Research on design of customer portrait system for E-commerce

Lingjie Cai; Jun Li

Detalles de publicación y publicación

Applied Mathematics and Nonlinear Sciences

Detalles de la revista

Formato: Revista
eISSN: 2444-8656
Primera edición: 01 Jan 2016
Calendario de la edición: 2 veces al año
Idiomas: Inglés

E-commerce often refers to commodity trading activities on the Internet, including e-commerce platforms, customers, merchants, advertisers and so on. Domestic e-commerce platforms are represented by Taobao, JD.COM, Suning.cn, Dangdang and other websites with their own characteristics. Global e-commerce platforms are represented by Amazon and eBay. In recent years, cross-border e-commerce such as Koala overseas purchase has been constantly emerging. The number of online shopping customers in China reached 812 million in the first half of 2021, an increase of 29.65 million compared with the end of 2020, of which the number of people purchasing mobile phones was about 650 million, accounting for 80% of all customers (as indicated by the statistics in August 2021); this situation is attributed to the improvement of the social income level and the continuous upgrading of consumption. GMV, an e-commerce website, has also been growing continuously, reaching a 12% increase over 20 years. E-commerce websites first provide customers with shopping services. Then, customers will generate static information data and dynamic behavior data, and then the e-commerce platform will analyze the customer data by portrait, which will not only enhance their experience, but also grasp its own operation and increase the visits and revenue, thus forming a closed loop where the customer portrait is an extremely important part [1, 2].

At present, most methods for building customer portraits rely on a single machine learning algorithm [3–6]. Due to the limitations of a single algorithm, it is not possible to further explore the hidden information among data features, which makes the traditional learning methods inadequate in many cases, such as in predicting customer demographic attribute tags according to customer network behaviour data [7,8]. After pre-processing, the features of customer network behaviour data are often sparse and high-dimensional, which makes it difficult for a single learning method to fully discern the relationship between features, and its prediction effect is usually mediocre.

The deep internal relations among multi-dimensional features are explored by studying the method of feature extraction; low-dimensional features and representative features are extracted by using various machine learning algorithms; and the framework of the prediction algorithm in customer portrait is designed by adopting ensemble learning methods with different strategies, so as to improve the prediction effect of customer portrait labels. Automatically extracting customer features through an algorithm model reduces the time and cost involved in manually processing features, and greatly improves the efficiency of feature processing. Realising the efficient and accurate prediction of customer portrait labels can improve the enterprise recommendation system and boost the marketing revenue, and at the same time, advances customer’s satisfaction and loyalty to the enterprise’s services. To sum up, it is believed that the study of the algorithm of customer portrait based on the data of customer network behaviour is of great practical significance and application value.

Overview of the customer portrait system

2.1

Customer portrait technology

According to the different needs, the multi-dimensional collection and analysis of massive data by customer portrait technology are also different [9–11]. Through the integrated analysis of customer portrait technology, the data are labelled, and the system users can have a clear and intuitive knowledge of the various dimensional characteristics of the customers. Customer portrait technology is more than simply mining and analysing customer data since studying customer behaviour is different from traditional market research, and can largely avoid the interference of subjective factors in the process of market research, allowing the data to speak, be closer to the actual situation of customers and improve the accuracy of investigation. In addition, customer portrait technology can narrow the distance between enterprises and customers, instead of leaving enterprises in a simple state of understanding the customers [12–14]. Through corporate customer portrait technology, the market positioning and the response speed to the market can be upgraded to a higher level [15, 16].

Customer portrait system can effectively help marketers find market changes, gain feedback from the market situation in time, adjust product strategies, continuously increase product competitiveness and maintain a strong momentum [17, 18]. Because the transform of products must be caused by the change of market, with the evolution of product life cycle and the change of market demand, the customer’s family structure, asset status and demand level will change, so the product needs to be adjusted in time according to the market and customer situation, otherwise it will eventually decline and fall. Moreover, compared with the traditional marketing model, the results of customer analysis have changed from static to dynamic, and the working mechanism of marketers has changed from post-event remedy to pre-warning, which has greatly improved the work efficiency [19–21].

2.2

Data mining technology

The amount of data to be processed in data mining is often large. In order to mine valuable information from this low-density and noisy information, certain mining algorithms and tools must be used to filter the source data accordingly and clean it into a unified format [22]. Then, data mining and analysis must be conducted, and thereafter a judgement should be made according to the characteristics of the mined data. There are many tools and algorithms for data mining, while the processing methods are different. The main process of data mining is shown in Figure 1.

2.2.1

Data acquisition

This stage is mainly to acquire data sets. Some data sets are ready-made and can be downloaded directly, while others need to be actively collected. This stage determines the scope and quality of the data set that the project will eventually study. The quality of data will directly affect the quality of mining and the final result.

2.2.2

Data collation

Since the acquired data are incomplete, random and noisy with high probability, this stage is particularly important, and it is also thus necessary to carry out preliminary data processing and normalise the data. It mainly includes three steps: data cleaning, data conversion and data integration.

2.2.3

Model establishment

At this stage, it is necessary to select the appropriate data mining model according to the data characteristics and optimise it according to the project characteristics. The selection of an appropriate model can achieve a better mining effect. At the same time, after selection of a suitable model, it may be found that the pre-processed data do not meet the requirements of the model, necessitating a return to the previous stage.

2.2.4

Analysis and summary

At this stage, the results of data mining are analysed to assess whether the expected goals and tasks have been completed, and whether there is any room for improvement; this would enable the application of the results of data mining in solving practical problems.

2.3

Cluster analysis

After understanding the process of data mining, it is necessary to choose the appropriate data mining methods to model the data according to the problems to be solved. There are many methods of data mining that are suitable for specific problems. Generally, several methods and models need to be tried, so as to select a relatively better model. Clustering is a very common choice in data mining.

The purpose of clustering analysis is to group together data of the same kind. There is little difference between the same kind of data and huge difference between classes, and clustering analysis does not need to know the structure of clustering objects in advance. It classifies the similarity between data variables, and in the process of classification, constructs a symmetric similarity matrix, which is usually a data matrix or a difference matrix. In this paper, the data matrix is used as the data structure of clustering analysis.

The data matrix is a structure with a combination of object and attribute. For example, if there is a customer with j attributes, then the data matrix is expressed by an i*j matrix, as shown in Eq. (1).

1 $\begin{aligned} (\begin{matrix} x_{11} & x_{12} & \dots \dots & x_{1 j} \\ \dots \dots & \dots \dots & \dots \dots & \dots \dots \\ x_{i 1} & x_{i 2} & \dots \dots & x_{i j} \end{matrix}) \end{aligned}$

The basis of clustering analysis is the similarity between objects, and whether objects are similar or not can be abstracted by ‘distance’. The smaller the distance between two objects, the more similar the two objects are; otherwise, the greater the difference between two objects, the greater the distance between them. For example, each object has one attribute, and x and y are the calculated values of two objects, respectively. So, x and y can be identified by vectors, as shown in Eq. (2).

2 $\begin{aligned} {\begin{cases} x = (x_{1}, x_{2}, \dots \dots, x_{j}) \\ y = (y_{1}, y_{2}, \dots \dots, y_{j}) \end{cases} \end{aligned}$

In this paper, the similarity between two individuals is defined by quoting the Euclidean distance, including the distance calculation of the subsequent customer index system, which adopts this method uniformly. The formula of Euclidean distance is shown in Eq. (3).

3 $\begin{aligned} d (x, y) = \sqrt{\sum_{i = 1}^{j} {(x_{i} - y_{i})}^{2}} \end{aligned}$

With the deepening of research in recent years, clustering algorithms are gradually being divided into five categories, namely partition method, hierarchical method, density-based method, grid-based method and model-based method. However, actually the data mining algorithm not only depends on the which data type to choose, Whether the type is applicable or not can also be the criterion of selection, and the users of the algorithm often combine the ideas of various clustering methods, so it is impossible to completely divide the adopted algorithm into a certain clustering method. Among the many clustering algorithms, the FCM algorithm is one of the most commonly used.

Design of customer portrait system

3.1

Overall scheme design of customer portrait system

According to the method of establishing the customer portrait system, this paper uses the data from the business analysis system of an e-commerce company [23–25]. First, we investigate the sample data needed for modelling according to the requirements of the portrait system [26]. We then process and clean the obtained customer data, and establish the data model by analysing and mining the customer behaviour data, in order to establish the customer portrait label and store the generated tag data [27, 28]. Finally, the customer portrait system is used to support the marketing of operators. The process is shown in Figure 2.

3.2

Customer portrait functional architecture

The basic functional framework of the customer portrait system consists of data collection layer, data model label layer and data basic management layer [29, 30]. The functional framework of the customer portrait is shown in Figure 3:

Functional architecture of the customer tag library

Among them, the data collection layer realises data pre-processing, customer log processing, text parsing processing, etc., and provides basic data for the preliminary processing of the customer portrait system. The tag layer mainly contains basic attributes, domain attributes, social attributes and purchase attributes of the customers [31]. The basic data management layer includes data storage of customer portrait, metadata management of the system, label life cycle management, query mechanism and update mechanism of the customer portrait system [32, 33].

Construction of customer portrait model based on the FCM algorithm

With the individualisation of the customers’ demand for products and the increase of customers, a large number of customer portraits will accumulate in the personalised recommendation system with different demands, which can cause a certain interference and affect the accuracy of system’s recommendation. Based on this, it is necessary to divide the related portraits, and cluster the portraits of value among the customer groups, so as to find customers with similar degree in the group and divide them, and recommend personalised information services according to different customer groups. Combined with the design of the system described above, this paper mainly divides the customer portrait indicators into four dimensions: basic attribute dimension, domain dimension, social acceptability dimension and purchasing power dimension. The basic attributes include e-commerce customer id, contact information, receiving address, purchase amount, etc. Dimensions include the areas where customers buy products or services, such as tourism, food, medicine, clothing, furniture, etc. Social acceptability of customers includes the usage and frequency of social software. Purchasing power is mainly determined according to the disposable income of customers and their willingness to consume.

4.1

Clustering of customer portrait model based on the FCM algorithm

The construction of customer portraits belonging to multiple categories in the same dimension, and the interests of many customers in the recommendation system, cross and overlap with each other. From the attribute characteristics, customers who are in the same attribute have the same membership degree. From another attribute characteristic, customers in this membership degree represent another category of customers. In this paper, the Fuzzy C-means (FCM) algorithm based on partitioned clustering is selected. Given the data set M = (M1, M2, …, Mi), the data set M is divided into c fuzzy classes, and the clustering centre $c_{i} (i = 1, 2, \dots m)$ of each fuzzy class is obtained by iteratively updating the membership degree, in which each sample j belongs to a fuzzy class i with a membership degree of $μ_{i j}$ .

Supposing that the customer portrait set is $U = {u_{1}, u_{2}, \dots, u_{m}, u_{m}}$ , in which each portrait u has n attributes, namely ${u_{i 1}, u_{i 2}, \dots, u_{i m}}$ , the corresponding attribute matrix is then given by: 4 $\begin{aligned} U^{*} = (\begin{matrix} u_{11} & \dots & u_{1 n} \\ ⋮ & ⋱ & ⋮ \\ u_{m 1} & \dots & u_{m n} \end{matrix}) \end{aligned}$

Through continuous iteration, the best cluster of the fuzzy classification matrix and the best matrix cluster of cluster centre are obtained, in which the values of the principal membership matrix are shown in Eq. (5): 5 $\begin{aligned} \sum_{i = 1}^{k} μ_{j i} = 1, (j = 1, 2, \dots, N); \sum_{i = 1}^{K} \sum_{j = 1}^{N} μ_{j i} = N \end{aligned}$ where $μ_{j i} \in [0, 1]$ to minimise the objective function. Therefore, the definition of an objective function and its constraint is as follows: 6 $\begin{aligned} J (U, V) = \sum_{i = 1}^{c} \sum_{j = 1}^{n} μ_{i j}^{m} x_{j} - {c_{i}}^{2} \end{aligned}$

To divide the collection of the customer portrait into k categories (2 < k < n), it is necessary to update the membership degree and the clustering centre iteratively, so as to obtain the minimisation objective function, the corresponding fuzzy classification matrix U and the clustering centre matrix W. We combine these formulas to obtain Eq. (7), and derivate the variables μ_ji and c_i, respectively, thus obtaining Eq. (7): 7 $\begin{aligned} J (U, V) & =\sum_{i = 1}^{c}\sum_{j = 1}^{n}μ_{i j}^{m}x_{i}-{c_{i}}^{2}+λ_{1}(\sum_{i = 1}^{c} μ_{i 1} - 1)+\dotsλ_{n}(\sum_{i = n}^{c} μ_{i n} - 1) \\ μ_{i j} & ={(\sum_{k = 1}^{c} {(\frac{x_{j} - c_{i}}{x_{j} - c_{k}})}^{\frac{2}{m - 1}})}^{- 1} \\ c_{i} & =\frac{\sum_{j = 1}^{n} (μ_{i j}^{m} x_{i})}{\sum_{j = 1}^{n} (μ_{i j}^{m})} \end{aligned}$ where J(U,V) represents the sum of squares of all customer portraits to the cluster centre of the class to which they belong. The vector c_i represents the cluster centre of ith class is an N-dimensional vector. ${‖ x_{j} - c_{i} ‖}^{2}$ represents the distance from customer portrait u_j to the cluster centre v_i. $\sum_{i = 1}^{c} \sum_{j = 1}^{n} μ_{i j}^{m} {‖ x_{j} - c_{i} ‖}^{2}$ represents the sum of squares of the weighted distances between the customer portrait u_j and each cluster centre C_i, and the weight is the membership degree of u_j that is subordinate to the ith class: the parameter m can change the relative membership degree of the customer portrait for strengthening the contrast of u_j in various membership degrees.

The best classification method in the FCM algorithm can be solved through the iterative operation and the infinite nature of the classification matrix U. When $u_{i} \neq u_{i} u_{i}$ , the FCM algorithm process can be assumed to be convergent by default, and the optimal classification centre matrix can be obtained according to the convergence.

The iterative steps are as follows:

Step 1:

Select k to be classified, and select 2 ≤ k ≤ m: Take the initial fuzzy classification matrix U⁰:

Step 2:

For U^(τ), calculate the cluster centre matrix.

8 $\begin{aligned} V^{τ} = {(V_{1}^{(τ)}, V_{2}^{(τ)}, \dots, V_{K}^{(τ)})}^{T} \end{aligned}$

Among them: 9 $\begin{aligned} V_{i}^{(τ)} = \sum_{j = 1}^{m} {(r_{i j}^{τ})}^{q} u_{j} / \sum_{j = 1}^{m} {(r_{i j}^{τ})}^{q} \end{aligned}$

Step 3:

Make auxiliary correction of the fuzzy classification matrix U in Step 1: 10 $\begin{aligned} μ_{i j}^{(τ + 1)} = {[\sum_{c = 1}^{k} {(\frac{u_{j} - v_{i}^{(τ)}}{u_{j} - v_{c}^{τ}})}^{\frac{2}{q - 1}}]}^{- 1} \end{aligned}$

where j = 1,2,…,m;c = 1,2,…,k.

Step 4:

For U and R^(τ+1), if the accuracy of determination is > 0, then: 11 $\begin{aligned} max {| μ_{i j}^{(τ + 1)} - μ_{i j}^{(τ)} |} \leq ε \end{aligned}$

If U^τ+1 and V^(τ) are the result of the calculation, stop iteration; otherwise τ = τ + 1, and return to Step 2. Take the obtained classification matrix and central matrix to divide the customer portrait. The solution methods of fuzzy classification cluster centre matrix and optimal cluster centre are as follows:

Fuzzy classification clustering centre matrix

Let the obtained optimal fuzzy classification matrix be: 12 $\begin{aligned} U^{*} = (\begin{matrix} μ_{11}^{*} & \dots & μ_{1 n}^{*} \\ ⋮ & ⋱ & ⋮ \\ μ_{k 1}^{*} & \dots & μ_{k n}^{*} \end{matrix}) \end{aligned}$ where $\forall u_{v} \in U$ , in column V, if 13 $\begin{aligned} μ_{i v}^{*} = max_{1 \leq j \leq k} (r_{j v}^{*}) \end{aligned}$

Then the object u_v belongs to category i, and is used to divide the group customer portrait.

The best clustering centre matrix clustering

Let the obtained optimal cluster centre matrix be: 14 $\begin{aligned} V^{*} = (\begin{matrix} V_{1}^{*} \\ V_{2}^{*} \\ ⋮ \\ V_{k}^{*} \end{matrix}) = (\begin{matrix} v_{11}^{*} & v_{12}^{*} & \dots & v_{1 n}^{*} \\ v_{21}^{*} & v_{22}^{*} & \dots & v_{2 n}^{*} \\ ⋮ & ⋮ & ⋮ \\ v_{k 1}^{*} & v_{k 1}^{*} & \dots & v_{k n}^{*} \end{matrix}) \end{aligned}$ where $\forall u_{v} \in U$ , if 15 $\begin{aligned} ‖ u_{v} - V_{l}^{*} ‖ = min_{1 \leq j \leq K} (‖ u_{V} - V_{j}^{*} ‖) \end{aligned}$

Then the object u_v is classified into the i class, and the group customer portrait is divided.

4.2

Improved FCM clustering algorithm

From the above algorithms, it can be seen that the FCM algorithm has the following defects: the initial clustering centre needs to be set, the sample data are unbalanced, the initial clustering centre is sensitive and the observation samples do not consider the influence of clustering results in the objective function. Therefore, in this paper, a method of adjusting the influence value of the density function is proposed, which is used to solve the problem that the clustering reliability of the FCM algorithm is reduced due to the sample base, thus improving the stability of the FCM algorithm and increasing the convergence of clustering centre.

Assuming that there are m sample points in the sample space data set, the Euclidian distance between µ_i and µ_j is represented by n_ij, $I_{i j} = ‖ μ_{i} - μ_{j} ‖$ , in which 1 ≤ i, j ≤ m; additionally, the density range value of sample I_ij < 1 T is assumed to be φ, and thus the density function centred on x_j can be expressed as: 16 $\begin{aligned} y_{j} = \sum_{i = 1}^{m_{j}} \frac{1}{I_{i j}}, i \neq j \end{aligned}$ where m_i must satisfy the number of sample points surrounding µ_j and sample points µ_j within the range of I_ij < 1. It can be concluded from the formula that within a certain range, the more points around, the more sample points surrounding µ_j, and the closer the relationship of µ_j, the larger the value of y_j. of the sample. The formula density function value ρ_j can be obtained by data principal component analysis and dimension reduction in y_j: 17 $\begin{aligned} ρ_{j} = y_{j} / \sum_{i = 1}^{m} y_{j} \end{aligned}$ where $ρ_{j}$ represents the degree of influence in classification about the surrounding sample of µ_j, when the larger the value of $ρ_{j}$ , the greater the degree of µ_j belonging to a certain category, and vice versa. By introducing sample influence value $ρ_{j}$ to improve the traditional FCM algorithm, the improved objective function is obtained. Similarly, the fuzzy classification matrix in the formula for $μ_{ij}$ and the cluster centre matrix c_i are derived by Lagrange definition, so that Eqs (18) can be obtained: 18 $\begin{aligned} J (U, C) & =\sum_{i = 1}^{c}\sum_{j = 1}^{n}ρ_{j}μ_{i j}^{m}x_{j}-{c_{i}}^{2} \\ μ_{i j} & ={(\sum_{k = 1}^{c} I_{i j} / I_{k j})}^{2 / (1 - m)} \\ c_{i} & =\frac{\sum_{j = 1}^{n} (ρ_{j} μ_{i j}^{m} x_{i})}{\sum_{j = 1}^{m} (ρ_{j} μ_{i j}^{m})} \end{aligned}$ where the value of $ρ_{j}$ is mainly determined by the density of distribution around its sample space, the influence value of the sample clustering centre in the formula is adjusted after the influence factor $ρ_{j}$ is introduced into the FCM algorithm and the membership degree in the sample data is indirectly adjusted because of the change in the iterative process. When $ρ_{j} = 1 / n$ is a fixed value, the influence on all samples in the sample space is consistent, which is the traditional FCM algorithm.

4.3

Analysis of experimental results

4.3.1

Selection of data set

In this paper, data from 120 million customers of Taobao are taken as the sample data, and based on the customer portrait index system, the original data are screened, cleaned and processed from the basic attribute dimension of e-commerce customers in the multi-level customer portrait model. Since the raw data are very important to e-commerce companies, the raw data will not be listed here, and a part of data is randomly selected and normalised according to the norm to obtain new data as shown in Table 1.

Table 1

Normalised data

	ρ1	ρ2	ρ3	ρ4
X1	0.38433	0.15642	0.38430	0.19573
X2	0.70499	0.15145	0.70497	0.18946
X3	0.19408	0.50566	0.19506	0.64245
X4	0.39654	0.30788	0.39696	0.37284
X5	0.10490	0.04545	0.10485	0.056646
X6	0.18035	0.04179	0.05144	0.18031

19 $\begin{aligned} ^{'} ρ_{i j} = 1 - \frac{ρ_{j max} - ρ_{i j}}{ρ_{j max} - ρ_{j min}} = \frac{ρ_{i j} - ρ_{j min}}{ρ_{j max} - ρ_{j min}} \end{aligned}$ where $^{'} ρ_{i j}$ refers to the value of customer x_i on attribute p_j, the maximum value is ρ_jmax, the minimum value is ρ_jmin and the normalised value is $^{'} ρ_{i j}$ .

4.3.2

Clustering results of the algorithm

The improved FCM algorithm is used to cluster and divide multi-level customer portraits. The weight coefficient is 2, the clustering data has been set to 4 and 10% of the experimental standard data set was randomly selected for analysis. Considering that the weight ratio of index ρ2 is only 0.1, the attribute characteristic value of e-commerce customers was reduced by three dimensions. They are the basic attribute dimension, purchasing power dimension and customer social dimension, which are represented by ρ1, ρ3 and ρ4, respectively, and the clustering effect is shown in Figure 4.

As can be seen from Figure 4, clustering is carried out according to the basic attribute dimension, purchasing power dimension and customer social dimension, and finally it is divided into four types of customer portrait clustering. Owing to the importance of raw data, it will not be listed and displayed here. According to Figure 4, the improved FCM algorithm proposed in this paper can perform better grouping in space, where the first to fourth categories of customers represent the four categories of key customers, customers to be developed, ordinary customers and worthless customers, respectively. The attribute types of customers mainly include customer loyalty, average profit ratio and profit growth rate, and the classification results as shown in Table 2 can be obtained.

Table 2

Results after clustering

Cluster centre	Proportion of customers (%)	Proportion of profit (%)	Profit growth rate (%)	Loyalty	Customer type
Class I	17.22	14.24	295.61	0.601	Key development
Class II	25.17	69.77	48.8	0.295	Key maintenance
Class III	38.01	10.90	98.59	0.223	Ordinary customers
Class IV	19.57	5.09	–49.87	0.035	Worthless

Table 3

Matrix table of evaluation criteria

Actual/predictive	Negative number	K constant
Negative number	m	n
K constant	i	j

According to the results after clustering in Table 1, it can be concluded that the customers with a large degree of membership with cluster centre v2 account for 25.17% of the total customers, and the sum of profits created is 69.77%. These e-commerce customers belong to important e-commerce enterprises. For such e-commerce customers, enterprises should not only meet their basic needs but also provide personalised value-added services according to various individual customer needs, strengthen the daily maintenance between ordinary and e-commerce customers, and improve the loyalty of e-commerce customers. Although the profit share of the key development customers in Table 1 is not very high, the profit growth rate has reached 295.b1%, with high loyalty and great potential value for enterprises. For this kind of e-commerce customers, enterprises should deeply understand their development needs and provide timely support and services. Ordinary customers account for the highest proportion of customers, reaching 38.44% of the total number. The profit growth rate of these customers is lower than that of the key development customers, but higher than that of the other types of customers, and they are regarded as high-quality customers of enterprises. For such customers, companies should do personalised marketing on providing convenient and professional services. For worthless customers, companies should deeply understand the reasons for their choices and turn them into ordinary customers through various forms of personalised customisation.

4.3.3

Evaluation of the improved FCM algorithm

In this experiment, the improved FCM clustering algorithm is used to divide the normalised data into multi-level customer portrait groups, and the accuracy of the recommendation results is measured by calculating the difference between the predicted scores and the actual evaluation data of the target customers. Mean absolute error (MAE) is mainly used to evaluate the accuracy of the recommendation algorithm. When the MAE value is small, the recommendation accuracy of the algorithm is proved to be high, and vice versa. Let the customer forecast score set be expressed as M = {m_i, i = 1,2, ⋯n}; the actual score set in the real data set is N = {n_i, i = 1,2, ⋯n}, where n is the number of items between target customers; then, the MAE calculation formula is as follows: 20 $\begin{aligned} M A E = \frac{1}{N} \sum_{i = 1}^{n} | m_{i} - n_{i} | \end{aligned}$

In order to verify the accuracy of the algorithm’s recommendation in this paper, indices of precision and recall are introduced for comparison.

21 $\begin{aligned} p r e c i s i o n & =\frac{j}{p + j} \\ =\frac{j}{i + j} \end{aligned}$

In the above three evaluation indexes, the higher the MAE value, the worse the recommendation quality. The smaller the accuracy and recall rate, the worse the recommendation quality, and on the contrary, the better the recommendation quality.

On the data set, the improved algorithm in this paper is compared with other algorithms. The results of MAE, accuracy and recall are shown in Table 4.

Table 4

Comparison of algorithm performance

	Precision	Recall	MAE
K-Means	0.61	0.35	0.95
Improvement K-Means	0.64	0.34	0.91
FCM	0.66	0.36	0.89
Improvement FCM	0.71	0.32	0.82

MAE, mean absolute error

From the above table, we can see that the improved FCM clustering algorithm and traditional K-Means; improved K-Means and traditional FCM algorithm have greatly improved the accuracy and recall rate of group customer image classification, and the MAE value has decreased compared with traditional K-Means, improved K-Means and traditional FCM, which greatly improved the accuracy of classification in the e-commerce customer portrait.

Conclusion

In this paper, massive e-commerce sales data are classified according to their characteristics as needed, and four dimensions of customer portrait indicators are finalised. Then, the appropriate algorithm is selected to build the customer portrait system model. Aiming at the problems in the FCM algorithm, such as sensitive initial clustering centre, little consideration about the influence of clustering results in objective function, etc., an improved FCM algorithm with multiple adjustment density function influence values is proposed. Finally, the improved FCM algorithm is compared with the traditional FCM algorithm, traditional K-Means algorithm and improved K-Means algorithm. Based on applied data sets, three performance indexes, MAE, accuracy and recall rate, are selected. The results show that the improved algorithm has excellent performance, which is applied to study the customer portrait system. It can effectively classify e-commerce customers and help e-commerce enterprises to formulate more accurate customer marketing strategies, thus promoting the development of e-commerce enterprises.

Figuras y tablas

Comparison of algorithm performance

	Precision	Recall	MAE
K-Means	0.61	0.35	0.95
Improvement K-Means	0.64	0.34	0.91
FCM	0.66	0.36	0.89
Improvement FCM	0.71	0.32	0.82

Normalised data

	ρ1	ρ2	ρ3	ρ4
X1	0.38433	0.15642	0.38430	0.19573
X2	0.70499	0.15145	0.70497	0.18946
X3	0.19408	0.50566	0.19506	0.64245
X4	0.39654	0.30788	0.39696	0.37284
X5	0.10490	0.04545	0.10485	0.056646
X6	0.18035	0.04179	0.05144	0.18031

Results after clustering

Cluster centre	Proportion of customers (%)	Proportion of profit (%)	Profit growth rate (%)	Loyalty	Customer type
Class I	17.22	14.24	295.61	0.601	Key development
Class II	25.17	69.77	48.8	0.295	Key maintenance
Class III	38.01	10.90	98.59	0.223	Ordinary customers
Class IV	19.57	5.09	–49.87	0.035	Worthless

Matrix table of evaluation criteria

Actual/predictive	Negative number	K constant
Negative number	m	n
K constant	i	j

Referencias

[1] Luo Y, Xiong H. Customer Portrait Analysis on Differential Density Clustering [J]. Hunan Electric Power, 2017. Luo Y, Xiong H. Customer Portrait Analysis on Differential Density Clustering [J]. Hunan Electric Power, 2017.Search in Google Scholar

[2] Ke Fangyuan, Zhong Yongjie, Hu ruiruirui. Research on the precision service of power supply business hall based on customer portrait [J]. Power demand side management, 2016, 018(0z1):109-110. Ke Fangyuan, Zhong Yongjie, Hu ruiruirui. Research on the precision service of power supply business hall based on customer portrait [J]. Power demand side management, 2016, 018(0z1):109-110.Search in Google Scholar

[3] Kriksciuniene D. Analytical methods for tracking customer portrait: intelligent interpretation of CRM philosophy. 2012. Kriksciuniene D. Analytical methods for tracking customer portrait: intelligent interpretation of CRM philosophy. 2012.Search in Google Scholar

[4] Pitner, Tomáš, Kriksciuniene D, Sakalauskas V. Tracking customer portrait by unsupervised classification techniques [J]. Transformations in Business and Economics, 2012, 11(3):167-189. Pitner, Tomáš, Kriksciuniene D, Sakalauskas V. Tracking customer portrait by unsupervised classification techniques [J]. Transformations in Business and Economics, 2012, 11(3):167-189.Search in Google Scholar

[5] Sun J, Zhao H, Mu S, et al. Purchasing Behavior Analysis Based on Customer’s Data Portrait Model [C]//2019 IEEE 43rd Annual Computer Software and Applications Conference (COMPSAC). IEEE, 2019. Sun J, Zhao H, Mu S, Purchasing Behavior Analysis Based on Customer’s Data Portrait Model [C]//2019 IEEE 43rd Annual Computer Software and Applications Conference (COMPSAC). IEEE, 2019.10.1109/COMPSAC.2019.00058Search in Google Scholar

[6] Mi Feng. Method for accurately building customer portrait of mobile communication data service:, 2010. Mi Feng. Method for accurately building customer portrait of mobile communication data service:, 2010.Search in Google Scholar

[7] Kiseleva E S, Yeryomin V V, Elenav. Y, et al. The psychological portrait as a tool to improve the subjective well-being of the client in the context of personal sales [J]. Shs Web of Conferences, 2016, 28:01122-. Kiseleva E S, Yeryomin V V, Elenav. Y, The psychological portrait as a tool to improve the subjective well-being of the client in the context of personal sales [J]. Shs Web of Conferences, 2016, 28:01122-.10.1051/shsconf/20162801122Search in Google Scholar

[8] Xuesheng, Lai, Lili, et al. Personalized Product Service Recommendation Based on User Portrait Mathematical Model [C]//2018. Xuesheng, Lai, Lili, Personalized Product Service Recommendation Based on User Portrait Mathematical Model [C]//2018.Search in Google Scholar

[9] Paterson K. Debit Rewards: A Farewell Portrait? [J]. U.s.banker, 2011. Paterson K. Debit Rewards: A Farewell Portrait? [J]. U.s.banker, 2011.Search in Google Scholar

[10] Qiu H, Ying T U, Ding Q. Construction of power customer portrait and its credit evaluation and electricity fee risk control based on tag library system [J]. Telecommunications Science. Qiu H, Ying T U, Ding Q. Construction of power customer portrait and its credit evaluation and electricity fee risk control based on tag library system [J]. Telecommunications Science.Search in Google Scholar

[11] Xiang-Qian Y U, Wang L X, Yun-Bing L I, et al. Application Research on the Customer’s Portrait of Electric Power Marketing Informatization [J]. Computing Technology and Automation, 2017. Xiang-Qian Y U, Wang L X, Yun-Bing L I, Application Research on the Customer’s Portrait of Electric Power Marketing Informatization [J]. Computing Technology and Automation, 2017.Search in Google Scholar

[12] Liang M A, Tao L T, Xie J K. Customer demands management based on the customer portraits [J]. Power Demand Side Management, 2016. Liang M A, Tao L T, Xie J K. Customer demands management based on the customer portraits [J]. Power Demand Side Management, 2016.Search in Google Scholar

[13] Yao L F, Li-Li H E. Model of Group User Portrait Based on Cloud Model Theory [J]. Computer Systems & Applications, 2018. Yao L F, Li-Li H E. Model of Group User Portrait Based on Cloud Model Theory [J]. Computer Systems & Applications, 2018.Search in Google Scholar

[14] Yue M, Feng H. Optimization and practice of requirement analysis based on prototype portrait in software development process [J]. Journal of Computational Methods in Sciences and Engineering, 2021(13):1-9. Yue M, Feng H. Optimization and practice of requirement analysis based on prototype portrait in software development process [J]. Journal of Computational Methods in Sciences and Engineering, 2021(13):1-9.10.3233/JCM-214973Search in Google Scholar

[15] Paterson, Ken. Debit Rewards: A Farewell Portrait? [J]. U.S. Banker, 2011. Paterson, Ken. Debit Rewards: A Farewell Portrait? [J]. U.S. Banker, 2011.Search in Google Scholar

[16] Xuesheng, Lai, Lili, et al. Personalized Product Service Recommendation Based on User Portrait Mathematical Model [C]//2018 International Symposium on Communication Engineering & Computer Science (CECS 2018). 0. Xuesheng, Lai, Lili, Personalized Product Service Recommendation Based on User Portrait Mathematical Model [C]//2018 International Symposium on Communication Engineering & Computer Science (CECS 2018). 0.Search in Google Scholar

[17] Liang X, Wang X. On the Hadoop-based User Portrait System of Digital Library [J]. Information Research, 2019. Liang X, Wang X. On the Hadoop-based User Portrait System of Digital Library [J]. Information Research, 2019.Search in Google Scholar

[18] En X U, Zhiwen Y U, He D U, et al. User Portrait System Based on Mobile Sensing Data [J]. Journal of Zhengzhou University (Natural Science Edition), 2019. En X U, Zhiwen Y U, He D U, User Portrait System Based on Mobile Sensing Data [J]. Journal of Zhengzhou University (Natural Science Edition), 2019.Search in Google Scholar

[19] Yan F, Bao Y, Yan C, et al. An optimization method of Voiceprint Recognition based on user portrait [C]//2018 2nd IEEE Conference on Energy Internet and Energy System Integration (EI2). IEEE, 2018. Yan F, Bao Y, Yan C, An optimization method of Voiceprint Recognition based on user portrait [C]//2018 2nd IEEE Conference on Energy Internet and Energy System Integration (EI2). IEEE, 2018.10.1109/EI2.2018.8582456Search in Google Scholar

[20] Qiu N, Yang X, Wang H. USER PORTRAIT REPRESENTATION LEARNING SYSTEM AND METHOD BASED ON DEEP NEURAL NETWORK:, WO2018000281A1 [P]. Qiu N, Yang X, Wang H. USER PORTRAIT REPRESENTATION LEARNING SYSTEM AND METHOD BASED ON DEEP NEURAL NETWORK:, WO2018000281A1 [P].Search in Google Scholar

[21] Min, Huang. Research on Recommendation Technology Based on User Portrait [C]//0. Min, Huang. Research on Recommendation Technology Based on User Portrait [C]//0.Search in Google Scholar

[22] Tkachev N S, Mishchenko P V, Zanekina D D. The System for Determining the Portrait of a User of Telecom Services [C]//International Scientific-Technical Conference on Actual Problems of Electronics Instrument Engineering. 0. Tkachev N S, Mishchenko P V, Zanekina D D. The System for Determining the Portrait of a User of Telecom Services [C]//International Scientific-Technical Conference on Actual Problems of Electronics Instrument Engineering. 0.Search in Google Scholar

[23] Wu T, Yang F, Zhang D, et al. Research on Recommendation system based on user portrait [C]//2020 IEEE International Conference on Artificial Intelligence and Information Systems (ICAIIS). IEEE, 2020. Wu T, Yang F, Zhang D, Research on Recommendation system based on user portrait [C]//2020 IEEE International Conference on Artificial Intelligence and Information Systems (ICAIIS). IEEE, 2020.10.1109/ICAIIS49377.2020.9194941Search in Google Scholar

[24] Gu H, Wang J, Wang Z, et al. Modeling of User Portrait Through Social Media [J]. 2019. Gu H, Wang J, Wang Z, Modeling of User Portrait Through Social Media [J]. 2019.10.1109/ICME.2018.8486595Search in Google Scholar

[25] Yuan Kai, Chen Jinlai, Bi shuchao, et al. User portrait based mobile terminal intelligent message pushing method, server and system:, 2015. Yuan Kai, Chen Jinlai, Bi shuchao, User portrait based mobile terminal intelligent message pushing method, server and system:, 2015.Search in Google Scholar

[26] Qiu N, Wang H. USER PORTRAIT BASED SKILL PACKAGE RECOMMENDATION DEVICE AND METHOD [J]. 2017. Qiu N, Wang H. USER PORTRAIT BASED SKILL PACKAGE RECOMMENDATION DEVICE AND METHOD [J]. 2017.Search in Google Scholar

[27] Anderson E C, Dalke G W. Method and system for auto rotating a graphical user interface for managing portrait and landscape images in an image capture unit [J]. US, 2001. Anderson E C, Dalke G W. Method and system for auto rotating a graphical user interface for managing portrait and landscape images in an image capture unit [J]. US, 2001.Search in Google Scholar

[28] Liu C, Du J. A User Portrait Construction Method for Financial Credit [M]. Springer, Singapore, 2020. Liu C, Du J. A User Portrait Construction Method for Financial Credit [M]. Springer, Singapore, 2020.10.1007/978-981-15-8458-9_4Search in Google Scholar

[29] Miller A B. Epson’s Gemini Professional Portrait Printing System [J]. Studio photography & design, 2002, 5(3):p. 18-19. Miller A B. Epson’s Gemini Professional Portrait Printing System [J]. Studio photography & design, 2002, 5(3):p. 18-19.Search in Google Scholar

[30] Zhang L, Liu K, Li X Y, et al. Enable Portrait Privacy Protection in Photo Capturing and Sharing [J]. Computer ence, 2014, 16(1):33-35. Zhang L, Liu K, Li X Y, Enable Portrait Privacy Protection in Photo Capturing and Sharing [J]. Computer ence, 2014, 16(1):33-35.Search in Google Scholar

[31] Zhang J. Railway passenger user portrait system [J]. Railway Computer Application, 2018. Zhang J. Railway passenger user portrait system [J]. Railway Computer Application, 2018.Search in Google Scholar

[32] Yang, ZHAO, Lu, et al. Application and Research of User Portrait in Digital Library [C]//0. Yang, ZHAO, Lu, Application and Research of User Portrait in Digital Library [C]//0.Search in Google Scholar

[33] Zhang H. Research on Digital Library Precision Recommendation Service Innovation Based on User Portrait Analysis. Zhang H. Research on Digital Library Precision Recommendation Service Innovation Based on User Portrait Analysis.Search in Google Scholar

Applied Mathematics and Nonlinear Sciences

Research on design of customer portrait system for E-commerce

Lingjie Cai y
Lingjie Cai y
Jun Li
Jun Li

Publicado en línea: 23 Dec 2022

Volumen & Edición: AHEAD OF PRINT

Páginas: -

Recibido: 08 Mar 2022

Aceptado: 10 Apr 2022

Detalles de publicación y publicación

Applied Mathematics and Nonlinear Sciences

Vista previa del PDF

Artículo

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Figuras y tablas

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Comparison of algorithm performance

Normalised data

Results after clustering

Matrix table of evaluation criteria

Referencias

Artículos Recientes

Planifique su conferencia remota con Sciendo

Applied Mathematics and Nonlinear Sciences

Research on design of customer portrait system for E-commerce

Lingjie Cai yLingjie Cai yJun LiJun Li

Publicado en línea: 23 Dec 2022

Volumen & Edición: AHEAD OF PRINT

Páginas: -

Recibido: 08 Mar 2022

Aceptado: 10 Apr 2022

Detalles de publicación y publicación

Applied Mathematics and Nonlinear Sciences

Vista previa del PDF

Artículo

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Figuras y tablas

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Comparison of algorithm performance

Normalised data

Results after clustering

Matrix table of evaluation criteria

Referencias

Artículos Recientes

Planifique su conferencia remota con Sciendo

Lingjie Cai y
Lingjie Cai y
Jun Li
Jun Li