Personal productivity monitoring through smartphones

3.1. Design and development of prodmapp

The first step of our approach is to design and develop a non-intrusive smartphone app – Prodmapp. The primary purpose of Prodmapp during this study is to non-obtrusively monitor the user’s routine activities such as visit to different locations, physical activities, social apps usage and save this sensing data for further analysis. In the following sub-sections, we briefly discuss the various sensors that are used in Prodmapp.

Location We integrate Google Places API [16] in Prodmapp that uses location sensor to keep track of user’s visited locations. It returns the user’s current location name, address, longitude, latitude, likelihood, and datetime after every 15 minutes. The reason behind choosing 15 minutes delay in location detection of the user is to reduce the battery consumption while still getting enough location instances.

Activity In order to detect user’s physical activities, we integrate Google Activity Recognition API [1]. The API returns the detected activity (i.e., running, walking, in vehicle etc), their likelihood and datetime. The data is collected after every 15 minutes. The reason behind choosing 15 minutes delay in activity detection of the user is to detect the activity parallel to location of the user.

Social apps usage To detect the social apps usage of users, we use the built-in android class UsageStatsManager [39]. It returns the names of social apps that a user is using such as Facebook, Twitter, Instagram, Snapchat, Youtube, Messenger and Gmail after every 15 minutes. The reason behind choosing 15 minutes delay in detecting the social apps usage of the user is that we are interested in finding out the social apps used by the user at a particular location which itself is detecting after every 15 minutes. In this work, we limit our analysis to the widely used social apps i.e., Facebook, Twitter, Instagram, Messenger, Snapchat, Youtube and Gmail.

OPEN IN VIEWER

Fig. 2.

Graphical user interfaces of prodmapp.

Ground truth We design a questionnaire in Prodmapp to collect the ground truth from users in order to quantify their productivity. Whenever the user visits a new location in a day, Prodmapp pops up a notification on notification drawer (See Fig. 2b). When user clicks on that notification a list will appear which display the visited locations of user (See Fig. 2c). When user clicks on a particular location a dialog box will appear that asks user a question that “Whether his visit at that particular location is productive or not?” (See Fig. 2d). If user answers “Yes” than a productivity score 1 is added against that location, otherwise 0 is added. There is no need for a user to respond to question immediately and user can answer at any time within 24 hours.

We started to gather the ground truth from users after every 5 minutes but we found that it was burdensome for them and they were ignoring the pop-up notifications. We then reduce the frequency of the notifications so that we may get ground truth instead of no responses. Also, we learnt that it’s not a good idea to force the user to give responses at the same time when notification pops up and keeping in view the memorability issues, we kept the number of notifications low. Currently, Prodmapp asks users about their productivity after every 15 minutes and at the same time gathers their location, activities and social apps usage data.

Prodmapp GUI and data storage We design the GUI of Prodmapp extremely simple to keep the things easy for users during data collection process. It displays just one button on main activity (See Fig. 2a). When user press that button, Prodmapp starts running in the background as a service, and tracks user physical activities, visited locations, social apps usage, generate timely notifications to gather ground truth and store all this data in the user’s smartphone in CSV file. We choose Prodmapp to run as background service on user’s smartphones because we want to collect the sensing data non-intrusively without burdening user to launch the app every time in order to track their location, activities and social apps usage.

In this study, we have collected the data from 10 users through Prodmapp. The users were the hostelite students of three different universities of Islamabad Pakistan that spend most of their time in university and hostel. Out of 10 users, 6 students were from FAST University, 2 from NUST University and 2 from TMUC University. The age range of the users were 19 to 26 years. The detailed profile of each user is shown in the Table 2.

The reason to choose student population for this experiment was that they were easily accessible to us and willing to participate in this experiment voluntarily. We could easily access the participated students daily to ensure that our app is up and running on their smartphones and gathering the sensing data. Moreover, to be productive is the major challenge for every student.

At the start of the data collection, we requested 18 users to fill an online Google Form that contained few general questions about them such as name, age, email, cell number, smartphone owns and workplace/university name. After analyzing the information provided by users through these forms, we have selected 13 users for the data collection, but 3 users faced some issues in their smartphones after installing Prodmapp and politely refused to be a part of data collection process. The remaining 10 users provided us complete 15 days’ data and there were no big issues in their data. The number of data samples collectively of all 10 users were 1407.

During the selection of users for data collection, only those users were selected in the experiment that owns latest model Samsung smartphones having Android Nougat or higher installed. The reason behind selecting such users was the better sensors quality and good battery timing of Samsung smartphones. Before app deployment on the smartphones of selected users, we tested our app on various Chinese manufactured smartphones too such as OPPO, VIVO and QMobile. But these phones were not able to provide the accurate location and activities data due to low sensors quality. Even some of these phones were not able to support the Activity Recognition API.

We also observed that Prodmapp was draining the battery of the smartphone of some users at very fast rate. The reason was that initially Prodmapp was sensing the user’s location through GPS after every 5 minutes. We overcome this problem to large extent by increasing the location sensing time from 5 minutes to 15 minutes. After every 15 minutes, GPS enables, sense the location and then disables. This drastically improves the battery performance of user’s smartphone.

Aruba dataset In order to evaluate the performance of our proposed approach, we have used the publicly available Aruba dataset from CASAS smart homes, a project of Washington State University [7] to quantify the productivity of housewife. The dataset comprises of 6471 instances of daily life activities data obtained using 31 motion sensors, 3 light sensors, 5 temperature sensors and 3 door sensors. 11 daily life productive and non-productive activities were performed by housewife for 220 days (7 months). Out of 11 activities, the productive activities includes Housekeeping, Meal preparation, Eating, Wash Dishes, Enter Home, Work whereas non-productive activities are Sleeping, Relax, Leave home, Resperate, Bed to Toilet. Unlike our dataset, the location data was not included in this dataset.

In this step, we processed the collected raw data to remove noise and outliers. We performed data preprocessing in two steps:

In first step, we removed all those location and detected activities samples from the data files of all users where likelihood was below 70 % . After this step we were left with 1319 data samples collectively of all 10 users.

The second step of our data preprocessing phase was clustering location data to find the significant locations that a user visited during the study period. We used a well-known data clustering algorithm known as Density-based spatial clustering of applications with noise (DBSCAN) [13]. DBSCAN groups together points which in our case are locations that are close to each other based on distance measurement and a minimum number of points. DBSCAN algorithm requires two parameters; Eps and Minpts. Eps refers to the minimum distance between two points, and Minpts refers to the minimum number of points required to form a dense region. We chose the values of Eps and Minpts as 0.4 and 4 respectively.

The reason behind using DBSCAN algorithm is that we do not need to specify the number of clusters prior to clustering like other data clustering algorithms such as k-means, Gaussian mixture model (GMM) etc. In this study, clusters represent the number of locations that a user visited during a day.

We applied clustering to only location samples that were in stationary state (i.e., a user is still) instead of transition state (i.e., a user is walking in the street, or driving a car, and his location changing frequently) because we are interested to determine the locations where users usually spend most of their time like home, workplace, restaurants etc. To eliminate the transition states of a user prior to clustering, we calculate the speed between two location points, and if the speed was greater than 1 km/h, we exclude those points before clustering. Figure 3a showing the transition points of one user we map through GPS visualizer [17] on geographical map. Figure 3b demonstrating the location clusters with labels we obtained after clustering.

OPEN IN VIEWER

Fig. 3.

Identification of (a) transition points and (b) location clusters.

After data collection and preprocessing, the next step of our methodology was feature extraction. We have identified 8 GPS location and 1 social apps usage related features after analyzing the preprocessed data.

We skipped some features in this study i.e.: hour of the day and weekend/weekday difference as we have performed the experiment with students that have different routine each day. We did try to add time as 3rd dimension in dbscan, but number of clusters were meaningless.

The 8 GPS location features are number of clusters, time spent at home, time spent at workplace, unique places visit, total distance travelled, transition state, entropy, and normalized entropy. These features were inspired from [6,34] in which authors used them to learn the behavior of individuals and assess their depression state. In our work, we are using these features to infer the productivity of individuals by learning their daily routine.

In this study, Number of Clusters indicates the amount of location clusters found by DBSCAN clustering algorithm in data preprocessing phase. Time Spent at Home represents the amount of time in hours a user spent at his home during this study. To find this, we first identified the home cluster through clustering in data preprocessing stage. We used one heuristic to find the home cluster i.e., the location where user spend most of his time between 1:00AM to 7:00AM will be his home. Time Spent at Workplace shows the amount of time in hours a user spent at his workplace during this study. To find this, we first identified the workplace cluster through clustering in data preprocessing stage. We used one heuristic to find the workplace cluster i.e., the location where user spend most of his time from 10:00AM to 6:00PM on weekdays will be his workplace. Unique Places Visit indicates the number of different locations visited by user during this study.

Total Distance Travelled represents the total distance in kilometer covered by a user. Transition State shows the percentage a user was in transition/non-stationary state. We measured it by dividing the location samples of a user in walking, running and vehicle state by the total number of location samples. In our work, feature Entropy indicates the uniformity in time that a user spends in various location clusters that we determined in data preprocessing stage. We have calculated the entropy using Eq. (1) (1) Entropy [36] = − ∑ i p i log p i Where each i = 1 , 2 , … , N representing the location cluster and p i representing the percentage of time user spent at location cluster i. The higher the uniformity in time spend by a user means higher the entropy. For example: if a user spent 85 % of time at his home and 15 % at workplace, the entropy will be 0.423 whereas if a user spent 50 % of time at his home and 50 % at workplace, the entropy will be 0.693.

Normalized Entropy represents the distribution of the visited location clusters. The value of normalized entropy ranges between 0 and 1, where 0 means that all location data points lies in the same cluster, and 1 indicates that they are uniformly distributed across all the clusters. We have calculated the normalized entropy of each user by dividing the entropy by logarithm of the total number of clusters. The equation of calculating normalized entropy is shown in Eq. (2) (2) Normalized Entropy = Entropy / log N

Frequency of Social Apps Usage shows the number of times each user interacted with various social apps collectively such as Facebook, Twitter, Instagram, Youtube, Snapchat, Messenger, Skype and Gmail. We have calculated this feature by adding the number of social apps used by the user at each visited location.

Table 3 shows the smartphone sensors used in the extraction of each feature.

We determine the relationship between the extracted features and productivity score of the user. In this paper, we use Pearson correlation [30] to determine the correlation among features and user’s productivity. Pearson correlation coefficient is a well-known statistical test used to find the dependency of one variable on another. The results of correlation analysis are discussed in Section 4.

3.6. Prediction analysis

In the final step of our proposed approach, we train and evaluate a generalized prediction model to assess whether it is possible to predict the change in productivity of users with the variation of feature values. We train three classifiers i.e., logistic regression, SVM and KNN in order to compare their performance on our dataset. An overview of each of the classifier is discussed below:

1) Logistic Regression: Logistic Regression is a regression method that can be used in modelling binary dependent variables. It is a mathematical modelling approach used to define the relationship among X independent variables and Y binary dependent variable which is coded as 0 or 1 for two possible categories [11,40]. In this work, our target variable is also binary in nature i.e., 0 means Productive and 1 means Non-Productive.

2) Support Vector Machine (SVM): SVM has wide variety of applications in machine learning and human behavior prediction [10]. It is a supervised learning algorithm that transforms training examples to a higher dimensional space and builds a linear model. This model is then used to classify the new training examples.

3) K-Nearest Neighbour (KNN): It is one of the most simple, supervised machine learning algorithm. It can be used for both classification and regression problems. It finds the distance between a query and all the examples in the data, selects the specified number examples (K) closest to the query, then votes for the most frequent label [2].

4.1. Computation of productivity score of users

To determine the correlation between features and productivity, we first calculate the total productivity score of 15 days of each user that they provided through Prodmapp. For each user, we first calculate the productivity score by adding the total number of 1’s and 0’s. For each user, we then take the average of productivity score by dividing the productivity score by total number of 1’s and 0’s. Table 4 shows the productivity score, and its average of each user.

We calculate 8 GPS location features and 1 social apps usage related feature for each user. The correlation analysis among these features and the user’s average productivity score reveals that 6 out of 9 features are significantly correlated to the user’s average productivity score. The null hypothesis is that no correlation exists between average productivity score and features. We reject this hypothesis if p-value is less than 0.05. The results of correlation analysis is shown in Table 5.

Specifically, 5 out of 8 GPS location features i.e., entropy, total distance travelled, transition state, time spent at home and time spent at workplace shows strong correlation with user’s average productivity score with Pearson correlation coefficients r of 0.642, 0.647, − 0.806 , − 0.635 , and 0.656 respectively. We find no correlation of number of clusters, normalized entropy and unique places visit with user’s average productivity score because their p-values are greater than 0.05.

Moreover, 1 feature related to social apps usage i.e., Frequency of social apps usage showed strong negative correlation with user’s average productivity score with r = − 0.662 and p-value = 0.037 .

Prodmapp dataset We only use those features to train Logistic Regression, SVM and KNN that are significantly correlated with average productivity score of users i.e., time spent at home, time spent at workplace, total distance travelled, transition state and frequency of social apps usage. We compute the feature values for each day and for each user. For each user and for each day we calculate the target variable/label as follows: We assign label 1 if productivity score of a user of that particular day is greater than the average productivity score of that day plus one standard deviation, otherwise 0.

We split our dataset into 70 % training and 30 % test. We perform grid search to find the optimal value of cost parameter C of logistic regression classifier. The range of C we find through grid search is 0.001 to 1000. However, we test the performance of our classifier on C values 0.0001 to 100000. For smaller C values i.e., 0.0001 to 1 the accuracy of our classifier on test set decreases. For C values 1 to 100000 the accuracy roughly remains the same. Therefore, we choose the value of parameter C = 1 as its optimal value.

In case of SVM, we use the values of kernel and gamma as linear and scale respectively. For KNN, we use the value of k as 5.

We get the overall accuracy of logistic regression, SVM and KNN as approximately 77 % , 84 % and 80 % respectively. We evaluate the performance of all three classifiers through leave one out cross validation approach. The reason behind choosing this approach for evaluating our classifiers performance is that it provides less randomness and less bias. We get 77.9 % , 82.2 % and 81.7 % average accuracy of logistic regression, SVM and KNN respectively with leave one out cross validation approach which are very close to actual accuracy of classifiers. Therefore, we can say that our classifiers generalizes well.

Table 6 shows the sensitivity, specificity and F-score of all three classifiers.

Table 7 lists the confusion matrix of Logistic Regression, SVM and KNN where 1 and 0 represent the two classes, “Productive” and “Non-Productive” respectively.

Aruba dataset We train the same three classifiers i.e., logistic regression, SVM and KNN on Aruba dataset to evaluate the performance of our proposed approach. The dataset includes the daily life productive and non-productive activities executed in a day by housewife with regular frequency. We train these classifiers with the same parameter values that we use in Prodmapp dataset.

We get the overall accuracy of logistic regression, SVM and KNN as approximately 89 % , 89 % and 90 % respectively. We evaluate the performance of all three classifiers through 10-fold cross validation. We get 88.2 % , 82.6 % and 91.6 % average accuracy of logistic regression, SVM and KNN respectively with 10-fold cross validation which are very close to actual accuracy of classifiers. Therefore we can say that our proposed approach infers the productivity of housewife accurate.

Table 8 shows the sensitivity, specificity and F-score of all three classifiers on Aruba Dataset.

Moreover, Table 9 lists the confusion matrix of Logistic Regression, SVM and KNN where 1 and 0 represent the two classes, “Productive” and “Non-Productive” respectively.

The results show that the proposed approach is able to monitor and enhance individual’s productivity non-intrusively through learning their daily routine. We find the significant correlation between several behavioral features and user’s average productivity score. Features Time Spent at Home and Time Spent at workplace shows negative and positive correlation respectively which means that the productivity of the user will increase if the user spent more time at workplace instead of home. Similarly, feature Entropy shows positive correlation with user’s average productivity score so we can say that the productivity of the users would improve if they spent uniform time at various visited locations. Productivity of the individuals might decrease if they move, and use social apps very frequently because there is a strong negative correlation among features Transition State and Frequency of Social Apps Usage with user’s average productivity score. We evaluate the performance of our proposed approach on Aruba Dataset from CASAS project to quantify the productivity of housewife. Results of the experiment shows that our approach generalizes well as the classifiers achieve good prediction accuracy. With all these strengths of our study, several limitations are also associated with this work that needs to be overcome.

The data is collected from 10 users and all of them were the hostelite male students of three different universities of Islamabad, Pakistan. No female was involved in this study. Therefore, we don’t claim that the results of this study will generalize to other user groups.