the members in this set is not important to the problem, the authors

Represent the household using the experience measure.

The household form depicts the current viewing “thinking” of the household that can substantially affect the generation of clickstream data. Current form r _t household is the value of the state space R. In one embodiment of the present invention, these forms may consist of values such as “normal”, “channel flipping”, “status checking”, and “favorite surfing”.

  Thus, a complete signal consists of household members and form as follows:

The signal state evolves over time by a rate-of-change function λ that probabilistically defines signal state changes. The probability of the state changing from state i to j after a certain time t is then given by

There is an individual rate of change function for each individual, the household members themselves, and the evolution of the household form. In one embodiment of the present invention, the rate-of-change function for individual i is the given individual, the signal experience measure, the current time, and some external environment variables.

You may depend only on.

The number of individuals in a household, n _t , varies with time depending on the birth rate and mortality rate. Birth rates and mortality rates can represent events that result in the transfer and transfer of one or more individuals to a household, as well as showing the birth of new life or the death of a living life. Birth rates and mortality rates are calculated based on the current status of all individuals in the household. For example, in one embodiment of the present invention, a rate of change function may be calculated that describes the likelihood that a single person will move a housemate or spouse into a household.

  In one embodiment of the invention, these rate of change functions are determined empirically by matching the estimated probabilities with valid demographic, macroeconomic, and expected state change values from viewing behavior data. Formulated as a mathematical equation using parameters. In another embodiment, age is deterministically evolved in continuous state space [0, 120].

2.2.2 Description of Observation Model In general, an observation model consists of click stream information generated by the interaction of DSTB with one or more individuals. In a preferred embodiment of the present invention, only current and past channel change information is represented in the observation model. _Given a region of M channels, a channel change queue at time t _k of Y _k = (y _k ,..., Y _{k−B + 1} ) channels viewed in the past B discrete time steps is obtained. In a preferred embodiment of the invention, only the time at which channel changes are made and the changed channels are recorded to reduce overhead.

In the more general case, the viewing queue includes volume history etc. along with this current and past channel. In the above case, the viewing queue degenerates to the channel change queue.
Thereafter, the viewing queue changes from state i to state j at time t based on the signal state and some downloadable content D _t (indicated by p _{i → j} (D _t , X _t )). The probability of doing is determined. In a preferred embodiment, this downloadable content includes, for example, whether the program is an “action movie” or “home comedy”, and the program time zone, program start time, program broadcast Some program information detailing a description of the qualitative categories of programs currently available for each program, such as the current channel.

In the absence of a special form, an empirical method has been created to calculate the Markov chain transition probabilities. These probabilities depend on all members of the household and the current state of available programs. This method is validated using observed viewing behavior and the laws of large numbers of Varadarajan. Assuming P is a discrete probability measure that assigns probabilities to Ω = (ω ₁ ,..., Ω _K ), N independent copies of the experiment selecting the elements are obtained. At this time, the law of large numbers is as follows.

Here, ω ⁱ is an i-th result when elements are randomly extracted from Ω.

In one embodiment of the present invention, the method focuses on calculating probabilities for a size 1 channel queue (ie, Y _k = y _k ). The probability of observation, i.e. the probability that the two viewing queues will switch in the next discrete step, may be first calculated by determining the probability of switching to a particular channel within this category after determining the probability of the program switching category. The first step is to calculate the relative percentage of category changes caused by channel changes and / or program changes on the same channel, often offline. To perform this calculation, all possible member states X _i are discrete such that f (X _t ) = π _t for some possible X _t with some π _t ∈Ｘ. It is mapped to the state space Π. A certain number of categories Cε {C ₁ , C ₂ ,. . . , C _K }. Further, each viewer record represents a certain period Δt, and each of the three sets of viewing records V (k) = (π, C ¹ , C ² ), k = 1, 2,. . . , Nv, there are Nv viewer records that contain information about the discretization state (π) and the category about the start (C ¹ ) and end (C ² ) of each household. At this time, the following equation is calculated for each π∈Π and ∀C _I , C _J ∈C.

When optimal estimation system is operating in real time, the probability for the category transition from C _I to occur in a given time step to C _J, by determining first the probability of the current category change programs available as given Calculated from the following equation.

Here A operates with all possible category switching possible based on the current program available. Thereafter, this probability is converted into a necessary channel transition probability by the following equation.

Where n _t (J) is the number of channels with programs that fall into category J at the end of the current time step.

  Another probability measure used is to calculate the “popularity” of the channel instead of the transition between channels at each discrete time step. This method is used to provide this form by simply summing the transition probabilities for a given category.

Furthermore, this probability is converted into a necessary channel transition probability by using an instance of the multiplication law as shown in the following equation.

Where n _t (J) is the number of channels with programs that fall into category J at the end of the current time step.

  In one embodiment of the invention, some or all of the categories are themselves programs, given the highest level of granularity. In another example, it is preferable to have a broad category to maintain the number of probabilities that need to be preserved.

2.3 Optimal estimation using Markov chain observations In the conventional filtering theory summarized above, the observed values are distorted and damaged by the following equations, and are partial signal measurements.

Where tk is the observation time for the k-th observation,

Is the driving noise process, ie, continuous time-varying. However, in the case of the DSTB model described by the authors in the immediately preceding measure, Y is a discrete-time Markov chain whose transition probability depends on the signal. In this case, since the new state Y _k can depend on the previous state, the above standard theory is invalid. In this section, a new similar theory and system is presented that solves the problem when the observations are Markov chains. One universality that deserves system attention is that it may be allowed to transition only to a subset of all states depending on the state in which the Markov chain observations are currently placed. This is a useful feature in targeted advertising applications because much of the data immediately before the viewing queue may remain in the viewing queue after observation and after some new data is inserted. . Although this is clearly general, we will mention this in reviewing targeted advertisements to facilitate identification.

Assume that there is a Markov signal X _t with generator L and the initial distribution [nu. To be precise, the signal is defined to be a unique D _E [0, ∞) process that satisfies the following (L, ν) martingale problem.

and

Is a martingale of all φεD (L).

Depends on X _t and some extrinsic information D _t {1, 2,. . . Estimates the conditional distribution of X _t based on discrete-time Markov chain observation M} values. To make it easier to understand,

Is P (v _k = i) = 1 / M, where i = 1, 2,. . . , M and k∈Z are a series of independent random variables unrelated to observations and observations.

Is a finite state space of available events {1,. . . , M} and Y _k = (y _k , y _k−1 ,..., Y _{k−B + 1} ) at time t _k , where

Has a homogeneous transition probability p _{i → j} (D _t , X _t ) for transitioning from state i to state j at time t {1,. . . , M} transition between the values of ^B. Where D _t and X _t are the current state of the appropriate extrinsic information and the signal state at the time of a possible state change.

  To make it easier to understand,

And set the following formula.

here,

It is.

For convenience of description, it is defined as Z ₀ = 1. At this time, the following equation is shown by some mathematical calculations.

here,

Where g = 1 and g = f respectively, the denominator and numerator of (6) are both

Note that it is calculated from here,

It is. Here, the following equation is required.

However, if the set of functions is sufficiently rich,

With further calculations

Satisfies the following equation.

Where all t∈ [0, ∞) and φ∈D (L), where

and

It is.

2.4 Filtering approximation In order to use the above differentiation in a real-time computer system, an approximation must be made so that the resulting equations can be implemented in a computer architecture. In order to use a particle filter or a discrete spatial filter, a different approximation must be made. These approximations are clarified in the following sections.

2.4.1 Approximation of particle filter It is only necessary to approximate the following equation by equation (6).

Where independent signal particles each having the same law as the past signal

Is assumed to define the following weights.

Then, this equation becomes the following equation according to deFinetti's theorem and the law of large numbers.

2.4.2 If the state space of the approximate X _t of discrete space is set to be E, each

L _N and M _N are

When

and

To satisfy.

in the case of,

Is

And all the discrete states are in different cells, in other words it is only expanding and contracting with the age reaching α = αN → 0. here,

As

And intuitively justify this below.

Is

All i∈D _N and

The test function

Is substituted into (10), the following equation is obtained.

here,

It is.

  At this time, the above equation becomes as follows.

2.5 Improvement of Stochastic Grid Filter by Discrete Finite State Space US Pat. No. 7,188,048 describes in detail the general form of a REST filter. This method and system has proven useful for several applications, particularly the Euclidean space tracking problem and the discrete count measure problem. However, several improvements have been found in this method, which provides a dramatic reduction in memory and computational requirements for embodiments of the present invention. New methods and systems for REST filters are described herein, and signals can be modeled by discrete and finite state spaces. An example using an advertising model targeted for clarity is provided, but this method is used for any problem characterized by the environment described below.

2.5 Environmental description In some problems, the signal is zero or more

And zero or more forms

Consists of. For example, in targeted advertising, one embodiment of a signal model is of the form X _t = (X _t , R _t ), where X _t is the target experience measure (ie, specifically household) Only one form. In addition, each target and form is discrete and has only a finite number of states, and there are a finite number of targets and forms (thus, possible combinations of targets and forms are finite). A finite number of combinations need not be all possible combinations, only a finite number of legal combinations are required. For example, a finite possible type of household (meaning a household with specific demographics inside) may be derived from census information that depends on the region at a relatively granular level. Rather than having every possible combination of individuals (up to some maximum household member ^nMAX ), only combinations that may be found within a given geographic region are considered valid and included in the state space There is a need.

  In these limited problems, some of the target and / or morphological states may be unchanged during the short period of time when the optimal estimate is being made. In these cases, such state information remains constant, but other parts of the state information continue to fluctuate. In one embodiment of the household signal model, the age, gender, income, and education level of each individual in the household is constant because these values vary over time, but the DSTB estimation is performed over several weeks. It may be considered that there is. However, the current viewing state of the household form information changes in a relatively short time frame, and as a result, the state remains fluctuating in the estimation problem. The invariant part of the signal

The fluctuation part of the signal

It will be expressed as N possible immutable states (i th such state

) And M _i possible variable states (jth state) for the i th invariant state

Is represented).

2.5.2 Overview of REST Finite State Space System FIG. 2 depicts a preferred embodiment of a REST filter in a finite state space environment. A REST consists of a set of invariant state cells, each of which represents one possible set of objects and forms for the signal, along with their invariant state characteristics. Each invariant cell contains a set of variable state cells, each variable state cell representing a possible time-varying state of a given invariant cell. Implicitly, variable cells contain invariant state information of their parent invariant cells, meaning that each variable cell represents a specific potential state of the signal. The immutable cell itself represents only the collective container object and is used for convenience. The collection of variable and immutable cells may be stored on the computer medium in the form of an array, vector, list, or queue. Cells that do not have a particle count at a given time t are removed from such containers to reduce space and computational requirements, but a mechanism is needed to later reinsert such cells. .

  As shown in FIG. 3, each variable state cell has a number of particles.

including. This particle number represents the discretized amplitude of the cell. As already mentioned, this amplitude is used to calculate the conditional probability of a given state. Each variable cell also has a set of virtual clocks

including. These virtual clocks represent possible state changes from a given state cell. For each variable state cell, there are Q _{i, j} possible state transitions. In this environment, all valid state transactions occur within the same immutable state cell. To account for simultaneous changes in conditional delivery of REST filters, particle number delta

A temporary counter, entitled, is used to store the number of particles that are added or deleted for a given variable state cell when all cells have been sequentially processed. Status

A cell having a valid state transition from a variable state cell having is said to be its neighbor.

  As mentioned earlier, immutable state cells are containers that are used to simplify the processing of information. Number of particles in each invariant state cell

Is a set of the number of child variation state cell particles. Similarly, the virtual time clock of an invariant state cell is a collection of all clocks from the variable cell. This aggregation facilitates the evolution of the filter, since invariant states that do not have a current particle number can be skipped at various stages of processing.

2.5.3 Evolution of REST filter FIG. 4 depicts a typical evolution of a REST filter. This evolution method updates the conditional delivery of a filter over a period of time dt by moving particles between adjacent cells using virtual clock values. The movement of particles between adjacent cells is known as an event. (By the way, the authors frequently replace particle movements with new births and deaths so that more birth and mortality cancellations can be made.) Simulated together to reduce overhead. The number of events to simulate is based on the sum λ _t of all virtual clocks for all cells. FIG. 5 shows a method for determining the number of particles moving to each neighbor. When the event simulation is complete, the particle count can be updated and the virtual clock is sealed to indicate that the filter state has changed.

  Compared to the previous method described in US Pat. No. 7,188,048, new steps have been added to improve the effectiveness of the filter. Specifically, the adjustment of the number of cell particles is performed here prior to the push-down observation method, and a draft back routine is added prior to particle control. In some problems, some states may not be current signal states based on observation information. For example, a household must have at least one member currently watching whether a channel change has been recorded. In such a situation, all particles in the invalid state must be redistributed in proportion to the valid state. Therefore, it should be redistributed

If there are particles, all effective state cells are

Will receive the particles of the probability

Will receive further particles with When using this type of observation-based adjustment, the rate of change that defines the evolution of the signal would have to be changed appropriately to be consistent with the use of observation data in such a way.

  A driftback method has been added to improve the robustness of the REST filter. In this method, the function in the variable state cell based on the initial delivery ν of the signal

Use particles. The number of particles added to each cell depends on the time of day, the given cell, and the overall state of the filter. This method prevents the filter from converging to one or more invariant states if it cannot return from an illegal position.

2.6 Headend Estimation In order to maximize the advertising revenue of multiple service operators, it is very important to determine which commercials should be delivered to the aggregate DSTB. Pricing for a specific commercial slot as more information about the actual audience rating of the commercial can be obtained based on the conditional distribution of the DSTB-based asymptotically optimal nonlinear filter (or a conditional estimate derived from this distribution) Is dynamic, improving total revenue.

  To take advantage of this potential, an estimate of the total household, including the number of people in each demographic, etc. is made at the headend based on random sampling of conditional DSTB estimates. The following model includes a preferred embodiment of the present invention.

2.6.1 Headend Signal Model The headend signal model consists of characteristic information relating to potential TV viewers with a DSTB box connected to a particular headend and current TV viewers. A state space S representing such a set of properties for a single individual is defined. In one embodiment of the present invention, this space may be composed of an individual's age group, gender, and recent viewing history. In order to track individuals, let C ⁰ = φ be a household type with no individuals and Cn be a set of household types with n individuals.

At this time, the set of households is the union of households with n people

It becomes. In reality, there will be a maximum household N that we can handle, and we have a household state space.

Set to be. Here, N is a large number.

  In order to process the estimates returned from the DSTB by the random sample mechanism, we also want to track the current channel for each DSTB. That is, each DSTB state including a potential household audience rating, viewing status, and current channel is selected from the following equation.

Here, there are M channels that can be changed by DSTB.
We are interested in how many DSTBs are in state dεD, not a single DSTB or which DSTB is in a particular state. Therefore, the number of DSTBs in each category dεD is counted so that the signal X is tracked according to a finite counting measure.

  In an embodiment of the present invention, it is possible to keep track of the possible number of DSTBs in each category to minimize computational requirements. In such a case, size 0 atoms are used so that the total number will again add up to the maximum number of DSTBs. For example, assume that there are 1 million DSTBs. At that time, 100,000 atoms (each consisting of a = 10 DSTBs) will be delivered to D. M (D) represents the counting measure for D;

Is assumed to denote a subset of M (D) with exactly 100,000 atoms. The signal will evolve mathematically according to the martingale problem.

Where t → M _t (f) is

Is a martingale for each continuous, bounded functional f, and L is an operator that should be determined primarily from the DSTB rate and the natural premise that the household acts independently.

Households that provide demographics in open mode are not considered part of the signal.
2.6.2 Headend Observation Model In this specification, two observation models are described. One is a model related to random sampling of DSTB, and the other is a model related to distribution statistics.

In the random sample observation model, the channel and audience rating are considered with X being the author's signal as in the previous section, and V _k represents random sampling at time t _k of the sampling process. Precisely, assuming there are M DSTBs for a particular headend, a DSTB that at least one person is currently watching will provide a sample with a constant probability of 5%. Shall. At this time, V _k is a matrix with random rows, each row consisting of M inputs with exactly one non-zero input corresponding to the index of the particular DSTB providing the sample. The number row is the number of DSTBs that provide the sample. The position of the non-zero input is obviously evident in the row and will be chosen uniformly for possible permutations that reflect the actual sampling to be performed.

  here,

It is assumed to be the column vector of any conditional distribution audience estimate of M stand DSTB at time t _k and the corresponding channel. At this time, this observation process is as follows.

Here, V _k performs random sampling, and h is a function that provides information selected for feedback.

In the case of the collective advertisement distribution statistical model, the authors have a time series function H _kj that provides the number of various advertisements distributed in the past from time t _{k to} t _j . There is often a small amount of noise W _kj , which is due to the fact that some DSTBs may not return any information due to a temporary malfunction (ie, “stop observation”) and normal delivery This is based on the fact that the estimated audience rating used to determine that is not guaranteed to be correct.

  The second observation information from the collective distribution statistics is as follows:

Here, j can vary in the spot segment reporting period, t _k is the reporting period.

2.6.3 Head-end filter The signal related to the head-end has a DSTB probability distribution.
2.7 Headend Commercial Selection In certain embodiments of the present invention, there may be other information that is also used to estimate the total audience rate. For example, aggregate (and possibly delayed) ad serving statistics can be used in the “exposure mode” where households choose to provide their status information (demographics, psychological aspects, etc.) in exchange for certain compensation along with DSTB's estimated audience ratings. Is also provided.

  In this setting, the commercial contract is modeled as a graph of incremental profits in terms of contract content, available resources, and future signal conditions. The authors refer to these graphs as “contract graphs” that arrive at a rate that depends on the contract content, signal conditions, and economic environment. Part of the contract includes the following items:

The number of times a commercial can be aired, possibly in thousands (possibly including maximum and minimum limits),
Day / week time range for commercials,
Target demographics for commercials,
Specific channels or programs that can broadcast commercials,
The customer who created the contract,
Some of the above items are optional.

  Random arrival of the contract graph is indicated by “contract graph process”. In addition, the allocation of resources to the contract graph process (not necessarily the maximum value that can be allocated to the contract), given the state (current and future) and environment, the allocated resources are available resources, ie It is called a “feasible choice” if it does not exceed available commercial spots across various categories. By the way, based on the fact that these limited resources are reduced when a person accepts a contract, the potential benefit of the present versus the future is modeled by an “utility” function. This utility function shows the trend of available contract graphs (random arrivals now and in the future) and returns a convincing number indicating the profit in terms of money and other forms. Since future movements in the contract graph are irregular, the utility function simply yields the maximum profit without taking into account the deviation from the expected profit in order to avoid the risk of a large margin due to maximization. Is not possible.

  In order to make an optimal commercial selection, it is necessary to define the following models, described as Headend Signal Model, Headend Observation Model, Contract Generation Model, and Utility (Profit) Model.

2.7.1 Contract Model The resulting commercial contract is modeled as a marked point process on the contract graph. Contract arrival rates depend on past contracts signed and external factors such as economic conditions.

Let l denote the Lebesgue measure. Where C denotes the space of a possible contract graph with a certain topology, {η _t , t ≧ 0} denotes the counting measure probability process for the arrival of the contract graph up to time t, and ξ is Let Poisson measure on C × [0, ∞) × [0, ∞) with average measure ν × l × l. Further, λ (c, η _{[0, t)} , t) is such that when η _{[0, t)} records the arrival of the contract graph from time 0 to t (but not including time t), a new contract is created. Suppose that the ratio (with respect to ν) will comprise the contract graph cεC at time t. At this time, modeling is performed by the following stochastic differential equation.

The above contract details may be changed upon acceptance of the contract. As a result, the contract content is modeled so that it can evolve over time depending on the external environment.

2.7.2 Description of Utility Function For the sake of clarity, let R (D _S ) be a resource that can be used now and in the future based on the program information D _S that can be downloaded at time s.

  The authors cannot accept all the contracts that arise and must decide whether to accept the contract without considering the future. The authors present acceptable choices as feasible choices so that future contracts and future observation information are not used to determine each resource allocation. In light of the notation in the previous section, nt is assumed to represent the number of various types of contracts that have arrived by time t (including time t), and

Where Q represents the set of all potential customers and {l _S , s ≧ 0} is the selection process, ie allocate resources to each contract c. At this time, if l _s ≦ R (D _s ) for each s ≧ 0, {l _S , s ≧ 0} is an acceptable choice, and l _s uses both future contracts and observation information. No, i.e., for each s ≧ 0

Can be measured. Here, γ _t ^(l) represents the profit obtained up to time t by the acceptable selection l. For simplicity, let Λ be the set of all such acceptable choices.

  The utility function J is changed to balance the current profit and the future profit so as to obtain a very high profit in a specific contract at a risk of an unreasonable profit or a low profit. In order to start effortlessly, the weight of future earnings will be reduced exponentially. In addition, the authors will include a decentralized state to avoid being overly aggressive. One embodiment of the resulting utility function is:

However, a small constant λ and a> 0. At this time, the purpose of the commercial selection process is to maximize E [J (X, l)] for l∈Λ. Such an objective can be solved using one or more asymptotic optimal filters.

  The foregoing description of the present invention has been presented for purposes of illustration and description. Furthermore, the description is not intended to limit the invention to the form disclosed herein. Accordingly, variations and modifications corresponding to the above teachings and skill and knowledge of the related art are within the scope of the invention. Furthermore, the foregoing embodiments describe the best mode known for the practice of the invention, and other persons skilled in the art will recognize that the invention may be used in such and other embodiments for specific applications of the invention. Alternatively, it can be used in various modifications required for usage. The following claims should be construed to include other embodiments to the extent permitted by the prior art.

1 is a schematic diagram of an advertising system targeted according to the present invention. 2 shows a REST structure according to the present invention. 2 shows a cell structure for a filter cell according to the invention. 4 is a flowchart illustrating a filter evolution process according to the present invention. FIG. 6 is a block diagram illustrating a process of simulating an event according to the present invention.

Claims (32)

A method of using a target resource by a user of a user equipment device in a communication network, comprising:
Developing an observation model based on input by one or more users regarding the user equipment device;
Incorporating into the observation model a signal representative of at least a user configuration of one or more users of the user equipment device with respect to time;
Estimating the user configuration at a target time by approximation of a conditional distribution of given measurement data by the signal;
Using the estimated user configuration for resources related to the user equipment device.   The method of claim 1, wherein the input is a temporal user input click stream and the observation model models the click stream as a Markov chain.   The method of claim 2, wherein the observation model considers programming related information of network content indicated by at least a portion of the input.   The observation of the Markov chain further comprises processing the Markov chain using a mathematical model capable of transitioning to only a subset of the entire set of states, the subset depending on the current state of the Markov chain The method of claim 3.   The method of claim 1, wherein the modeling comprises modeling the observation model as a Markov chain with a k-step Markov chain.   6. The method of claim 5, wherein the transition function of the observed Markov chain depends on the position of the signal to be estimated.   The method of claim 1, wherein the signal is determined to represent the user configuration and another factor that provides the user input.   The method of claim 1, wherein the model of the signal can represent the user configuration as comprising multiple users.   The method of claim 1, wherein the model of the signal can represent a change in the user configuration.   The method of claim 9, wherein the change is a multiple user change associated with the user equipment device.   The method of claim 1, wherein the modeling includes defining a filter to obtain a probabilistic estimate of the signal based on the observation model and the measurement data.   The method of claim 11, wherein the modeling comprises defining a non-linear filter to obtain a probabilistic estimate of the signal based on the observation model.   The method of claim 12, wherein the modeling further comprises constructing an approximate filter that approximates an operation of the nonlinear filter.   The method of claim 13, wherein the approximate filter is a particle filter.   The method of claim 13, wherein the approximate filter is a discrete spatial filter.   The method of claim 1, wherein using comprises providing information based on the user configuration to a network platform that functions to insert resources into a content stream of the network.   The method of claim 16, wherein the information identifies demographics for one or more users of the user equipment device.   The method of claim 17, wherein the platform functions to aggregate user configuration information associated with a plurality of user equipment devices and to select one or more resources to insert based on the aggregated information. .   The platform functions to process information from a plurality of user equipment devices as an observation model and apply a filter on the observation model to estimate a composite configuration of network viewers at the target time. Item 17. The method according to Item 16.   The method of claim 17, wherein the platform functions to select a resource to insert based on the composite configuration and additional information that affects the delivery value of a particular resource.   The method of claim 16, wherein the information identifies one or more appropriate resources to deliver to the user equipment device based on the user configuration.   The method of claim 1, wherein using comprises selecting resources at the user equipment device for distribution to the one or more users.   The method of claim 1, wherein the using comprises reporting a fitness of resources delivered at the user equipment device for the one or more users. An apparatus used for a target resource of a user of a user equipment device in a communication network,
A port operable to receive input information relating to input by one or more users relating to the user equipment device;
Providing an observation model based on the input, modeling the observation model as dependent on a signal representing at least a user configuration of one or more users of the user equipment device with respect to time, and determining the user configuration at a target time as a signal A processor that operates to determine the state and use the determined user configuration for resources related to the user equipment device;
A device comprising:   25. The apparatus of claim 24, wherein the processor is operative to define a non-linear filter to obtain an estimate of the signal based on the observation model and the measurement data.   26. The apparatus of claim 25, wherein the processor is operative to construct an approximate filter that approximates the operation of the nonlinear filter.   27. The apparatus of claim 26, wherein the non-linear filter is one of a particle filter and a discrete spatial filter.   25. The apparatus of claim 24, further comprising a port for transmitting information to use for a target resource of another network platform, wherein the information is based on the determined user configuration. A method of using a target resource in a broadcast network,
Comprehensively analyzing the flow of data corresponding to a series of user inputs;
Applying logic to match the pattern described by this flow to characteristics associated with the user's audience classification;
A method comprising:   30. The method of claim 29, wherein the step of comprehensively analyzing includes building an observation model, and the series of user inputs is modeled as a Markov chain.   Applying the logic comprises using a non-linear filter model to extract signal estimates and distributions from the series of user inputs and using the signal to obtain the characteristics. 30. The method according to 29.   30. The method of claim 29, wherein applying the logic comprises performing an approximate filter to approximate the operation of the nonlinear filter.