CN104331816B - Knowledge based learns big data user's purchase intention Forecasting Methodology with secret protection - Google Patents

Abstract

The invention discloses big data user's purchase intention Forecasting Methodology of knowledge based study and secret protection, its step is as follows：(1) normalized is done to a large amount of historical datas and a small amount of current data；(2) data division group and training sample set is built；(3) user's purchase intention probability of each group of statistics；(4) calculating group label；(5) training set is trained using improved SVMs；(6) structure forecast function；(7) data input to be predicted is predicted the outcome into anticipation function.The present invention uses improved SVMs; current a small amount of data group probabilistic information and substantial amounts of history data set probabilistic information are incorporated in structural risk minimization learning framework; the study of different times knowledge is realized by constructing between data similarity distance; so as to construct a kind of knowledge based study and the Forecasting Methodology of user's purchase intention of Privacy Preservation Mechanism, the problem concerning study of large sample is applicable to.

Description

Big data user purchase intention prediction method based on knowledge learning and privacy protection

Technical Field

The invention belongs to the technical field of marketing, relates to a pattern recognition technology, and discloses a big data user purchase intention prediction method based on knowledge learning and privacy protection.

Background

The invention belongs to the technical field of marketing, relates to a pattern recognition technology, and discloses a big data user purchase intention prediction method based on knowledge learning and privacy protection.

The consumer is a guide for various business activities of the enterprise, and the purchasing intention of the consumer is the basis of purchasing behavior and can be used for predicting the behavior of the consumer. From the marketing perspective, after an enterprise masters the purchase intention of a consumer, the purchase of raw materials can be reasonably arranged, the structure of a product is adjusted, and a production plan of the product is formulated; after the marketing personnel master the purchase intention of the consumer, the relevant commodities can be recommended to the consumer in a targeted manner, so that the sales volume is increased; after shopping willingness of consumers is mastered in shopping malls and supermarkets, commodities can be purchased purposefully, and rotation of the commodities is accelerated. Therefore, the purchase intention of research users is important content of marketing research, and has important theoretical and practical significance for correctly guiding the consumption of residents and guiding enterprises to make scientific and reasonable production and marketing strategies.

There are two general methods for measuring and calculating the purchase intention of a user: the first type is that the purchasing intention of the user is measured by a direct inquiry method, such as a Choice Based model, which sets eight attribute indexes, determines a weight for each index, then the user scores the eight indexes, and finally the purchasing intention of the user for the product is calculated according to the weights and the scores. The second type is that an information technology, such as Wang duckweed, is used for establishing a classification model of the purchasing intention of a customer by using a decision tree and a neural network method to predict the purchasing tendency of the customer (the data mining technology is used for predicting the purchasing tendency of the customer, namely the method and the empirical research, the intelligence science, 5 months 2005); wu Hua et al introduces several attribute variables and constructs a stochastic model describing the purchasing behavior or the predicted purchasing probability of the customer (analysis of influence factors of the purchasing behavior of the customer and prediction of the re-purchasing probability, the report of management engineering, 1 month in 2005). However, the two methods have defects, the first method is simple and easy to implement, but the weight of the index is not easy to control, the measurement and calculation precision is low, the historical data used in the method is not enough to reflect all characteristics of the current data, and the second method is provided that the time interval between the front and back purchase of a customer obeys Gamma distribution, which has great limitation in practical use; and the two methods are not suitable for the situation of big data, particularly disclosing the purchasing intention of the user in the using process of the model, and not considering the confidentiality of personal information of the user.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: firstly, the existing user purchase intention prediction method has low prediction accuracy in the scene of a large amount of historical related data and a small amount of latest data; secondly, the existing method is not suitable for a big data scene, and the time required for training the model is long; third, the existing methods cannot effectively protect the privacy of the personal information of the user.

The technical scheme of the invention is as follows: a big data user purchase intention prediction method based on knowledge learning and privacy protection is characterized in that an improved support vector machine technology is used, a small amount of current data set purchase intention probability information and a large amount of historical data set purchase intention probability information are merged into a structure risk minimization learning framework, knowledge learning is achieved by constructing data similar distance items in different periods, and user purchase intentions are predicted, and the method comprises the following steps:

the method comprises the following steps: carrying out normalization processing on a large number of historical data samples and a small number of current data samples to obtain an initial sample set (x)_i,y_i)_i＝1,…,NWherein x is_iIs a feature vector of the sample, y_i∈ { +1, -1} is the class label of the sample, N is the total number of samples;

step two: dividing historical data samples and current data samples into groups, wherein the number of the data in each group is approximately the same, and constructing a training sample set D ═ D₁,…,D_n,D_n+1,…,D_n+dThe previous n groups are historical data samples, and the next d groups are current data samples;

step three: counting probability p of user's purchase intention in each group_iCalculated as equation (1):

wherein, | S_iL is the number of samples contained in each group;

step four: compute group tag l_iCalculated as equation (2):

step five: training the training set by using an improved support vector machine, wherein the form of a training model is as follows (3):

wherein,is composed ofKernel space mapping function of, w_hAndweight vectors, C, for the historical data sample and the current data sample, respectively_hAnd C_cRespectively, the historical data sample and the current data sample regularization parameters, ξ_iAndrelaxation variables, ξ, respectively, for historical data samples_jAnd ξ_i ^*Respectively, the relaxation variable of the current data sample, λ is a normal number as the balance parameter, l_iFor the group label of each group obtained in step four,_i(i ═ 1, …, n) and'_i(i ═ n +1, …, n + d) is the approximation accuracy of each set of samples in the historical data sample and the current data sample, respectively, and the calculation formula is as follows:

wherein p is_iΔ is a constant ranging from 0.01 to 0.1 for the probability calculated in step three;

equation (3) above can be converted to a quadratic programming form as follows:

wherein

L_h＝[l₁,...,l_n],L_c＝[l_n+1,...,l_n+d],

Is a kernel function, x_i、x_jSolving the formula (3) and the formula (5) to obtain the optimal solution w of w and b_c ^*，b_c ^*Wherein w is_c ^*Can be represented by formula (6):

b_c ^*can be represented by formula (7):

step six: constructing a prediction function, the prediction function having the form of equation (8):

wherein, w_c ^*And b_c ^*Obtaining an optimal solution for the step five;

step seven: normalizing user data to be predicted and inputting the normalized user data into the prediction function to obtain a prediction result of the purchase intention of the user, and if:

the beneficial effects of the invention are as follows: 1) the invention acquires the relevant knowledge by using a large amount of historical data, assists the knowledge learning only containing a small amount of current data, establishes the prediction model based on the improved support vector machine, inherits the advantage of the maximum interval of the support vector machine based on the empirical risk minimization framework, and achieves the effect of better prediction of the current data. 2) The sample labels are replaced by group probabilities when training the historical data samples and the current data samples, so the spatial complexity of the invention is (O (N)²) Time complexity of (O (N))³) Wherein N is the number of groups of the historical data and the current data which are grouped together, is more suitable for being used in a big data environment than the traditional SVM (the space complexity of the traditional SVM is O (M))²) Time complexity of (O (M))³) (where M is the number of samples taken of the historical data and the current data), where N is the number of samples taken<<And M. 3) The sample label contains very important privacy information, and the group probability is used for replacing the sample label when the historical data sample and the current data sample are trained, so that the advantage of privacy protection provided for the original data is realized, and the privacy information of a user can be effectively protected from being leaked in the training process.

Drawings

Fig. 1 is a general flowchart of the big data user buying desire prediction method based on knowledge learning and privacy protection of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings in conjunction with specific examples.

In the embodiment, survey data of a certain digital product is selected as a research object, wherein historical data is survey data of digital products of other models of the product brand half a year ago and comprises 10000 samples in total, and current data is survey data of digital products of the same brand and new models released in the recent day and comprises 100 samples in total. All data includes 40 attributes in total, including: AGE, SEX, MARITAL, JOB, TRAVTIME, STATEMOD, DOMESTIC, BRAND, MODEL NUMBER, AMUNT, PRICE, RETURN, etc.

The method comprises the following steps: carrying out normalization processing on a large number of historical data samples and a small number of current data samples;

in order to improve the accuracy of the prediction method, the data needs to be normalized to the [0,1] or [ -1,1] interval. The embodiment adopts a normalization method: linearly converting all data of each characteristic value into a [0,1] interval according to the maximum value and the minimum value of each characteristic value, wherein the conversion formula is as follows:

wherein x is a characteristic value before normalization, max (x) and min (x) respectively represent that the maximum value and the minimum value are taken for x, and x' is a characteristic value after normalization. After the normalization of the characteristic values is completed, all the characteristic value groups are combinedForming a feature vector, wherein the feature vector x is expressed as x ═ x₁,...,x₄₀}^TWherein x is₁,...,x₄₀Each represents a specific value normalized by the 40 attributes.

The initial training sample set contains 10000 historical data samples and 50 current data sample samples, denoted as (x)_i,y_i)_i＝1,…,NWherein x is_iIs a feature vector of the sample, y_i∈ { +1, -1} is the class label of the sample, N ═ 10050, and the test sample set is the remaining 50 current data samples, denoted as (x)_i,y_i)_i＝1,…,MM is 50, y in the test set_iUsed to verify the accuracy of the method;

step two: dividing historical data samples and current data samples into groups according to the principle of 10 samples in each group, dividing the historical data samples into 100 groups, dividing the current data samples into 5 groups, and constructing a training sample set D ═ D₁,…,D_n,D_n+1,…,D_n+dThe previous n groups are historical data samples, the next d group is current data samples, n is 100, and d is 5;

step three: counting in each group the probability p of the user's willingness to purchase_iThe calculation formula is as follows:

probability of purchase of user's will

Wherein, | S_iI is the number of samples contained in each packet, S in this embodiment_i|＝10；

Step four: compute group tag l_iCalculated as follows:

step five: training the training set by using an improved support vector machine, wherein the form of a training model is as follows (4):

wherein,is composed ofAs a kernel space mapping function, w_hAndweight vectors, C, for the historical data sample and the current data sample, respectively_hAnd C_cRespectively, the regularization parameters of the historical data sample and the current data sample, C in this embodiment_hAnd C_cIn grid 2^-8,2^-7,2^-6,2^-5,2^-4,2^-3,2^-2,2^-1,2⁰,2¹,2²,2³,2⁴,2⁵,2⁶,2⁷,2⁸,2⁹,2¹⁰Search for the optimal value, ξ_iAndrelaxation variables, ξ, respectively, for historical data samples_jAnd ξ_i ^*Respectively, the relaxation variables of the current data sample, where λ is a balance parameter, which is a normal number, and in this embodiment, the balance parameter λ is in the interval {2 }^-6,2^-5,2^-4,2^-3,2^-2,2^-1,2⁰,2¹,2²,2³,2⁴,2⁵,2⁶,2⁷,2⁸,2⁹,2¹⁰,2¹¹Search for the optimal value, l_iTo calculate the group tag using equation (3),_iand'_iRespectively calculating the approximation precision of each group of samples in the historical data samples and the current data samples according to the following formula:

wherein p is_iCalculated by formula (2), Δ is a constant ranging from 0.01 to 0.1, in this embodiment Δ is 0.1;

the above equation can be converted into a quadratic programming form as follows:

wherein

L_h＝[l₁,...,l_n],L_c＝[l_n+1,...,l_n+d],

Is a kernel function, x_i、x_jFeature vectors for the ith and jth samples respectively,

in this embodiment, the kernel function is a Gaussian kernel function K (x)_i,x_j)＝exp(-r||x_i-x_j||²) Where r is the kernel-width parameter, in this example r is in the interval {2 }^-15,2^-13,...,2³Search for the optimal value, x_i、x_jThe feature vectors of the ith and jth samples respectively, | · | | computationally²Representing the euclidean distance.

The optimal solution w of w, b can be obtained by solving the equations (3) and (5)_c ^*，b_c ^*Wherein w is_c ^*Can be represented by formula (7):

b_c ^*can be represented by formula (8):

step six: constructing a prediction function, the prediction function being of the form:

wherein, w_c ^*，b_c ^*The optimal solution is obtained;

step seven: inputting a test set comprising 50 samples into a prediction functionIn the step (A), if:

the accuracy and the running time of the method are shown in table 1, and the results of the method are compared with the prediction results obtained by training a sample set which is obtained by combining 10000 historical data samples, 50 current data samples, 10000 historical data samples and 50 current data samples and using the sample set as a training set of a common Support Vector Machine (SVM) (the three methods are SVM respectively)_{10000 historical data}、SVM_{50 current data}And SVM_{10050 mixed data}) The experimental platform is MATLAB 2009 (a). As can be seen from the data in Table 1, the other three methods are less accurate than the method of the present invention because of the SVM_{10000 historical data}Only historical data is used for training, and changes among data in different periods are ignored; SVM_{50 current data}Only 50 current data samples are used for prediction, and a better prediction model is obtained because training samples are too few; SVM_{10050 mixed data}The difference between data in different periods is not considered, and in addition, the three methods all disclose label information of the sample in training, and privacy protection is not performed on important information of purchase intention of the user.

Table 1: method and SVM of the invention_{10000 historical data}、SVM_{50 current data}、SVM_{10050 mixed data}Comparison of discrimination accuracy (%) and running time (in seconds)

	Accuracy (%)	Run time (seconds)
			The invention	94％	215
SVM10000 historical data	88％	90681
			SVM50 current data	64％	213
SVM10050 mixed data	90％	91190

The above examples are intended to be illustrative of the present invention and are not to be construed as limiting the invention. Those skilled in the art can make various other modifications and alterations without departing from the spirit of the invention in light of the teachings of the present disclosure, and such modifications and alterations are intended to be included within the scope of the invention.

Claims (1)

1. The big data user purchase intention prediction method based on knowledge learning and privacy protection is characterized by comprising the following steps of:

the method comprises the following steps: carrying out normalization processing on a large number of historical data samples and a small number of current data samples to obtain an initial sample set (x)_i,y_i)_i＝1,…,NWherein x is_iIs a feature vector of the sample, y_i∈ { +1, -1} is the class label of the sample, N is the total number of samples;

step two: dividing historical data samples and current data samples into groups, wherein the groups are divided into groupsThe data numbers are approximately the same, and a training sample set D ═ D is constructed₁,…,D_n,D_n+1,…,D_n+dThe previous n groups are historical data samples, and the next d groups are current data samples;

step three: counting probability p of user's purchase intention in each group_iCalculated as equation (1):

wherein, | S_iL is the number of samples contained in each group;

step four: compute group tag l_iCalculated as equation (2):

step five: training a support vector machine improved by a training set, wherein the form of a training model is as follows (3):

wherein,is composed ofAs a kernel space mapping function, w_hAndweight vectors, C, for the historical data sample and the current data sample, respectively_hAnd C_cRespectively, the historical data sample and the current data sample regularization parameters, ξ_iAndrelaxation variables, ξ, respectively, for historical data samples_jAnd ξ_i ^*Respectively, the relaxation variable of the current data sample, λ is a balance parameter, which is a normal number, l_iFor the group tag calculated by equation (2),_iand'_iThe approximation accuracy of each group of samples in the historical data samples and the current data samples is respectively, and the calculation formula is as follows:

wherein p is_iCalculated from equation (1), Δ is a constant ranging from 0.01 to 0.1;

the above equation can be converted into a quadratic programming form as follows:

s.t.f^Tβ＝0 (5)

wherein

Is a kernel function, x_i、x_jFeature vectors for the ith and jth samples respectively,

the optimal solution w of w, b can be obtained by solving the equations (3) and (5)_c ^*，b_c ^*Wherein w is_c ^*Can be represented by formula (6):

b_c ^*can be represented by formula (7):

step six: constructing a prediction function, the prediction function having the form of equation (8):

wherein, w_c ^*And b_c ^*The optimal solution obtained in the step five;

step seven: normalizing user data to be predicted and inputting the normalized user data into the prediction function to obtain a prediction result of the purchase intention of the user, and if:

。