Research on Precision Marketing Strategy of Rural Tourism Combining Big Data and Cloud Computing Technology

Under the mobile cloud computing environment, create a cloud service platform for rural leisure tourism industry, provide precise interaction and docking channels for the demand side, the provider and the regulator, establish supply-side product chains and demand-side product chains, and carry out product supply and demand analyses, hot sales analyses and prediction analyses on the basis of this. Relying on the cloud service platform to realize the e-commerce of leisure tourism products.

Service providers obtain information on actual customers and potential customers of leisure tourism products through the platform, a large amount of industrial information data is accumulated on the cloud service platform of rural leisure tourism industry, and industry managers, service providers, and leisure tourists share data resources, and they actually participate in the product design and innovation of service enterprises, playing the role of “cooperative producers”, which is the basis for customer interface innovation.

High-quality service products formed based on new service concepts, new service ideas, new service processes, new customer interfaces, etc. are delivered to consumers through new service delivery systems. Leisure tourism product e-commerce has changed the business transaction mode and process, the internal organization and staff capacity of the service providing enterprise should also be changed, and the service provider should organize, manage and coordinate all the links.

New technologies such as mobile cloud computing and big data analysis can inject new momentum into the rapid development of the rural service industry, and are platforms, tools, methods and means to improve service quality. The innovative design adopts the mode of “cloud platform + Internet + mobile terminal” to provide a platform environment for the supply side, demand side and regulatory side of the rural leisure tourism industry to fully dock, exchange and interact.

The basic idea of the density-based approach is to add the density of points in a single region to clusters similar to it whenever it is greater than a certain threshold. The traditional K-mean algorithm uses the Euclidean distance as a similarity measure, and the random selection of the initial center does not reflect the distribution of the data, while the k objects that are furthest away from each other are more representative. Therefore, in order to avoid the interference of noise points in the low-density area, this paper selects the k farthest away from each other as the initial clustering center from the high-density area.

Definition 1: The distance formula of 2 objects is: (2)dij=∑k=1p(xik−xjk)2${d_{ij}} = \sqrt {\sum\limits_{k = 1}^p {{{\left( {{x_{ik}} - {x_{jk}}} \right)}^2}} }$

where d_ij is the distance between 2 p-dimensional objects i and j. x_ik is the data of the object i in k dimensions.

Definition 2: ε neighborhood is the area within the radius ε of a given object. For example, the ε neighborhood of a 1 point p is N_ε(p) = {o ∈ D ∣ dist(p, o) ≤ ε}.

Definition 3: Core object: an object is said to be core if the ε neighborhood of 1 object (p) contains at least the minimum number MinPts of objects.

The improved density-based centroid initialization algorithm is described as follows: input: a dataset containing n object, the number of clusters k the radius of the neighborhood ε and the minimum number of objects contained in the neighborhood MinPts. Output: k initial centroid objects.

The steps of the algorithm are as follows: 1)

Calculate the distance d_ij of all n objects.

In the traditional K-means algorithm, the time complexity of the process of attributing each object to the class of its nearest center is O(nkd), where n is the number of objects, k is the number of clusters, and d is the number of dimensions of the data objects. The time overhead of this algorithm is more considerable in practical applications where the amount of data is large and the number of dimensions of data objects is large.

Since the K-means algorithm uses the Euclidean distance to measure the similarity between the objects, this paper considers the theory of the relationship between the sum of the two sides of the triangle is greater than the third side of the triangle to simplify the calculation process, the algorithm in one iteration process is as follows: 1)

Calculate the distance d(ci,cj)$d\left( {{c_i},{c_j}} \right)$ from any 2 cluster centers, where i = 1, 2, ⋯, k, j = 1, 2, ⋯, k.

The time complexity of this improved algorithm is O(nvd)$O\left({nvd} \right)$, where v is the average number of computations for 1 object during an iteration, 1 ≤ v ≤ k. i.e., the number of computations for an object to the centroid during an iteration is computed 1 in the best case, and k in the worst case, and the improvement in the efficiency of the algorithm is obvious if the sample set (n)$\left( n \right)$ is large.

In real life when we are faced with the choice of goods, we will often first consult the people around us who have experience in purchasing, according to their advice to make a suitable choice for their own purchases, or will first consult the opinions of people with similar interests to help them make a suitable choice. User-based collaborative filtering algorithms are mainly used to mine and calculate user information, and find similar users according to the results of information analysis. Similarity can be obtained by obtaining the explicit or implicit behavior of the user such as ratings, retweets, saves, tags, comments, clicks, page stay time and whether to buy, etc., and calculated after processing, the closer the similarity of the results of the calculation, we believe that the similarity of the interests of similar users [36]. The algorithm process is: 1)

Transform the collected user information into a two-dimensional matrix for digital representation, and perform noise reduction and normalization on the data.

The basic formula of user-based collaborative filtering algorithm is shown in (3): (3)P(u,i)=∑v∈S(u,K)∩N(i)wuvrvi$P(u,i) = \sum\limits_{v \in S(u,K) \cap N(i)} {{w_{uv}}} {r_{vi}}$

User-based collaborative filtering algorithms are suitable for cases with a small number of users due to their own characteristics, and when the set of users is too large, the computation of similarity through the user matrix can be costly.

The collaborative filtering algorithm based on commodities is relatively more used in practical applications, it does not calculate the commodity attributes, but calculates the similarity between commodities through user ratings, which simply means that it recommends similar commodities to the commodities that the user has liked. The basic formula is shown in (4): (4)Puj=∑j∈N∑(u)∩S(i,K)Wjiruj${P_{uj}} = \sum\limits_{j \in N} {\sum\limits_{{{(u)}^{ \cap S(i,K)}}} {{W_{ji}}} } {r_{uj}}$

The recommendation calculation process is roughly similar to the user-based collaborative filtering algorithm, the only difference is that after the establishment of the user-item scoring matrix this algorithm calculates the similarity between the goods, by obtaining the user’s historical data to determine the goods that the user has liked, and recommend the list of similar goods of this product TOP-N to the target user.

When the item data of a system is much smaller than the user set data and the number of items is not likely to change too much, it is more suitable to use the collaborative filtering algorithm based on items.

Because the number of tourist attractions is not easy to change and is much smaller than the number of users, and the resulting attractions - evaluation index matrix data is dense and there is basically a common value between the variables, for the attraction similarity selection of Euclidean distance calculation. The algorithm is based on the user-item matrix data for calculation, after determining the weight of each item of the evaluation index system, a 549×12-dimensional matrix of attraction-evaluation indicators can be constructed by crawling the relevant data from the network: (6)Jmn= J1 J2 ⋯ J549( A1 A2 ⋯ A12 B1 B2 ⋯ B12 ⋯ ⋯ ⋯ ⋯ N1 N2 ⋯ N12)${J_{mn}} = \begin{array}{*{20}{c}} {{J_1}} \\ {{J_2}} \\ \cdots \\ {{J_{549}}} \end{array}\left( {\begin{array}{*{20}{c}} {{A_1}}&{{A_2}}& \cdots &{{A_{12}}} \\ {{B_1}}&{{B_2}}& \cdots &{{B_{12}}} \\ \cdots & \cdots & \cdots & \cdots \\ {{N_1}}&{{N_2}}& \cdots &{{N_{12}}} \end{array}} \right)$

Where each row represents the data of an attraction and the columns represent the data of the evaluation indicators that have been constructed, multiplying the data of each column with the weight of that indicator to obtain the Attraction-Weight Indicator Matrix M_mn: (7)Mmn=Jmn×( T1 T2 … T12)${M_{mn}} = {J_{mn}} \times \left( {\begin{array}{*{20}{l}} {{T_1}}&{{T_2}}& \ldots &{{T_{12}}} \end{array}} \right)$

The similarity of the attraction data results in the matrix M_m, the choice of Python language cyclic assignment matrix for each line X, Y, respectively, two by two to calculate its Euclidean distance and select TOP 30 as the attractions of similar attractions, similar attractions list is recorded as L₁.

Construct a user-attraction rating matrix based on user ratings of attractions: (8)(U−S)mn= J1 J2 ⋯ Jm U1 U1 ⋯ Un( s11 s12 ⋯ s1m s21 s22 ⋯ s2m ⋯ ⋯ ⋯ ⋯ sn1 sn2 ⋯ snm)${\left( {U - S} \right)_{mn}} = \begin{array}{*{20}{c}} {\begin{array}{*{20}{c}} {}&\quad{{J_1}}&{{J_2}}& \cdots &{{J_m}} \end{array}} \\ {\begin{array}{*{20}{c}} {{U_1}} \\ {{U_1}} \\ \cdots \\ {{U_n}} \end{array}\left( {\begin{array}{*{20}{c}} {{s_{11}}}&{{s_{12}}}& \cdots &{{s_{1m}}} \\ {{s_{21}}}&{{s_{22}}}& \cdots &{{s_{2m}}} \\ \cdots & \cdots & \cdots & \cdots \\ {{s_{n1}}}&{{s_{n2}}}& \cdots &{{s_{nm}}} \end{array}} \right)} \end{array}$

Here each user is a m-dimensional vector, where s_m represents the rating value of the nrd user for the mth attraction, and the similarity between users is calculated by applying the Euclidean distance formula (5) to the user’s feature vectors. Among them, when user U_m is a user who has been rated by the system, the similarity is calculated by the feature vector Um(sm1,sm2,…..smmmm)${U_m}\left( {{s_{m1}},{s_{m2}}, \ldots ..sm{m_{mm}}} \right)$ of user U_m in the user-attraction rating matrix and the feature vectors of other users to obtain the similar user TOP-N. When the user is a user who has not rated the attractions in the system, the similarity is calculated by extracting the user characteristics based on the user information, and the weighted average of the ratings of the attractions by the similar users are calculated as new user’s rating is added to the (U−S)mm${\left( {U - S} \right)_{mm}}$ matrix. The nearest neighboring users of the user are derived, and the weighted average is calculated based on the ratings of the neighboring users for the unrated attractions to get the predicted ratings of the user and derive the list of attractions for TOP-N L₂.

Assuming that the number of attractions in the final recommended list L is s, and the number of predicted user ratings greater than or equal to 3 in the filtering judgment list L₂ is n. When n > = s, take the first s attractions in list L₂ and fill them into list L to generate the recommended list. When n < s, the first n attractions in list L₂ are filled into list L, and the remaining n − s values in list L₁ are filtered out of the n attractions in list L, the similarity value ranking and user rating ranking in the first n − s attractions, and the n attractions together form the final recommendation list L, which is recommended for the user.

In this paper, with the help of improved K-means clustering analysis to classify the behavior and preference patterns of tourists’ tourism consumption, the clustering of 35 variables, marking the performance of the research object in the 35 indicators of tourists’ behavior. In addition, the clustering object is 255 tourists who have been or are traveling and consuming in Hangzhou. According to the questionnaire statistics and the control of tourists’ behavioral stages, the K value is taken as 4, so that the sample size is distributed more evenly. After the mean clustering and ANOVA test, the results of the division of tourist behavior under 35 indicators are shown in Table 1.

The degree of difference of F-test is significant, meanwhile, the significance level of all 35 indicators is lower than 0.05, which represents the higher validity of data analysis when K=4, and the results of K-means clustering in this group are established. In addition, the significant indicators between the four types are obviously different, and the sample size of each cluster set is divided more evenly, and does not need to be divided again.

After improved K-means clustering, the visitor behavior and preference under are mainly reflected in the following four types: 1)

Personalized push sensitive type. For personalized marketing push content response sensitive tourists accounted for 27.7% of the total sample size, this type of tourists in the current digital environment accounted for a higher proportion than the other three types, the characteristics of which are mainly manifested in the start of the trip before the start of the trip, easy to because of the personal history of the network traces caused by the customization of tourism, leisure-related product recommendations, and then form a payment conversion. Although this category of tourists passively obtain promotional content before the trip, as long as it meets their personalized preferences, it is able to form a high marketing conversion.

The basic attributes of the population in the questionnaire corresponding to the four types of tourist behavior and preference types were counted, and the results are shown in Table 2.

In the personalized push-sensitive tourist groups, the ratio of men and women is basically equal, and mainly concentrated in the middle-aged groups in the stages of 26~45 years old and 46~65 years old, and from the income point of view is concentrated in the middle- and high-income groups. It can be proved that the customized push type online marketing meets the preference of the main consumer group in the current market.

In terms of consumption share, personalized accurate push and technological interactive experience marketing promotion can play a significant consumption promotion role for leisure and entertainment programs in the countryside (38.16% and 23.41% respectively). Meanwhile, the interactive marketing of social media and self-media as well as positioning push marketing can effectively boost the consumption of food and beverage (more than 25% of total consumption) and the revenue of attraction programs (about 30% of total consumers on average).

It can be seen that the digital marketing system planned for the above four types of tourists’ preferences meets the core needs of tourism consumers and can bring about local consumption growth in leisure, accommodation and shopping.

Using squared Euclidean distance and again for systematic clustering, the classification results of tourists’ behaviors and preferences in the digital era are shown in Figure 1.

Figure 1.

The result of the tourist behavior clustering

The systematic clustering can verify the validity of K-mean rapid clustering and intuitively determine the similarity and dissimilarity between tourists’ consumption behaviors in the digital era. From the Ward’s linkage clustering map, the tourists’ behaviors can be roughly divided into 2 major categories, and the refinement can be divided into 4 small types, which is consistent with the results of the improved K-means clustering, and the number of sample distributions divided into 4 categories is also consistent with the results of the rapid clustering analysis. Therefore, it can be verified that the behavior of tourists in the digital era can be divided into four main types: personalized push-sensitive, social media and self-media active access to information, technological interaction and experience, and reliance on travel path customization + positioning push project type.

In addition, a cluster scatter plot was plotted using python for cluster analysis of the collected data, and the results obtained are shown in Figure 2. Looking at the 4 types of tourist behavior:

Figure 2.

Cluster scatter diagram

Type I - personalized push-sensitive tourists are mainly manifested as accepting customized content and channel push based on preferences, so it is necessary to optimize the potential tourists’ consumption data mining and precise promotion calculations before the trip to activate and expand the consumption demand.

Type 2 - social media and self-media active access to information type and type 3 - technology interactive experience type of tourists need to match the multi-channel digital media integrated marketing, facilitating interactive experience, content identification, and other forms of user preferences to enhance the marketing Conversion rate.

As for the fourth type of tourists who rely on official channels for information, the cloud-based big data tourism service information platform can be upgraded and shared to push customized travel routes and peripheral projects to tourists based on their preferences and rural traffic and scenic area conditions to stimulate consumption.

In summary, based on the classification of tourist behavior research, tourism digital marketing system should be divided into the following three main levels, front-end strategy: tourism consumption data mining and cloud computing. Middle-end strategy: digital media marketing integration. Back-end strategy: upgrading and sharing cloud-based big data tourism service information platform.

Crawl the user comments of a travel website as experimental data. According to the acquired user travel text data for LDA theme analysis, extract the user’s tourism interest theme feature words, some user interest feature words are shown in Table 3. Calculate the interest similarity between users, and number the users using USER for easy representation.

The user’s comment data is processed and analyzed to extract the user’s attribute word-emotion word pairs about the attractions, the attribute word-emotion word pairs of the attractions are subjected to emotion scoring and emotion polarity determination to obtain the user’s scores of the attribute word-emotion about the attractions, the attribute word-emotion scores of the same attribute facet in the same user’s comment are summed and then averaged to obtain the user’s scores of the attribute facets of the attractions and the scores of the emotions of The sentiment scores of each user on different attribute facets of attractions are calculated, and then the sentiment scores are used to calculate the sentiment similarity between users on attraction-attribute facets. After obtaining the user’s interest similarity and emotion similarity, the two similarities are integrated according to 0.5 and 0.5 to give weight, and the integrated user’s comprehensive similarity is shown in Table 4.

After integration through the integrated similarity to find the nearest neighbor users similar to the target use, find the attractions played by similar users, calculate the attribute sentiment score of each attraction, and then use the Euclidean distance to calculate the similarity between the attraction’s attribute sentiment score and the target user’s attraction attribute sentiment characteristics, the attractions with higher similarity will be recommended, and the set of attractions recommended by similar users is shown in Table 5.

In order to validate the recommendation effect of the recommendation method in this paper, the recall rate, accuracy rate and integrated metric F1 value are used as the judging criteria. Recall rate: indicates the probability of recommending the attractions that the user is interested in to that user. That is, the ratio of the attractions that the user is interested in the list of recommended attractions to all the attractions that the user is actually interested in. Accuracy rate: indicates the probability of whether the user likes or dislikes the recommended attractions, i.e., the ratio of the attractions in the recommended attractions list that the user is interested in to all the attractions in the recommended list. The F1 value is used to evaluate the accuracy and recall rate together. The larger the values of recall, accuracy and F1 value, the better the recommendation effect.

In order to verify the recommendation effect of the recommendation algorithm proposed in this paper, the recommendation algorithm proposed in this paper and the traditional collaborative filtering algorithm based on user attraction ratings and the recommendation algorithm based on similar users are compared and analyzed, and Fig. 3 shows the accuracy curve of the three recommendation algorithms. Under the selection of different attraction recommendation list length, the number of attractions in the recommendation list gradually increases, while the number of attractions of interest to the user in the recommendation list increases less, so the accuracy rate decreases after the rise. According to the results of the experiment, this paper’s recommendation algorithm based on user ratings has a greater improvement in accuracy compared to the collaborative filtering algorithm based on user attraction ratings, and it is also slightly higher than the recommendation algorithm for similar users in terms of accuracy.

Figure 3.

The comparison of recommendation accuracy of the algorithms

Fig. 4 shows the recall graphs of the three recommendation algorithms. The recall rates of the three algorithms show an increasing trend as the length of the recommended attractions list increases. When the number of recommended attractions increases, the proportion of attractions in the list of attractions recommended to the user that the user is interested in to all the attractions that the user is actually interested in increases gradually, so the recall rate also increases gradually. The user rating based recommendation algorithm in this paper has an increase in recall compared to the other two algorithms, so the user rating based recommendation algorithm proposed in this paper has a certain advantage in terms of recall.

Figure 4.

The comparison of recall rate of the algorithms

The F1 value curves of the three recommendation algorithms are shown in Fig. 5, and the recommendation algorithm of this paper and the recommendation algorithm based on similar users are better than the traditional collaborative filtering algorithm based on attraction ratings in terms of the comprehensive F1 value. And the F1 value of this paper’s recommendation algorithm based on user image model is also slightly higher than that of the recommendation algorithm based on similar users. In summary, the recommendation algorithm based on user image model in this paper has better recommendation effect.

Figure 5.

The comparison of F1 of the algorithms