ICACT20220017 Question.1
Questioner: 211120098@fzu.edu.cn 2022-02-14 오전 3:50:51 |
ICACT20220017 Answer.1
Answer by Auhor patrick.hosein@sta.uwi.edu 2022-02-14 오전 3:50:51
Chrome  Click!! |
| Why you chose a relatively simple model like logistic regression model and what are its benefits？ |
I tested many different models such as decision tree, support vector machine (svm) and naive bayes classifier, and I even included more complex models in my testing like random forest and gradient boosting algorithm. However, the accuracy metric obtained from these models were either lower or only marginally better than the logistic regression. I chose to use the logistic regression model because it offers the benefit of interpretability. Remember, the model is being used to help the Account Managers select the best product recommendation to upsell to the customers. The managers would most probably want to know the main reasons why the particular upsell product is being recommended, so they can use this information in developing a relevant and persuasive strategy when targeting the customers. With a logistic regression, it is easy to discover which variables impact on the recommended output made by the model, and the weight of this impact. This can be easily interpreted using the coefficients of the regressors in the model. |
ICACT20220017 Question.2
Questioner: 211120098@fzu.edu.cn 2022-02-14 오전 10:01:20 |
ICACT20220017 Answer.2
Answer by Auhor patrick.hosein@sta.uwi.edu 2022-02-14 오전 10:01:20
Chrome Click!! |
| Why is the threshold t added and how does it improve the model？ |
I wanted to account for misclassification costs in the model, since in the reality of the upsell business there is a disproportionate cost associated with making a false positive versus a false negative error. In general there are two approaches I could have taken, the first being a cost-sensitive approach such as rebalancing the data by class weighting, under-sampling, or over-sampling techniques. In this case, observations from the less costly class are given higher importance during training. However, I chose a cost-insensitive approach by adding a threshold. In this case, probability estimates, outputted by the classifier that learned from the training set as given, are used to obtain a probability cut-off (the mentioned threshold t) in order to compute cost-optimal decisions. I favored this approach because I did not want to modify the dataset as there was already not a very large data sample. Implementing an optimized threshold in the model will minimize misclassification costs. Therefore, although the accuracy of the model will be reduced, the business loss of the upsell process will be minimized, and therefore the upsell process will be more profitable if the model with the added threshold is used. |
ICACT20220017 Question.3
Questioner: 211120098@fzu.edu.cn 2022-02-14 오전 10:06:26 |
ICACT20220017 Answer.3
Answer by Auhor patrick.hosein@sta.uwi.edu 2022-02-14 오전 10:06:26
Chrome Click!! |
| How do you plan to predict the probability of upsell within a defined future
time period in future work,any ideas or suggestions? |
I have the idea of using a rolling time window, for example, using a time period of every quarter or every 6 months. In this initial study, the dataset taken for an entire yearly period was used in the training set (2018) and it was tested on data for the following year (2019). However, there is the propensity for new products to be added to the market and existing products to become obsolete. It is therefore crucial to ensure that the time window used for training and testing captures a representative picture of the products to be upsold. The prediction of the probability of upsell will be similar to the one used in this study, where training data from a historical time window will be used to develop the model, which will then be tested on data from the next time window. This will be done on a continuous basis at the end of each time window so a dynamic, rather than static, model will be implemented. |
ICACT20220017 Question.4
Questioner: tomayoon@ieee.org 2022-02-14 오전 10:11:04 |
ICACT20220017 Answer.4
Answer by Auhor patrick.hosein@sta.uwi.edu 2022-02-14 오전 10:11:04
Chrome Click!! |
| First of all, thank you for the good research paper and presentation contents. I would like to know more about the research environment used in this study, and also in detail about the operating systems, development languages and software libraries you used. |
Thank you very much for your kind feedback. I used the software language R to conduct the research and analysis, on a Windows 10 laptop. The generateFilterValuesData function from the FSelectorRcpp package in R was used to compute a rank of the variables by assigning an importance value to each feature with the use of information gain. All missing numerical values were imputed using the KNN (k-nearest neighbours) method with the caret package in R. Model training was also done with the caret package. The expected cost to be optimized was evaluated using the makeCostMeasure function in the mlr package in R. |
ICACT20220017 Question.5
Questioner: 211120098@fzu.edu.cn 2022-02-14 오후 12:43:07 |
ICACT20220017 Answer.5
Answer by Auhor patrick.hosein@sta.uwi.edu 2022-02-14 오후 12:43:07
Chrome Click!! |
| Can you provide youtube link. |
I do not have a YouTube link for this presentation currently but I will try to provide one as soon as I possibly can. |
ICACT20220017 Question.6
Questioner: jypark9740@g.skku.edu 2022-02-14 오후 10:55:07 |
ICACT20220017 Answer.6
Answer by Auhor patrick.hosein@sta.uwi.edu 2022-02-14 오후 10:55:07
Chrome Click!! |
| Thank you for your informative papers. During the presentation, we had a total of 34 successful upcoming sales, and 34/80 = 42.5%. What conditions are 80 customers? |
From the period February 2018 to February 2019, there were a total of 270 customers who were approached by the telecommunications service provider (from a pool of 3117 customers in total). Yes, we note that the telecommunications company's approach rate is very low. The traditional semi-manual approach for upselling to customers, via the Account Managers, was employed for the upsell process. For the model training, these 270 customers were randomly split into a training set and test set using a 70%/30% partition. Hence, 190 customers were in the training set and the remaining 80 customers belonged to the test set. Out of these 80 customers who were in the test set, only 34 customers engaged in a successful upsell, hence the success rate from the traditional semi-manual approach was 34/80 = 42.5%. Retrospectively, if our model was used to target these customers, there would have been a higher number of successful upsells when approaching 80 customers (from the pool of 3117 customers in total). We report on the 80 customers to make a direct comparison between the model performance and the traditional semi-manual approach used by the telecommunications service provider. |
ICACT20220017 Question.7
Questioner: tamnd17@viettel.com.vn 2022-02-18 오전 1:24:09 |
ICACT20220017 Answer.7
Answer by Auhor patrick.hosein@sta.uwi.edu 2022-02-18 오전 1:24:09
Chrome Click!! |
| Thank you for the information in this research. Can you provide more about what are the telecommunications products you use? |
We developed the recommender system for voice products only. The legacy products include Avaya/Nortel PBX, Cisco PBX, Mitel PBX, Single Lines, T1s and Trunk Lines. Upgraded products include hosted PBX and SIP lines. Note that PBX is an acronym for Private Branch Exchange, and SIP is an acronym for Session Initiation Protocol. |
ICACT20220017 Question.8
Questioner: tamnd17@viettel.com.vn 2022-02-18 오전 1:30:13 |
ICACT20220017 Answer.8
Answer by Auhor patrick.hosein@sta.uwi.edu 2022-02-18 오전 1:30:13
Chrome Click!! |
| In the case of complex telecommunications packages such as mixed service packs, do you have a way to assess the level of impact on revenue and other packages when customers change the package? |
This is a good question but unfortunately I do not have the answer for you, as revenue information was not provided for the study. The revenue could vary a lot for different types of sales so it isn't easy to assess a general ballpark figure. I am sure the telecommunications service provider can have the sales data for each customers' service packages, but this data was not provided for the study. With regards to the complexity of mixed service packages, I acknowledge that there are a variety of services offered to the corporate customers, such as voice, network, security and cloud services. Independent recommender systems for each of the different services are being developed (currently we have one for voice services and one for network services). However, we currently do not account for interaction or hybridization of packages. |
