Customer Service Recommendation to the Restaurants based on Predictive Analysis

Customer Service Recommendation to the Restaurants based on Predictive Analysis

Tejaswini Dhupad Information Technology Savitribai Phule Pune University Pune, India

Ayushi Deshmukh Business Analytics Deakin University Melbourne, Australia

Girimal Dhupad

Computer Science Savitribai Phule Pune University

Pune, India

AbstractOver the recent years, using the electronic networks has been increased and the internet has changed into one of the main channels of obtaining the information, hence today the decision making process of the consumer is highly driven by the information that has been catered to them over the internet. To effectively monetize the customer base and maximize revenues, carriers have become increasingly interested in how they can leverage customer analytics to ensure that the best offer is presented to the right customer at the most appropriate time across any channel. Therefore, predictive analysis can be used to identify the service strategy for the new and established restaurants to implement profitable business model.

1. INTRODUCTION

   To determine the factors that explain customer satisfaction in the full service restaurant industry, the data was gathered from the Bangalore food assist (BFA) in association with zomato, who was interested in precisely identifying if restaurants should provide online order or book table service to their patrons. BFA has provided with a sample of 40,000 established restaurants data which are represented using address, rate, votes, neighbourhood, cuisines, online_order, book_table etc. The restaurants are rated according to the quality of service they provide and customer satisfaction level. These services are valued by customers and can be used as an added advantage over its competitors. BFA is trying to analyze if online order and book table service should be appropriate for Bangalore restaurants to be inherited in their business model. The findings suggested that the customer satisfaction was greatly influenced by the customer review and other factors (such as cuisines, meal-type, restaurant-type, restaurant name, address etc). These services impact the attractiveness of the restaurant and its customer rating. The study tests the existing model and enhances the literature on to implement profitable business model for restaurants.

   A. Duncan's multiple range test

   When the analysis of variance test on an accuracy measure produces evidence to reject the null hypotheses, H~ and H~, we can accept the alternative hypothesis-that all of the mean accuracies are not equal. However, we still do not know which of the means are significantly different from which other means, so we will use Duncan's multiple range test to separate significantly different means into subsets of

   homogeneous means. For the difference between two subsets of means to be significant it must exceed a certain value. This value is called the least significant range for the p means, Rp, and is given by

   Rp = rp V/~, (1)

   where the sample variance, s 2, is estimated from the error mean square from the analysis of variance, s~, r the number of observations (rows), and rp the least significant studentized- range for a given level of significance, and the degrees of freedom [1].

   The below table predicts the classification of attractive restaurants in the given neighbourhood[2]

   Figure 1:Classification of attractive restaurants in the given neighbourhood

2. DATA EXPLORATION AND PREPARTION IN RAPIDMINER

   Based on the business problem identified and stated in executive summary, identified attributes are considered as significant: rate(numerical), votes(numerical), average_cost(numerical), neighbourhood(categorical), online- order, book-table for answering attractive restaurants in a neighbourhood.

   The dataset provided by BFA is showing few errors while

   predicting the book table attribute(label). As seen from the plot the restaurants providing book table services are too high on average as compared to restaurants not proving the service when plotted against average cost. This implies for unbalanced data in the data-set. To deal with this problem, we have made use of SMOTE operator in the model while predicting the book_table label.

   The vote attribute which is used as one of the predictors to identify attractive restaurants in the neighbourhood has few extreme values (i.e. outliers) as clearly seen below.

   Outliers

   Outliers

   Figure 3: Scatter plot for votes vs neigbourhood

   As seen from the bar chart of the attractive restaurants are found in Koramangala 7th block, BTM followed by Jayanagar. This result was formulated using two key strategies: rate per dollar charge (normalized) attribute must be higher than the average rate per dollar charge and votes gained by the restaurants must be greater than the threshold votes to classify restaurants into Attractive restaurant and Non-Attractive restaurant within Attractiveness attribute. Missing values are dealt by using impute missing value operator (sub-process KNN schema). Reason behind using KNN schema was to predict appropriate closest k neighbours for all missing values.

   Figure 5: Attractive restaurants in the neighbourhood

3. DISCOVERING RELATIONSHIPS AND DATA TRANSFORMATION IN RAPIDMINER

   As per the BFA interest in determining if the restaurants must or must not provide online booking services in Bangalore. Considering – online order and book table as the labels. Out of all the 17 attributes most prominent once are address, name, phone, cusines, menu_item, (weighted against the labels). Follow below graphs for more insights about attributes relationships:

   Figure 6: Histogram for featuring relationship between book table and average_cost

   The outliers are tackled by normalizing the data to eliminate extreme values and gain appropriate results.

   The outliers are tackled by normalizing the data to eliminate extreme values and gain appropriate results.

   Figure4: Impute missing value and outlier

   As the scatter plot interprets that restaurants with average rating more than 4.5 has more density of votes (no online order service available). Hence, we can consider rate and votes as prominent attributes to interpret online_order column.

   Figure 7: Relationship between online order, rates and votes

   The high weights signify that the provided predictor attributes (numerical and non-text) are more important for predicting the label attributes: online order, book table while building the model. In the models, the attributes which weight more than

   0.3 are considered as prominent, as they have some relation with the label attribute. Furthermore, the correlation between numerical attribute is also weighted. For example, book_table and average_cost have high correlation between them i.e. 0.617.

   Book table Online order

   K

   Figure 8: Model to find correlation between numeric variables

   Figure 9: Model for weighing categorical variables

4. CREATE A MODEL(S) IN RAPIDMINER

   1. K-NN Classification:

      Based on the Data retrieved from BFA, required predictors are identified by weight (address, average_cost, cuisines, dish_liked, name, online_order, phone, rate, votes). In the designed model missing values of rate and average_cost are taken care by replace missing value operator by averaging the values and role is set as Online_order as a label attribute. Sampling type is chosen as stratified sample as the label is nominal attribute, as this will ensure both the classification of online_order are selected proportionall in training and testing data set. The sampling data is split into two segments of 70%(training) and 30%(testing). The model is designed using K-NN classification technique. The distance type and mixed measure are taken as Mixed Measure and Euclidean distance respectively, as it is most optimal and less biased. In K-NN model k value is decided based on accuracy and kappa of the model. The testing data is applied to the model to get a good accuracy and kappa of 93.73 and 0.875 respectively.

      SMOTE upsampling is done to handle unbalanced data. It creates a balance between restaurants who provide book-table service and those who doesnt provide. Moreover, it is applied on training data and not on validation data-set, to keep the accuracy end results as practical to true outcome as possible . For SMOTE upsampling the number of neighbourhood parameter is kept as 5.

      SMOTE upsampling is done to handle unbalanced data. It creates a balance between restaurants who provide book-table service and those who doesnt provide. Moreover, it is applied on training data and not on validation data-set, to keep the accuracy end results as practical to true outcome as possible . For SMOTE upsampling the number of neighbourhood parameter is kept as 5.

      Figure 10: Model for established restaurants for label book tabel.

      The value of k is calculated using a loop parameter with 10 iterations, best k and kappa value is found ,accordingly that k value was picked. As per the observation as the k value increases the accuracy is decreasing but kappa is increasing. As more kappa accounts for occurring of correct predictions by chance. Hence it is one of the robust measure of calculating the performance of the model.

      Figure 11: Loop parameter to calculate k value

      As seen from the graph plot of accuracy and k, by using elbow rule k=11 is good value for the model.

      As seen from the graph plot of accuracy and k, by using elbow rule k=11 is good value for the model.

      Figure 12: Elbow plot

   2. Decision Tree:

      To compare the results, alternative model has been built using decision tree, in which restaurants data-set is passed and label is set to online_order. We have not dealt with missing values, as the decision tree algorithm is well capable of handling missing values. The data is split into two sub- segments of 70% and 30%. With 70% dataset model is built and the 30% data is passed as a testing data to the trained model to observe and track its performance. The performance parameters were observed to get accuracy and kappa.

      Within the decision tree depth parameter is set to 10, as we are limiting the number of nodes in a branch to 10. This is giving us more appropriate results by saving cost as compared to other integer values. However, the pruning parameter is set to 0.1 to get a good fit of training data and testing data as well. If the prune value increases, the tree size decreases and may give us overfitting model which will not be good for testing data.

      Figure 13: Decision tree model

5. EVALUATE AND IMPROVE THE MODEL(S) IN RAPIDMINER

   As it can be analyzed from the table below that decision tree is the best fit as it has comparatively low accuracy and kappa, but it gives more accuracy on validation data. Moreover, K- NN is a lazy classification technique which learnings all its training data. The accuracy is achieved more using cross validation (with number of folds=10).

   Table 1: Comparison table

   Even the accuracy and kappa are increased with the use of cross validation along with it specificity has increased exponentially which means that the data is less biased.

   Cross validation model sub-process are applied as:

   Figure 14: Cross validation model

   The precision of the model(one of the other most important validation factor) increases with Cross Validation, this can be analyzed from the confusion matrix as follows:

   Figure 15: Decision tree with 70:30 split -Confusion matrix

   As it can be analyzed using a confusion matrix of decision tree the predicted no (i.e. book table service is not provided) was 2983 cases and prediction yes (service provided) was 3235 cases when true positive. But, after cross validation values are 15169 and 15182 which are significantly higher. Thus, it is an upgrade performance over already existing models.

   Figure 16: Cross Validation with decision tree -Confusion Matrix

   ROC is a plot which provides insights about stability and feasibility of the model (BRADLEY, 1997)[1]. Higher lift and more area under the curve, better the model.

   AUC value close to 1 is considered as a good model. Accuracy of the model is low as compared to other still it is a better one because we should not solely focused on gaining high accuracy while checking the performance evaluation table. Performance evaluation other major factors like class

   precision, class recall and kappa should also be taken into consideration.

   Hence, the cross validation decision tree is best model on considering all these statistics and also after analysing the Below ROC plot:

   Figure 17: ROC curve and AUC

   Figure 18: Performance of cross validation decision tree.

6. DEPLOYMENT IN RAPIDMINER

   For deploying the model and validating it in the Zomato score data, following steps needs to be implemented in RapidMiner:

   Use read CSV operator in a new process. By using import configuration wizard, read the csv file that needs to be scored.

   Check missing values in data and ensure that data doesnt contain any missing values.

   Apply all the pre-requisites applied on the training data.

   For applying the entire constructed model, use apply model operator.

   Use write model operator in the process to back the scored data.

   Connections: Connect the Read CSV to input of Apply Model operator.

   Connect apply model to write model.

   Run the model.

   Figure 19: Honest testing using test data.

   The honest testing is working good on build model as it is giving us 85.33% accuracy and 0.667 kappa.

   Figure 20: Performance Vector

7. CONCLUSION

As asked from BFA end that what strategy must be implemented by new restaurants to provide online order and table booking services.

To do so, we have analyzed the established restaurants whose votes are more significant than the threshold (i.e. 24).Have

even considered attributes such as address, name, phone, cuisines and discarded customer review attributes because, it will help in identifying the restaurants which are well established and provide good customer satisfactory services such as online order and book table. Furthermore, it will also provide insights about strategic approach towards implementing these services thus giving a competitive edge over the rivals. This strategy is beneficial and can be achieved only by new restaurants which have crossed the threshold vote parameter bar. As these services are more aligned towards customer satisfaction, the new restaurants must only take these services when they have established themselves in the industry and are ready to compete with the old restaurants.

Another BFAs query was regarding the strategy for already established restaurants:

By analyzing the data, the classification of date is being done based on the customer ratings. Restaurants which have more average rating are doing good in their business as compared to restaurants with low rating. The Lower rated restaurants provide basic online order or book table services but the services are not up to the mark, as they are facing late delivery or wrong food parcel getting submitted. Hence, restaurants with less average rating should focus more on solving the above problem while providing online delivery and book table services as they will built a good customer review platform and can compete with their competitors in terms of more valued services. This strategy is beneficial only for established restaurants who havent marked the br of good ratings and are willing to expand their business.

REFERENCE

1. BRADLEY, A. E., 1997. The use of the area under the ROC curve in the evaluation of machine learning algorithms. Pergamon, Volume 30.

2. Hoffman, K., Kelley, S. and Rotalsky, H. (1995), "Tracking service failures and employee recovery efforts", Journal of Services Marketing,

   Vol. 9 No. 2, pp. 49-61

3. Syed Saad Andaleeb and Carolyn ConwaySam and Irene Black School of Business, Penn State Erie, The Behrend College, Erie, Pennsylvania, USA Customer satisfaction in the restaurantindustry: an examination of thetransaction-specic model.