Research in online commerce is heavily focused on identifying and understanding the factors that drive behaviors of users, since such findings can have immediate and direct impact on the sales volumes. One of the long-established principles in the field have been that the more time a prospective customer spends in a visit, the more likely is the purchase decision. Meanwhile, the annual growth of mobile commerce (20-35%) is even higher than of ecommerce in general (15-20%), and many established beliefs about web users' behavior are shifting. In our paper we build two simple models with datasets available through Kaggle repository, to explore the effect of visit time on purchase decisions and online revenues. The results suggest that the time spent on mobile app did have moderate effect on the sales (R2 = 0.249), while time spent on website was not significant at all. The F1-score in the classification model for the visitors' buy / no buy decision was also rather low, at 0. Correspondingly, we discuss some other related factors that could be used to enhance the quality of the models and presumably improve the online sales. Thus our results might be of interest to both machine learning specialists and to electronic commerce or marketing practitioners.
Every year the number of Internet users is increasing, which leads to an increase in the volume of
online purchases (e-commerce) by 15-20% annually [
In conventional stores, it is generally very difficult to get information about the number of visitors,
the time spent in the store, where the customers come from, etc. On the contrary, modern web analytic
systems allow obtaining such information easily for e-shops. The main metric for online stores is the
number of sales, that is, user reaching the final product order or check-out page [
Visit duration has long been considered a key performance metric in e-commerce, enhancing
conversion rate and reflecting loyalty to the e-tailor. In a study from 2015 [
The results of the above and similar research works have been considered helpful in improving
(optimizing) e-commerce websites, to achieve longer average visit durations and presumably enhance
sales. As we mentioned above, the “time on site” metric has been important in online commerce
already two decades, the reasoning being as follows. The longer a user has spent in the online store,
the more interested she is in the products of this site [
The research objective of our paper is investigating how the time spent by users in online stores affects their income. For this end, we test two hypotheses:
1. Does the time spent by the user on online store website affect the purchases (income from a customer)? 2. Does the time spent by the user in m-commerce mobile application affect the income?
In our study, we relied on two openly available datasets that we describe in the following Section. In Section 3, we test the hypotheses using SPSS Statistics software2 and report the results. Finally, we discuss the findings and make the conclusions. 2. 2.1.
The first dataset that we used was Online Shopper’s intentions, openly accessible through Kaggle
[
The second dataset that we used in the study was the Linear Regression E-commerce Dataset, also
from the Kaggle repository [
User behavioral models are subdivided into short-term and long-term models. The construction of
the long-term models is mainly required for their further application (for example, for use in
recommender systems). Long-term behavior describes relatively stable user preferences based on an
extensive history. The short term models describe a user’s current activity and his / her actual
intentions. Each type of models has its own advantages and shortcomings – for instance, long-term
models can collapse when a web resource is updated, which happens rather often nowadays [
Since the datasets that we employ in our research involve relatively extensive user history, they should correspond to long-term behavioral model. As a particular method we used for the analysis of the two datasets, logistic and linear regressions were chosen, in accordance to the data structure, particularly the binary dependent variable of the Online Shopper’s intentions dataset.
Logistic regression is a form of multiple regression, whose general purpose is to analyze the
relationship between multiple independent variables (also called regressors or predictors) and a
dependent variable. Binary logistic regression is applied when the dependent variable is binary – i.e. it
can take only two values. The independent variables are generally interval or rational scale, which is
exactly our case for the duration times. Through the use of logistic regression, it is possible to
estimate the probability that an event will occur for a particular subject (sick / healthy, loan repayment
/ default, etc.) [
Linear regression, which we naturally employed for the Linear Regression E-commerce Dataset, is used in statistical analysis and machine learning to describe the effect that one or several independent variables (factors, regressors) have on a dependent (explained) variable. All variables are rational or interval scale, which both the time and the annual revenue conform to, and the effect is expressed as a linear dependence function. The quality of linear regression model is an extensive research topic, but we in our current study used the widely recognized R-squared and significance metrics.
First, we analyzed the Online Shopper's intention dataset, in which three independent variables are involved: 1) the time of working with the administrative part of the site, 2) the time of working with the information part of the site, and 3) the time of working with the product pages. In the dataset, 12112 records (98.2%) turned out being valid. The dependent binary variable, the site's income, takes the value TRUE or FALSE: whether the user made a purchase during this session or not. The regression model was built using the stepwise Forward LR method.
When a model was built, with a cut-off value of 0.5 (default) the percentage of cases correctly classified into observed categories was 84.4. However, the classes in the dataset are unbalanced: only 1875 (15.5%) visitors had made a purchase (TRUE), while 10,237 (84.5%) did not (FALSE). So we varied cut-off values (0.45, 0.4, 0.35, 0.3, 0.25, 0.2, etc.), obtaining the resulting Precision and Recall as presented in Table 3. Notably, at the cut-off value level of 0.15, the Accuracy of our model began to decline rapidly (by 11.6% compared to the cut-off value of 0.2). The lowest threshold of the cut-off value adjustment that we performed was 0.12, as Precision had achieved 1 at this level, the Accuracy however being already rather poor, at 15.5. The maximum F1-score value (0.342) was found for the cut-off value of 0.14.
In Fig. 1 we present the Precision-Recall curve for the model, corresponding to the considered cutoff values. In Fig. 2 we also show the ROC-curve – although it’s more suited for the case of the balanced classes, we decided to include it for the illustrative purposes. The area under the curve is 0.681.
As the quality of the classification model is relatively low, we can conclude that our hypothesis “the time spent by the user on the website of the online store affects the purchase (income)” supported much. The three considered time values failed to considerably improve over the “null” predictor that would just predict TRUE for 15.5% of the classified cases. 3.2.
The next dataset that we analyzed was the Linear Regression E-commerce Dataset, for which we used the pairwise linear regression method. The multiple regression method was inappropriate because the independent variables are highly collinear with each other, as can be seen in Fig. 3. We built the first model in which we used the Time on App as the independent variable (the time the user spent in the application during the year), and the Yearly Amount Spent (the sum of the user's purchases for the entire year) as the dependent variable.
The results of the linear regression analyses for the two independent variables are presented in Table 4. The R2 = 0.249 for the Time on App is not very high, but it is considerably higher than R2 < 0.01 for the Time on Website variable, which was also not statistically significant. In the regression equation for Time on App, the constant was not statistically significant (p = 0.608), which is in line with the conceptual reasoning that zero time spent in online store cannot result in any revenue from such a user. Still, we employ the regression analysis that includes the constant, for the sake of comparability. In Fig. 4 we show the linear regression plot for the Time on App model.
The linear regression analysis results presented in Table 4 suggest that the Time on App variable much better explains the dependent variable than the Time on Website. Hence, it can be assumed that time spent on an app is more likely to result in a purchase than time spent on a website. So, the hypothesis 1 (impact of time on websites) should be rejected, as the classification model in the dataset 1 was weak and Time on Website was not significant in the dataset 2. The hypothesis 2 (impact of time on mobile apps) can be confirmed, as Time on App was significant in the dataset 2, even though the R2 = 0.249 was low.
The low share of variance explained by the time in the online store revenue suggests the effect of other important factors. Among the factors related to online store visits, we can note the week day of the visit, whether the visit is a repeated one, the user’s region, etc. These deserve further explor to improve the models for the dataset 2. In turn, our classification model for dataset 1 (F1 = 0.342) may foremost suffer from lack of consideration of repeated visits. That is, in one visiting session the user could obtain all the necessary information for the purchase decision-making, while the actual purchase can be made during the next, shorter visit.
However, the general outcome of the study is consistent with some related research, e.g. in [
We would also like to note that our study was very much limited in the number of online stores that were affected. In both Online Shopper’s intentions and the Linear Regression E-commerce datasets, the data for only one (though different) online store is presented. Correspondingly, a study of much larger scale is necessary before generalizing our conclusions.