| dc.description.abstract | Ecommerce websites have a huge number of customers browsing their sites per hour, other larger stores have millions per minute. But, just a small fraction of these customers' activities results to purchase. Numerous marketing studies have made efforts to understand this weak relationship, but one common method in their approach is the use of customer responses from surveys and questionnaires. Scholars have described this method as an obstructive method, given that customer responses may often differ from their actual behaviour. We were motivated to Our motivation for this master thesis is the call for marketing studies using the clickstream dataset. Clickstream data is a raw log file that keeps a record of all customer's activities on a website, it keeps track of the customer's actual behaviour. This research was conducted using clickstream data of 20,692,840 million customer events. The behavioural patterns of 1,204,158 customers were extracted from 3,055,272 unique browsing sessions through a method called ‘session-level aggregation’. The “effects of customers' online behavioural patterns on their purchase decision” was investigated using the extracted behavioural features, and the following findings were concluded:
Customers are less likely to purchase when they: view more product pages, evaluate more unique products/brands and remove more products from carts (indicating an active formation of consideration set). But they are more likely to purchase when they: spend more time on browsing sessions, and add more products to carts. Our findings also indicate that customers are less likely to purchase when they evaluate product alternatives with popular brands, and are more likely to purchase when they evaluate product alternatives with non-popular brands. We reported that customers alternative evaluation does not depend on product category. Five behavioural segments of online customers were identified, and behavioural characteristics of each segment was used to classify the stage of the customers in their buying decision process. Finally, five machine learning models was used to predict customers purchase and non-purchase sessions - extreme gradient boosted method (XGB) was reported to be the best predictive model. Explainable Artificial Intelligence (XAI) was explored on XGB model - it was discovered that time spent on browsing sessions, is most important variable that predicts customers purchase and non-purchase. | |