Method and apparatus for calibrating user activity models used by mobile devices

Abstract

Systems, computer-implemented methods, and computer program products that can facilitate calibrating a user activity model of a user device node are described. According to embodiments, a method for calibrating a user activity model used by a mobile device can include receiving sensor data from a sensor of the mobile device. Further, a first weight can be applied to a first likelihood of a first occurrence of a first activity, where the first likelihood is determined by a first estimator of the user activity model by applying a preconfigured criterion to the sensor data. The method can also include performing an action based on a determination of the first occurrence of the first activity, the determination based on the first weight and the first likelihood of the first occurrence of the first activity.

Description

[0001] Related Applications

[0002] The subject patent application claims priority to pending U.S. Application No. 16 / 582,241, filed September 25, 2019, entitled “METHOD AND APPARATUS FOR CALIBRATING A USER ACTIVITY MODEL USED BY A MOBILE DEVICE,” by Karanpreet Singh et al., which claims priority to U.S. Provisional Application No. 62 / 857,330, filed June 5, 2019, entitled “METHOD AND APPARATUS FOR CALIBRATING A USER ACTIVITY MODEL USED BY A MOBILE DEVICE,” by Karanpreet Singh et al. These applications are incorporated by reference in their entirety for all purposes.

[0003] Copyright Notice

[0004] A portion of the disclosure of this patent document contains material that is subject to copyright protection. The copyright owner has no objection to the facsimile reproduction by anyone of the patent document or the patent disclosure, as it appears in the Patent and Trademark Office patent file or records, but otherwise reserves all copyright rights whatsoever. TECHNICAL FIELD

[0005] One or more embodiments relate generally to the field of human-computer interaction, and more particularly to methods, apparatuses, and systems for calibrating a user activity model used by a mobile device.

[0006] BACKGROUND

[0007] The material discussed in the background section should not be assumed to be prior art merely because of its discussion in the background section. Similarly, the material discussed in the background section should not be assumed to be pertinent art only because of its discussion in the background section. The subject matter discussed in the background section is merely for

[0008] Human activity monitoring devices are becoming increasingly popular. Different devices can use different methods to interpret data collected from the device sensors. However, problems arise when the model used to interpret the sensor data is based on a sample of the mainstream population.

[0009] SUMMARY

[0010] The following presents a summary to provide a basic understanding of one or more embodiments of the application. This summary is not intended to identify key or critical elements or delineate any scope of certain embodiments or any scope of any claims. Its sole purpose is to present concepts in a simplified form as a prelude to the more detailed description that is presented later. In one or more embodiments described herein, devices, systems, methods, and computer-implemented methods that can facilitate calibrating a user activity model of a user device are described.

[0011] According to embodiments, a method for calibrating a user activity model used by a mobile device can include receiving sensor data from a sensor of the mobile device. Further, applying a first weight to a first likelihood of a first occurrence of a first activity, wherein the first likelihood is determined by a first estimator of the user activity model by applying preconfigured criteria to the sensor data. The method can also include performing an action based on a determination of the first occurrence of the first activity, the determination based on the first weight and the first likelihood of the first occurrence of the first activity. In a variant, the first activity can be an activity of a user of the mobile device. The variant can also include facilitating an assessment of a physical characteristic of the user of the mobile device, and selecting the first weight based on the assessment of the physical characteristic of the user. In some implementations, wherein selecting the first weight can include selecting the first weight to improve an accuracy of the first likelihood of the first occurrence of the first activity for the user of the mobile device. The variant can further select the first weight based on training data of the first activity. In additional or alternative embodiments, determining the first occurrence of the first activity can include comparing the first likelihood of the first occurrence of the first activity to a second likelihood of a second occurrence of a second activity. In the embodiments discussed above, applying the first weight to the first likelihood can include modifying the first likelihood. Further, receiving the sensor data can include receiving data from at least one of an accelerometer, a magnetometer, or a gyroscope.

[0012] In another embodiment, a mobile device can include a sensor, a processor, and a memory that can store executable instructions that, when executed by the processor, facilitate performance of operations including receiving sensor data from the sensor and applying a first weight to a first likelihood of a first occurrence of a first activity, wherein the first likelihood is determined by a first estimator of a user activity model by applying preconfigured criteria to the sensor data. The operations can also include performing an action based on a determination of the first occurrence of the first activity, the determination based on the first weight and the first likelihood of the first occurrence of the first activity.

[0013] In variations of the above embodiments, the first activity can be an activity of a user of the mobile device. Further, the operations can further include facilitating an assessment of a physical characteristic of the user of the mobile device, and selecting the first weight based on the assessment of the physical characteristic of the user. In some embodiments, selecting the first weight can include selecting the first weight to improve an accuracy of the first likelihood of the first occurrence of the first activity for the user of the mobile device. In one or more embodiments, selecting the first weight can be further based on training data for the first activity. Further, determining the first occurrence of the first activity can include comparing the first likelihood of the first occurrence of the first activity to a second likelihood of a second occurrence of a second activity. In additional or alternative embodiments, applying the first weight to the first likelihood includes modifying the first likelihood. In some embodiments, the sensors can include one or more of an accelerometer, a magnetometer, or a gyroscope.

[0014] In another embodiment, a computer-readable recording medium has program instructions executable by various computer components to perform operations including receiving sensor data from sensors of a mobile device, and applying a first weight to a first likelihood of a first occurrence of a first activity, where the first likelihood is determined by a first estimator of a user activity model by applying a preconfigured criterion to the sensor data. In some embodiments, the operations can further include performing an action based on a determination of the first occurrence of the first activity based on the first weight and the first likelihood of the first occurrence of the first activity. Further, in this embodiment, the operations can further include facilitating an assessment of a physical characteristic of a user of the mobile device, and selecting the first weight based on the assessment of the physical characteristic of the user, where the first activity is an activity of the user.

[0015] In some implementations, selecting the first weight can include selecting the first weight to improve an accuracy of the first likelihood of the first occurrence of the first activity for the user of the mobile device. Additionally, in one or more embodiments, selecting the first weight can be further based on training data for the first activity. BRIEF DESCRIPTION OF DRAWINGS

[0017] The included drawings are for illustrative purposes and are in no way limiting of the embodiments presented herein. These drawings are intended to provide examples of possible structures and process steps for the disclosed technology. These drawings are not meant to limit the scope of the embodiments in any way.

[0018] Figure 1 An example of a system according to one or more embodiments is shown that can calibrate an activity model used by a device based on weights generated by a personal weight determiner.

[0019] Figure 2A more detailed view of the general model component and the operation of the personal weight applier and personal weight determiner is shown in accordance with one or more embodiments.

[0020] Figure 3 An example formula that can describe the modification of gradient boosting machine (GBM) model parameters is depicted in accordance with one or more embodiments.

[0021] Figure 4 An example formula that can describe the use of a loss function to select weights to be applied to an estimator is depicted in accordance with one or more embodiments.

[0022] Figure 5 A flowchart of a single-user output cross-validation (CV) process for adjusting GBM model weights and model evaluation is depicted in accordance with one or more embodiments.

[0023] Figure 6 A table including example features that can be used for activity classification is included in accordance with one or more embodiments. To illustrate different concepts, the following Figure 6 and Figure 7 discuss daily and sports activity datasets.

[0024] Figure 7 A comparison between the baseline of a GBM model with single-user output CV accuracy and other ML models is depicted in accordance with one or more embodiments.

[0025] Continuing with this example, Figures 8A-8B and Figure 9 depict charts 810-880 and table 900, respectively, showing accuracy before and after using the generated weights, as described herein.

[0026] Figure 10 A chart is depicted in accordance with one or more embodiments showing the average increase in overall single-user output CV accuracy for each class after adjusting the baseline GBM.

[0027] Figure 11 A chart is depicted that depicts a comparison of receiver operating characteristic (ROC) curves for subjects #7 and #8 for the "biking" and "walking" classes to illustrate aspects of one or more embodiments.

[0028] Figure 12 and Figure 13 depict charts and tables, respectively, showing illustrative data from another example dataset to illustrate additional aspects of one or more embodiments.

[0029] DETAILED DESCRIPTION

[0030] This section describes applications of methods and apparatuses in accordance with one or more embodiments. These examples are provided merely for the purpose of increasing the overall understanding of the present disclosure. As will be apparent to one of ordinary skill in the art, the techniques described herein can be practiced without some or all of these specific details. In other instances, well known process steps have not been described in detail in order to avoid unnecessarily obscuring the present disclosure. Other applications are possible, and the following examples should not be taken as limiting in scope or setting.

[0031] In the following detailed description, references are made to the accompanying drawings that form a part hereof, and in which are shown by way of illustration specific embodiments. While these embodiments are described in sufficient detail to enable those skilled in the art to practice the disclosure, it should be understood that these examples are not limiting, and that other embodiments can be utilized and that logical and mechanical changes can be made without departing from the spirit and scope of the disclosure.

[0032] One or more embodiments can be implemented in a number of ways, including as a process, an apparatus, a system, a device, a method, a computer readable medium such as a computer readable storage medium containing computer readable instructions or computer program code, or as a computer program product including a computer usable medium having computer readable program code embodied therein.

[0033] The drawings in the description below, which are included by illustration only, refer to preferred embodiments. It should be noted that alternative embodiments of the structures and methods disclosed herein will be readily apparent to those skilled in the art, and the present disclosure is intended to include all such alternatives as can be reasonably inferred from the description below.

[0034] Generally speaking, one or more embodiments can improve the accuracy of the use of human computer interaction (HCI) techniques, specifically, the selection of an activity that can be occurring in an HCI interaction by a device based on sensor data of the device. As described further below, one or more embodiments can modify the output of a trained model without retraining an estimator of the model, and in some cases described herein, the embodiments described herein can significantly improve the accuracy and F-1 score for identifying activities and activity categories for a particular type of user.

[0035] Figure 1 An example of a system 100 in accordance with one or more embodiments is shown that can calibrate activity models used by a device 105 based on weights generated by a personal weight determiner 175.

[0036] Device 105 can take any of a variety of forms, including but not limited to a cellular telephone, a personal computer, a personal digital assistant, a smart watch, and any other device having a sensor 170 capable of sensing different conditions. In this regard, it will be appreciated that while components of touch-sensitive device 105 are shown as being within a single housing, this is optional, and these components can be located in separately packaged components, such as external sensors configured to provide data to device 105, such as heart rate monitors, pacing sensors, step pace sensors, and other similar sensor components that can be external to the housing of device 105.

[0037] Device 105 can include various I / O components, including but not limited to touch- sensitive system 110, display system 120, audio system 160, and sensors 170, which in this example are coupled to signal processing unit 125 via interface unit 115. Signal processing unit 125 can receive signals, which can be in digital form, from interface unit 115 and prepare the signals for further processing. Signal processing unit 125 can perform at least one of sampling, quantization, and encoding processes to convert such analog signals to digital signals. Signal processing unit 125 can provide the digital signals to processor 130 and other system components.

[0038] In one or more embodiments, display system 120 can output images using display 122, touch- sensitive system 110 can receive touch inputs using touch-sensitive surface 112, and audio system can output audio using audio sensor 162 (e.g., a microphone and / or connected to a microphone) and audio output 164 (e.g., a speaker or connected to a speaker).

[0039] Device 105 can also have a processor 130, such as a microprocessor, microcontroller, or any other type of programmable control device, or a pre-programmed or dedicated processing or control system. Device 105 can further include a memory system 140 used by processor 130. Memory system 140 can provide programming and other forms of instructions to processor 130, and can be used for other purposes. Memory system 140 can include read-only memory, random access semiconductor memory, or other types of memory or computer-readable media, which can be permanently installed or detachably installed to device 105. In addition, device 105 can also access another memory system 140 separate from touch-sensitive device 105 through communication system 180. In one or more embodiments, a database 165 can also be provided to store programs and other data, such as generated personal weights.

[0040] The communication system 180 can take the form of any optical, radio frequency, or other electrical circuit or system that can convert data into a form suitable for transmission through an optical, radio frequency, or other form of wired or wireless signal. The communication system 180 can be used for a variety of purposes, including but not limited to transmitting and receiving instruction sets and exchanging data with remote sensors or memory systems.

[0041] According to one embodiment of the present application, at least some of the functionality of the general model component 132, the personal weight applier 134, the personal weight determiner 175, the interface unit 115, the signal processing unit 125, the database 165, and other components discussed below can be program modules that control or communicate with other well-known hardware components, or are components for executing software. In one or more embodiments, program modules can be included in the device 105 in the form of operating systems, application programs modules, or other program modules, and can be physically stored in various well-known storage devices. In addition, program modules can be stored in remote storage devices that can communicate with the touch sensitive device 105 through the communication system 180. Such program modules can also include, but are not limited to, routines, subroutines, programs, objects, components, data structures, etc., for performing particular tasks or performing particular abstract data types as described below in accordance with the present application. Such program modules can also be represented in hardware configurations suitable to perform the functions associated with such modules.

[0042] To further describe the functionality and capabilities of one or more embodiments, the general model component 132, the personal weight applier 134, and the personal weight determiner 175 are discussed below by way of example.

[0043] Figure 2 A more detailed view of the operation of the general model component 132, as well as the personal weight applier 134 and the personal weight determiner 175, in accordance with one or more embodiments is shown. Repetitive description of like elements employed in other embodiments described herein is omitted for sake of brevity.

[0044] A method that can be used to identify activities combines sensor data with models that can interpret the data. For example, when a device is in the pocket of a person who is sitting, example sensor output can include the angle of the device as measured by a gyroscope sensor, the stillness of the device as measured by an accelerometer, the lack of touch on the touch interface of the device, and other combinations of data that are known and discoverable through experimental use. Based on this example sensor data, the device can determine that the user of the device is likely currently sitting, and based on this determination provide functionality such as turning off location detection sensors, providing notifications through custom vibrations, and other actions associated with the determination.

[0045] In some implementations, the device 105 can determine occurrences of different activities by employing the general model components 132. Included among these components are individual estimators 215A-D that can utilize some or all of the data from the analysis sensors 170 and detect a particular activity or combination of activities associated with the individual estimators 215A-D. For example, estimator 215A can be configured to determine the likelihood that the device 105 is in a user's pocket, for example, by analyzing light sensor, accelerometer, and gyroscope data. Alternatively, estimator 215A can be configured to recognize a combination of activities, for example, that the device 105 is in a pocket and the user is sitting, associated with the sensor data discussed in the example introduced above. In another alternative, the two activities can be recognized by different estimators 215C-D, and the results can be grouped into an estimator group 217 that provides a single value for the combination.

[0046] In some cases, the general model components 132, which are trained with data designed to accurately measure most users, can be inaccurate for a small number of users. For example, when detecting the activity of "going from sitting to standing up," the data collected for the standard model can not accurately apply to children, disabled persons, or elderly persons, for example, the motion speed and mechanics of most people can change significantly based on the youth, disability, or age of the person. In certain cases, another example activity that the standard model can not accurately assess is the "running" activity. Different users have different concepts of running, and in certain cases, for example, due to the speed and intensity of the motion, the running activity of an elderly person is assessed as walking.

[0047] One reason that the inaccuracies described above can occur is that the models used to analyze the sensor 170 data to determine a likely activity are not tailored to the specific circumstances of a particular user. To improve the accuracy of the device's determination of a likely activity, one or more embodiments can receive an indication from the standard model regarding a particular activity, for example, the likelihood that the device user is currently walking, and based on a personalized assessment of the device user, as detailed below, a weight can be applied to the value, for example, making the activity more likely, less likely, or having the same likelihood. The activity trigger component 230 of the general model components 132 can then assess the changed value to determine whether the modified likelihood is sufficient to trigger the activity output 250. In an example, the activity output 250 can result in performing an action associated with walking, for example, footstep detection, turning on a location determination sensor, and other activities associated with walking.

[0048] In one or more embodiments, the weighting of the output from the estimators 215A-D can also be referred to as tuning, calibration, adjustment, boosting, and other similar terms. As noted above, the estimators can generate output (e.g., the likelihood of activity occurring), and as described herein, this output can also be referred to as an estimator parameter. As used herein, the weights can be referred to as individual weights, individual weight, estimator weight, and other similar terms. The terms described in this paragraph are merely examples of equivalent terms, and other terms used herein can have equivalent or similar meanings without specific mention.

[0049] It should also be noted that, as used in the various example embodiments described herein, the non-limiting example model used by the estimators can be a gradient boosting machine (GBM), such as a machine learning (ML) method. Given the description herein, one of skill in the relevant art(s) (or arts) will understand the methods behind the training of standard estimators (e.g., GBM ML models). As discussed further herein, in one or more embodiments, the data collected using the sensors 170 can be used to determine the weights (W JP ) applied to change the results of the GBM. Although GBM models are discussed herein, one of skill in the relevant art(s) (or arts) will understand, upon being given the description herein, that other models can also be calibrated based on one or more embodiments.

[0050] In one or more embodiments, to address some of the situations mentioned above, the individual weight determiner 175 can receive sensor 170 data and select weights 225A-C to modify the output values of the general model components 132, including the estimators 215A-215D. In this approach, one or more embodiments can use a method based on transfer learning, where a standard model has been trained on a large dataset and provided on the device 105, and once the user owns the device, the standard results can be changed based on a smaller, personalized dataset. To generate this dataset, one or more embodiments can perform one or more of the following operations: collect data from routine, normal use (e.g., walking frequently), or specifically prompt the user to perform specific activities at specific times, such as sitting, running, driving, and other activities.

[0051] In one or more embodiments, once one or more estimators 215A-D generate likelihoods of occurrences of different activities from the general model component 132, the activity trigger component 230 can evaluate the one or more likelihoods of activities identified by the estimators 215A-D and determine whether to trigger the occurrence of an event associated with one or more activities, e.g., activity outputs 250. In other words, the activity trigger component 230 can evaluate multiple estimators 215A-D by using an ensemble algorithm like random forest. In this algorithm, an average of the relevant estimator outputs is determined, e.g., the model in the ensemble. Once the outputs are aggregated, the activity trigger component 230 can make a determination of the activity that is triggered. In this context, the weighting of the estimators 215A-D output by one or more embodiments can be referred to as an enhanced ensemble approach.

[0052] Returning to this example, for the fast moving person of the example, both the walking estimator 215A and the running estimator 215B can generate a likelihood that the respective activity is occurring. In a simple determination, the activity trigger component 230 can select the highest likelihood and compare that value to a threshold to determine the walking or running activity. In other approaches, a combination of other sensors 170 can also provide relevant data, e.g., an accelerometer can determine the intensity at which the individual is moving.

[0053] In one or more embodiments, the personal weight applier 134 can apply weights to the output of the personal estimators prior to these estimates being evaluated by the activity trigger component 230. Thus, in the example where the model determines a 25% likelihood of running and a 70% likelihood of walking, for a person (e.g., a child or a disabled person) that is determined (through analysis of sensor data by the personal weight determiner 175) to be experiencing false negatives of running (e.g., being assigned a 25% value by mistake), the personal weight determiner 175 can apply a weight 225B that indicates that the running estimator 215B can have a false low value, and the personal weight applier 134 can apply the weight 225B and raise the determined likelihood of running from 25% to 75%, thereby beneficially adjusting the application of the general model component 132.

[0054] It is important to note that in using one or more embodiments of this approach, the estimators 215A-D are not modified, which is beneficial because in some cases, the estimators 215A-D cannot be changed on the device 105. With this approach, in some cases, one or more embodiments can improve system accuracy for a particular user of the device 105 without having to change the installed model. Among the additional benefits of not modifying the estimators 215A-D, retraining aspects of the standard model in the device 105 can require a large amount of computational resources and time, e.g., more than what is available to the device 105 (which can be a smart watch).

[0055] In yet another benefit of one or more methods described herein, in some cases, retraining of the standard device model can not be feasible because only a limited dataset is available for retraining. For example, a GBM can be trained based on data available from many users for "running" and "walking" activities. However, this dataset can not be representative of every human behavior in real life. This can result in reduced accuracy of activity recognition.

[0056] Turning now to additional details regarding the sensors 170, these components can include, but are not limited to:

[0057] • Piezoelectric bending elements

[0058] • Piezoelectric films

[0059] • Accelerometers (e.g., Linear Variable Differential Transformer (LVDT), potentiometer, variable reluctance, piezoelectric, piezoresistive, capacitive, servo (force balanced), MEMS)

[0060] • Displacement sensors

[0061] • Velocity sensors

[0062] • Vibration sensors

[0063] • Gyroscopes

[0064] • Proximity sensors

[0065] • Electric microphones

[0066] • Hydrophones

[0067] • Capacitor microphones

[0068] • Electret condenser microphones

[0069] • Dynamic microphones

[0070] • Ribbon microphones

[0071] • Carbon granule microphones

[0072] • Piezoelectric microphone

[0073] • Optical microphone

[0074] • Laser microphone

[0075] • Liquid microphone

[0076] • MEMS microphone

[0077] Analysis of data from the sensor 170 can be performed by different system components using various functions, including, but not limited to, the personal weight determiner 175, including, but not limited to:

[0078] • Mean

[0079] • Standard deviation

[0080] • Standard deviation (normalized by overall amplitude)

[0081] • Variance

[0082] • Skewness

[0083] • Kurtosis

[0084] • Sum

[0085] • Absolute sum

[0086] • Root mean square (RMS)

[0087] • Crest factor

[0088] • Dispersion

[0089] • Entropy

[0090] • Power sum

[0091] • Centroid (center of mass)

[0092] • Coefficient of variation

[0093] • Zero-crossings

[0094] The personal weight determiner 175 can also use other methods to determine weights, including, but not limited to, basic heuristics, decision trees, support vector machines, random forests, naive Bayes, elastic matching, dynamic time warping, template matching, K-means clustering, K- nearest neighbor algorithm, neural networks, multilayer perceptron, multinomial logistic regression, Gaussian mixture models, and AdaBoost.

[0095] Figure 3An example formula 300 is depicted in accordance with one or more embodiments, which can describe modification of GBM model parameters. Repetitive description of like elements employed in other embodiments described herein is omitted for purposes of brevity.

[0096] Figure 4 An example formula 400 is depicted in accordance with one or more embodiments, which can describe selection of weights 225A to be applied to estimator 215A using a loss function. Repetitive description of like elements employed in other embodiments described herein is omitted for purposes of brevity.

[0097] Figure 5 A flowchart 500 is depicted in accordance with one or more embodiments for a single-user output cross-validation (CV) process for adjusting GBM model weights and model evaluation. Repetitive description of like elements employed in other embodiments described herein is omitted for purposes of brevity.

[0098] In this section, two publicly available datasets are used to discuss flowchart 500. The first dataset is referred to as the "Daily and Sports Activity Dataset" and the second dataset is referred to as the "PAMAP2 Dataset: Physical Activity Monitoring." In one or more embodiments, these datasets can be used to compare single-user output cross-validation (CV) and F-1 scores for baseline GBM and adjusted GBM. Flowchart 500 shows a flowchart for a single-user output CV process for adjusting GBM weights and model evaluation, at block 520, a baseline GBM single-user output CV is computed by training a GBM using (N-1) training user data. At block 520, to adjust GBM weights, data for the Nth user is split into set A and set B. First, at block 530, GBM weights are adjusted on set A, then, at block 540, the adjusted GBM is used to make predictions on set B at block 550, and vice versa, with blocks 535 and 545 making predictions. Using this approach, one or more embodiments can compute a single-user output CV using an adjusted GBM, where a portion of the adjustment data is used as a validation set to select a final model based on validation set accuracy.

[0099] Figure 6 A table 600 is included that provides example features that can be used for activity classification in accordance with one or more embodiments. To illustrate different concepts, the Daily and Sports Activity Dataset is discussed below in connection with Figure 6 and Figure 7 other embodiments described herein is omitted for purposes of brevity.

[0100] The example daily and athletic activity dataset has 19 different activities performed by 8 different subjects. The data was collected using accelerometers, gyroscopes, and magnetometers attached to different parts of the subjects' bodies.

[0101] Figure 7 A comparison 700 between the baseline of the GBM model with single-user output CV accuracy and other ML models is depicted in accordance with one or more embodiments. Repetitive description of like elements employed in other embodiments described herein is omitted for brevity.

[0102] To illustrate aspects of the different embodiments discussed herein, data from an accelerometer sensor 170 attached to an arm / wrist is shown in graph 710 for four different activities from the dataset: running, cycling, resting, and walking. In this example, the data is collected at a sampling frequency of 25 Hz with a one second time lag, e.g., a total of 25 samples per second, for generating one instance of the features in graph 710. Figure 6 Different features 600 calculated using these samples are shown. These features are calculated using accelerometer data around the X, Y, and Z axes.

[0103] First, in Figure 7 For comparison, a baseline GBM model with single-user output CV accuracy is depicted in comparison to other types of ML models. It is noted that for the "resting" and "running" categories, each model has high accuracy. However, in this example, the GBM outperforms the other models for the "cycling" and "walking" categories. In one or more embodiments, these can be handled differently because the ways in which users walk and cycle are more varied than the ways in which users walk and run. In some cases, one or more embodiments can generate weights based on these types of factors, resulting in an improvement in accuracy. This is also because the ways in which different users walk and cycle can be different from other users.

[0104] Continuing with this example, Figures 8A-8B and Figure 9 Graphs 810-880 and table 900 are depicted, respectively, showing accuracy before and after using the generated weights, as described herein. Repetitive description of like elements employed in other embodiments described herein is omitted for brevity.

[0105] As Figures 8A-8B depicted in the example data for the eight subjects, embodiments of adjusting the algorithm can be used for the daily and athletic activity dataset to improve the accuracy of single-user output CV. Figures 8A-8B and Figure 9Table 910 in Figure 9 depicts the average improvement in accuracy for each subject before and after applying one or more embodiments described herein. For example, it can be seen that the accuracy for each subject improved, with the accuracy for the "biking" and "walking" categories for subject #7 significantly improving from 65% and 73% to 88% and 94%, respectively. Additionally, for subject #8, the accuracy for the "biking" activity improved from 83% to 90%.

[0106] Figure 10 Depicted is a graph 1000 showing the average improvement in overall single-user output CV accuracy for each category after tuning the baseline GBM according to one or more embodiments. For the sake of brevity, repeated descriptions of similar elements employed in other embodiments described herein are omitted.

[0107] As shown, the example baseline accuracies for the "cycling," "resting," "running," and "walking" categories are 90%, 98%, 96%, and 94%, respectively. It should be noted that after applying one or more of the methods described herein, these accuracy values ​​improve to 96%, 99%, 99%, and 97%, respectively. Thus, in this example, according to one or more embodiments, in some cases, an error reduction of over 50% can be achieved by tuning the GBM on a particular user's data.

[0108] Figure 11 A graph 1110 is depicted that depicts a comparison of receiver operating characteristic (ROC) curves for subjects #7 and #8 for the "cycling" and "walking" categories to illustrate aspects of one or more embodiments. For the sake of brevity, repeated descriptions of similar elements employed in other embodiments described herein are omitted.

[0109] As shown in graph 1110, for subject #7 "cycling" category, after treatment according to one or more embodiments, the area under the curve (AUC) value significantly improved from 0.886 to 0.982 after adjustment. It should also be noted that for this data, according to one or more embodiments, Figure 9 Table 900 shows a comparison of the F-1 scores of the baseline GBM and the adjusted GBM, where the overall F-1 score improves from 0.9456 to 0.9758.

[0110] To illustrate additional aspects of one or more embodiments, Figure 12 and Figure 13 Depicted are a chart 1200 and a table 1300, respectively, showing illustrative data from another example data set. For the sake of brevity, repeated descriptions of similar elements employed in other embodiments described herein are omitted.

[0111] This second example dataset, named PAMAP2, is populated with accelerometer data for three different activities (e.g., cycling, resting, and walking) performed by nine different subjects. In this example, the dataset is collected at 100 Hz with a one second lag, e.g., showing 100 samples of data.

[0112] For this second example, a process similar to the process shown in flowchart 500 of FIG. 4 is used to train the baseline GBM to find single-user output CV accuracy. Figure 5 Figure 12 Graphs 1210-1280 of FIG. 12 show the average improvement in single-user output CV accuracy for each class of subject and the overall improvement in CV accuracy based on adjusting the GBM according to one or more embodiments described herein.

[0113] As another example, Figure 13 Table 1300 is depicted with a comparison of F-1 scores for the baseline GBM and the GBM adjusted according to one or more embodiments. It should be noted that an overall F-1 score improvement from 0.9307 to 0.9619 is shown, as well as a significant improvement in subject F-1 score for subject #7, e.g., from 0.8117 to 0.9671.

[0114] In additional illustrations of features of one or more embodiments, Figure 12 It is further indicated in graph 1270 that the "walking" accuracy for subject #7 improved from 55% to 95%. Graph 1200 also shows the overall single-user output CV improvement for each class, with baseline accuracies of 92%, 96%, and 88% for the "cycling," "resting," and "walking" classes, respectively. Additional benefits of one or more embodiments are illustrated by the improvement in accuracy for these classes from baseline values of 94%, 97%, and 95%, respectively.

[0115] ​The one or more embodiments described above can be implemented in the form of program instructions, which can be executed by various computer components, and can be stored on a computer-readable recording medium. The computer-readable recording medium can include program instructions, data files, data structures, etc., individually or in combination. The program instructions stored on the computer-readable recording medium can be specifically designed and configured for one or more embodiments, or can be known and available to those skilled in the computer software field. Examples of the computer-readable recording medium include the following: magnetic media, such as a hard disk, a floppy disk, and a magnetic tape; optical media, such as a compact disc read only memory (CD-ROM) and a digital versatile disc (DVD); magneto-optical media, such as a floptical disk; and hardware devices, such as a read only memory (ROM), a random access memory (RAM), and a flash memory, which are specially configured to store and execute program instructions. Examples of the program instructions include not only machine language codes created by a compiler, etc., but also high-level language codes that can be executed by a computer using an interpreter, etc. The above hardware devices can be changed to one or more software modules to perform the operations of one or more embodiments, and vice versa.

[0116] Although one or more embodiments have been described above in connection with specific limitations (e.g., detailed components and limited embodiments and drawings), these are provided only to help the general understanding of the present invention. One or more embodiments described herein are not limited to the above-described embodiments, and those skilled in the art will understand that various changes and modifications are possible in accordance with the above description.

[0117] Therefore, the spirit of one or more embodiments should not be limited to the above-described embodiments, and the entire scope of the appended claims and their equivalents will fall within the scope and spirit of the present invention.

Claims

1. A method for calibrating an activity model, the method comprising: receiving sensor data from a sensor of the electronic device, wherein the sensor is operable to detect a plurality of different activities; collecting training data for the electronic device associated with the plurality of different activities from the sensor; generating a weight for each of the plurality of different activities based on the training data for the electronic device; For the electronic device, inputting the sensor data into a general model of the electronic device, the general model outputting a likelihood result for each of the plurality of different activities, wherein the general model is trained based on sensor data from a plurality of different electronic devices detecting the plurality of different activities; applying, for the electronic device, each weight to each of its corresponding activities of the plurality of different activities to adjust its corresponding likelihood result from the trained model without modifying the general model and without retraining the general model; determining that the first activity among the plurality of different activities be performed by the electronic device by selecting a highest value among the adjusted likelihood results; and An action is performed on the electronic device based on the determination of the first activity.

2. The method according to claim 1, wherein The first activity is an activity of a user of the electronic device.

3. The method according to claim 2, further comprising: facilitating an assessment of a physical characteristic of the user of the electronic device; and A first weight is selected based on the evaluation of the physical characteristic of the user.

4. The method according to claim 3, wherein: Selecting the first weight includes selecting the first weight to increase the accuracy of the first likelihood outcome of the first activity for the user of the electronic device.

5. The method according to claim 3, wherein Selecting the first weight is also based on training data for the first activity.

6. The method according to claim 4, wherein: Selecting the first weight is also based on training data for the first activity.

7. The method according to claim 4, wherein: The determination of the first activity includes comparing the first likely outcome of the first activity to a second likely outcome of a second activity.

8. The method according to any one of claims 1 to 6, wherein Receiving the sensor data includes receiving data from at least one of an accelerometer, a magnetometer, or a gyroscope.

9. An electronic device comprising: sensor; processor; and a memory storing executable instructions that, when executed by the processor, facilitate performance of operations comprising: receiving sensor data from the sensor, wherein the sensor is operable to detect a plurality of different activities; collecting training data for the electronic device associated with the plurality of different activities from the sensor; generating a weight for each of the plurality of different activities based on the training data for the electronic device; For the electronic device, inputting the sensor data into a general model of the electronic device, the general model outputting a likelihood result for each of the plurality of different activities, wherein the general model is trained based on sensor data from a plurality of different electronic devices detecting the plurality of different activities; applying, for the electronic device, each weight to each of its corresponding activities of the plurality of different activities to adjust its corresponding likelihood result from the trained model without modifying the general model and without retraining the general model; determining that a first activity among the plurality of different activities be performed by the electronic device by selecting a highest value among the adjusted likelihood results; and An action is performed on the electronic device based on the determination of the first activity.

10. The electronic device according to claim 9, wherein The first activity is an activity of a user of the electronic device.

11. The electronic device according to claim 10, further comprising: facilitating an assessment of a physical characteristic of the user of the electronic device; and A first weight is selected based on the evaluation of the physical characteristic of the user.

12. The electronic device according to claim 11, wherein Selecting the first weight includes selecting the first weight to increase the accuracy of the first likelihood outcome of the first activity for the user of the electronic device.

13. The electronic device according to claim 11, wherein Selecting the first weight is also based on training data for the first activity.

14. The electronic device according to claim 12, wherein: Selecting the first weight is also based on training data for the first activity.

15. The electronic device according to claim 12, wherein The determining of the first activity includes comparing the first likely outcome of the first activity with a second likely outcome of a second activity.

16. The electronic device according to any one of claims 9 to 14, wherein: The sensor includes one or more of the following: accelerometer, magnetometer, or Gyroscope.

17. A computer-readable recording medium having program instructions, the program instructions being executable by various computer components to perform operations, the operations comprising: receiving sensor data from a sensor of the electronic device, wherein the sensor is operable to detect a plurality of different activities; collecting training data for the electronic device associated with the plurality of different activities from the sensor; generating a weight for each of the plurality of different activities based on the training data for the electronic device; For the electronic device, inputting the sensor data into a general model of the electronic device, the general model outputting a likelihood result for each of the plurality of different activities, wherein the general model is trained based on sensor data from a plurality of different electronic devices detecting the plurality of different activities; applying, for the electronic device, each weight to each of its corresponding activities of the plurality of different activities to adjust its corresponding likelihood result from the trained model without modifying the general model and without retraining the general model; determining that the first activity among the plurality of different activities be performed by the electronic device by selecting a highest value among the adjusted likelihood results; and An action is performed on the electronic device based on the determination of the first activity.

18. The computer-readable recording medium according to claim 17, wherein The operations further include: facilitating assessment of a physical characteristic of a user of the electronic device; and A first weight is selected based on the evaluation of the physical characteristic of the user, wherein the first activity is an activity of the user.

19. The computer-readable recording medium according to claim 18, wherein Selecting the first weight includes selecting the first weight to increase the accuracy of the first likelihood outcome of the first activity for the user of the electronic device.

20. The computer-readable recording medium according to claim 18 or 19, wherein Selecting the first weight is also based on training data for the first activity.

Images