A Novel Intelligence Recommendation Model for Insurance Products with Consumer Segmentation

1 Introduction

Financial product recommendations were financial marketing problems, the purpose is to provide products and services met the personalized needs according to the characteristics of financial customers. Traditional financial marketing used cross-selling concepts and methods. Starting in the 1990s, the domestic and foreign scholars have made certain definitions to crossselling. Jarrar and Neely[1] believed that cross-selling is a new way of marketing, CRM is used to find a variety of purchase requirements of current customers, then selling a wide range of products and services for them accordingly. Kamakura[2] put forward that cross-selling was referred to the efforts to make customers buy products at the same company. Because the company has established relationships with the customer, cross-selling products and services to the cost of the current customers are much lower than to find a new customer. Important cross-selling forecast model with latent trait model[3], acquisition model[4], survival analysis[5] and NPTB model[6]. Basic data for the customer of this kind of model makes the demographic characteristics and behavior characteristics, uses the method of statistics or econometrics model, the results can be achieved certain cross-selling opportunities.

But there are massive amounts of customer data and information in modern context, such cross-sell products recommended results from research are a bit too simple, and not easy to large scale operations. Personalized recommendation methods and recommended system, established a link between the user and the product information of digging discoveries have been made by each user’s potential interest in, so as to implement product recommendations. Personalized recommendation attracted much attention as an independent field of study since the 1990’s[⁷]. “Collaborative filtering” [8] was first invented and applied in 1992 by researchers. Then 1997 ACM international conference put forward a automated collaborative filtering system called Grouplens[9], can realize the recommendation of the movie and news. In 1997 e-commerce recommendation system was defined by Resnick and Varian[10], then it was being widely used. Later, recommendation system theory and application research has been further deepened at home and abroad. Zhao et al.[11] put forward to increase two process of the personalized recommendation algorithm. Ronald Yager et al. studied the rough set theory to recommend improvements of the system rules. Liu[12] added customer life cycle theory to the study of product recommendations.

The financial sector has studied the application of personalized product recommendations. As early as 1996, Fletcher[13] and others studied Britain’s major Banks marketing system and came to the conclusion that IT technology be applied into the marketing strategy to keep market advantage of the bank. Scholars study in our country, bank and insurance are both important industry of application products recommend. Li et al.[14] using collaborative filtering to construct a system of bank product recommendations. Wang and Wu[15] studied the application of collaborative filtering in cross-selling, using collaborative filtering for a set of simulated customer purchase information to forecast. Wang et al.[16] studied the application of clustering in implementation of cross selling life insurance.

This paper makes recommendation of insurance products using known customer buying behavior, having two major aspects of the different to previous research results. First of all, this paper selects the association rules as a way of product recommendation. Because similar collaborative filtering recommendation according to the user to complete the personalization recommendation, need to know the user’s purchasing behavior and the assessment for each product, the initial recommendation effect is poorer. As for the insurance products, we can directly get the customer’s purchasing behavior information, but could not get the product ratings of customers. Secondly, set up the model of product recommendations based on customer segmentation, determining the classification of the customer first, and then we will recommend products to customers. At the same time of lower association rules analysis cost, recommend more targeted products for customers and improve the accuracy of the recommendation for insurance products.

2 Theoretical backgrounds

3 The proposed method

In this part, this paper proposes a recommendation model for insurance products based on consumer segmentation. Firstly, we need to do the data preprocessing to the daily transaction records of insurance company. Then, the clustering analysis is used to determine the classification characteristics of the policyholders. Next, customer segmentation model is constructed using Neural Network, analyzing the influence degree of factors on the classification results made by policyholder personal information and vehicle information etc. Then, we use Apriori algorithm to obtain association rules and the propensity to buy of consumers. The product recommendation model is built. Finally, the precision, recall rate and F values which are widely used in the search engine evaluation are adopted to evaluate the performance of recommendation. We compare the product recommendation based on customer segmentation with general product recommendation, and come to the conclusion. The framework of the proposed method is illustrated in Figure 1, and the details will be discussed in the following subsections.

Figure 1

The framework of the proposed method

4 Empirical analysis

First of all, the experiments use SimpleKMeans and FarthestFirst which are the clustering algorithms in the WEKA software. The results of cluster analysis are listed in Table 2.

We can get specific distribution of four types of customers from the scatter diagram. With reference to the customer value theory, we can know that the result of K-means algorithm is more in line with the actual situation. So we use the result of K-means algorithm as the result of the customer classification, and do further research on the influencing factors of classification. The result of K-means clustering is shown in Figure 2.

Figure 2

Result of K-means clustering

Policyholders can be divided into four categories according to premium and the number of claims. This accords with customer value matrix. Policyholders with high premium and less claims, who are the most valuable, onl.y account for 1% of the total number of customers. More than 70% of the customers are regular customers with low premium.

In the process of classification to the customer, we use some algorithms such as Back- propagation Neural Network, Multiple logistic regression, C4.5 of decision tree, support vector classification by software WEKA. At the same time, some algorithms such as Price sensitive function, Oversampling, Sub-sampling, Bagging, Boosting are used to dispose the unbalanced data. The experimental results are shown in Table 3 and Figure 3.

Figure 3

The comparison of classification performance

Figure 4

The comparison of performance between two recommendation model

From the perspective of the overall accuracy of classification, the accuracy of the Neural Network classification is above 70%, which is better than the other algorithms. After using some algorithms of unbalanced data processing, Price sensitive function and Bagging have a good effect.

Through the product recommendation model based on customer segmentation, the number of valuable rules can be seen in the Table 4.

After the valuable rules are applied in the test sets, some evaluation indicators are shown in Table 5.

As can be seen from Table 5, first of all, under the condition of the same confidence and support the more rules means the more similar purchase characteristics these customers of the same class have. So it is more convenient for insurance companies to make product recommendations. Secondly, when recall value is high, the recommendation result contains the more hidden products. The product recommendation result of Class I customer is generally superior to other classes, and only the recommendation effect of Class IV is not ideal. Customers of Class IV who make biggest contribution to the insurance company are the most valuable customer resources, however, the recommendation result of this class is not ideal. To investigate its reason, on the one hand the quantity of Class I customer is minimum, which accounts for only about 1% of all the customers. On the other hand, Such customers’ purchase characteristics may be to purchase a large amount of product portfolios.

To contrast, for general product recommendation model, it is need to do association analysis of all the customer purchase information directly. The methods and related parameters setting are as the same as the recommendation model based on Consumer Segmentation. Table 6 describes the contrast between two models in terms of data size.

The experimental results are in line with the expectations of our study, product recommendation based on customer segmentation can obtain more accurate results. Recall rate is increased by 4.6%, accuracy is increased by 6.3% and F value is increased by 5.5%. This shows that the model based on customer segmentation make insurance product recommendation more targeted. targeted.

Compare the different categories of customers, product recommendation for customers except Class IV get more accurate results. The reasons for this result are discussed above. Although the recommendation result of Class IV is not ideal, but the insurance company still need to spend a lot of power and resources to maintain the good relationship with the customers of Class IV, to avoid the loss of such class of customers. the loss of such class of customers.

Figure 5

The comparison of performance between four classes of customers

5 Conclusion

This paper proposes a recommendation model for insurance products based on consumer segmentation. On the foundation of customer segmentation model constructed by the clustering analysis and Neural Network, we use Apriori algorithm to build the product recommendation model. Through the experiments on real policyholder data by the insurance company, we mainly obtain the following conclusions. Firstly, the insurance premiums and claims can clearly reflect the value of policyholders when constructing the consumer segmentation model. Policyholders with high premium and less claims, who are the most valuable, only account for 1% of the total number of customers. More than 70% of the customers are regular customers with low premium. Secondly, the information of policyholders from insurance company can be divided into two types, personal insurance information and vehicle information. We use these two types of information to do the classification forecast in the experiments, and the results we obtain are good. After using some algorithms of unbalanced data processing, Price sensitive function and Bagging have a good effect. Finally, through the comparison in the experiments product recommendation based on customer segmentation can obtain more accurate results. This shows that the model based on customer segmentation not only makes the consumption of association rules analysis reduced, it also makes insurance product recommendation more targeted.

In addition, there are also some problems to do further discussion. Firstly, due to the limitation of data size, this paper chooses the premiums and claims on behalf of the customer’s current value and potential value in process of constructing customer segmentation model. To a certain extent, it affects the accuracy of the classification model. So in future studies we should attempt to add more segmentation indicators to distinguish different categories of customers more effectively. Secondly, the method of products recommendation this paper propose only use the classic Apriori algorithm, and what we need to do is to improve it according to the actual situation of this study. So that it can lead the recommendation model has better recommendation results.

Finally, in the experiments of customer classification, we use the default parameters in the WEKA when apply Back-propagation Neural Network, Multiple logistic regression, C4.5 of decision tree, support vector classification. Future research will focus on looking for the standard procedures or methods to make the parameters of model achieve the overall optimization.