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 .
This office action is in responsive to RCE filed on 11/26/2025. Claims 1-20 remain pending in the application. Claims 1, 8, and 15 are independent.
Claim Objections
Applicant's amendment to claims and clarification in Pages 8-9 of the Remarks correct some of previous objections; therefore, some of previous objections are withdrawn. The remaining objections are shown below.
Claims 1 and 14 are objected to because of the following informalities:
in Claim 1, lines 4-5, "… by determining a Bayesian a priori probability for each occurrence of a specific category of a feature variable …" appears to be "… by determining Bayesian a priori probability for each occurrence of a specific category of a feature variable …" according to Claims 8 and 15;
in Claim 14, lines 3-4, "… by assuming a Gamma distribution as the prior for a distribution of the one or more factors" appears to be "… by assuming a Gamma distribution as a prior for a distribution of the one or more factors" according to Claim 7.
Appropriate correction is required.
Claim Rejections - 35 USC § 101
35 U.S.C. 101 reads as follows:
Whoever invents or discovers any new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof, may obtain a patent therefor, subject to the conditions and requirements of this title.
Claims 1-20 are rejected under 35 U.S.C. 101 because the claimed invention is directed to abstract idea without significantly more. The claim(s) recite(s) "calculating model parameters of a cyclic boosting model" (Claims 1-20), "regularizing/regularize/regularizes one or more factors in one or more categories of training data by determining a Bayesian a priori probability for each occurrence of a specific category of a feature variable, wherein the regularizing improves a numerical stability of the cyclic boosting model (Claims 1-20), wherein the one or more factors each correspond to a strength of a factor contributing to the predicted value (Claims 3, 10, and 17), wherein the training data comprises one or more of: one or more historic sales patterns, one or more prices, one or more promotions, one or more weather conditions and one or more factors influencing demand of an item sold in a given store on a specific day (Claims 4, 11, and 18)", "calculating/calculate/calculates a global average from all the observed target values across all bins and features, wherein the bins and features are defined by a matrix" (Claims 1-20), "initializing/initialize/initializes each factor of the one or more factors to 1, wherein each factor corresponds to a single bin and a single feature" (Claims 1-20), "calculating/calculate/calculates iteratively, for each feature and corresponding bin, partial factors and aggregate factors, wherein a partial factor is multiplied by an aggregate factor in each iteration of a multiple of iterations, and wherein calculation of the partial factor in each iteration includes a learning rate to reduce dependency on a sequence of features" (Claims 1-20), "calculating/calculate/calculates a predicted value of a target variable for each of the multiple of iterations until a stopping criteria is met, wherein the calculated predicted values follow a Poisson or Poisson-Gamma distribution (Claims 1-20), wherein the stopping criteria comprises a mean absolute deviation or a mean squared error (Claims 2, 9, and 16)", "a modification to the feature variable to identify and calculate a corresponding change in the predicted value" (Claims 6, 13, and 20), and "regularizing/regularize one or more factors in one or more categories of training data by assuming a Gamma distribution as a prior for a distribution of the one or more factors" (Claims 7 and 14) which can be reasonably considered as mathematical concepts/algorithms/calculations or mental processes (i.e., which "can be performed in the human mind, or by a human using a pen and paper"). This judicial exception is not integrated into a practical application because the claim(s) recite(s) additional elements/limitations of "computer" (Claims 1-7), "processor" (Claims 1-14), "memory" (Claims 1-14), "server" (Claims 8-14), "non-transitory computer-readable medium embodied with software" (Claims 15-20), "automated machinery" (Claims 1-20), "at least one sensor" (Claims 1-20), "scanning product data associated with observed target values" (Claims 1-20), "rendering/render/renders, for display on a user interface, a visualization comprising a two- dimensional visualization of one or more deviations between the calculated predicted values and corresponding actual value" (Claims 1-20), "receiving/receives/receives a selection from a user interface module of a list of one or more sample input variables, one or more sample input variable definitions and one or more sample input variable sequence" (Claims 5, 12, and 19), and "receive a modification" (Claims 6, 13, and 20) which only amount to "apply it" with the use of generic computer components or insignificant extra solution activity. None of the additional elements/limitations, taken alone or in combination, integrate the abstract idea into a practical application. The claim(s) does/do not include additional elements/limitations that are sufficient to amount to significantly more than the judicial exception because (a) the additional elements/limitations of "scanning product data associated with observed target values" (Claims 1-20), "receiving/receives/receives a selection from a user interface module of a list of one or more sample input variables, one or more sample input variable definitions and one or more sample input variable sequence" (Claims 5, 12, and 19), and "receive a modification" (Claims 6, 13, and 20) are well-understood, routine and conventional (WURC) activity similar to "receiving or transmitting data over a network" (see MPEP 2106.05(d), "Receiving or transmitting data over a network, e.g., using the Internet to gather data, Symantec, 838 F.3d at 1321, 120 USPQ2d at 1362 (utilizing an intermediary computer to forward information); buySAFE, Inc. v. Google, Inc., 765 F.3d 1350, 1355, 112 USPQ2d 1093, 1096 (Fed. Cir. 2014) (computer receives and sends information over a network)"); and (b) the additional element/limitation of "rendering/render/renders, for display on a user interface, a visualization comprising a two- dimensional visualization of one or more deviations between the calculated predicted values and corresponding actual value" (Claims 1-20) is also well-understood, routine and conventional (WURC) activity similar to "presenting offers and gathering statistics" (see MPEP 2106.05(d), "Presenting offers and gathering statistics, OIP Techs., 788 F.3d at 1362-63, 115 USPQ2d at 1092-93"). Thus, none of the additional elements/limitations, taken either alone or combined, amount to significantly more than the abstract idea.
Allowable Subject Matter
Claims 1-20 would be allowable if rewritten or amended to overcome the rejection(s) under 35 U.S.C. 101, set forth in this Office action.
The following is a statement of reasons for the indication of allowable subject matter:
In regard to independent Claims 1, 8, and 15, prior arts of records, either singularly or in combination, do not teach or suggest the combination of claimed elements including "a computer-implemented method for calculating model parameters of a cyclic boosting model, comprising: regularizing, by a computer comprising a processor and a memory, one or more factors in one or more categories of training data by determining Bayesian a priori probability for each occurrence of a specific category of a feature variable, wherein the regularizing improves a numerical stability of the cyclic boosting model; scanning, by automated machinery having at least one sensor, product data associated with observed target values; calculating, by the computer, a global average from all the observed target values across all bins and features, wherein the bins and features are defined by a matrix; initializing, by the computer, each factor of the one or more factors to 1, wherein each factor corresponds to a single bin and a single feature; calculating iteratively, by the computer, for each feature and corresponding bin, partial factors and aggregate factors, wherein a partial factor is multiplied by an aggregate factor in each iteration of a multiple of iterations, and wherein calculation of the partial factor in each iteration includes a learning rate to reduce dependency on a sequence of features; calculating, by the computer, a predicted value of a target variable for each of the multiple of iterations until a stopping criteria is met, wherein the calculated predicted values follow a Poisson or Poisson-Gamma distribution; and rendering, for display on a user interface, a visualization comprising a two-dimensional visualization of one or more deviations between the calculated predicted values and corresponding actual values", "a system for calculating model parameters of a cyclic boosting model, comprising: a server, comprising a processor and a memory, configured to: regularize one or more factors in one or more categories of training data by determining Bayesian a priori probability for each occurrence of a specific category of a feature variable, wherein the regularizing improves a numerical stability of the cyclic boosting model; scan, by automated machinery having at least one sensor, product data associated with observed target values; calculate a global average from all the observed target values across all bins and features, wherein the bins and features are defined by a matrix; initialize each factor of the one or more factors to 1, wherein each factor corresponds to a single bin and a single feature; calculate iteratively, for each feature and corresponding bin, partial factors and aggregate factors, wherein a partial factor is multiplied by an aggregate factor in each iteration of a multiple of iterations, and wherein calculation of the partial factor in each iteration includes a learning rate to reduce dependency on a sequence of features; calculate a predicted value of a target variable for each of the multiple of iterations until a stopping criteria is met, wherein the calculated predicted values follow a Poisson or Poisson-Gamma distribution; and render, for display on a user interface, a visualization comprising a two-dimensional visualization of one or more deviations between the calculated predicted values and corresponding actual values", or "a non-transitory computer-readable medium embodied with software for calculating model parameters of a cyclic boosting model, wherein the software, when executed: regularizes one or more factors in one or more categories of training data by determining Bayesian a priori probability for each occurrence of a specific category of a feature variable, wherein the regularizing improves a numerical stability of the cyclic boosting model; scans, by automated machinery having at least one sensor, product data associated with observed target values; calculates a global average from all the observed target values across all bins and features, wherein the bins and features are defined by a matrix; initializes each factor of the one or more factors to 1, wherein each factor corresponds to a single bin and a single feature; calculates iteratively, for each feature and corresponding bin, partial factors and aggregate factors, wherein a partial factor is multiplied by an aggregate factor in each iteration of a multiple of iterations, and wherein calculation of the partial factor in each iteration includes a learning rate to reduce dependency on a sequence of features; calculates a predicted value of a target variable for each of the multiple of iterations until a stopping criteria is met, wherein the calculated predicted values follow a Poisson or Poisson- Gamma distribution; and renders, for display on a user interface, a visualization comprising a two-dimensional visualization of one or more deviations between the calculated predicted values and corresponding actual values" when interpreted as a whole.
KENG et al. (US 2021/0110429 A1, filed on 03/21/2018) discloses in ABSTRACT that and ¶¶ [0004]-[0013] (1) generating an output analytic for a promotion; (2) receiving historical data related to one or more products and a plurality of previous promotions; (3) receiving at least one received input parameter for the promotion from a user, at least one of the input parameters comprising a macroscopic objective of the promotion; (4) determining, using an optimization machine learning model trained or instantiated with an optimization training set, at least one determined parameter for the promotion which optimizes at least one of received input parameters, wherein the optimization training set comprising received historical data; (5) forecasting, using a promotion forecasting machine learning model trained or instantiated with an forecasting training set, at least one output analytic of the promotion, wherein the prediction training set comprising the received historical data, the at least one received input parameter and the at least one determined parameter; (6) outputting the at least one output analytic to the user; (7) the promotion forecasting machine learning model comprises at least one of an average price model and a regression model, wherein the average price model comprises a Random Forest model to predict an average effective discounted price of the promotion based on a category of products, and the regression model incorporates covariates to predict demand; (8) the regression model is used to determine the discounted price prediction on a per-product basis on a group of products in the same brand or subcategory, or both.; (9) the regression model incorporates indicator variables for the one or more products, determining the indicator variables comprising, for each product, normalizing absolute units by a mean for periods with no promotion, and where such mean is not available, normalizing by the mean of the product's entire history; (10) the promotion forecasting machine learning model comprises a first Ridge Regression model combined with a second Ridge Regression model, wherein the first Ride Regression model comprising at least one training set feature different than the second Ridge Regression model; (11) the historical data comprises one or products in a similar category or brand; (12) the plurality of previous promotions are aggregated in a stacked relationship; (13) the historical data comprises transaction history for the product and one or more other products in the same product category, wherein the transaction history comprising at least one of date sold, product, units sold, price sold; (14) the at least one output analytic comprises one of promotion lift, cannibalization, halo effect, pull forward, and price elasticity of demand; and (15) determining a confidence indicator to indicate the reliability of the forecast, wherein determining the confidence indicator comprises: (a) determining if the forecast is in a predetermined scope; and (b) determining, using an accuracy machine learning model trained or instantiated with an accuracy training set, the confidence indicator, wherein the accuracy training set comprising previous forecasts and their respective actualized values. KENG further discloses in ¶¶ [0020]-[0021] that (1) generating at least one output analytic for promotional materials; (2) receiving historical data related to one or more products and a plurality of previous promotional materials; (3) receiving one or more input parameters related to the promotional materials from a user; (4) selecting, using a machine learning model trained or instantiated with a selection training set, a configuration and a layout for the one or more products on the promotional materials, wherein the selection training set comprising the historical data and the one or more input parameters, and the selection comprising: (a) assigning a prominence weight to each of the one or more products; (b) normalizing the prominence weight for each of the one or more products; (c) determining a block structure for the promotional materials based on the prominence weight of each of the one or more products; and (d) determining a location for each of the products on the promotional materials based on the prominence weight of each of the one or more products; (5) outputting the promotional materials based on the selection of the configuration and layout; and (6) selecting, using the selection machine learning model, the one or more products to be promoted on the promotional materials. KENG also discloses in ¶¶ [0022]-[0023] that (1) generating of at least one output analytic for per-store unit demand; (2) receiving historical data related to one or more products, wherein the historical data comprising historical inventory level of the one or more products at a retail store; (3) forecasting, using a demand machine learning model trained or instantiated with a demand training set, a demand for the one or more products at the retail store, wherein the demand machine learning model comprising a first model for predicting the total unit demand for the retail store and a second model for predicting the demand in the retail store for the one or more products, wherein the demand training set comprising the historical data, and the forecast comprising multiplying the prediction of the total unit demand for the retail store for a predetermined time period by the prediction of the demand in the retail store for the one or more products; (3) outputting the at least one output analytic to the user; and (4) the forecast further comprises adding a covariate for a stock out condition. KENG further teaches in ¶¶ [0045]-[0055] with FIGS. 1 and 3 that (1) promotions can come in different types and be directed to different targets; (2) a common problem with developing promotions is determining optimal analytical aspects of the promotion and using such analytics to forecast demand.; (3) at block 51, historical data related to one or more products and a plurality of previous promotions is received by the input module 26; (4) at block 52, at least one input parameter of the promotion is received from the user via the input interface 16 to the input module 26, wherein the input parameter received from the user will reflect the macroscopic objective of the promotion; e.g., product uplift, product sales target, or product demand targets; (5) the input parameter received from the user can also include input parameters that constrain the promotion; such as, transaction history, which product is to be promoted, how the product is to be promoted, or other limitations; (6) at block 54, the machine learning module 30 passes the received input parameters through a machine learning model; (7) the machine learning model can use time series approaches that primarily use historical data as basis for analytically estimating future behavior, wherein time series approaches can include, e.g., ARIMAX, AR, Moving Average, Exponential smoothing, or the like; (8) the machine learning model can use regression based approaches that use a variety of factors (including past data points) to predict future outcomes with an implicit concept of time (through the data points); (9) regression based approaches can include, e.g., linear regression, random forest, neural network, or the like; at block 56, the input module 26 determines, via the machine learning module 30, at least one other input parameter of the promotion, wherein the input parameters can include, e.g., product selection, promotion mechanics, time period for promotion, other causal factors, or the like; (10) the input module 26 determines values for the other input parameters which optimize the forecast based on the constraints of the received input parameters; (11) the historical data can include transaction history, e.g., historical data relating to the product; (12) the historical data can include, e.g., date sold, product, units sold, price sold, or the like; (13) the transaction history can also include historical data relating to similar products, e.g., products in the same product category; (14) the other causal factors can include, e.g., budget, vendor subsidy, seasonality, distribution, appearance, stock, SKU age, star buy, feature promotion, points promotion, shelving, or the like; (15) at block 58, output analytics related to the promotion are determined by the output module 28 via the machine learning module 30; (16) the output analytics can include, e.g., forecasted demand, forecasted price, baseline demand, baseline price, forecast without promotion mechanics, inventory forecasting (such as at a warehouse or store level), or the like; (17) at block 60, other secondary analytics can be derived from the output analytics by the output module 28; and (18) the user, via the input interface 16, can adjust the input parameters in the input module 26 after a forecast has been provided, in order to determine which input parameters arrive at a forecast for a desired output. KENG further teaches in ¶¶ [0056]-[0128] with FIG. 1 that (1) the machine learning module 30 can use machine learning techniques with the machine learning model (called a promotion forecasting model) to forecast output analytics; (2) the promotion forecasting model can be trained using input parameters related to past promotions; (3) the promotion forecasting model can be instantiated with data, such as transaction history, provided by a user; (4) multiple promotion forecasting models can be used such that their results may be averaged or weighted accordingly; (5) use a regression model to determine forecasted lift as a result of a promotion, wherein the regression is performed on a daily basis to account for the fact that many promotions do not align with weekly boundaries, treat each promotion equally (without regard to the type), and consider the number of promotions (being the difference in the number of promotions between the preceding and promotion periods); (6) an L2 regularized regression (ridge regression) can be used to achieve an intended robust result, wherein the L2 regularized regression in the case of ordinary linear regression can include putting a Gaussian prior on the coefficients; (7) without the regularization, the coefficients may not produce reasonable results as sometimes the indicator coefficients can have too much weight associated with them; (8) with regularization, estimate appear to be more objectionably reasonable; (9) once the regression model is determined by the machine learning module 30, intermediates values can be determined, and the promotion lift can then be determined by the prediction of the actual sales during the promotion period produced by the model and what the total estimated sales would have been if the current promotion did not run as predicted by the model; (10) price elasticity analytics can be forecasted as a measure of the sensitivity of unit sales to changes in price; (11) the model for price elasticity that is used is the multiplicative model which can model demand using the regression model; (12) the promotion forecasting model can use at least one of an average price model and a regression model; (13) the regression model can predict demand using covariates such as, e.g., average price from the average price model, relevant additional promotion mechanics, and any other relevant causal factors; (14) a reason for the double-pronged promotion forecasting model can be two-fold: (a) first, price (or effective price) is generally the primary driver of sales so translating promotion mechanics into an average price can work particularly well; and (b) second, causal factors generally affect the promotional demand more than trend behavior due to because when a promotion occurs, the promotion is typically a step function (not on sale to on sale), and additionally, relevant seasonal terms or long term trends can be easily encoded within a regression problem; (15) any combination of promotion mechanics may be stacked together as a combined promotion, and the machine learning module 30 can use new promotion data to encode each one of the promotion mechanic types into several columns; (16) the machine learning module 30 can similarly repeat the column creation for other types of promotions; (17) if two promotions of the same type are stacked, the machine learning module 30 can aggregate the duplicated rows (e.g. take an average), which, in some cases, works out in the regression (half-way between the two promotions); (18) the average price model can be used to approximately predict the average effective discounted price as an outcome variable from the promotion mechanics, and the prediction can use a machine learning model, such as a Random Forest mode; (19) the machine learning module 30 can use a machine learning model that can effectively learn the mapping function for mapping promotion mechanics to a predicted promotion price, while being able to interpolate between unseen corner cases, as long as the given category has enough training data; (20) once the average price model has been fit on an entire category, the regression promotion forecasting model can be used by the machine learning module 30 to train models; (21) an ensemble approach can be used by the machine learning module 30 because such approach tends to be the most empirically accurate; (22) due to the nature of Random Forest models, these models can be better at interpolating rather than extrapolating data, and accordingly, these models are useful for "memorizing" past promotion information and are particularly advantageous when there is a lot of previous data (e.g., many different price points) and a forecasted prediction resembles past behavior; (23) generally, Ridge Regression can be interpreted as a simple linear regression with a zero-mean normally distributed Bayesian Prior; (24) the machine learning module 30 can "cap" the forecast at either a) mean+ 3 * standard deviation, or b) a previous maximum; (25) the machine learning module 30 can impose a non-zero mean Bayesian Prior that can help fill in missing or sparse data when making an estimate with respect to a single SKU; and (26) the machine learning module 30 can implement a MAP estimation for a Bayesian prior using a Ridge Regression with Bayesian Prior. KENG also teaches in ¶¶ [0129]-[0134] with FIG. 1 that (1) for each one of the forecasts determined by the machine learning module 30, the confidence module 40 can provide a confidence indicator to indicate the reliability of such forecast; (2) the confidence module 40 then uses a confidence machine learning model that is trained on previous SKU forecasts to determine the accuracy of previous SKU forecasts versus their actualized values; (3) the confidence module 40 can be used for cutting off forecasts if the confidence in the prediction is too low; e.g., where the forecast is below a given confidence score; and (4) confidence can be determined by the confidence module 40 using other metrics, e.g., model fit metrics, mean absolute percentage error, or the like;
Aguilar-Palacios et al. ("Forecasting Promotional Sales Within the Neighbourhood", IEEE Access, Vol. 7, 2019, pp. 74759-74775) discloses in Abstract and Section I of Page 74759-74760 that (1) introduce a novel interpretable machine learning method specifically tailored to the automatic prediction of promotional sales in real-market applications; (2) present fully automated weighted k-nearest neighbors (kNN) where the distances are calculated based on a feature selection process that focuses on the similarity of promotional sales; (3) the method learns online, thereby avoiding the model being retrained and redeployed; (4) it is robust and able to infer the mechanisms leading to sales as demonstrated on detailed surrogate models; (5) the proposed method significantly improves the accuracy of the forecast in many diverse categories and geographical locations, yielding significant and operative benefits for supply chains; (6) ideally, a forecasting algorithm for promotions should be adaptable, interpretable, accurate, and stable; (7) adaptability requires that the algorithm performs well under distinct categories, different geographical locations and market changes, all factors which impact product sales; (8) our variant of kNN searches for the closest promotions using a distance that is scaled by the relevance of each feature; (9) each stock keeping unit (SKU) is treated independently and the past promotions are used to train the algorithm: (10) to produce a forecast, the inverse of these distances constitute the weights of a weighted average that aggregates the past sales; and (11) it is lightweight and interpretative and the results can be easily modified by an operator in a cost efficient framework, overcoming the classical view of a demand forecasting being costly. Aguilar-Palacios further discloses in Section II of Pages 74760-74762 that (1) a promotion can be described as a paired biset (x, y), where y [Symbol font/0xCE] R is the response variable or KPI, and x is a vector of the parameters that describe the promotion itself and determine the response variable to an extent, as there are other factors that are not known; (2) the temporal intrinsic relationship among these variables, in such a way that parameters in x are all known before the promotion starts, whereas the KPI is only known at the end of the promotional period; (3) the parameters that define a promotion and referred here as features, and they contain information in different data formats; (4) The part of x that contains n1 continuous or metric features information are random real processes where each follows a probability density function; (5) the non-metric counterpart are n2 random categorical processes where each follows a mass density function; (6) the data that we observe from historical promotions comes from the underlying joint distribution p(x, y), which contains the description of the features and response variables; (7) our motivation is to be able to estimate the values of y given the parameters in x, which are known or even can be controlled when the promotion is planned; i.e., in probabilistic terms this corresponds to know the conditional or posterior probability of the KPI for a given value of the context, given by p(y|x); (8) in the case of having complete knowledge of the statistical distribution, we could use exact probabilistic data models to estimate the effect of the features vector on the response variable, as follows: p(y|x) = p(x|y)p(y)/p(x); (9) since we do not hold the full knowledge of the distributions, this motivates us to use a parametric model whose parameters are adjusted with the data available at the prediction time and in an online manner; (10) estimate y given x through the following equation: y = ŷ + ε = Γ(x) + ε, where Γ(x) is an ideal estimator of y and ε represents the unknown information that is not captured in the context x or retained by the estimating function Γ; (11) to represent the full set of N observed historical promotions, introduce the following matrix notation X = [x1, x2, …, xN ]T and y = [y1, y2, …, yN ]T so that set Ɗ is formed by concatenating as per x and y, i.e., Ɗ = {(x1; y1), …, (xN ; yN)}, generally, Ɗ = {X, y}, and X is often called the input data matrix, or simply the data matrix; and (12) define an operator Ψ, in such a way that given a set of feature observations and a vector of free parameters θ controlling its adjustment, it yields a function g that approaches to the ideal estimator Γ. Aguilar-Palacios also discloses in Section III of Pages 74762-74765 that (1) the promotional data collected by retailers can be divided into four categories: numerical, binary, time-date variables and categorical; (2) estimate y in an interpretative manner but also favoring direct modifications of the estimate; (3) the weighted arithmetic mean is a fairly simple manner of calculating the performance of a promotion at time k based on the combination of p ≤ M historical realisations; (4) to approximate the behavior of the expert mathematically, Ψ(∙) employs a matrix of distances M and vector of variable importance v; (5) quantify the closeness between the features of promotion k and the remaining M – 1 promotions; (6) define a feature importance vector v that determines the importance of each variable by scaling the columns of M;(6) the closeness of each observation in X' to xk is therefore the product M(k)v; (7) define the weights w from Eq.(11) so the largest contributions come from the most similar promotions; (8) formulate the problem of finding v can be tackled as minimizing the norm forecast error; (9) propose a non-optimal approach based on the intuition that similar sales are driven by similar features; (10) setting the weight wi to a measure derived from the difference in sales; (11) solve a non-negative least squares (NNLS) problem established in equation (17) or (19), where the parameter λ controls the amount of regularisation that prevents a variable to massively dominate the NNLS solution; (12) find the most relevant variables related to the sales of a promotion k; (13) to gain wider insight into the features and also, to improve stability and accuracy, we repeat this process using a bootstrap aggregating (bagging) approach; (14) repeat B times the process described above and average the results to obtain the feature importance vector v, as outlined in Algorithm 1; (15) once we have the feature importance vector v, forecasting a new upcoming promotion k + 1 is a matter of two steps: (a) first, matrix M(k+1) is built using Eq.(12) to represent the feature distances between k + and the M observations from the training set; and (b) the second step is to calculate weights w that scale the historical KPIs by combining M(k+1) and v; and (16) bagging, despite its robustness, can take a longer training time than expected for an interactive application so we looked for a faster alternative without significantly decreasing accuracy. We found a sub-sampling strategy that yields similar results in a fraction of the bagging time, summarized in Algorithm 2.
Ilin et al. ("Cognitively motivated learning of categorical data with Modeling Fields Theory", International Joint Conference on Neural Networks (IJCNN), June 2012, pp. 1-8) discloses in Abstract and Section I of Page 1 that (1) Modeling Fields Theory (MFT) is utilized as the basic methodology for learning categorical data, represented by large binary vectors; (2) accelerated MAP allows simultaneous learning and selection of the number of models; (3) the key element of accelerated MAP is a steady increase of the regularization penalty combined with gradual decrease of the model vagueness; (4) Modeling Fields Theory (MFT) is a cognitively inspired mathematical framework providing a generic way of finding an optimal match between a set of models with uncertain parameters, representing a priori knowledge, and the sensor input data; (5) the match is obtained by maximizing the similarity between the models and the data, and a key feature of MFT is the vague-to-crisp process, which starts the matching by assigning almost equal association weights between the models and the data so that the process results in efficient computation and helps avoid local maxima; (6) MFT has been recently extended to models capable of handling categorical data by using binary representation, where binary vectors provide a generic way of representing categorical data, utilized in a vast variety of applications, such as image processing, text processing, and information fusion; (7) MFT provides a framework for unsupervised learning of such data by assigning each input vector to a model in a way that maximizes the total similarity; and (8) based on Maximum a Posteriori (MAP) estimation to address the model selection aspect of the problem by defining a new MFT based algorithm that starts with a large number of models and turns off the unnecessary models in the course of its execution, and introduce a constructive algorithm to decrease the number of classes during learning. Ilin further discloses in Section II of Pages 1-2 that (1) the MFT considers, within a cognitive framework, a general problem of finding the best match between a set of models {M} that depend on a set of parameters {S} and a set of data inputs {X}; (2) MFT attributes this efficiency to a vague-to-crisp process of simultaneous data association and parameter estimation; (3) the process is expressed mathematically as the maximization of the total similarity shown in eqn. (1), where the similarity between data element xn and a model Mh is measured by a function l(xn, Mh), and the total similarity is given as a product over all the data and a summation over all the models, and the quantities rh are the relative weights of the models; (4) the mathematical form of the models reflects the prior knowledge and the model parameter values are estimated based on the evidence given by the data; (5) the maximization of eqn. (1) is achieved by iterative evaluation of the model parameters and the special quantities referred to as association weights as shown in eqn. (2) with α is a constant; (6) the association weights are computed based on the current values of parameters and the parameters are recomputed in the estimation step based on the current association weights; (7) each step increases the value of the objective function and it can be shown that it converges to a (possibly local) maximum of L(X, M(S)); (7) the models can be described in probabilistic terms as probability density functions depending on parameters Sh; and (8) in this case the similarity between the data element and the model is given by the likelihood of observing the data given the model, and the optimization criterion in eqn.(1) becomes the total data likelihood. Ilin also discloses in Section III of Pages 2-3 that (1) the objective function for a probabilistic version of MFT is the likelihood of the data given the set of probability distributions or probability density functions for each of the models; (2) the probability of the data point xn conditional on model h is the measure of similarity between the data point and the model; (3) in the case of binary data, each input xn is a D-dimensional binary vector, and such data can be modeled as coming from a multivariate Bernoulli distribution as shown in eqn. (5); (4) one way to seamlessly incorporate model selection into MFT consists of using a modified objective function which penalizes overly complex data description, and this method is referred to as regularization; (5) from the Bayesian perspective it is equivalent to defining a priori probability distribution for the parameters of the models that favors fewer models; (6) in the case of Bernoulli distribution the beta function often serves as the prior, and the modified objective function takes the form shown in eqn. (6), wherein λ is the regularization constant; (7) the first term in the derivative of the objective function shown in eqn. (8) is the same as in the maximum likelihood estimation, and the second term comes from the beta distribution; and (9) the computational scheme proposed in this contribution consists of two modifications to the basic MFT algorithm: (a) the first modification is the introduction of a variable parameter b which starts with small values and gradually increases after each iteration of the algorithm, and such strategy is referred to as acceleration in this work; and (b) the second modification is the use of a special form of gradient ascent in the maximization step, where the gradient ascent employed here takes into account only the sign of the derivative (8), and each parameter is updated by a constant increment in the direction of the derivative so that such update scheme provides better control over the rate of change of the parameter values and is related to the key idea of controlling the vagueness of the models, and this algorithm is referred to in this contribution as accelerated MAP.
LEI et al. (US 2019/0188536 A1, filed on 12/18/2017), hereinafter LEI'536 discloses that a product forecasting system 870 that forecasts future product demand for one or more products at one or more retail stores (LEI’536, 870 in FIG. 8; ¶ [0095]), wherein the system use one or more trained models generated from one or more different algorithms and one or more feature sets, and may ultimately combined the forecast from multiple trained models to arrive at a final demand forecast, where trained models can include trained linear regression models or supervised machine learning techniques, such as decision or regression trees, Support Vector Machines ("SVM") or neural networks (LEI’536, ¶¶ [0032]-[0039]). LEI’536 further discloses that the system use the automatically determined one or more feature sets to generate one or more trained models generated from one or more different algorithms in order to determine a sales forecast or a demand forecast to achieve a more accurate prediction and a better understanding of the impact a promotion has on demand with less processing cycles (LEI’536, ¶¶ [0030] and [0064]). LEI’536 also discloses that a specialized point of sale ("POS") terminal 100 generates the transactional data and historical sales data, e.g., data concerning transactions of each item/SKU at each retail store, used to forecast demand (LEI’536, 100 in FIG. 1; ¶ [0026]), wherein (1) historical sales/performance data may include, e.g., a number of units of an item sold in each of a plurality of past retail periods as well as associated promotion data; (2) historical item sales data is received for all items for all stores for a particular class/category of products, or for only a single item of interest; and (3) all the valid data points are pooled to form a training dataset D with N data points at a given aggregated level, wherein the aggregate levels are typically picked to be low enough to capture the low level details of the merchandise, but also high enough that the data pool is rich enough for a robust estimation of the promotion effects (LEI’536, 202 in FIG. 2; 502 in FIG. 5; ¶¶ [0031], [0041]-[0044], and [0066]-[0068]). LEI’536 further teaches that inventory system 820 stores inventory and provides transportation logistics to deliver items to stores 801-804 using trucks 810-813 or some other transportation mechanisms, wherein the inventory system 820 uses input from forecasting system 870 to determine levels of inventories and the amount and timing of the delivery of items to stores 801-804 for a specialized inventory control (LEI’536, FIG. 8; ¶ [0096]), wherein each retail location 801-804 sends sales data and historic forecast data to forecasting system 870, and the sales data includes inventory depletion statistics for each item, or SKU/UPC for each sales period, typically days, in the previous sales cycles (i.e. weeks), typically 4-7 weeks of inventory cycles (LEI’536, ¶ [0099]). LEI’536 also teaches that historical item sales data received for all items for all stores are for a particular class/category of products; e.g., the class/category can be "yogurt", "coffee" or "milk", wherein each class has one or more subclasses, all the way down to the SKU or Universal Product Code ("UPC") level, which would be each individual item for sale; e.g., for the class of yogurt, a sub-class could be each brand of yogurt, and further sub-classes could be flavor, size, type (e.g., Greek or regular), down to an SKU which would correspond to every individual different type of yogurt item sold; e.g., the determined feature set generated from the functionality of FIG. 2 is for a given product category at a given location, such as yogurt in the Baltimore, MD area (LEI’536, FIGS. 3 and 4A; ¶¶ [0041], [0046], and [0063]). LEI’536 further discloses that historical sales/performance data may include, e.g., a number of units of an item sold in each of a plurality of past retail periods as well as associated promotion data (i.e., for each retail period, which promotions were in effect for that period); e.g., a time-period specific promotion effects for each week during a 13 week sales period (LEI’536, FIG. 7A-B; ¶¶ [0031] and [0088]-[0089]). LEI’536 also discloses that (1) for each round/iteration, the promotion effects/factors for each promotion/feature is determined during each sales period that the promotion is in effect (e.g., each week of training dataset/each bin); (2) each round/iteration can use linear regression, SVM, neural networks, etc.; (3) after each round/iteration a set of model parameters are generated that describe the training dataset/bin used; and (4) to forecast future demand, for each data point x, M(i) is iteratively applied to the input to produce the final results y as follows: y = sum( f(M(i), x) * w'(i) ), where y is the forecasted demand, and f is the function to create the forecast, corresponding to the model (LEI’536, FIGS. 5-6 and 7A-B; ¶¶ [0068]-[0072], [0076], [0079]-[0083], and [0088]). LEI’536 further teaches that the iterative process will be completed and the optimized feature set is determined when the early stopping metric or the maximum number of iterations has been reached; and an algorithm is trained using the training data set from sales history data and using the features of feature test set S (i.e., both the mandatory and optional features) to generate a trained algorithm/model (LEI’536, 208-216 in FIG. 2; ¶¶ [0048]-[0052]). LEI’536 also teaches that use the optimized feature sets as input to forecasting algorithms to generate forecasting models, wherein for each training dataset D(i) extracted/pooled from historical sale/supply chain data, one of the following machine learning algorithms are used to produce the model M(i): linear regression, Support Vector Machine ("SVM"), and Artificial Neural Networks ("ANN"); and to forecast future demand, for each data point x, M(i) is iteratively applied to the input to produce the final results y as follows: y = sum( f(M(i), x) * w'(i) ), where y is the forecasted demand, and f is the function to create the forecast, corresponding to the model (LEI’536, FIG. 5; ¶¶ [0065]-[0076]). LEI’536 further discloses that (1) FIG. 6 shows promotion effects for each promotion 1-10, which are product/location specific but not time period specific; (2) FIGS. 7A-B show a comparison of predictions for each week during a 13 week sales period and for a given store/SKU, wherein (a) row 701 provides a baseline demand; (b) row 702 provides seasonality; (c) rows 702-712 provide an indication (as indicated by an "X"), for each promotion, whether that promotion was active during the corresponding week; (d) row 713 indicates actual sales during the corresponding time period; (e) for the prediction of promotion effects, row 714 indicates the predictions of sales for each week from round A, in which all data points are used using known methods of using all available data; (f) rows 715-719 indicates the predictions/estimated using each of rounds 1-5 for each time period; and (g) row 720 is the average prediction from Rounds 1-5 (LEI’536, FIGS. 6 and 7A-B; ¶¶ [0083]-[0089]). LEI’536 further discloses that the demand mode is as follows sales = (base demand)*(seasonality)*(promo 1 effect)*(promo 2 effect)* ... (promo 10 effect), wherein base demand can be calculated as moving average, simple exponential smoothing, etc. (LEI’536, ¶¶ [0073] and [0081]-[0082]).
Palinginis et al. (US 2019/0156357 A1, filed on 10/24/2018), hereinafter Palinginis discloses that the predicted demand may be presented to a user through a graphical user interface that enables the user to interactively and iteratively select or refine one or more parameters of the promotion in order to view or adjust factors relevant to the proposed promotion; and graphical user interface 500 may be presented on a computer system as a user-friendly, explicit visualization of the predicted demand values generated by of the blended models (Palinginis, 116 in FIGS. 1 and 5A-B; ¶¶ [0044] and [0086]). Palinginis further discloses that (1) graphical user interface 500 may include a parameter input screen 510 configured to enable the user to select or define promotion parameters the user would like to model, e.g., a plurality of parameters for defining a data set, such as sales data related to the SKUs of a particular department/store, target product, and terms of the proposed promotion; (2) one or more promotion analyses may be displayed in a summary interface 520 or a more detailed promotion analysis display 530 for review by the user, e.g., the user may be able to select a promotion identifier 522 and be presented with a graphical representation of predicted demand from the model represented in a business significant series of predicted demand values; (3) also presenting one or more interactive features for enabling the user to drill down into the analysis, see the data from different perspectives, and/or provide other utilities for visualizing, manipulating, or exporting the predicted demand values; e.g., (a) time selector 532 may indicate a plurality of time segments, such as by week or combined for the entire period of the promotion, that may be selected for viewing the performance of the promotion in the various segments; (b) presentation selector 534 may provide a selection of visualization settings for the data, such as waterfall view or table view; (c) metric selector 536 may provide a selection of business metrics against which the modeled promotion can be displayed, such as incremental adjusted margin, incremental sales, and incremental variable contribution of each particular component of the promotion as represented in predicted demand values (Palinginis, 118 in FIGS. 1 and 5A-B; ¶¶ [0044] and [0087]-[0091]). Palinginis also discloses that generating predicted demand mode using cluster-level regression model and item-level correction model (i.e., blended clustering-based demand models) to calculate one or more profitability factors related to the predicted demand values, e.g., baseline, uplift, discount, vendor fund, cannibalization, pull forward, halo effect, and total incremental value, among others (Palinginis, FIGS. 2-3 and 8-9; ¶¶ [0045]-[0075] and [0109]).
Mersov et al. (US 2014/0278981 A1, published on 09/18/2014), hereinafter Mersov discloses that collect user behavior data combined with sales data relating to past campaigns so as to provide sufficient training data sets (Mersov, ¶¶ [0007] and [0086]), wherein (1) continuous or real-valued context variables may be discretized in order to facilitate optimization of media buying; e.g., in statistics and machine learning, discretization (sometimes referred to as "quantization") refers to the process of converting or partitioning continuous attributes, features or variables into to discretized or nominal versions of the same quantities and is a form of binning, as in making a histogram; (2) discretization can be accomplished in different ways; e.g., data may be discretized into partitions of K equal lengths/width (equal intervals) or K% of the total data (equal frequencies); (3) alternatively, continuous or real-valued context variables can be discretized using a percentile based approach where the data for each variable is sorted and assigned to sequential buckets, each bucket having an equal number of examples assigned; and (4) alternatively, continuous data can be discretized using an MDL method, in which best bids are defined recursively by information gains; other possibilities include CAIM, CACC, Ameva, and others (Mersov, ¶ [0089]).
Elbsat (US 2015/0316907 A1, published on 11/05/2015), hereinafter Elbsat discloses systems and methods using historical values to train a model for prediction (Elbsat, ABSTRACT), wherein binning module 602 may be configured to collect or receive historical data and to organize the historical data into various bins (e.g., sets, groups, categories, etc.), and organize the data into bins based on one or more binning parameters associated with the data, and organizes the historical data into n different bins, where n is the total number of bins (Elbsat, FIG. 6; ¶¶ [0101], [0107], and [0138]).
Ayala et al. (US 2007/0094168A1, published on 04/26/2007) discloses a system and a method relating to an artificial neural networks (Ayala, ¶ [0004]), wherein a training setup interface may provide a number of options for user selection to configure the training operation, such as an identification of a training algorithm, training algorithm parameters, learning rates, visualization parameters, and stop criteria; e.g., the organizing technique panel may provide a user with an opportunity to specify the details of the weight adjustment formula and its constituent functions and parameters; the visualization panel of the interface may allow the user to specify such parameters as a time delay, neuron interconnections visualization, a connection type, a minimal distance between connections, and whether to show diagonal grid connections; the stop criteria panel of the interface may allow the user to specify a number of iteration steps, a threshold error, and a number of patterns or iterations, after which training is stopped, wherein stopping the training process may be directed to, for example, pausing the calculations to allow the viewing of intermediate or temporal results, to test for convergence issues, or for a debugging opportunity (Ayala, ¶ [0076]).
Chen et al. (US 2009/0024450 A1, published on 01/22/2009) discloses a system and a method relating to retail sales management (Chen, ¶ [0002]), which includes (1) determining a demand distribution for a product provided by a retail presence, and (2) evaluating a probability of a lost sales occurrence, including (a) determining a predetermined time period and a probability of no sales over the predetermined time period, (b) determining if no sales have occurred over a time period corresponding in length to the predetermined time period; and (4) if the probability of no sales is below a threshold, determining an estimate of lost sales (Chen, FIG. 3; ¶¶ [0007] and [0021]). Chen further discloses the prediction engine 102 can determine an appropriate probability distribution P (e.g., Poisson, geometric, etc.) as the demand distribution which fits sample statistics (e.g., the sample mean, variance, and potentially the k-th order centered-moments) in the observed point-of-sale (POS) sample data (Chen, ¶¶ [0006] and [0022]-[0032]).
Boyle et al. (US 2018/0349795 A1, published on 12/06/2018) discloses a system and a method relating to designing a product to meet one or more goals using artificial intelligence (Boyle, ¶¶ [0001] and [0022]), wherein (1) an optimization goal is received via design tool 208 (e.g., GUI 700/800/90), where the goal components may have target segments including target business line, target product type, client segment , and seasonality, etc.; (2) the optimization goal is used to identify one or more base option candidates from among possible base options, wherein base option candidates (selected from among base options in: a catalog of products, a group of human-curated base options, and/or a machine-selected group of base options) are presented to a user/designer in a ranked order associated with their desirability with respect to the optimization goal by evaluating a divergence between actual performance and predicted performance of the base option as a product; (3) a selection of a selected base option is received; (4) one or more trained models to evaluate alternative features for the selected based option are selected (e.g., a particular category of trained model may be selected based at least in part on the optimization type; a particular segment may be selected based at least in part on the optimization goal component); (5) eligible alternative features are identified for the selected base option; (6) the selected trained model(s) are used to score and rank one or more sets of features; and (7) an ordered list of the alternative features is generated and provided based on the ranking of alternative features (Boyle, FIGS. 1-16; ¶¶ [0023]-[0025], [0041]-[0047], [0049], [0051], [0071]-[0088], and [0106]-[0132]).
However, closest arts of records, as discussed above, singly or in combination do not teach or suggest at least following features "regularizing one or more factors in one or more categories of training data by determining Bayesian a priori probability for each occurrence of a specific category of a feature variable, wherein the regularizing improves a numerical stability of the cyclic boosting model; scanning product data associated with observed target values; calculating a global average from all the observed target values across all bins and features, wherein the bins and features are defined by a matrix; initializing each factor of the one or more factors to 1, wherein each factor corresponds to a single bin and a single feature; calculating iteratively, for each feature and corresponding bin, partial factors and aggregate factors, wherein a partial factor is multiplied by an aggregate factor in each iteration of a multiple of iterations, and wherein calculation of the partial factor in each iteration includes a learning rate to reduce dependency on a sequence of features; calculating a predicted value of a target variable for each of the multiple of iterations until a stopping criteria is met, wherein the calculated predicted values follow a Poisson or Poisson-Gamma distribution; and rendering, for display on a user interface, a visualization comprising a two- dimensional visualization of one or more deviations between the calculated predicted values and corresponding actual values" when combining with all other limitations of these claims as a whole.
Response to Arguments
Applicant's arguments filed 11/26/2025 have been fully considered but they are not persuasive.
Applicant argues on Pages 9-13 of the Remarks regarding 101 rejections that the claimed invention utilizes, among other things, scanning, by automated machinery having at least one sensor, and accordingly integrates any abstract idea into a practical application in light of ¶¶ [0031] and [0096] of the specification of the claimed invention.
In response, examiner respectfully disagrees. In order to overcome 101 abstract idea rejections, the addition elements/limitations (i.e., non-abstract idea elements/limitations) must be integrated with other abstract idea elements/limitations in a meaningful manner (i.e., not just "apply it") so that the improvement of a practical application/technology, described in the speciation (e.g., ¶¶ [0031] and [0096]) of the claimed invention, is reflected in the claim(s) as a whole. There are one improvement of a practical application/technology described in ¶ [0031] and one improvement of the practical application/technology described in ¶ [0096]. ¶ [0031] describes that "based on the forecasted demand (or other retail volume) and the identification and calculated factors of features, retailers accessing one or more planning and executions systems 130, inventory system 140, and/or transportation network 150 may initiate an action to adjust inventory levels at various stocking locations, production of products of manufacturing equipment, proportional or alternative sourcing of one or more supply chain entities 160, and the configuration and quantity of packaging and shipping of products and taking into account the current inventory or production levels at one or more supply chain entities 160". In other words, "initiate an action to adjust inventory levels based on the forecasted demand" can prevent "inventory shortage" (i.e., advantage or improvement). Although the "scanning" process can monitor the "current inventory or production levels", however, without reciting "initiate an action to adjust current inventory or production levels" when the "current inventory or production levels" (i.e., observed values) are less than the "forecasted demands" (i.e., predicted values), the improvement recited in ¶ [0031] is not reflected in the claim. ¶ [0096] describes that "Cyclic boosting system 110 regularizes the factors
f
j
k
across bins k for each feature j to improve the numerical stability of cyclic boosting model 204 during training". In other words, "regularizing factors" can improve "stability of cyclic boosting model" during training so that "the forecasted demand" (i.e., predicted values) obtained from "cyclic boosting model" are more accurate. However, the improvement described in ¶ [0096] is caused by the "regularizing" process alone, and not caused by integrating the "regularizing" process and the "scanning" process. Since the "regularizing" process is also an abstract idea element/limitation (i.e., not an additional element/limitation), it cannot be used to integrate with other abstract idea elements/limitations into a practical application. Therefore, the "scanning" process in the current claim is an insignificant solution activity and is not integrated with any abstract idea elements into a practical application. In order to overcome 101 abstract idea rejection (i.e., in order for the "scanning" process becoming a significant solution activity), " 'initiate an action to adjust current inventory or production levels' (i.e., initiate an action to adjust the observed values) based on comparison of the 'current inventory or production levels' (i.e., observed values) and the 'forecasted demands' (i.e., predicted values)" must be recited in the claim so that both advantages/improvements of the practical application described in ¶¶ [0031] and [0096] (i.e., provide more accurate control/adjustment of 'current inventory levels' based on the 'forecasted demands' to prevent 'inventory shortage') can be reflected in the claim when integrating the "scanning" process with other limitations as a whole.
Conclusion
Any inquiry concerning this communication or earlier communications from the examiner should be directed to HWEI-MIN LU whose telephone number is (313)446-4913. The examiner can normally be reached Mon - Fri: 9:00 AM - 6:00 PM EST.
Examiner interviews are available via telephone, in-person, and video conferencing using a USPTO supplied web-based collaboration tool. To schedule an interview, applicant is encouraged to use the USPTO Automated Interview Request (AIR) at http://www.uspto.gov/interviewpractice.
If attempts to reach the examiner by telephone are unsuccessful, the examiner’s supervisor, Mariela D. Reyes can be reached on (571) 270-1006. The fax phone number for the organization where this application or proceeding is assigned is 571-273-8300.
Information regarding the status of published or unpublished applications may be obtained from Patent Center. Unpublished application information in Patent Center is available to registered users. To file and manage patent submissions in Patent Center, visit: https://patentcenter.uspto.gov. Visit https://www.uspto.gov/patents/apply/patent-center for more information about Patent Center and https://www.uspto.gov/patents/docx for information about filing in DOCX format. For additional questions, contact the Electronic Business Center (EBC) at 866-217-9197 (toll-free). If you would like assistance from a USPTO Customer Service Representative, call 800-786-9199 (IN USA OR CANADA) or 571-272-1000.
/HWEI-MIN LU/Primary Examiner, Art Unit 2142