Systems, media, and methods for identifying and correcting retailer virtual inventory

Abstract

A method includes calculating a model probability and a list of ranked locations and receiving a selection of a plurality of ranked locations. The method also includes receiving actual item count data from at least a subset of the selected locations and comparing the actual item count data to corresponding inventory records. The method further includes outputting a difference indication based on a difference between the actual item count data and the corresponding inventory record.

Description

BACKGROUND

[0001] Virtual inventory is a common source of shelving that is empty, loss of sales for a retailer, and frustration for its customers. Virtual inventory occurs when a retailer's inventory system reports that an item is in stock at a store, but the item is not actually available for purchase by a consumer at the store at the present time. There are various potential root causes for virtual inventory, for example, the store received less inventory than was allocated, an item was returned to inventory but later damaged and discarded, a consumer moved the item to another location so that it is effectively lost, and theft. SUMMARY

[0002] In one embodiment, a method includes computing model probabilities and a ranked list of locations and receiving a selection of a plurality of the ranked locations. The method also includes receiving actual item count data from at least a subset of the selected locations and comparing the actual item count data to corresponding inventory records. The method further includes outputting a discrepancy indication based on a discrepancy between the actual item count data and the corresponding inventory records.

[0003] In another embodiment, at least one non-transitory computer-readable medium is encoded with instructions that, when executed, configure at least one processor to compute model probabilities and a ranked list of locations and receive a selection of a plurality of the ranked locations. The non-transitory computer-readable medium also includes instructions to receive actual item count data from at least a subset of the selected locations and compare the actual item count data to corresponding inventory records. The non-transitory computer-readable medium further includes instructions to output a discrepancy indication based on a discrepancy between the actual item count data and the corresponding inventory records.

[0004] In yet another embodiment, a system includes at least one computing device including one or more processors and at least one memory coupled to at least one of the one or more processors. The at least one memory includes instructions that configure the at least one computing device to compute model probabilities and a ranked list of locations and receive a selection of a plurality of the ranked locations. The at least one computing device is also configured to receive actual item count data from at least a subset of the selected locations and compare the actual item count data to corresponding inventory records. The at least one computing device is also configured to output a discrepancy indication based on a discrepancy between the actual item count data and the corresponding inventory records. BRIEF DESCRIPTION OF DRAWINGS

[0005] Figure 1 is a schematic diagram illustrating virtual inventory identification and correction operations in accordance with various embodiments herein;

[0006] Figure 2is a schematic diagram illustrating data splits for classification model training, tuning, and testing according to various embodiments herein;

[0007] Figure 3 is a measure of feature importance for a light gradient boosting machine model according to various embodiments herein;

[0008] Figure 4A is a plot illustrating feature dependence for number of days since last sale according to various embodiments herein;

[0009] Figure 4B is a plot illustrating feature dependence for number of units in stock according to various embodiments herein;

[0010] Figure 5A an example of an in-store audit for counting products on a shelf and correcting any discrepancies in virtual inventory in a store system is illustrated according to various embodiments herein;

[0011] Figure 5B is a schematic diagram illustrating a hierarchy of product flows according to various embodiments herein;

[0012] Figure 6A an example of an actual weighted value payoff matrix is illustrated according to various embodiments herein;

[0013] Figure 6B an example of an actual unweighted value payoff matrix is illustrated according to various embodiments herein;

[0014] Figure 6C an example of a randomized weighted value payoff matrix is illustrated according to various embodiments herein;

[0015] Figure 6D an example of a randomized unweighted value payoff matrix is illustrated according to various embodiments herein;

[0016] Figure 7 a payoff matrix is illustrated according to various embodiments herein;

[0017] Figure 8 a flowchart for identifying / correcting virtual inventory is illustrated according to various embodiments herein;

[0018] Figure 9 a front-end user interface for a software application is illustrated according to various embodiments herein; and

[0019] Figure 10 is a block diagram illustrating computing hardware used according to various embodiments herein. DETAILED DESCRIPTION

[0020] Virtual inventory can include the impact of potentially depriving consumers of the opportunity to purchase items. In some cases, consumers may substitute with another item that may have a different profit margin, selling price, or other business attributes. In other cases, consumers may simply leave the store, causing the retailer to potentially lose all the potential sales that the customer would have otherwise made. Worse still, consumers may choose to stop shopping at that retailer altogether and switch to a competitor.

[0021] Product manufacturers may also be concerned about preventing virtual inventory. Both manufacturers and retailers can benefit when consumers purchase items. In implementations, retailers typically reorder products based on automated algorithms triggered when reported inventory falls below a threshold; they may also consider volatile or seasonal demand. Without automated algorithms to provide a rollback rhythm for reordering, instances of virtual inventory could prevent automated inventory reordering over extended periods.

[0022] Retailers may face the challenge that manual audits to correct virtual inventory are both time-consuming and expensive, both in terms of hourly wages and the lost opportunity to better utilize workers who might be performing other revenue-generating duties. Conducting continuous manual audits can also be inefficient, as not all items are equally likely to be in virtual inventory, and given limited resources and time, not all items are equally valuable to correct from virtual inventory. Therefore, identifying and correcting virtual inventory benefits manufacturers, retailers, and consumers alike.

[0023] This paper describes methods, media, and systems for identifying and correcting virtual inventory. In the identification operation, a classification model can identify and categorize items that may be in a virtual inventory state in the future (hours, days, weeks, etc.) at a given store on a given day. In the correction operation, the classification model can prioritize items, item categories, and stores requiring intervention based on estimated net values ​​and the results of the identification operation. Interventions may include correcting the retailer's inventory system or reordering. The algorithmic processes used for the identification and correction operations are flexible, and illustrative implementation examples are provided in this paper.

[0024] In embodiments, the identification operation can be accomplished using a classification model. Classification model inputs can include point of sale (POS) trends, such as how many days have passed since the item was last sold at the store and / or item characteristics, such as physical dimensions. Classification model inputs can also include planogram (POG) characteristics, such as the height of the shelf where the item is located and / or store characteristics, such as total sales volume, which are used to compare item sales at the store to similar stores. The classification model can output a virtual inventory probability score between 0 and 1 (or can utilize any other suitable range), where as a non-limiting example, a probability closer to 1 indicates a higher likelihood that the item is in a virtual inventory state, such as the next day in this non-limiting example. Class labels can be used to indicate whether a given item is predicted to be in a virtual inventory state (in this example, these class labels are binary, such that yes = 1, no = 0), and can be computed based on a selected probability threshold, which can be static or dynamic. For example, an adjustable probability threshold can be adjusted at a category or item level. As a non-limiting example, increasing the threshold will decrease the number of positive predictions but increase the accuracy of those predictions, while lowering the threshold will increase the number of positive predictions (increase sensitivity or recall) but decrease the accuracy. In some embodiments, an adjustable probability threshold allows flexibility around business constraints or preferences regarding the cost and benefit of correcting predicted cases of virtual inventory. In some embodiments, the ranked positions are determined at least in part by a return on investment matrix estimation calculation. In various embodiments, the at least one processor is further configured to output instructions to replenish inventory for items that are not available or that are indicated by the model as likely not available. In some embodiments, the ranking of positions is based at least in part on an overall value of fixing discrepancies within a category. In various embodiments, the at least one processor is further configured to monitor performance of the classification model based at least in part on received inputs including auditor intervention to correct inventory quantities, where the monitoring further includes utilizing performance metrics for model degradation, and where the at least one processor is further configured to retrain the model based on detected model degradation. In some embodiments, the model probability includes a probability that an item is not present at a location on a particular date.

[0025] Figure 1Embodiments of the identification operation are illustrated. The classification model can receive input information for model training 102, where the input information can include data such as store, shelf, and item attribute data 104, comparison data to similar stores 106, sales data 107, retail annual audit data 108, prioritized intervention audits 110, and / or any other suitable type of data. In some embodiments, the retail annual audit data 108 can refer to any suitable amount of time other than annual, such as daily, weekly, monthly, etc. In embodiments, the correction operation of the classification model can include translating item-level probabilities and other data into a revenue matrix of dollar value data. In some embodiments, this can include historical sales volume and dollar amounts, seasonality, inventory availability, and / or substitution estimates, by way of non-limiting example. As also described herein, converting item-level probabilities received from the identification operation into a revenue matrix input of dollar values can optionally include sales data (e.g., average dollar sales expected from the item at the store on the day), business data (e.g., hourly wage spent on intervention), expected store foot traffic data, customer substitution behavior data, basket data, loyalty loss clause data, etc. The revenue matrix input can capture the overall impact of virtual inventory. Based on the revenue matrix or other ranking, targeted interventions can be used for high value stores, item categories, or items. High value stores and item categories can be audited, and the retailer inventory system can be corrected if necessary. This correction data can then be fed back into the classification model to retrain and improve the classification model. After the retailer inventory is corrected, an automated reordering algorithm can be triggered to restock the store.

[0026] Figure 1 The correction operation is also illustrated. Virtual inventory probability scores can be calculated at the store, item, and day level, and some virtual inventory probability scores can be set to zero using business logic 112. If the virtual inventory situation is not corrected, the expected missed sales dollar value can be used to prioritize items that need intervention to fix virtual inventory 114. The overall value of fixing virtual inventory for a category can be used to rank stores for audit 116.

[0027] If available, model input data to train the classification model to identify virtual inventory can be provided by a major store audit of actual inventory conducted by the retailer or by another party. The major store audit can then be utilized to provide ground truth to train the classification model. The classification model can be trained on a major model dataset from the major store audit up to the date of the subsequent annual store audit. Each record in the training data can be assigned a target class label, 0 or 1 in this example. In this non-limiting example, 1 is the virtual inventory status and applies in the following circumstances: (1) the reported item inventory is greater than 0 on the day before the subsequent store audit; (2) the adjusted (corrected) inventory is 0 at the end of the subsequent audit day; (3) and 0 units of the item were sold on the subsequent audit day. The requirement to utilize 0 units of the item sold on the audit day makes the 0 item inventory at the end of the subsequent audit day independent of sales on that day. Otherwise, the records in the training data can be assigned the target class label 0. Examples are provided in Tables 1 and 2. In Table 1, the inventory is corrected by an audit from 19 units to 12 units and the item at that store is not in a virtual inventory status. In Table 2, the inventory has been corrected by an audit from 13 units to 0 units and the item in that store is in a virtual inventory status.

[0028]

[0029] Table 1: Example of items not in virtual inventory status on August 3, 2022, regarding virtual inventory for the next day 0 .

[0030]

[0031] Table 2: Example of items in virtual inventory status on August 3, 2022, regarding virtual inventory for the next day =1 .

[0032] An example embodiment of the classification model can use a binary classification algorithm based on a Light Gradient Boosting Machine (LGBM) model. Comparative testing of different binary classification algorithms using a tool such as AutoML in Databricks, which runs dozens of models tuned in different ways and ranks them based on a selected metric on validation or test data, indicates that the LGBM model has better performance. However, other embodiments can utilize sklearn (a python-based open source machine learning package), binary classification algorithms including logistic regression, random forest, XGBOOST, and / or any other suitable algorithm.

[0033] Turning to Figure 2A classification model can be trained on an available historical dataset 202, where the historical dataset can be based on audit data and classification model inputs. The classification model can be trained with hyperparameters 204 that are best tuned. Hyperparameters, such as learning rate of an algorithm as a non-limiting example, can be external to the classification model inputs and can be used to control the learning process of the classification model. In some embodiments, hyperparameter tuning 206 can be performed on a training subset of 80% of the historical dataset, such as a tuning data subset 212, using a python-based hyperopt package. The hyperopt package in this embodiment is based on Bayesian optimization. The training subset of the historical dataset, such as the tuning data subset 212, can be further split into a 60% tuning data subset 212 and a 20% validation data subset 214. The 60% tuning data subset 212 can be used to train the classification model to identify virtual inventory. The 20% validation data subset 214 can be used to test the classification model trained on the 60% tuning data subset 212. The tuned classification model can then be estimated on a 20% holdout test data subset 216 as part of model testing 208, where the holdout test data subset 216 can be the historical dataset 202 that was not used for tuning or validation of the classification model. Out-of-time (OOT) testing can also be performed using an OOT dataset 218 from a time period that the classification model has never seen. This test can be performed to achieve similar performance of the model on the holdout test data subset 216 on the test OOT dataset 218. The impact of each input feature on the virtual inventory probability can be computed using a python-based Shapley Additive Explanations (SHAP) explainability package or a similar feature importance computation method. SHAP is based on game theory concepts and Shapley values can be used to estimate the contribution of a given feature to the virtual probability prediction of the classification model.

[0034] Turning to Figure 3 , a graphical illustration depicts the impact of different input features in an embodiment. The y-axis 302 can be used to represent different input features and the x-axis 304 represents the impact on the virtual inventory probability with respect to each of the different input features by means of average SHAP values. As can be seen from this plot, the number of days since last sale and the inventory units are the most important features contributing to the virtual probability prediction of the classification model. Feature dependence that can be used to illustrate the impact of individual input features on the prediction made by the classification model can also be computed with the SHAP explainability package or a similar feature importance computation method.

[0035] Figure 4A and Figure 4B the impact of different input features on the virtual inventory probability is represented on the x-axis 404, x-axis 410 Figure 3An example graphical illustration of feature dependence in terms of SHAP values on y-axis 402, y-axis 408 for two most important features of a product (i.e., days since last sale and units in stock). Units in stock = 0 408 means target score = 0 (if the system knows that the stock is zero, it is not “virtual”). The lower the number of reported units in stock (e.g., reported units in stock 412), the greater the likelihood that the target score = 1. The greater the number of days since last sale (e.g., days 406), the greater the likelihood that the target score = 1.

[0036] Figure 5A An example of an in-store audit 500a is illustrated for counting products on shelves and correcting any instances of virtual inventory in a store system. The implementation can be flexible with respect to how the intervention is conducted. For example, a store employee can check all items within a category that has been flagged as valuable to correct in that store on that day. This can require additional time compared to spot checking only specific items, but there can be items in virtual inventory that were not flagged by the model, and those items can be subsequently corrected. A hybrid approach between spot checking and full category checking can involve an employee spot checking flagged items, and then additionally checking any “holes” in the POG or shelves to see if those items are truly out of stock or in virtual inventory. Figure 5B A hierarchy of product flow from manufacturers to rapid deployment centers to individual stores is shown. In Figure 5B Different categories of intervention can include a distribution center level 502, a rapid deployment center (RDC) level 504 that receives products from manufacturers (such as 3M ® ), and store level allocation 506 based on estimated current demand. The estimated value of the intervention can be consumed by the RDC, for example via an application programming interface (API), and this can trigger a micro-reallocation. This can allow the business to impose its own logic and / or revenue calculations on top of the method output. In some embodiments, the intervention can be an algorithm at the RDC. A variety of metrics can be considered to evaluate the performance of the classification model, including:

[0037] Accuracy: The percentage of cases in which the classification model virtual inventory probability prediction matches the actual virtual inventory. The interpretation can depend on the relative ratio of actual virtual inventory (PI) cases to non-PI cases (e.g., a model that always predicts non-PI is 92% “accurate” if 8% of cases are actual PI).

[0038] Area under the curve (AUC): The area under the receiver operating characteristic (ROC) curve for true positive rate versus false positive rate, which provides a measure of how well the classification model separates classes independent of the classification threshold.

[0039] Kolmogorov-Smirnov test (KS): A statistic comparing the model probability distributions for actual PI vs. actual non-PI cases. A high KS value means that the classification model is doing a good job at separating the classes, independent of the threshold as well.

[0040] d-prime: A sensitivity index obtained from signal detection theory that is a measure of separation between the model probability distributions for actual PI vs. actual non-PI cases, taking into account their dispersions.

[0041] β: The optimal threshold obtained from signal detection theory, taking into account the costs / benefits of correct rejections, false alarms, misses, and hits: a benefit matrix. For PI, this can be at the item level, where the business provides a benefit matrix.

[0042] In some embodiments, a classification model with a higher d-prime value can be preferred, as a higher d-prime value can correspond to a more clear separation of the predicted probability distributions of the two classes. Because the problem is imbalanced (there are many more cases of PI=0 than PI=1), some metrics achieve good recognition of PI=0 cases. However, good recognition of PI=1 cases is equally or more important from a business perspective. Embodiments of the model can be tuned using d-prime as an optimization metric, and evaluated using the d-prime OOT score. In this example, the d-prime-based optimization achieves better class separation than some alternative optimization metrics.

[0043] For corrective action, some embodiments can use a benefit matrix to estimate the net benefit of a classification model recognizing a case of a potential virtual inventory correction. The estimated value of an intervention can be calculated based on the average duration (in days) until the virtual inventory would be corrected (e.g., until an automated re-order or until an internal store audit) under normal business conditions, minus the time taken to restock after the inventory system is corrected, multiplied by the average dollar value of daily sales that would have been expected if the item had been in stock at the store. Wage costs and opportunity loss of an employee checking the item can be subtracted. As a non-limiting example, item-level estimates of intervention can be aggregated to sub-categories or categories (e.g., across all masking tape) or stores.

[0044] One non-limiting example of determining an estimated value of an intervention is: estimated value of intervention = ((average number of days until the virtual inventory would be corrected - time to restock) x (expected average dollar value of daily sales of the item if in stock)) - wage and opportunity costs, although any suitable approach can be utilized. The benefit matrix estimate can then be used to rank stores for prioritization of inventory audits.

[0045] Some modifications to the revenue matrix approach can provide improved efficiency. As a non-limiting example, instead of a simple average daily sales, it can be an average sales by day of the week. According to this approach, products that sell a lot on the weekends can be made to bubble up more frequently on Fridays for intervention, as compared to items with higher revenue that are less likely to sell on Mondays.

[0046] Here, the results of an illustrative revenue matrix calculation are shown based on POS average daily sales and typical time until virtual inventory is resolved without intervention. This illustrative scenario is based on an in-store audit to correct virtual inventory. Figure 6A to Figure 6D The positive correlation between the predicted actual values 600a (weighted) and the inferred actual values 600b (unweighted) is illustrated as compared to the random shuffle baseline 600c (weighted), 600 (unweighted). The values are based on the day before a subsequent store audit to provide the primary data set. The predicted revenue values show sufficient positive correlation with the inferred actual revenue values (600a, 600b) for all three key categories considered as compared to the random shuffle (600c, 600d) where there is no correlation. Thus, this confirms the approach herein for prioritizing stores and categories for intervention.

[0047] Turning to Figure 7 , a matrix is depicted to showcase additional data, demonstrating that more complex revenue matrix calculations are possible. Starting with the POS average 700, then considering customer substitution behavior 702, and finally adding basket and loyalty loss estimates 704. Thus, some embodiments can identify virtual inventory by a classification model, then correct it by an in-store audit prioritized based on revenue values.

[0048] Turning to Figure 8 , an illustrative flowchart 800 for identifying / correcting virtual inventory, which can be run daily or at any other suitable frequency. At block 802, the flowchart begins and can proceed to block 804 to process the latest sales data and inventory data from stores. At block 806, input data to a classification model can be received to predict virtual inventory probabilities for individual items. The virtual inventory probabilities can be utilized in a revenue matrix estimate calculation to determine a ranked list of recommended stores for auditing. The ranked list of recommended stores can also be displayed to a user, such as an auditor.

[0049] At block 808, the user can then select the particular store that they wish to prioritize for auditing. At block 810, the auditor can then travel to the particular store and conduct an inventory audit. Details of the audit process are described herein. Information gathered during the store audit by the auditor can be input into the system to save the data. At block 812, a comparison of the information from the store audit to the last known inventory record can be made to reveal inventory discrepancies. If there are discrepancies (condition yes), at block 814, updates can be recommended to the store’s inventory system, and the flowchart proceeds to block 816. Otherwise, if there are no inventory discrepancies (condition no), the flowchart proceeds directly to block 816 to save the final results of the audit for additional analysis.

[0050] At block 818, performance metrics / key performance indicators (KPIs) can be calculated to measure the performance of the classification model and the impact of auditor intervention on correcting the inventory system. At block 820, the model degradation of the classification model can be monitored via the performance metrics and / or KPIs. If the model has degraded (condition yes), at block 822, a recommendation to retrain the classification model based on the monitoring results can be provided, and the flowchart can terminate at block 824. Otherwise, if the model has not degraded (condition no), the flowchart can proceed directly to terminate at block 824. In some embodiments, at block 824, the flowchart can loop back and return to block 802 for further execution.

[0051] The classification model can run on a software application. As a non-limiting example, the software application can provide the auditor with store ordering and store their audit result data. Other methods of passing and recording information, such as API calls, can also be used. Figure 9 An example of a front-end user interface of the software application described herein is illustrated. The auditor can take one or more pre-audit photos 900 that include shelves of items in the store. The photos can show items that are out of stock, some of which can constitute virtual inventory, and other defects, such as misplacement of items, which can also be corrected by the auditor. The auditor can then count how many of each item exist and the location of each of the items within the store. The auditor can also collect additional metadata, including but not limited to one or more post-audit photos 902 to record the completion of the audit.

[0052] After the audit is completed, the software application can produce a summary of the results. As a non-limiting example, the counts found by the auditor can be flagged as the audited count column 904 in a table titled "Inventory Discrepancy Summary." In many cases, the audited count matches the store inventory system record, which is the reported inventory column 906, such that the discrepancy, represented by the discrepancy column 908, can be zero units. In other cases, there can be a non-zero discrepancy, and the audit can help the store correct the inventory. Specifically, when the reported inventory is greater than zero and the audited count is zero (and no items were sold that day), the item can be corrected from the virtual inventory's state. The first row depicts an example of a virtual inventory, where the reported inventory is corrected from six units to zero. The corrected inventory will then fulfill the replenishment to zero inventory, and also optionally fulfill the replenishment to low inventory or follow other retailer logic. The replenishment (e.g., from a deployment center) can occur after the audit is completed. In some embodiments, the user can be provided instructions to (i) replenish the inventory for the unavailable item and / or (ii) automatically replenish the inventory for the unavailable item. After the out-of-stock item is replenished, it is expected that the purchases by consumers at that store will resume, providing positive results to consumers, retailers, and manufacturers, among others.

[0053] Reference is now made to Figure 10 , a block diagram illustrates computing hardware, such as exemplary computing device 1000, through which embodiments of the present disclosure can be implemented. The computing device 1000 described herein is merely one example of a suitable computing device and does not imply any limitation on the scope of any presented embodiments. Nothing described herein should be interpreted as creating any type of dependency or requirement for any element or elements to be present in any embodiment. In various embodiments, the computing device 1000 can include, but is not necessarily limited to, a desktop computer, a laptop computer, a server, a client, a tablet computer, a smart phone, a computing cloud, or any other type of device that can utilize data. In one embodiment, the computing device 1000 includes at least one processor 1002 and memory including non-volatile memory 1008 and / or volatile memory 1010. The computing device 1000 can include one or more displays, display hardware, and / or output devices 1004, such as, for example, AR / VR / MR / XR hardware, monitors, speakers, headphones, projectors, wearable displays, holographic displays, and / or printers. The output devices 1004 can also include, for example, displays and / or speakers, devices that emit energy (radio, microwave, infrared, visible light, ultraviolet, x-ray, and gamma-ray), electronic output devices (Wi-Fi, radar, lasers, etc.), and (any frequency) audio, among others.

[0054] The computing device 1000 can also include one or more input devices 1006, which can include, by way of example, any type of mouse, keyboard, disk / media drive, memory stick / thumb drive, memory card, pen, touch input device, biometric scanner, gaze and / or blink tracker, tracker, voice / audio input device, motion detector, camera, ruler, and any device capable of measuring data such as motion data (e.g., accelerometer, GPS, magnetometer, gyroscope, etc.), biometric data (e.g., blood pressure, pulse, heart rate, perspiration, temperature, voice, facial recognition, motion / pose tracking, gaze tracking, iris or other type of eye recognition, hand geometry, oxygen saturation, glucose level, fingerprint, DNA, dental record, weight, or any other suitable type of biometric data, etc.), video / still images, and audio (including human-audible and human-inaudible ultrasonic sound). The input devices 1006 can include any type of device capable of receiving data from another device, from real-world captured visual and / or audio data, and object detection data, etc. The input devices 1006 can include a camera (with or without audio recording), such as a digital and / or analog camera, still camera, video camera, thermal imaging camera, infrared camera, camera with charge-coupled display, night vision camera, three-dimensional camera, webcam, and audio recorder, etc.

[0055] The computing device 1000 typically includes non-volatile memory 1008 (e.g., ROM, flash memory, etc.), volatile memory 1010 (e.g., RAM, etc.), or a combination thereof. The network interface 1012 can facilitate communication with other data sources, such as a database 1018, over a network 1014 via wires, wide area networks, local area networks, personal area networks, cellular networks, and satellite networks, etc. Suitable local area networks can include wired Ethernet and / or wireless technologies such as, for example, wireless fidelity (Wi-Fi). Suitable personal area networks can include wireless technologies such as, for example, IrDA, Bluetooth, wireless USB, Z-Wave, ZigBee, and / or other near-field communication protocols. Suitable personal area networks can similarly include wired computer buses such as, for example, USB and FireWire. Suitable cellular networks can include, but are not limited to, technologies such as LTE, WiMAX, UMTS, CDMA, and GSM. The network interface 1012 can be communicatively coupled to any device capable of sending and / or receiving data via one or more networks 1014. Thus, the network interface 1012 can include a communication transceiver for transmitting and / or receiving any wired or wireless communication. For example, the network interface 1012 can include an antenna, modem, LAN port, Wi-Fi card, WiMax card, mobile communication hardware, near-field communication hardware, satellite communication hardware, and / or any wired or wireless hardware for communicating with other networks and / or devices.

[0056] A computer-readable medium 1016 includes one or more computer-readable media each of which is non-transitory. The computer-readable medium can reside in, for example, the input device 1006, the non-volatile memory 1008, the volatile memory 1010, or any combination thereof. The readable storage media can include a tangible medium that can retain instructions associated with the device or system and / or used by or of the device or system. Computer-readable media (also referred to as non-transitory computer-readable media) includes, as non-limiting examples, RAM, ROM, cache, optical media, EPROM / flash memory, CD / DVD / BD-ROM, hard drives, solid state memory, optical or magnetic storage devices, magnetic diskettes, drives, or any combination thereof. Non-transitory computer-readable media also can include a system or device, for example, a magnetic, optical, semiconductor, or electrical type, for example. Non-transitory computer-readable media does not include a carrier wave, a propagated signal, or any other form of a transitory computer-readable medium.

[0057] The computing device 1000 can include one or more network interfaces 1012 to facilitate communications with one or more remote devices, which can include, for example, client and / or server devices. The network interface 1012 also can be described as a communication module, as these terms are used interchangeably. The database 1018 is depicted as being accessible over the network 1014 and can reside in a server, cloud, or any other configuration to support remote access to data and storage of data in the database 1018.

Claims

1. A method, the method comprising: Calculate the model probability and the sorted list of positions; Receive selections from a plurality of the sorted positions; Receive actual item count data from at least a subset of the selected locations; Compare the actual item count data with the corresponding inventory records; as well as A difference indication is output based on the difference between the actual item count data and the corresponding inventory record.

2. The method of claim 1, further comprising outputting instructions to replenish inventory for unavailable items or items indicated by the model as potentially unavailable.

3. The method of claim 1, wherein the ranking of positions is based at least in part on the overall value of the differences within the repair category.

4. The method of claim 1, further comprising monitoring the performance of the classification model using computational performance metrics based at least in part on received inputs, including auditor intervention to correct inventory quantities.

5. The method of claim 4, wherein the monitoring further comprises utilizing a performance metric for model degradation.

6. The method of claim 5, further comprising retraining the model based on the detected model degradation.

7. The method of claim 1, wherein the model probability includes the probability that an item is not present at a location on a specific date.

8. At least one non-transitory computer-readable medium storing instructions, which, when executed, configure at least one processor to: Calculate the model probability and the sorted list of positions; Receive selections from a plurality of the sorted positions; Receive actual item count data from at least a subset of the selected locations; Compare the actual item count data with the corresponding inventory records; as well as A difference indication is output based on the difference between the actual item count data and the corresponding inventory record.

9. The at least one non-transitory computer-readable medium according to claim 8, wherein the sorted position is determined at least in part by a payoff matrix estimation calculation.

10. The at least one non-transitory computer-readable medium of claim 8, wherein the at least one processor is further configured to output instructions to replenish inventory for unavailable items or items indicated by the model as potentially unavailable.

11. The at least one non-transitory computer-readable medium of claim 8, wherein the ordering of positions is based at least in part on the overall value of the differences within the repair category.

12. The at least one non-transitory computer-readable medium of claim 8, wherein the at least one processor is further configured to monitor the performance of the classification model using computational performance metrics based at least in part on received inputs, including auditor intervention to correct inventory quantities.

13. The at least one non-transitory computer-readable medium of claim 12, wherein the monitoring further comprises utilizing a performance metric for model degradation.

14. The at least one non-transitory computer-readable medium of claim 13, wherein the at least one processor is further configured to retrain the model based on the detected model degradation.

15. The at least one non-transitory computer-readable medium of claim 8, wherein the model probability includes the probability that an item is not present at a location on a particular date.

16. A system comprising: At least one computing device, the at least one computing device comprising one or more processors; as well as At least one memory coupled to at least one of the one or more processors, wherein the at least one memory includes instructions that configure the at least one computing device to: Calculate the model probability and the sorted list of positions; Receive selections from a plurality of the sorted positions; Receive actual item count data from at least a subset of the selected locations; Compare the actual item count data with the corresponding inventory records; as well as A difference indication is output based on the difference between the actual item count data and the corresponding inventory record.

17. The system of claim 16, wherein the at least one computing device is further configured to output instructions to replenish the inventory for unavailable items or items indicated by the model as potentially unavailable.

18. The system of claim 16, wherein the ranking of positions is based at least in part on the overall value of the differences within the repair category.

19. The system of claim 16, wherein the at least one computing device is further configured to monitor the performance of the classification model using computational performance metrics based at least in part on received inputs, including auditor intervention to correct inventory quantities.

20. The system of claim 19, wherein the monitoring further includes utilizing performance metrics for model degradation.

Images