Online Shopping purchasing behaviour was collected from a website over a 1 year period, and marked whether the visit ended in generating revenue or not.
A model was built using the given information to determine if the visit will end with an item being purchased.
- Assumptions
- Key Insights from modeling the data
- Data
- Key features that drove analysis
- Modeling Pipeline
- Final model
- Conclusions and focus points
- Additional improvements
- Data is representative of the whole population
- The values of bounce rate and Page values provided is for the final page the customer views before exiting the site
- It is more important to capture visits that will generate revenue, rather than those that do not. For example – Target marketing campaigns, discount offers and other promotions etc.
The dataset can be found here
- The lower the bounce rate, the better
- A drop in bounce rates by 0.04 increases the odds of a customer making a purchase by 54%
- An increase in Page Value increases the odds of Revenue
- 43 percent of all visits generate revenue for pages whose Page Value score is greater than 0
- Odds of a purchase can increase by up to 8 times
- November has the most favorable traffic ratios, with conversion opportunities in December
- Odds of a purchase increase by over 250%
- Whereas December immediately drops the odds by 12%
As mentioned above, the dataset can be found in the UCI ML repository, along with the data dictionary.
- The class distribution is unbalanced with a ratio of 5.6:1 (False:True)
- Data set was split into three categories of Training, validation and holdout of 70%:15%:15%
- Sampling strategy employed was stratified sampling to maintina the imbalance ratio.
This section contains information on key features that drove the analysis and subsequent modeling.
Visits which generated revenue had a marked increase in the number of informational pages and product related pages visited.
The higher the number of informational and product pages visited, the more likely the visit ended with a purchase
Visitors entering sites with lower bounce rates, or higher page values, were more likely to generate revenue
While most sales happen in November, there is a large rise of traffic in the months of March, May and December, but sales are not proportionately rising
There is definitely a potential conversion opportunity in the months of March, May and December.
The data was subject through to a standard processing pipeline.
Numerical columns were subject to correlation analysis to filter out highly correlated variables
Categorical variables were subject to Chi Square tests against the Revenue variable
To visualize potential clustering of the preprocessed data, it was projected into a low dimensional space using tSNE and plotted
Clustering algorithms like KMeans and DBSCAN could not form any significant groupings on the dataset
The preprocessed data was classified using Logistic regression to establish a baseline, and to find significant features
Upon evaluating multiple feature selection methods, it was found that using the significant features, combined with weighting the desired Revenue group gave the best result
In a similar fashion, multiple classifiers were evaluated, with metrics and trade offs analyzed for each algorithm
While F1 scores were comparable, Precision and Recall values were deviant between each algorithm. The final choice of algorithm was Logistic Regression to also retain interpretability.
The Logistic Regression model allowed us to retain comparable F1 scores, and study the effect of significant features on Revenue
-
Page Values have the most positive effect on generating Revenue, increasing odds by upto 8 times for each 18 units. Similarly, the most deleterious effect is from Bounce Rates, where an increase in 0.04 reduces the odds by nearly 45%.
-
In November, the odds of a purchase can increase by 275% compared to the prior months. Whereas in December, the odds can drop by 12%
- Pages with lower bounce rates were shown to be a strong indicator of whether the visit generates revenue
- A drop in bounce rates by as much as 0.04 increases the odds of a customer making a purchase by 55%
- Analysing features of pages with lower bounce rates, can provide further insights
- Higher Page Value is a strong indicator of whether the visit generates revenue
- 43 percent of all visits generate revenue for pages whose Page Value score is greater than 0
- An increase of 19 units in Page Value, increases the odds of a customer making a purchase by 8 times
- November generates most favorable traffic ratios, but months such as December have a large amount of traffic as well
- Odds of a purchase in November increase by 275% compared to the prior months
- Traffic in December can be targeted for conversion to generate more Revenue
Further improvements can be made for model generalization and improving metrics
- More complex feature engineering
- Focus on non-linear classifiers
- Increase granularity of data collection (e.g. weekly level)
- Collect more samples
- Acquire different dimensions and features (for example, page level information)