DETAILED ACTION
Notice of Pre-AIA or AIA Status
The present application, filed on or after March 16, 2013, is being examined under the first inventor to file provisions of the AIA .
Response to Amendment
This action is in response to submission filed 08 July 2025 for application 16/230,620. Claims 1, 8, and 15 have been amended. Claims 3, 4, 10, 11, 17, 18, 23, and 24 are canceled. Claim 25 has been newly added. Currently claims 1, 2, 5-9, 12-16, 19-22, and 25 are pending and have been examined.
The 35 USC § 101 rejection has been reconsidered and withdrawn in view of the amendments made and arguments presented.
Response to Arguments
Regarding Applicant’s arguments, filed 08 July 2025, see pages 17 and 18, with respect to the rejections under 35 U.S.C. 103, Applicant believes (on Page 17) no amendments are necessary at this time with regard to the § 103 rejections and submits that the cited references, individually or in combination, do not teach at least the claim features of compiling…, generating view block-level features…, generating a two-dimensional (NxF) view block-level feature vector, for each return path data household, and implementing the neural network.
Examiner’s Response: Examiner disagrees that the cited references do not teach those limitations because as shown in the detailed rejection below the combination of cited references teach each and every element of the limitations. It is unclear as to which of these limitations Applicant would like further clarification on in Pages 17 and 18 as no arguments have been presented as to why Applicant thinks the references do not teach these limitations. It will be helpful if arguments were provided for Examiner to consider and respond more appropriately. Please see below for more responses to other arguments presented.
Regarding Applicant’s arguments, filed 08 July 2025, see page 18, with respect to the rejections under 35 U.S.C. 103, Applicant also maintains the arguments set forth in the response to office action filed May 23, 2024, with respect to the Sullivan and Zheng references.
Examiner’s Response: Applicant’s comment has been noted but there is no office action filed May 23 2024. The Final rejection is dated February 23 2024 and the Remarks filed by the Applicant to that office action is dated May 23 2024. Examiner has addressed the arguments from Remarks dated 23 May 2024 in the Response to Arguments section (Pages 2-7) of the Non-Final rejection dated 13 February 2025.
Regarding Applicant’s arguments, filed 08 July 2025, see pages 18 and 19, with respect to the rejections under 35 U.S.C. 103, nevertheless, without acquiescing in the Examiner's rationale for rejecting the claims under § 103, Applicant has amended the claims as indicated above and submits that the cited references, individually or in combination, do not teach at least the amended claim features of: " compiling the respective return path data into a set of view blocks that are contiguous in time and generated by the return path data household over a particular observation period; based on panel data reported from meters that monitor media devices associated with panelist households, training a neural network to process the view block-level feature vector of each return path data household generated from the return path data to predict demographic classification probabilities for the return path data households, wherein the training comprises: creating view blocks from the panel data for the panelist households, generating view block-level feature vectors for respective ones of the panelist households from the view blocks created for the respective panelist households, applying the view block-level feature vectors for the respective ones of the panelist households to the neural network according to a training procedure that adjusts internal parameters of the neural network to reduce an error between the predicted demographic classification probabilities output by the neural network and actual demographics known for the panelist households based on the panel data, and shuffling an order of the view blocks of the return path data households that are fed into the neural network during each of a plurality of training epochs of the neural network," as recited in claim 1 and as similarly recited in claims 8 and 15.
Examiner’s Response: Examiner disagrees that the cited references do not teach those limitations because as shown in the detailed rejection below the combination of cited references teach each and every element of the limitations. It is unclear as to which of these limitations Applicant would like further clarification on in Pages 18 and 19 as no arguments have been presented as to why Applicant thinks the references do not teach these limitations. It will be helpful if arguments were provided for Examiner to consider and respond more appropriately. Please see below for more responses to other arguments presented.
Regarding Applicant’s arguments, filed 08 July 2025, see pages 19 and 20, with respect to the rejections under 35 U.S.C. 103, Applicant argues specifically on Page 19 (Paragraph 2) that as one particular example, claim 1 as now amended recites "compiling the respective return path data into a set of view blocks that are contiguous in time and generated by the return path data household over a particular observation period." In Paragraph 3 Applicant states, This demonstrates that the view blocks in the present application are contiguous in time. Applicant specifically argues in Paragraph 4 that by contrast, Sullivan does not appear to acquire time-series data that is contiguous in time. Applicant continues to argue on Page 20 that thus, in practice, the tuning data in Sullivan is "discrete"-the entire process does not partition tuning events into blocks based on temporal sequences or extract them in a contiguous in time pattern. This can also be inferred from the "decision tree model" used in Sullivan. As is well- known to those skilled in the art, decision tree models typically rely on discretized or static features (i.e., statistics of independent tuning events) and lack the capability to model temporal continuity. They are generally ineffective at processing the "time-series structure" of view blocks. Therefore, under Sullivan's technical solution, it only acquires "time information for distinct tuning events" rather than time-series information organized sequentially. Moreover, due to the use of a decision tree model, Sullivan has no technical need to acquire time-series data during data collection. Consequently, based on Sullivan, those skilled in the art would neither conceive of nor find it necessary to acquire "a set of view blocks that are contiguous in time," as required in the present application. Furthermore, since Sullivan does not acquire any view block sequences that are contiguous in time, a person skilled in the art could not directly utilize any models requiring time-series information (such as the neural network model in Morfi) without compliant input data. Thus, the discrete data obtained by Sullivan is incompatible with the neural network input requirements of Morfi. Sullivan also provides no technical teaching regarding converting discrete data into continuous view blocks. Therefore, a person skilled in the art would not conceive of combining Sullivan with Morfi.
Examiner’s Response: Firstly, Applicant’s specific arguments are very useful in helping the Examiner understand Applicant’s perspective and are therefore appreciated. Applicants arguments have been fully considered but they are not persuasive and Examiner disagrees that Sullivan2 (US 20180152762 A1) does not teach the compiling limitation.
Firstly, Sullivan (US 20170064358 A1) is not relied upon to teach that limitation but instead Sullivan2 (US 20180152762 A1) is used to teach it.
Secondly, Sullivan2 clearly shows in Figure 3 that view block 326 is joined together continuously with 324 showing that they are contiguous in time corresponding to the limitation of “compiling the respective return path data into a set of view blocks that are contiguous in time”. Hence although Applicant argues that Sullivan is discrete this Figure clearly shows that there is a contiguous time pattern. Also, please see rejection below for a more detailed mapping.
Thirdly, although Applicant argues that Sullivan uses “decision tree model” which typically rely on discretized or static features, Examiner disagrees because paragraph [0086] of Sullivan very clearly states that alternative examples utilize other forms of machine learning e.g., neural networks etc and not just decision tree model.
Lastly, although Applicant argues that since Sullivan does not acquire any view block sequences that are contiguous in time, a person skilled in the art could not directly utilize any models requiring time-series information (such as the neural network model in Morfi) without compliant input data. Thus, the discrete data obtained by Sullivan is incompatible with the neural network input requirements of Morfi. Examiner disagrees because as explained above and shown in the detailed rejection below both Sullivan and Morfi use neural networks and is therefore it would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the combined teachings of Sullivan and Morfi. The motivation for combining is shown in the detailed rejection below.
Hence, the cited references still teach that limitation as explained above and shown in the detailed rejection below.
Regarding Applicant’s arguments, see pages 20-22, with respect to the rejections under 35 U.S.C. 103, Applicant specifically argues on Page 20 (Paragraph 5) that moreover, as another specific example, Applicant submits that Morfi does not teach or suggest "a merge layer configured to, for each return path data household, generate a merged feature vector by merging features of the generated one-dimensional feature vector with the household-level feature vector for the return path data household" as recited in claim 1.
Applicant continues to argue on Page 22 (Paragraph 3) that Notably, Morfi does not teach or suggest any "merge layer" that combines 1D feature vectors with household-level features. For at least this reason, Morfi does not teach or suggest the neural network structure comprising time-distributed dense layers, recurrent neural network layers, and merge layers as described in the present application and as claimed.
Lastly, Applicant argues on Page 22 (Paragraph 4) that For at least these reasons, the rejection of claim 1 should be withdrawn. And for largely the same reasons, the rejections of independent claims 8 and 15 should also be withdrawn. Further, for at least the reason that each of the remaining claims depends from one of claims 1, 8, and 15, the rejections of these dependent claims should also be withdrawn.
Examiner’s Response: Applicant’s arguments, with respect to the feature mentioned above as recited in independent claim 1 (and similarly in independent claims 8 and 15) have been fully considered but are not persuasive. Examiner disagrees that the cited references do not teach that limitation because firstly, Morfi is not relied upon to teach “merge layer” but Zhang is used to teach it. Secondly, Sullivan2 is used to teach “for each return path data household, generate a merged feature vector by merging features of the generated one-dimensional feature vector with the household-level feature vector for the return path data household”. Hence, the combination of Zhang and Sullivan2 is used to teach the limitation of “a merge layer configured to, for each return path data household, generate a merged feature vector by merging features of the generated one-dimensional feature vector with the household-level feature vector for the return path data household” as shown in the detailed rejection below.
Thirdly, although Morfi does not teach “merge layer” it clearly teaches “wherein the neural network includes: a time distributed dense layer configured to, for each return path data household, reduce the view block-level features into a compressed set of features less in number than the view block-level features wherein the time distributed dense layer includes a set of weights to map the view blocks of the view block-level features into the compressed set of features, a recurrent neural network layer configured to, for each return path data household, generate a one-dimensional feature vector by processing the compressed set of features” because at least Page 4, Section 2.2.1 states Neural Network Architecture and Page 5, Paragraph 1 states Next we apply time distributed dense layers to reduce feature-length dimensionality. The dimensions of each prediction are Tx1. Note: Tx1 corresponds to one-dimensional feature vector. Reduce feature-length dimensionality corresponds to compressed set of features. Examiner also notes that reduce feature-length dimensionality corresponds to reducing the view block level-features into a compressed set of features less in number than the view block level-features. Hence, as explained above and shown in the detailed rejection below the combination of Morfi, Zhang, and Sullivan 2 teach those limitations.
Lastly, since the combination of cited references teach each and every limitation of claim 1 the rejection is maintained. For the same reason the rejections of independent claims 8 and 15 are also maintained. The rest of the claims are dependent from one of the independent claims and are therefore also rejected for the same reasons.
Regarding Applicant’s arguments, see pages 22 and 23, with respect to the rejections under 35 U.S.C. 103, Applicant argues specifically on Page 22 that furthermore, in the present office action, the Examiner again alleged that it would have been obvious to one of ordinary skill in the art to combine various portions of five references to arrive at the invention of claim 1, and more than 5 references to arrive at the invention of each of the dependent claims. Even assuming, for the sake of example, that the cited portions of the five references teach the various elements or portions thereof, as alleged by the Examiner (which Applicant disputes for at least the reasons provided above), the Examiner's combination and application of the five references in this manner involves dissecting the claim elements and portions thereof down into such small components that the relationship between them, and the overall meaning of the claim, is entirely lost. This type of approach does not adhere to the requirement that the claim must be considered as a whole, and instead evidences impermissible hindsight which must be avoided during examination under M.P.E.P. § 2142.
Indeed, Applicant's claims recite a specific combination of elements- namely, the view block-level features and household-level features that are processed by a neural network that is specifically configured to predict demographic classification probabilities for the return path data households. Given these specific limitations, and given the deficiencies of the cited references discussed above, it would not have been obvious to one of ordinary skill in the art to combine various portions of five references to arrive at the invention of claim 1 without the benefit of hindsight.
Applicant continues to argue on Page 23 that in the present case, the § 103 rejection mirrors this same approach, in which hindsight bias drives the picking and choosing of dissected elements (in this case, from five cited references in rejecting the independent claims, and from more than five cited references in rejecting the dependent claims), which prevents the claim from being considered as a whole, as required under M.P.E.P. § 2142.
For at least these additional reasons, the obviousness rejections should be withdrawn.
Examiners response: Applicant’s arguments have been fully considered but they are not persuasive. Examiner respectfully disagrees that the obviousness rejections should be withdrawn because Claims are rejected under 35 USC § 103 as the combination of cited references teach every element of the amended claims as shown in the detailed rejection below and explained above. As mentioned in the previous Non-Final rejection dated 13 February 2025, per MPEP, it must be recognized that any judgment on obviousness is in a sense necessarily a reconstruction based upon hindsight reasoning. But so long as it takes into account only knowledge which was within the level of ordinary skill at the time the claimed invention was made, and does not include knowledge gleaned only from the applicant’s disclosure, such a reconstruction is proper. Hence, the rejection is maintained.
Claim Rejections - 35 USC § 112
The following is a quotation of 35 U.S.C. 112(b):
(b) CONCLUSION.—The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the inventor or a joint inventor regards as the invention.
The following is a quotation of 35 U.S.C. 112 (pre-AIA ), second paragraph:
The specification shall conclude with one or more claims particularly pointing out and distinctly claiming the subject matter which the applicant regards as his invention.
Claims 1, 2, 5-9, and 12-14 are rejected under 35 U.S.C. 112(b) or 35 U.S.C. 112 (pre-AIA ), second paragraph, as being indefinite for failing to particularly point out and distinctly claim the subject matter which the inventor or a joint inventor (or for applications subject to pre-AIA 35 U.S.C. 112, the applicant), regards as the invention.
Claim 1 recites the limitation "… the computing system …" in on Page 3 (line 6). There is insufficient antecedent basis for this limitation in the claim. Claims 2 and 5-7 depend on claim 1 and therefore inherit the same rejection.
Claim 8 recites the limitation "… the computing system …" in on Page 6 (last but 3rd and 2nd lines). There is insufficient antecedent basis for this limitation in the claim. Claims 9 and 12-14 depend on claim 8 and therefore inherit the same rejection.
Claim Rejections - 35 USC § 103
The following is a quotation of 35 U.S.C. 103 which forms the basis for all obviousness rejections set forth in this Office action:
A patent for a claimed invention may not be obtained, notwithstanding that the claimed invention is not identically disclosed as set forth in section 102, if the differences between the claimed invention and the prior art are such that the claimed invention as a whole would have been obvious before the effective filing date of the claimed invention to a person having ordinary skill in the art to which the claimed invention pertains. Patentability shall not be negated by the manner in which the invention was made.
The factual inquiries for establishing a background for determining obviousness under 35 U.S.C. 103 are summarized as follows:
1. Determining the scope and contents of the prior art.
2. Ascertaining the differences between the prior art and the claims at issue.
3. Resolving the level of ordinary skill in the pertinent art.
4. Considering objective evidence present in the application indicating obviousness or nonobviousness.
This application currently names joint inventors. In considering patentability of the claims the examiner presumes that the subject matter of the various claims was commonly owned as of the effective filing date of the claimed invention(s) absent any evidence to the contrary. Applicant is advised of the obligation under 37 CFR 1.56 to point out the inventor and effective filing dates of each claim that was not commonly owned as of the effective filing date of the later invention in order for the examiner to consider the applicability of 35 U.S.C. 102(b)(2)(C) for any potential 35 U.S.C. 102(a)(2) prior art against the later invention.
In the event the determination of the status of the application as subject to AIA 35 U.S.C. 102 and 103 (or as subject to pre-AIA 35 U.S.C. 102 and 103) is incorrect, any correction of the statutory basis for the rejection will not be considered a new ground of rejection if the prior art relied upon, and the rationale supporting the rejection, would be the same under either status.
Claims 1, 5, 8, 12, 15, 19, and 25 are rejected under 35 U.S.C. 103 as being unpatentable over Sullivan et al (US 20170064358 A1) in view of Sullivan et al (US 20180152762 A1) hereafter referred to as Sullivan2 and further in view of Campbell (WO 2008150575 A2), Morfi et al (Deep Learning on Low-Resource Datasets, 2018) and Zhang et al (Model and forecast stock market behavior integrating investor sentiment analysis and transaction data, 2017).
Regarding claim 1
Sullivan teaches: A demographic estimation system comprising: a network interface; a processor; and memory having stored thereon machine-readable instructions that, when executed by the processor, cause performance of operations comprising ([0051] From time to time (periodically, aperiodically, randomly, when the STB 110 is filled with data, etc.), the STB 110 communicates the collected tuning data 108 to the AME 104 via the network 106 (e.g., the Internet, a local area network, a wide area network, a cellular network, etc.) via wired and/or wireless connections (e.g., a cable/DSL/satellite modem, a cell tower, etc.). [0090] In this example, the machine readable instructions comprise a program for execution by a processor such as the processor 1112 shown in the example processor platform 1100 discussed below in connection with FIG. 11. The program may be embodied in software stored on a tangible computer readable storage medium such as a CD-ROM, a floppy disk, a hard drive, a digital versatile disk (DVD), a Blu-ray disk, or a memory associated with the processor 1112, but the entire program and/or parts thereof could alternatively be executed by a device other than the processor 1112 and/or embodied in firmware or dedicated hardware. [0091] As mentioned above, the example processes of FIGS. 3-5 may be implemented using coded instructions (e.g., computer and/or machine readable instructions) stored on a tangible computer readable storage medium such as a hard disk drive, a flash memory, a read-only memory (ROM), a compact disk (CD), a digital versatile disk (DVD), a cache, a random-access memory (RAM) and/or any other storage device or storage disk in which information is stored for any duration (e.g., for extended time periods, permanently, for brief instances, for temporarily buffering, and/or for caching of the information)):
generating features from return path data reported from set-top boxes associated with return path data households, wherein the generating comprises, for the respective return path data reported by each return path data household ([0003] collect tuning data from set-top boxes of panelist households. [0075] the illustrated example constructs feature matrices associated with the respective training group and testing group of the panelist households. An example feature matrix constructed by the decision tree trainer 208 includes rows associated respective panelist households and columns associated with respective household features. Additionally or alternatively, some columns of example feature matrices are associated with other household characteristics (e.g., a total number of minutes consumed by the household, a number of minutes consumed by the household per predetermined time-period segments (e.g. per quarter-hours of the day), a number of STBs within a household, etc.). Note: Tuning data corresponds to Return path data and panelist household corresponds to return path data household),
and generating a two-dimensional (NxF) view block-level feature vector representing a record of all view blocks of the return path data household over the particular observation period, wherein N is a total number of view blocks in the set of view blocks and F is a total number of the view block-level features ([Fig. 6] Note: Figure 6 shows a matrix corresponding to two-dimensional. The horizontal line at the bottom shows 5pm to 10pm corresponding to the particular observation period. The rectangles correspond to wherein N is a total number of view blocks in the set of view blocks. The channels in the first column correspond to F is a total number of the view block-level features);
collecting panel data from meters located at panelist households remotely from the computing system by communicating over the network interface with the meters, wherein the meters are configured to monitor media exposure and media devices at the panelists households ([0022] To enable the AMEs to collect such consumption data, the AMEs typically provide panelist households with meter(s) that monitor media presentation devices (e.g., televisions, stereos, speakers, computers, portable devices, gaming consoles, and/or online media presentation devices, etc.) of the household. [0045] FIG. 1 is a block diagram of an example environment 100 that includes a household 102, an AME 104, and a network 106. In the example environment 100, the AME 104 predicts and/or estimates household characteristics (e.g., demographic characteristics) of the household 102 (e.g., a non-panelist household). The network 106 of the illustrated example connects, among other things, the household 102 and the AME 104. The AME 104 of the illustrated example collects tuning data 108 associated with the household 102. The AME 104 processes the tuning data 108 to determine estimated household characteristics for the household 102. In the illustrated example, the example AME 104 estimates demographic characteristics of the household 102 to estimate a composition and/or size of an audience consuming media (e.g., television programming, advertising, movies, etc.) to produce media ratings);
based on the panel data, training a neural network to process the view block-level feature vector of each return path data household generated from the return path data to predict demographic classification probabilities for the return path data households, wherein the training comprises ([0003] Audience measurement entities may also collect tuning data from set-top boxes of panelist households. [0022] To enable the AMEs to collect such consumption data, the AMEs typically provide panelist households with meter(s) that monitor media presentation devices (e.g., televisions, stereos, speakers, computers, portable devices, gaming consoles, and/or online media presentation devices, etc.) of the household. [0056] For example, for a tuning event of the tuning data 108 associated with “Premier League Live” on NBCSports at 7:30 A.M. on Sunday, the distribution calculator 124 collects demographics data associated with panelists who viewed the same channel (i.e., NBCSports) at substantially the same time (e.g., 7:32 A.M. on Sunday) and calculates a demographic distribution for those panelists (e.g., 20% are 18-45 year-old females, 40% are 18-45 year-old males, 10% are 46-64 year-old females, 20% are 46-64 year-old males, 5% are 65+ year-old females, and 5% are 65+ year old males). As a result, a demographic distribution represents probabilities or likelihoods that a consumer of media (e.g., one of the members 112, 114, 116 of the household 102) matches particular demographic dimensions of interest. [0086] alternative examples of the household estimator 210 utilize other forms of machine learning (e.g., neural networks, support vector machines, clustering, Bayesian networks, etc.) to estimate the demographics of the household 102. In such examples, the decision tree trainer 208 and/or another machine learning trainer constructs the corresponding machine learning classifier (e.g., neural networks, support vector machines, a clustering mechanism, Bayesian networks) utilized to estimate the demographics of the household 102):
creating panel view blocks from the panel data for the panelist households ([0056] For example, for a tuning event of the tuning data 108 associated with “Premier League Live” on NBCSports at 7:30 A.M. on Sunday, the distribution calculator 124 collects demographics data associated with panelists who viewed the same channel (i.e., NBCSports) at substantially the same time (e.g., 7:32 A.M. on Sunday). Note: Tuning event corresponds to panel view block),
generating panel view block-level feature vectors for respective ones of the panelist households from the panel view blocks created for the respective panelist households ([0056] and calculates a demographic distribution for those panelists (e.g., 20% are 18-45 year-old females, 40% are 18-45 year-old males, 10% are 46-64 year-old females, 20% are 46-64 year-old males, 5% are 65+ year-old females, and 5% are 65+ year old males)),
applying the panel view block-level feature vectors for the respective ones of the panelist households to the neural network according to a training procedure that adjusts internal parameters of the neural network to reduce an error between the predicted demographic classification probabilities output by the neural network and actual demographics known for the panelist households based on the panel data ([0056] As a result, a demographic distribution represents probabilities or likelihoods that a consumer of media (e.g., one of the members 112, 114, 116 of the household 102) matches particular demographic dimensions of interest. For example, a person who views “Premier League Live” on NBCSports at 7:30 A.M. on Sunday is 20% likely to be 18-45 year-old female, 40% likely to be a 18-45 year-old male, 10% likely to be a 46-64 year-old female, 20% likely to be a 46-64 year-old male, 5% likely to be a 65+ year-old female, and 5% likely to be a 65+ year old male. [0058] [0058] Based on the tuning data 108, the demographics distributions associated with respective tuning events and/or the average demographics distribution of the panelists, the characteristic estimator 126 estimates household characteristics of the household 102 such as (1) a number of members of the household 102 (e.g., three household members 112, 114, 116) and (2) the demographics of each of the estimated household members (e.g., the demographics of each of the members 112, 114, 116). Thus, to measure a size and composition of media audiences, the characteristic estimator 126 of the example AME 104 analyzes the tuning data 108 of the household 102 and the demographics and consumption data of the panelist households to estimate the household characteristic of the household 102. [0069] [0069] To account for the disproportionate consumption of media by some demographic constraints, the score calculator 206 calculates scores for the respective demographic constraints. [0086] While the household estimator 210 of the illustrated example utilizes a decision tree ensemble to estimate household characteristics of the household 102, alternative examples of the household estimator 210 utilize other forms of machine learning (e.g., neural networks, support vector machines, clustering, Bayesian networks, etc.) to estimate the demographics of the household 102. In such examples, the decision tree trainer 208 and/or another machine learning trainer constructs the corresponding machine learning classifier (e.g., neural networks, support vector machines, a clustering mechanism, Bayesian networks) utilized to estimate the demographics of the household 102. Note: To account for the disproportionate consumption of media by some demographic constraints, the score calculator 206 calculates scores corresponds to adjusts internal parameters to reduce an error),
and shuffling an order of the view blocks of the return path data households that are fed into the neural network during each of a plurality of training epochs of the neural network ([0052] As illustrated in FIG. 1, the AME 104 includes a tuning event database 120, a panelist database 122, a distribution calculator 124, and a characteristic estimator 126. [0058] Based on the tuning data 108, the demographics distributions associated with respective tuning events and/or the average demographics distribution of the panelists, the characteristic estimator 126 estimates household characteristics of the household 102. [0090] Further, although the example programs are described with reference to the flowcharts illustrated in FIGS. 3-5, many other methods of implementing the example characteristic estimator 126 may alternatively be used. For example, the order of execution of the blocks may be changed, and/or some of the blocks described may be changed, eliminated, or combined. 0097] At block 322, the decision tree trainer 208 trains decision tree ensemble(s) associated with respective household characteristic(s) using consumption data, demographic data, and/or other household characteristic data of the panelist households. Alternatively, at block 322, the decision tree trainer 208 and/or another machine learning trainer may construct a machine learning classifier other than a decision tree classifier (e.g., neural networks, support vector machines, a clustering mechanism, Bayesian networks) based on the data of the panelist households),
and implementing the neural network to process the view block-level feature vector of each return path data household generated from the return path data to predict the demographic classification probabilities for the return path data households ([0086] alternative examples of the household estimator 210 utilize other forms of machine learning (e.g., neural networks, etc.) to estimate the demographics of the household 102. In such examples, the decision tree trainer 208 and/or another machine learning trainer constructs the corresponding machine learning classifier (e.g., neural networks) utilized to estimate the demographics of the household 102. [0003] collect tuning data from set-top boxes of panelist households. [0056] For example, for a tuning event of the tuning data 108 associated with “Premier League Live” on NBCSports at 7:30 A.M. on Sunday, the distribution calculator 124 collects demographics data associated with panelists who viewed the same channel (i.e., NBCSports) at substantially the same time (e.g., 7:32 A.M. on Sunday) and calculates a demographic distribution for those panelists (e.g., 20% are 18-45 year-old females, 40% are 18-45 year-old males, 10% are 46-64 year-old females, 20% are 46-64 year-old males, 5% are 65+ year-old females, and 5% are 65+ year old males). As a result, a demographic distribution represents probabilities or likelihoods that a consumer of media (e.g., one of the members 112, 114, 116 of the household 102) matches particular demographic dimensions of interest. [0022] To enable the AMEs to collect such consumption data, the AMEs typically provide panelist households with meter(s) that monitor media presentation devices (e.g., televisions, stereos, speakers, computers, portable devices, gaming consoles, and/or online media presentation devices, etc.) of the household. Note: Tuning data corresponds to Return path data);
and assigning one or more demographic categories to respective ones of the return path data households based on the predicted demographic classification probabilities ([0056] For example, for a tuning event of the tuning data 108 associated with “Premier League Live” on NBCSports at 7:30 A.M. on Sunday, the distribution calculator 124 collects demographics data associated with panelists who viewed the same channel (i.e., NBCSports) at substantially the same time (e.g., 7:32 A.M. on Sunday) and calculates a demographic distribution for those panelists (e.g., 20% are 18-45 year-old females, 40% are 18-45 year-old males, 10% are 46-64 year-old females, 20% are 46-64 year-old males, 5% are 65+ year-old females, and 5% are 65+ year old males). As a result, a demographic distribution represents probabilities or likelihoods that a consumer of media (e.g., one of the members 112, 114, 116 of the household 102) matches particular demographic dimensions of interest. For example, a person who views “Premier League Live” on NBCSports at 7:30 A.M. on Sunday is 20% likely to be 18-45 year-old female, 40% likely to be a 18-45 year-old male, 10% likely to be a 46-64 year-old female, 20% likely to be a 46-64 year-old male, 5% likely to be a 65+ year-old female, and 5% likely to be a 65+ year old male. [0058] Based on the tuning data 108, the demographics distributions associated with respective tuning events and/or the average demographics distribution of the panelists, the characteristic estimator 126 estimates household characteristics of the household 102 such as (1) a number of members of the household 102 (e.g., three household members 112, 114, 116) and (2) the demographics of each of the estimated household members (e.g., the demographics of each of the members 112, 114, 116)).
However, Sullivan does not explicitly disclose: compiling the respective return path data into a set of view blocks that are contiguous in time and generated by the return path data household over a particular observation period, wherein each view block is associated with a respective different time interval and aggregates a set of one or more tuning events corresponding to the time interval and included in the respective return path data into a plurality of view block-level features represented by a view block-level feature vector, the view block-level features comprising: a channel change rate determined based on a ratio of (i) a number of channel changes that occurred during the time interval to (ii) a duration of the time interval, for each of a plurality of stations to which a set-top box of the return path data household is configured to tune, a respective total number of minutes that the station was viewed, and a list of stations of the plurality of stations that were visited during the time interval; for each return path data household, aggregating household-level features into an H-dimensional (1xH) household-level feature vector, wherein the household-level features are different from the view block-level features, wherein H is a number of household-level features, and wherein the household-level features comprise a total number of view blocks reported for the return path data household over the particular observation period and a total number of tuners known to be included in the return path data household; wherein the neural network includes: a time distributed dense layer configured to, for each return path data household, reduce the view block-level features into a compressed set of features less in number than the view block-level features, wherein the time distributed dense layer includes a set of weights to map the view blocks of the view block-level features into the compressed set of features, a recurrent neural network layer configured to, for each return path data household, generate a one-dimensional feature vector by processing the compressed set of features, and a merge layer configured to, for each return path data household, generate a merged feature vector by merging features of the generated one-dimensional feature vector with the household-level feature vector for the return path data household.
Sullivan2 teaches, in an analogous system: compiling the respective return path data into a set of view blocks that are contiguous in time and generated by the return path data household over a particular observation period, wherein each view block is associated with a respective different time interval and aggregates a set of one or more tuning events corresponding to the time interval and included in the respective return path data into a plurality of view block-level features represented by a view block-level feature vector, the view block-level features comprising ([0025] The example audience measurement module 122 is provided with the example RPD model generator 208 to generate a model of RPD tuning information based on predicted extensions of the durations of panel tuning segments reported in the collected panel tuning information. For example, FIG. 3 is a schematic representation of panel tuning information 302 corresponding to media played on multiple media sets 304, 306, 308, 310, 312, 314 over a period of time. [0026] Thus, during the represented period, the first media set 304 is associated with two panel tuning segments 320, 322 separated by a gap in time indicating the first media set 304 was turned off between the two tuning segments 320, 322. Separate panel tuning segments are not necessarily spaced in time. For example, the second media set 306 includes two panel tuning segments 324, 326 in which the second panel tuning segment 326 immediately follows the first panel tuning segment 324 indicating the audience member changed the channel or station to which the RPD device 114 was tuned. Note: See Figure 3 showing the time interval on x axis corresponding to the observation period and 320+322 corresponds to the aggregating a set of one or more tuning events into a view block. Also, see 326+324 showing that they are contiguous in time corresponding to another example for compiling the respective return path data into a set of view blocks that are contiguous in time and generated by the return path data household over a particular observation period):
for each return path data household, aggregating household-level features into an H-dimensional (1xH) household-level feature vector, wherein the household-level features are different from the view block-level features, wherein H is a number of household-level features, and wherein the household-level features comprise a total number of view blocks reported for the return path data household over the particular observation period and a total number of tuners known to be included in the return path data household ([0006] FIG. 3 is a schematic representation of panel tuning information corresponding to media played on multiple different media sets over a period of time. Note: The rectangles correspond to a total number of view blocks reported for the return path data household over the particular observation period and Set1 through Set 6 corresponds to a total number of tuners known to be included in the return path data household);
for each return path data household, generate a merged feature vector by merging features of the generated one-dimensional feature vector with the household-level feature vector for the return path data household ([0006] FIG. 3 is a schematic representation of panel tuning information corresponding to media played on multiple different media sets over a period of time. Note: Fig 3 corresponds to a merged feature vector by merging features of the generated one-dimensional feature vector with the household-level feature vector wherein as claimed earlier the household-level features comprise a total number of view blocks reported for the return path data household over the particular observation period and a total number of tuners known to be included in the return path data household. The rectangles correspond to a total number of view blocks reported for the return path data household over the particular observation period and Set1 through Set 6 corresponds to a total number of tuners known to be included in the return path data household).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the demographic estimation system of Sullivan to incorporate the teachings of Sullivan2 to compile the respective return path data into a set of view blocks that are contiguous in time and generated by the return path data household over a particular observation period, wherein each view block is associated with a respective different time interval and aggregates a set of one or more tuning events corresponding to the time interval and included in the respective return path data into a plurality of view block-level features represented by a view block-level feature vector; for each return path data household, aggregating household-level features into an H-dimensional (1xH) household-level feature vector, wherein the household-level features are different from the view block-level features, wherein H is a number of household-level features, and wherein the household-level features comprise a total number of view blocks reported for the return path data household over the particular observation period and a total number of tuners known to be included in the return path data household; for each return path data household, generate a merged feature vector by merging features of the generated one-dimensional feature vector with the household-level feature vector for the return path data household. One would have been motivated to do this modification because doing so would give the benefit of representation of panel tuning information corresponding to media played on multiple media sets over a period of time as taught by Sullivan2 [0112].
Campbell teaches, in an analogous system: a channel change rate determined based on a ratio of (i) a number of channel changes that occurred during the time interval to (ii) a duration of the time interval ([0071] For example, relatively short tuning periods may be indicative of channel surfing. Note: Channel surfing corresponds to channel change rate because several channels are changed during a time interval),
for each of a plurality of stations to which a set-top box of the return path data household is configured to tune, a respective total number of minutes that the station was viewed ([0050] Without limitation, the linking variables 410 may include the number of sets in a household, time tuned total, time tuned to a particular channel, time tuned to a particular network (e.g., The Food Network.sup.®, ABC, NBC, etc.), time tuned to a particular channel genre, and/or time tuned by daypart (e.g., between 1 :00 to 6:00 A.M., between 4:00 to 6:00 P.M., etc.));
and a list of stations of the plurality of stations that were visited during the time interval ([0050] Without limitation, the linking variables 410 may include time tuned to a particular channel, time tuned to a particular network (e.g., The Food Network.sup.®, ABC, NBC, etc.), time tuned to a particular channel genre, and/or time tuned by daypart (e.g., between 1 :00 to 6:00 A.M., between 4:00 to 6:00 P.M., etc.)).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the combined teachings of Sullivan and Sullivan2 to incorporate the teachings of Campbell to use a channel change rate determined based on a ratio of (i) a number of channel changes that occurred during the time interval to (ii) a duration of the time interval for each of a plurality of stations to which a set-top box of the return path data household is configured to tune, a respective total number of minutes that the station was viewed and a list of stations of the plurality of stations that were visited during the time interval. One would have been motivated to do this modification because doing so would give the benefit of extracting one or more sessions of the received set-top box data that are deemed useful as taught by Campbell [0071].
Morfi teaches, in an analogous system: wherein the neural network includes: a time distributed dense layer configured to, for each return path data household, reduce the view block-level features into a compressed set of features less in number than the view block level-features ([Page 4, Section 2.2.1] Neural Network Architecture. [Page 5, Paragraph 1] Next we apply time distributed dense layers to reduce feature-length dimensionality. Note: Reduce feature-length dimensionality corresponds to reducing the view block level-features into a compressed set of features less in number than the view block level-features), the time distributed dense layer including a set of weights to map the first set of features into the second set of features ([Page 5, Paragraph 1] Next we apply time distributed dense layers to reduce feature-length dimensionality. [Page 6, Paragraph 2] Training is performed by updating the network weights),
wherein the time distributed dense layer includes a set of weights to map the view blocks of the view block-level features into the compressed set of features ([Page 5, Paragraph 1] Next we apply time distributed dense layers to reduce feature-length dimensionality. [Page 6, Paragraph 2] Training is performed by updating the network weights),
a recurrent neural network layer configured to, for each return path data household, generate a one-dimensional feature vector by processing the compressed set of features ([Page 4, Section 2.2.1] Neural Network Architecture. [Page 5, Paragraph 1] Next we apply time distributed dense layers to reduce feature-length dimensionality. The dimensions of
each prediction are Tx1. Note: Tx1 corresponds to one-dimensional feature vector. Reduce feature-length dimensionality corresponds to compressed set of features).
It would have been obvious to one of ordinary skill in the art before the effective filing date of the claimed invention to have modified the combined teachings of Sullivan, Sullivan2, and Campbell to incorporate the teachings of Morfi wherein the neural network includes a time distributed dense layer configured to, for each return path data household, reduce the view block level-features into a compressed set of features less in number than the view block-level features, wherein the time distributed dense layer includes a set of weights to map the view blocks of the view block-level features into the compressed set of features, a recurrent neural network layer configured to, for each return path data household, generate a one-dimensional feature vecto