Research on Precision Marketing and Smart Tourism Service Optimization of Online Marketing Driven E-commerce Platform Based on Big Data and Machine Learning

Research all kinds of e-commerce display terminals and data sources, broaden the study of factors affecting e-commerce users, and deepen the combing and collection of user data from all e-commerce terminals and internal and external data sources. Formulate data cleaning rules and data standardization rules, according to the rules of data cleaning and processing, including duplication, vacancies, errors, inconsistencies and other issues of data processing, to ensure data accuracy. Research all the user identification in the data, including: imei, idfa, idfv, ID, cookie, email, phone and other identification fields screening, study the user’s unified identification UID, build ID mapping system, mining and use of various types of associated ID, open the data barriers of various data sources, terminals, and businesses, and bring together the user’s data in various channels to form The data form with the user’s unified identity as the main key is shown in Figure 1.

Figure 1.

ID mapping system

Construct e-commerce-oriented user portrait model, divide demographic attributes, social attributes, behavioral attributes, interest attributes, ability attributes, and psychological attributes into six dimensions based on e-commerce elements, construct the user portrait feature layer, form the user’s performance in each dimension into a feature, and use the labeling rules and algorithms to realize the user marking, and label the user data. Build a complete labeling system, complete the construction of internal labels in a dimensional and hierarchical way, and study the two-level system of user labeling refinement division scheme.

The e-commerce user portrait is divided into 6 dimensions, 13 first-level labels, 92 second-level labels, and the model is represented as: 1Persona={ PO,SO,BE,IN,AB,PS }$$Persona = \left\{ {PO,SO,BE,IN,AB,PS} \right\}$$ where each dimension corresponds to the first level of labeling and the model is represented as: 2{ P0={ NAC,SOC }S0={ SOF,SPF }BE={ VIB,SHB,AFB }IN={ LOI,SHI }AB={ COA,PAA }PS={ LIS,PEC }$$\left\{ {\matrix{ {P0 = \left\{ {NAC,SOC} \right\}} \hfill \cr {S0 = \left\{ {SOF,SPF} \right\}} \hfill \cr {BE = \left\{ {VIB,SHB,AFB} \right\}} \hfill \cr {IN = \left\{ {LOI,SHI} \right\}} \hfill \cr {AB = \left\{ {COA,PAA} \right\}} \hfill \cr {PS = \left\{ {LIS,PEC} \right\}} \hfill \cr } } \right.$$

Each first-level label corresponds to a more refined second-level label, which is directly applied to the business and whose feature values have a high impact on the business.

Based on the e-commerce user profile, we will study the crowd classification problem from the perspective of precision marketing and gain insight into user classification. The first step is to study the main classification of the crowd, based on the e-commerce user profile behavioral attributes of the series of labels to divide the new user N, old user O two categories. The second step is to study the subcategorization problem of the crowd, further subdividing the old user O category based on the multi-class labels of the e-commerce user portrait, and applying the improved K-means algorithm to divide the old user O into two subcategories of stable user S and churn-prone user E. The third step is to study the subcategory population re-segmentation problem, combining the three categories of NSE users, using AP clustering method to obtain the fine-grained clusters, and completing the subcategory population re-segmentation.

Among them, the old user O category segmentation uses the improved K-Means algorithm, which is suitable for mining large-scale datasets, high efficiency, scalability, time complexity near linear, and suitable for the initial classification of users.The K-Means algorithm uses the distance as the similarity evaluation index, and the sum of squares of the errors from the sample points to the center of the category is used as the evaluation data for the goodness of the clustering, and the iterative method is used to make the overall categorization of the error sum of squares function is minimized. The details are as follows:

The old user data set is defined as (X₁, X₂,…,X_n), where each X_i is a feature vector of a user multidimension. The set of user group data is divided into two categories S = {S₁, S₂}, where the sum of squared distances of all elements in S_j to the category center U_k is ∑xi∈sj‖ xi−uk ‖2$$\sum\limits_{{x_i} \in {s_j}} {{{\left\| {{x_i} - {u_k}} \right\|}^2}} $$ where ||x_i – u_k|| is the Euclidean distance. The objective of the classification is to minimize the weighted sum of the sum of squared distances of the two categories, and the objective function can be expressed as: 3F(U)=argmin∑k=12∑xi∈∫j2‖ xi−uk ‖2$$F\left( U \right) = \arg \min \sum\limits_{k = 1}^2 {\sum\limits_{{x_i} \in {\smallint _j}}^2 {{{\left\| {{x_i} - {u_k}} \right\|}^2}} } $$

The binary variable z_ik ∈ {0,1} is introduced to test into which category the data points should be classified. z_ij is defined as follows. 4zik={ 1if‖ xi−uk ‖2=mink=∈{ 1,2 }min‖ xi−uk ‖20other$${z_{ik}} = \left\{ {\matrix{ 1 \hfill & {if{{\left\| {{x_i} - {u_k}} \right\|}^2} = \min _{k = \in \left\{ {1,2} \right\}}^{\min }{{\left\| {{x_i} - {u_k}} \right\|}^2}} \hfill \cr 0 \hfill & {other} \hfill \cr } } \right.$$

Clustering centers for: 5uk=∑i=1nzikxik/∑i=1nzik$${u_k} = \mathop \sum \limits_{i = 1}^n {z_{ik}}{x_{ik}}/\mathop \sum \limits_{i = 1}^n {z_{ik}}$$

Therefore the improved objective function is: 6F(z,U)=argmin∑k=12∑xi∫j2zik‖ xi−uk ‖2$$F\left( {z,U} \right) = \arg \min \sum\limits_{k = 1}^2 {\sum\limits_{{x_i}{\smallint _j}}^2 {{z_{ik}}} } {\left\| {{x_i} - {u_k}} \right\|^2}$$

Iterate the algorithm by minimizing the objective function as the optimization goal, the first iteration needs to be an optimization, from the set (X₁, X₂,…,X_n) consciously selected the distance between the two elements as the center of the two categories, the purpose is to pull away from the distance of the initial point, reduce the similarity between the clusters, to improve the efficiency and effectiveness of clustering, and then calculate the distance from each element in the set to the center of the two categories, and classify these elements into the distance to the the smallest one to the category for clustering. Recalculate the center of each category according to the results of the previous clustering. Repeat the above steps by constantly updating the centers U = {u₁, u₂} of the two clusters until the classification results no longer change.

On the basis of NSE classification, various types of users need to be subdivided to form segmented populations. Here the AP clustering algorithm is used, which converts the user data into network nodes, calculates the clustering centers through two kinds of message passing between the attraction and belongingness of each node of the network, and continuously updates the attraction and belongingness of each node through iteration until multiple high clustering centers are generated to get the crowd segmentation, the advantage of the AP clustering algorithm is that there is no need to specify the number of clusters, it is insensitive to the selection of the initial value, and the clustering result is stable and conforms to the crowd segmentation scenario. The details are as follows:

After NSE classification one of the classes of user data set is defined as (Y₁, Y₂,…,Y_n), and the initial similarity matrix is generated using Euclidean distance: 7s(i,j)=−‖ yi−yk ‖2$$s\left( {i,j} \right) = - {\left\| {{y_i} - {y_k}} \right\|^2}$$

Set the initial reference degree P, P_i = P(i) as the reference degree of Y_i, which refers to the reliability of using Y_i as the center of clustering. Generally set P as the median of the similarity values.

Calculate the attractiveness value between two user data: 8r(i,k)=s(i,k)−k′≠kmax{ a(i,k)+s(i,k) }$$r\left( {i,k} \right) = s\left( {i,k} \right) - _{k' \ne k}^{\max }\left\{ {a\left( {i,k} \right) + s\left( {i,k} \right)} \right\}$$ where r(i, k) describes the degree to which the user data object y_k is suitable as a clustering center for the user data object y_i, and a(i,k′) describes the degree to which the user data object y_i is suitable for selecting the user data object y_k as its clustering center.

Calculate the value of the degree of belonging between two user data: 9a(i,k)=min{ 0,r(k,k)+∑i≠i,kmax{ 0,r(i,k) } }$$a\left( {i,k} \right) = \min \left\{ {0,r\left( {k,k} \right) + \sum\limits_{i \ne i,k} {\max } \left\{ {0,r\left( {i,k} \right)} \right\}} \right\}$$ 10a(k,k)=∑i≠kmax{ 0,r(i,k) }$$a\left( {k,k} \right) = \sum\limits_{i \ne k} {\max } \left\{ {0,r\left( {i,k} \right)} \right\}$$

The decay coefficient is introduced to update the iteration attraction r(i,k) and attribution a(i,k), and the calculation is terminated when the clustering centers no longer change over a number of iterations, and the individual clustering centers and individual classes are determined to obtain the segmented population.

The three above types of NSE users have a life cycle relationship, and for specific user roles, there is an evolution and transition relationship between new, stable, and churn-prone users. Formulate NSE precision marketing strategy, and formulate three precision marketing strategies: “attracting new users”, stabilizing user “customer protection” strategy, and “recall” strategy of easily losing users on the basis of e-commerce user portraits, so as to comprehensively improve marketing results.

The NSE precision marketing strategy is mathematically expressed as: 11SalesEffect=NewUsers+StableUsers+EasilyLostUsers$$SalesEffect = NewUsers + StableUsers + EasilyLostUsers$$

Figure 2

Schematic diagram of support vector machine

Mutually exclusive feature bundling strategy improves the model training speed from the perspective of reducing the feature dimensionality, this algorithm binds mutually exclusive features in high dimensional features, thus reducing the number of features to improve the training speed. High-dimensional features are often mutually exclusive in sparse feature space, i.e., these mutually exclusive features cannot take non-zero values at the same time, so the mutually exclusive features can be reasonably bundled into a single feature, thus reducing the number of features and improving the computational efficiency, and if the two features are not completely mutually exclusive, i.e., the two features take non-zero values in order to strike a balance between the computational accuracy and efficiency, the degree of non-mutual exclusivity of features can be measured using the conflict ratio of the measure to evaluate whether to bundle features together or not.

The histogram algorithm is a substitute for XGBoost’s pre-sorting algorithm, designed to address the issue of excessive number of split points. The basic idea of this algorithm is to split the features into boxes: divide the continuous feature values into k box, and construct a histogram with width k. When searching for the optimal segmentation point, only the discrete values of the histogram are searched, thus effectively reducing the number of segmentation points and lowering the storage and computation costs. Compared to the pre-sorting algorithm of XGBoost, the time complexity of LightGBM is greatly reduced. In addition, the histogram-based algorithm saves approximately seven times as much memory as the pre-sorting algorithm.

LightGBM grows the tree by using a leaf-by-leaf growth strategy, which selects the leaf nodes with the highest splitting gain to grow. Most tree models grow the tree by a per-layer growth strategy, i.e., by traversing the data once to split all leaf nodes in the same layer, a strategy that can be easily optimized for multiple threads. However, in practice, many leaf nodes have very low splitting gains, and splitting all leaf nodes indiscriminately increases the computational overhead. LightGBM employs a grow-by-leaf method to select the leaf node that has the greatest splitting gain among all leaf nodes to split. In addition, LightGBM restricts the depth of the tree to prevent overfitting by growing too deep decision trees, so the maximum depth parameter of the tree and the maximum number of leaf nodes parameter should be controlled during the model tuning process to avoid overfitting.

For numerical data, due to the different ranges of values of each variable and different scale, it will cause bias to the analysis results, so for numerical data, it needs to be standardized, in this paper, the numerical variables in the dataset are standardized by using Z fractional standardization method, so that the value of each variable is distributed in the range of -1 to 1. The formula of the Z -fractional standardization method is as follows: 20yi=xi−x¯s$${y_i} = {{{x_i} - \bar x} \over s}$$

There are six variables in this dataset, which are “Chinese_subscribe_num”, “math_subscribe_num”, “add_friend”, “add_group”, “study_num”, and “city_num_l”. In the classification problem, the existence of sub-type variables will affect the calculation of label distance, so it is necessary to do one-hot encoding on these six sub-type variables, and the one-hot encoding will map the variables to the Euclidean space, so as to avoid the influence of the values of the variables on the calculation of distance.

On the basis of the analysis and establishment of the user portrait labeling system, for the prediction of user dynamic interest labels, this paper proposes a user interest prediction algorithm based on the deep forest model and improves the model by using feature selection weights. The method utilizes the multi-granularity scanning of the deep forest to learn high-dimensional representations at a lower cost, and the number of layers of the cascade forest becomes a parameter that is automatically set according to the data situation through adaptive decision, which ensures the model complexity and accuracy while breaking through the limitation of hyperparameters.

From the analysis in the previous section, it can be seen that increasing the width and depth of the network structure has a significant effect on improving the performance of the model. For the deep forest model, increasing the width of the network is realized by increasing the types of base learners integrating the model in the two-part structure, while increasing the depth of the network is realized by expanding the number of training layers in the cascade forest structure.

First the application of the model is presented. The final model is obtained through model training, and the next step is to apply the trained model to a real task. It is trained to achieve high-efficiency and high-quality attraction recommendations. Given a user u who inputs a time constraint of m to travel in the recommender system and a candidate list consisting of attractions S₁(u,l), the recommender system generates a TOP-k list of recommendations based on the recommendation formula 1, where the higher the result indicates that the attraction is more in line with the user’s travel preferences. The result of Recommendation Equation S₁(u,l) describes the extent to which the attraction meets the user’s travel preferences while satisfying the user’s time constraints. From there, a Top-k attraction is presented to the user as a recommendation list based on the recommendation score: 32S1(u,l)∝(Vu+Vm)T▯v1$${S_1}(u,l)\alpha {({V_u} + {V_m})^T}\unicode {x25AF}{v_1}$$

Where V_u denotes the user-level implicit semantic feature vector of user u, V_m denotes the time-level implicit semantic feature vector of user’s traveling, and v₁ denotes the attraction-level implicit semantic feature vector of all attractions. Next, the trained model is applied to the real attraction recommendation task, in which users generally do not choose attractions that they have visited before. Therefore, a TOP-k recommendation list is generated based on the demand of user u at time point m, and the set of P attractions that the user has visited before l_p and the candidate list of other attractions l according to the recommendation formula S₂(u,l) : 33S2(u,l)∝(Vu+Vm+Vlp)T▯v1$${S_2}\left( {u,l} \right) \propto {({V_u} + {V_m} + {V_{lp}})^T}\unicode {x25AF}{v_1}$$

Where V_u denotes the user-level implicit semantic feature vector of user u, V_m denotes the time-level implicit semantic feature vector of the user’s tour, v₁ denotes the attraction-level implicit semantic feature vector of all attractions, and V_lp denotes the sum of implicit semantic feature vectors of all attractions in the attraction dataset that the tourists have already visited. It is found that the cold-start problem of attraction recommendation can be effectively solved by the MLS2vec model. When a brand-new user comes into the recommendation system, a TOP-k recommendation list is generated based on the user’s travel time point m and the candidate list composed of attractions l using recommendation formula S₁(u,l) : 34S1(u,l)∝VmT▯v1$${S_1}(u,l) \propto {V_m}^T \unicode {x25AF} {v_1}$$

The user conversion rate of the e-commerce platform is analyzed using the precision marketing model constructed in this paper. Figure 3 shows the trend graph for the order. The data of the user order form is briefly analyzed, and the user order form includes key order information such as the date of placing an order, the region in which the order was placed, and the number of pieces of goods ordered. Since the goal of the contest is to analyze whether the target user placed an order in the tag month and the earliest order date, which is closely related to the information in the order table, analyzing the user order table is a top priority. The figure depicts the total number of orders placed by the target user on a daily basis between June 1, 2021 and April 30, 2022. As can be seen from the graph, there were several large fluctuations between July-August 2021, November-December 2021, December 2021-January 2022, and March-April 2022, which may be attributed to the use of the Precision Marketing Model coupled with the effects of the 618 Shopping Festival, the Double Eleven and Double Twelve Shopping Festivals, and the April Promotions. After the intervention of the Precision Marketing Model, the overall order trend after the start of the year 2022 is higher than the June to December 2021 period. The number of orders for 2021 and 2022 is 138 and 245 respectively, which represents a 77.62% improvement in the number of orders.

Figure 3.

Order trend

Figure 4 shows the number of orders and purchasing users month by month, where the total number of orders and users within each month from May 2021 to April 2022 are calculated using month as the statistical dimension. It can be seen that the trends in month-by-month orders and month-by-month purchasing users are heading in a similar direction and can be combined and analyzed together, with two more notable increases between May 2021 and January 2022, in June 2021 and November 2021, when the number of month-by-month orders and purchasing users were 25,424 and 17,678, respectively, for the month of June. The number of orders and users purchased month by month in November was 46,634 and 31,357 respectively. The reason analysis is similar to Figure 3, which is considered as the effect of the precision marketing model. the number of users and the number of orders showed a significant growth trend from the beginning of February 2022 to April 2022, considering the reason is that the target users are selected from the part of the users who have purchased the target category of goods in the three months prior to the examination of the time period of May 2022, and therefore it is reasonable to see that the users’ purchasing behaviors are active in these three months, presenting this trend of change with the The selection criteria of the target users are related.

Figure 4.

Monthly order and purchase number

Figure 5 shows the month-by-month average number of orders and products purchased, where the average number of orders and the average number of products purchased by users within each month from May 2021 to April 2021 are calculated using month as the statistical dimension. As can be seen from the figure, there is a small difference in the mean values between the months, with a distribution range of [1.33, 1.48] for the average number of orders and [2.03, 2.61] for the average number of products purchased.

Figure 5.

Monthly orders and purchase of products

Through the descriptive analysis of the research participants on the questionnaire situation as shown in Table 2, to understand the mean and variance of each dimension and the overall, it can be found that the research participants of the smart tourism experience, perceived value, behavioral intention are biased towards positive affirmation, the mean value of the smart tourism experience, perceived value, and behavioral intention are 3.7359, 3.6534, and 3.9455, respectively. The very small value of all the question items, the very large value, skewness, and kurtosis are all within reasonable limits.

Based on the previous literature analysis, it can be found that previous studies have shown that demographic variables affect the results of perceived value and behavioral intention. In this paper, the research object is tourists who have visited CL tourism resorts, whose gender, age, education, monthly income, and occupation may affect the significant difference of tourists’ smart tourism experience, perceived value, and behavioral intention; therefore, this paper adopts the independent samples t-test to conduct the analysis of the difference between the gender on the smart tourism experience, perceived value, and behavioral intention, and one-way analysis of variance (ANOVA) to conduct the analysis of the difference between the age, education, monthly income, and occupation on the differential analysis of smart tourism experience, perceived value, and behavioral intention. 1)

Table 3 shows the independent samples t-test of gender differences on the relevant variables regarding the effect of gender on each variable. The independent sample t-test was used to confirm whether gender makes a significant difference in the means of each research variable. From the results of the T-test, it can be seen that the hypothesized equal significant values of variance for smart tourism experience, perceived value, and behavioral intention are 0.0945, 0.4596, and 0.1655, respectively, with a significance of P>0.05, which indicates that gender differences do not make a significant difference in smart tourism experience, perceived value, and behavioral intention.

From the running results of the first-order dimensional structural equation model of the five segmented dimensions of smart tourism experience, perceived value, and behavioral intention, the path analysis table of the structural equation model of the segmented dimensions of smart tourism experience was compiled, as shown in Table 5, which shows the path of the influence of the five segmented dimensions of the smart tourism experience on perceived value and behavioral intention, and the specific results are as follows: the influence of the smart tourism marketing experience on behavioral intention and perceived value The standardized coefficients are 0.136 and 0.193 respectively, and the P-value is less than 0.05, which indicates that the smart tourism marketing experience has a positive influence on tourists’ behavioral intention and perceived value, and the hypotheses of H1a and H2a are valid.

The standardized coefficients of the influence of smart tourism service experience on behavioral intention and perceived value are 0.148 and 0.187, respectively, with P-value less than 0.05, indicating that smart tourism service experience has a positive influence on tourists’ behavioral intention and perceived value, and the hypotheses of H1b and H2b are valid. The standardized coefficients of the influence of smart maintenance and guarantee experience on behavioral intention and perceived value are 0.128 and 0.194 respectively, with P-value less than 0.05, indicating that the smart maintenance and guarantee experience has a positive influence on tourists’ behavioral intention and perceived value, and the hypotheses of H1c and H2c are valid.

The standardized coefficients of the influence of smart integrated management experience on behavioral intention and perceived value are 0.259 and 0.175 respectively, with P-value less than 0.05, indicating that smart integrated management experience has a positive influence on tourists’ behavioral intention and perceived value, and the hypotheses of H1d and H2d are valid.

The standardized coefficient of the influence of smart tourism product experience on behavioral intention is 0.069 respectively, and the P=0.318 value is greater than 0.05, indicating that the influence of smart tourism product experience on tourists’ behavioral intention is not significant, and the H1e hypothesis is not established. The standardized coefficients of the influence of smart tourism product experience on perceived value are 0.168 respectively, P value is less than 0.05, indicating that smart tourism product experience has a positive influence on the perceived value of tourists, and H2e hypothesis is established.