Feature Engineering:

Data exploration revealed outliers in the count data of 'nrservice' (mean = 4.4, max =207) and 'nrchatbot' (mean = 10.5, max =610). Rather than eliminating these outliers, we decided to apply the ‘winsorization’ technique to retain valuable information about the customers in the outlier groups.

Specifically we found that, customers who are in the outlier group in both ‘nrchatbot’ and ‘nrservice’ are marked by significantly higher positive call center service, longer service length, higher usage rates, and higher energy consumption (as shown in the higher mean value of  av_gas, av_elec, and av_bill) than the general customer base. Therefore, the outliers represent a specific subset of customers with unique characteristics, rather than erroneous. Using the winsorization technique to cap the outliers, we can retain valuable information about these customers' unique characteristics.

‘nrservice’ distribution and the comparison of customers behavior & characteristics in outlier vs non-outlier groups

Overview of the reasons behind the adoption of 3 models in this project

Initial analysis using Poisson Regression identified evidence of overdispersion  (result of dispertion test with p < 2.2e-16), where the variance of the count data significantly exceeded the mean. This violation of the equal mean and variance assumption led us to adopt the Negative Binomial Regression model, which includes a parameter to account for excess variance. However, due to zero-truncation in our dataset, we proceeded to employ a Truncated Count Model. This decision was supported by the fact that our dataset include zero-truncation, where observations with a count of zero are not included.

From 2000 observations, we observe 871 and 689 observations have a minimum value of 1 in ‘nrservice’ and ‘nrchatbot’ respectively.

To compare the fit of the three models, we adopt the AIC and BIC information criterion. The results consistently revealed the Truncated Count Model as the best fit for both 'nrservice' and 'nrchatbot'. For 'nrservice', AIC and BIC scores were lowest for the Truncated Count Model (AIC=7452.185, BIC=7524.997) when compared to both the Poisson (AIC=9726.4, BIC=9793.61) and Negative Binomial Models (AIC=8580.325, BIC=8653.137). Similar results were observed for 'nrchatbot', where the Truncated Count Model (AIC= 9777.1, BIC= 9849.9) outperformed the Poisson (AIC=20028.17, BIC=20100.98) and Negative Binomial Models (AIC=10989.33, BIC=11062.14). Thus, the finding clearly demonstrates the importance of proper model selection based on data characteristics and the usefulness of the Truncated Count Model for zero-truncated count data. The analysis of the truncated negative binomial models has provided substantial insights into the variables influencing customer service calls, 'nrservice' .

Summarization of results

• A unit increase in 'diff_chatbot' leads to an estimated 1.82% increase in the expected count of 'nrservice', which implies that customers who have more positive experiences with the chatbot tend to use the service calls more frequently. This highlights the intertwined nature of different service channels and the need for integrated management strategies that synergistically improve both.

• The variable 'days_since_first_service' has a negative coefficient, showing customers who start using the service call in the early periods make more 0.074% service calls for each additional days. This insight implies that older customers may be accustomed to and rely more on traditional service (calls).

• Conversely, the 'days_since_first_chatbot' variable positively impacts 'nrservice', suggesting that recent (new) chatbot users are more likely to have 0.04% more service calls for each additional day since the first chatbot take place. This could indicate that after an initial interaction with the chatbot, customers may have unmet needs or questions that lead them to seek support via a service call. Alternatively, it might indicate that chatbot users are more comfortable with technology and, thus, more likely to use multiple service channels.

• The 'years_since_customer' shows that longer-term customers make fewer service calls, decreasing by about 20.6% for each additional year. Customers may become less dependent on service calls as they become more familiar with the products.

• Additionally, for each extra member in the household ('hh_size'), 'nrservice' increases by around 20.5%, suggesting larger households may have more complex needs and require more customer service interactions.

• Lastly, a slightly positive impact exists with the 'av_bill' variable, where a unit increase leads to a 0.01% increase in expected 'nrservice'.

• Analyzing the truncated negative binomial model on chatbot interactions (‘nrchatbot’) shows several significant indicators.

• For every unit increase in the net positive experience in service calls (diff_service), there's approximately a 10% increase in the expected count of ‘nrchatbot’, which is consistent with the findings earlier on the adoption of multiple service channels.

• The 'days_since_first_service' and 'days_since_first_chatbot' also play roles but with marginal effects of a 0.01% increase in chatbot usage and a 0.12% decrease in usage, respectively. Customer satisfaction has a positive impact, with each unit increase in satisfaction predicting approximately a 14% increase in chatbot usage.

• Household size is a positive predictor, with each unit increase in household size leading to a 24% increase in chatbot usage.

• Years since the installation of solar panels has had a negative impact on chatbot usage. For each additional year since solar panel installation, chatbot usage decreases by approximately 7.4%. This could suggest that customers with older solar panel installations interact less with the chatbot, possibly because they are more familiar with the technology and have fewer questions.

• Interestingly, customers who have received a personalized solar email are about 4.2% less likely to interact with the chatbot, while those who receive newsletter email have about 0.19% more interactions with the chatbot than those who didn't.

• Finally, the average bill has a positive but marginally significant impact of a 0.01% increase in chatbot usage.

  Overall, this analysis suggests that proactive communication and enhanced onboarding support for new customers can reduce their need for further assistance. Investments should be targeted towards improving chatbot technology, especially considering its increased use with larger households, while ensuring high customer satisfaction. Lastly, personalized services based on customer traits and thorough solar panel installation education can further augment customer satisfaction while reducing dependence on customer service and chatbot interactions.