Understanding the predictability of user demographics from cyber-physical-social behaviours in indoor retail spaces

Much research on demographic prediction focuses on cyber behaviours investigate through Web pages browsed [2‐5], queries issued [2, 6], mobile Apps installed, or Web contributed or commented upon (e.g., images, tweets, comments, likes) [7‐9].

Murray and Durell [2] studied the problem of inferring demographics based on users Internet usage, including queries submitted and Web pages accessed. Although they experienced poor results, the researchers concluded that it is possible to infer anonymous Internet users’ demographics. Jones et al. [6] found that a sequence of queries can be mapped to demographics with a simple classifier. Bi et al [8] tried to enhance the predictability from query logs in terms of inferring user demographics. They found the accuracy of identifying ‘age’ and ‘gender’ are high. Hu et al. [3] inferred users’ age and gender from browsing behavior. They built a user-Web page bipartite graph by using the logs of a particular Web site, and applied smoothing techniques to overcome data sparseness, and found the proposed model achieves good performance in age and gender prediction. Kamvar and Baluja [10] studied users’ searching behaviour and the general category of Web content they click on Google’s mobile search interface. Kumar and Tomkins [11] studies users’ searching behaviour on Yahoo search engines and their browsing behaviour via a toolbar in Web browser. Goel et al. [4] conducted a large scale study of the correlation between users demographics and a Web content categories. The researchers found that browsing histories are a strong signal for inferring demographics. Li et al. [5] predicted user demographics from both unencrypted and encrypted Web traffic. They found that reasonable accuracy could be obtained predicting gender and education level.

Considering social media, Kosinski et al. [7] showed that Facebook likes can be used to accurately predict a wide range of personal attributes, including sexual orientation, ethnicity, religious and political views, personality traits, intelligence, happiness, parental separation, age, and gender. Filippova [12] showed that the comments users left on YouTube can be used to predict their gender.

You et al. [13] found it possible to predict gender from the images posted in online social networks. Culotta et al. [9] proposed a distant supervision method to inferring twitter use demographics by using audience measurement data from thousands of Websites. Together with textual content and the social network relationships, they produced good correlation across various demographic attributes. Finally, Seneviratne et al. [14] discovered that the mobile Apps installed on a person’s smartphone can be used to predict their gender.

Predicting user demographic has been tried based on where users are [15], how they move [16], and what they do [17‐20]. Zhong et al. [16] explored the problem of predicting user demographics, based on the data set of the third task of the Nokia Mobile Data Challence (MDC) [21], which includes users call logs, app logs, activity logs, media logs, and environment contexts. They proposed a contextual feature construction framework to extract a set of features to describe users, including temporal and movement behaviors, and found high accuracy can be obtained when using multi-task learning methods. Mohammady and Culotta [15] studied how to use county-level statistics to infer the demographics of Twitter users, and found that classifiers trained on county-level information can predict user-level attributes accurately. Wang et al. [17, 22] attempted to predict user demographics from a shopping purchase history. They proposed a multi-task representation neural model to learn a shared semantic representation across multiple prediction tasks. They found this multi-tasking learning method improved demographic prediction accuracy. Stach and Buhner [18] studied how to recognize the gender information based on users’ automotive driving data, and found that acceleration, gas pedal actuation and situation dependent driving speed are closely related to driver’s gender. James et al. [19] studied the differences in risk taking between male and female, and found that the differences on the majority of test tasks (e.g. smoking) are significantly different. Harris and Jenkins [20] further studied why women take fewer risks than man. Coluccia and Louse [23] studied the gender differences in spatial orientation, and men show better orientation performance than women.

There is some research utilizing an online social network to predict demographics, but there is little focusing on the social behavior in a physical grouping context, the focus of this paper. Mislove et al. [24] studied how the social relationship in social networks can be used to predict users’ college, year, and major finding a small fraction of training data with social network graph can lead to high accuracy. Dong et al. [25] studied how people’s social strategies can be discovered from their daily mobile communication, and then be further used to predict their demographics. They also found these social information can lead to high prediction accuracy in users’ age and gender. Culotta et al. [26] predicted Twitter users’ demographics based on whom they follow via a distantly labeled dataset by collecting audience measurement data for a large number of websites. They found this lead to competitive performance with a fully supervised method. Volkova and Bachrach [27] studied the problem from the aspect of neighbors in a social network. They specifically contrasted the emotions of a user’s neighbors in a social network, and found this contrast information was correlated with certain user demographic traits. Moreover, investigating the behaviour of individuals and their relationship in the physical space is also active in sociometric sensors, e.g. the SocioPatterns collaboration.¹ More recently, Sapiezynski [28] proposed a model to infer person-to-person proximity by using WiFi signals. Sekara and Lehmann [29] studied how to deploy the electronic datasets as a valid proxy for real life social interactions, and found that the strength of electronic signals can be used to distinguish between strong and weak friendship ties.

Some other information has been used to predict users’ demographics: predicting authors’ age and gender from their writing styles [30]; gender classification from videos analysis [31]; and gender/age estimation from images of faces [32, 33]. Due to the scope of this study, we will not detail existing research in this direction.

In this paper, we present a methodology to derive demographic information about visitors to a constrained, indoor environment based on large-scale, indoor WiFi sensing data. Such data are available to indoor space operators and cover a large cross-section of the visitors. Previous studies [34‐36] focused on more coarse spatio-temporal demographic sampling, usually covering extensive parts of cities and linking demographic information to social network check-ins and also cell-tower data, and given the higher sparsity of the data, the analysis are done on aggregate behaviours rather than modelling individual trajectories for prediction. The analysis of social media users is a distinct problem, with a highly biased population (users of twitter, Foursquare etc. are not evenly drawn from the general population). The predictions based on such data are usually coarse and the portability of these methods outside of countries with high Foursquare uptake is questionable. Recently, Goel et al. [4] also studied how to infer user demographic from users’ Web logs by treating it as binary classification problem, e.g. for users’ age, they got an accuracy of 0.8 with two coarse classes: <25 and >25 years old. In our study, we attempt fine grained classification, such as four classes for users’ age (18-24 years, 25-39 years, 40-54 years and >55 years). Finally, we contribute a perspective on how the users’ cyber, physical, and social behaviors combine to improve demographic prediction, a special focus of this paper.

The online questionnaire was targeted at the registered customers of an inner city shopping mall in Sydney, Australia. The owners of the mall conducted the questionnaire online in November 2014. It was administered in English via email. The questionnaire was sent after the customers visited the mall. Overall, more than 10,000 customers were sent the questionnaire and offered a chance to enter a prize draw if they completed the questionnaire.² A total of 547 (5% completion rate) questionnaires were returned.

The questionnaire contained four sections:

Demographic Attributes collecting customers’ age, education, income, parental status, and user type, see Table 1 for more detail.

Web Attributes enquiring about customers’ use of the Web via the mall’s WiFi system, including estimates of time spent online (further, online duration), frequency of use, frequency of searching on the Web, and the categories of Web sites visited (based on definitions by Brightcloud).³ See Table 2 for more detail.

Physical Attributes enquiring about customers’ physical activities in the mall, including visiting frequency and duration, common visiting days of week, and the most favored shop category (as defined by the operator of the mall). Note, these categories are different to those defined by Brightcloud. See Table 3 for more detail.

Social Attributes collecting information about the social grouping status of the visitors of the mall: coming alone, with child/children, with an adult, or in a group. See Table 4 for more detail.

The users tracked in the mall have opted-in to use the free public WiFi network provided by the mall operator. Users need to log in with their email address to instantiate a session and access the Internet. The sampling is done every 5 minutes on each Access Point (AP), and only the RSSI and history of online browsing and searching behaviours of devices connected to the AP are logged. The user’s device is only tracked when associated with an AP, and can only be associated with a single AP at a time. The precision of the indoor localisation has been analysed in our previous study [37]. Namely, two types of logs have been captured and studied: a WiFi Association Log (AL) capturing users’ physical movement, and a Web Browsing Log (BL) capturing users’ cyber activities.

We investigated the differences in the physical and online access duration across different demographics groups by age, education, income, parental status, and user type. Specifically, the physical duration is the duration of the visit in the mall (Table 3), and the online duration means the percentage of this visit spent online (Table 2).

The distribution of duration for each demographic group was found to generally follow a normal distribution. Based on this, we used an ANOVA test to examine whether different demographic groups present significant differences in physical/online mean duration (see Tables 2 and 3). All demographic groups show significantly different (specific) distributions, except for online duration by parental status and user type, see Table 5. We investigated the patterns in these behaviors and found that: (1) investigating the users by age, people under 55 tend to stay longer both in the mall and spend more time on the Web; (2) when stratifying by education, people with an honors degree tend to stay comparatively shorter both in the mall and on the Web; (3) by income, people earning over 180k tend to stay longer both in the mall and on the Web; (4) by parental status, people having kids tend to stay longer in the mall, but not on the Web; (5) by user type, Central Business District (CBD) workers tend to stay longer in the mall, but not on the Web. Overall, we observe the differences in physical visiting duration are likely larger than their counterparts in the online duration.

Next, we examined the correlation of demographic attributes with browsing/querying categories, physical shopping categories, and social grouping status.

First, we investigate the reported popularity of each content category (Web and Physical location categories), as well as social grouping favored by the visitors (Figure 1). Note, users may choose multiple categories for the corresponding questions in the questionnaire. It appears that: (1) mall visitors tend to browse for generic information (e.g. social network and communication) in the mall, but their search behavior is more targeted as they search for specific content (e.g. shopping, food & drink); (2) the popularity of physical and Web browsing categories is skewed towards distinctly different content; (3) people’s physical shopping behavior and Web browsing behavior dominate their querying behavior. (4) The majority (over 80%) of visitors come to the mall alone or with another adult, and less than 20% come with children or in a group.

Next, we consider the relationship between different demographic groups and physical/Web/social categories. We conduct a \(\chi^{2}\)-test to test for significant differences in category popularity across each group.⁴ The detailed \(\chi^{2}\) values and p-value of the significant associations are shown in Table 6. The corresponding distributions are shown in Figure 2, where the horizontal axis indicates a user group and the vertical axis shows the probability of how many users in this group answer ‘yes’ to a particular question. For example, in Figure 2(a), the blue dashed line with * markers shows the probability of users that answered ‘yes’ to ‘Social Network (Browsing)’ question across 4 different user age groups. Overall, people’s demographics are significantly associated with their CPS behaviors. There are 10 content categories significantly associated with Age, 6 with Education, 4 with Income, 7 with Parental Status and 13 with User Type. This might indicate that Age and User Type are relatively easy to predict from their CPS behaviors, but Education, Income and Parental Status may be less strongly associated with content categories and hence be more difficult to predict. We will experimentaly investigate this in Section 6.

For measuring the correspondence of logs and questionnaire responses, we extracted the corresponding attributes from both the AL and BL.

The following physical attributes have been extracted from AL:

The following Web attributes have been extracted from BL:

In contrast to the self-reported questionnaire data, two attributes can not be extracted from the logs reliably: online duration and the noted social grouping status. For online duration, the reason is that the time parameter reported in BL only contains the date, not the hours, minutes and seconds. This makes it impossible to compute how long the user is active on the Web. Similarly, the social grouping status, cannot be studied as there is no guarantee that the accompanying survey participant will also use the opt-in WiFi system. The random sample of visitors does not allow for a reliable analysis of their social behavior.

Here, we analyze how well the logs reflecting users’ self-declared behaviors from the questionnaire responses, by measuring how close the log-based attributes are to the corresponding questionnaire responses. As we observed two kinds of attribute values, numeric and categorical, we use different metrics to evaluate the association between them.

The compared self-declared questionnaire attributes (\(v_{s}\)), the corresponding log-based attributes (\(v_{l}\)), and the comparison results are shown in Table 7. The corresponding MAE and MSD results are small for all compared attributes, which indicates that the log-based attribute values are close to the values of the self-declared questionnaire attributes: Overall, log-based attribute values reasonably reflect the self-reported values. Logs also suffer less from biases of underestimation/overestimation of activities where capturing the exact frequency is hard if reporting occurs with a delay.

Here, we formulate the predictability of user demographics as a classification problem. Consider the parental status as an example. There are two classes, having kids and no kids, and the responses/attributes describing users CPS behaviors are the features for each user. We apply Support Vector Machine (SVM) as the classification model, tested with 5-fold cross validation, and prediction accuracy as the measurement metric - the proportion of the correctly classified participants over all tested participants.

We examine the following combinations of CPS behavior attributes available from both the questionnaire responses and WiFi logs: Note, for both questionnaire-based and logs-based cyber/physical behaviors, we do not know the Web content accessed, only the coarse category of the content. In other words, we know the category of the Web site, not the exact web site/page which may be distinct. Similarly, we know the category of the shops the users are interested in, not the names of the shops or the products they sell. This is different from other reviewed research (Section 2), in particular because this study focuses on users of a large heterogeneous mall, not of a single e-commerce site or a specific store.

We use two baselines for comparison of the accuracy of our predictions: mostPop always predicting an attribute of a user based on the demographic group for which this content is most popular, and a random model, predicting user class randomly.

Table 8 shows the accuracy of demographic prediction from questionnaire data, log data and baseline methods, and Table 9 shows the two-tailed paired t-test among these results. We find that (1) both questionnaire-based and log-based predictions significantly outperform random and mostPop baselines; (2) for both questionnaire-based and log-based features, the combination of all CPS features always significantly outperforms predictions based on features capturing a single category of behavior; (3) applied to all questionnaire participants, log-based features achieve comparable results to their questionnaire-based counterparts, although the log-based features do not cover the social behavior features. The reason for this might be that some of the log-based features contains more fine-grained information than their corresponding categorical counterparts in questionnaire data, e.g. WiFi usage frequency; (4) when classifying participants using cyber browsing/searching logs, the prediction accuracies significantly improve for all demographics except for parent status (see line (cyber-physical) cyber user of Table 8). Specifically, the accuracy of User Type prediction is boosted from 47.25% to 66.04%. In addition, this also indicates that capturing social grouping behaviors is necessary to predict parent status; (5) the ability to predict different demographic attributes varies. Figure 4 shows the improvement in prediction of demographic attributes based on either self-reported CPS features or logcaptured CPS features, compared with mostPop. For questionnaire-based features, the largest improvement is achieved on user type (35%), while the least improvement is obtained on education (5.1%); the improvement on the remaining demographic attributes, i.e., age, income and parent is similar, around 15%. This may indicates that while a visitors user type is relatively easy to predict, predicting attained education is hard, while remaining attributes are similarly predicatble. (6) the predictive capacity of individual CPS features differs by demographic group: