E-commerce platforms have become a prevalent means of selling and purchasing goods. They aim to provide efficient services for both retailers who sell goods and their customers. Optimisation of e-commerce platforms requires effective modelling of interactions among the retailers, customers and the platforms themselves. This thesis is concerned with two specific aspects for the contemporary e-commerce platform optimisation: (1) the cross-platform nature of the sales cycles that span Social Media and e-commerce platforms and require ways to assess the effects of Social Media activities on the sales and (2) the real-time involvement and session interventions to support customers' purchase decisions and provide incentives for customers to complete their purchases.
E-commerce platforms operate within a complex ecosystem of online services. They facilitate sales of products while the product retailers engage in Social Media Activities (SMAs) to drive E-commerce Platform Activities (EPAs), enticing consumers to search, browse and buy products. With the increased investment in social media campaigns, it is important to assess their effects on e-commerce outcomes, including product sales. That requires consideration of the sales cycle that may begin through SMAs by engaging with customers through social media posts and concludes with the selection and purchase of products through EPAs. In the past research, modelling such activities have been done in an ad-hoc manner, without analysing the process of the sales cycle that spread across the social media and e-commerce platforms in a systematic way.
In order to assess the relationship between SMAs and EPAs in a principled way, this thesis proposed the Multi-platform Process Analytics (MultiPA) framework, which incorporates two behaviour models to describe the activities in both social media and e-commerce platforms, which are aligned with the components of the sales cycle and support descriptive and predictive analytics to characterise the relationships. The dataset used for experiments includes the social media data from Weibo's platform and e-commerce platform data from JD.com in China.
By running the correlation analysis, post activity correlates with EPAs the most compared to the comment and repost activities. Through a detailed post content analysis based on LDA, the thesis identified several topics that are highly correlated with EPAs, which may help in the EPAs prediction. As guided by the MultiPA framework, the thesis further conducted the predictive analytics and made predictions of EPAs based on the frequency and content characteristics of SMAs in order to assess the level of alignment and congruency between SMAs and EPAs. The prediction results suggest that features derived from the topics of vendors' posts outperform the features derived from frequencies of SMAs. Both prediction results demonstrate that Social Media has selective success in predicting EPAs, i.e., Social Media can predict the top quantile of order and low quantiles of clickthrough and search. This approach, supported by the MultiPA framework is the first to provide a systematic and principled analysis of the cross-platform sales support. The prediction result provides insights to product vendors on how to devise the marketing messages that are more congruent with the sales activities on the e-commerce platform in order to achieve high sales conversions for e-commerce optimisation.
Considering live interventions during active customer sessions, it is essential to provide the intervening agents with information that can support the decision when and how to engage with the customer. The decision may involve multiple factors related to the anticipated customer behaviour. In particular, to increase the conversion rate through timely offer of deals, it is important to predict the user's purchase intent and to estimate the time available to intervene and offer a deal.
For modelling user behaviour with regards to these two problems, the thesis uses the real world user clickstream data from the e-commerce platform JD.com in China. The clickstream data includes different aspects of user visits, such as the time that user spent, products that user visited and so on. When modelling the user's intent to purchase, the thesis adopts the Purchase Decision Model (PDM) and proposed the Action Purchase Model (APM) to analyse the consumer actions in the clickstream data and constructs the multi-action motif features, which are assumed to be able to capture the progression of the consumer decision making process through the PDM. Experiment result shows that by training the Logistic Regression classifier based on the multi-action motif features, it can achieve results that are on par with the state-of-the-art LSTM sequence model in predicting user purchase in a session, while requiring less computational resources and delivering explainable results. Similarly, for modelling the session duration of the session, we evaluated the performance of traditional machine learning models with features and the sequence model LSTM, and found that the LSTM sequence model outperforms machine learning models with features.
Based on the modelling of user purchase intent and session duration, the thesis further proposed the Real-time User Modelling (ReTUM) framework which provides a systematic way of testing the machine learning predictors and simulating the agents' behaviour. Experiments demonstrate that the machine learning fusion of the user intent and session duration predictors can be applied successfully to learn the 'intervention classifier' that corresponds to the pre-defined agent's policy for live intervention. Similarly, ReTUM supports evaluation of the stopping criteria in terms of the expected gain that can be achieved for the specific agent's intervention practice. Once the system is operational, ReTUM can use the actual data from the agents' interventions to feed into the machine fusion of multiple factors and provide real-time support for the agents. In this way, it can reduce the loss of potential opportunities for sales and achieve the e-commerce optimisation.
|Date of Award||16 Nov 2019|
- Univerisity of Nottingham
|Supervisor||E Ch'ng (Supervisor), Natasa Milic-Frayling (Supervisor) & Hing Kai Chan (Supervisor)|