CN110060102B - Big data prediction method for user store location based on partial label learning - Google Patents

Abstract

The invention discloses a bias label learning-based method for predicting the positioning big data of a shop where a user is located, which comprises the following steps: 101, preprocessing the shopping state data of a user; 102, constructing a bias marking data set according to a candidate shop set corresponding to each sample; 103, carrying out feature extraction operation on the partial mark data set; 104, constructing a similarity graph according to the feature space; 105, carrying out probability propagation according to the similarity graph; 106 predict the shops with future behavioral interaction of the user from the candidate shop set of the partial mark data set by propagating the converged probabilities. According to the method, the historical data of the user are preprocessed, the characteristics are extracted, the bias mark data set is converted, the bias mark learning model is established, and the shops with behavior interaction in the future of the user are predicted from the candidate shop set corresponding to each user according to the bias mark data set of the position behavior of the user, so that the user can obtain more accurate personalized push service, and the shopping experience of the user is improved.

Description

Big data prediction method for user store location based on partial label learning

technical field

The invention belongs to the technical fields of partial mark learning and big data processing, in particular to big data prediction of store location where a user is based on a probability propagation model.

Background technique

Partial label learning is a weakly supervised learning in which the output space is associated with a set of candidate labels. Only one of the candidate labels is the true label, and the remaining labels are regarded as noise labels. In the process of partial labeling training, the true label of each training sample is submerged in the candidate label set, so it is impossible to directly obtain the input space from the data set to the output space of the learning algorithm similar to strong supervised learning. However, in real life, datasets with accurate unique label information are harder and harder to obtain. So we have to face the serious problem of how to learn from datasets that do not have uniformity and clarity. Recently, partial label learning has provided many effective methods to solve such problems, and has been widely used in many practical applications, especially in the problem of store location where users are located.

With the rapid popularization of Internet mobile payment, we enjoy the convenience of life brought by more and more intelligent positioning. For example, when a customer walks into a restaurant in the mall, the mobile phone will automatically pop up a coupon for the restaurant; when a customer walks into a clothing store in the mall, the mobile phone can automatically recommend the clothes you like in this store; when the customer passes by a jewelry store in the mall When going to the store, the mobile phone can automatically remind customers that a diamond ring that they have been thinking about for a long time is already in stock; when leaving the shopping mall parking lot, the mobile phone can automatically pay the parking fee with the customer's permission. The intimate services enjoyed by these customers are inseparable from the support of big data mining and machine learning behind them. The location analysis of the store where the customer is located not only implicitly brings the customer an artificial intelligence experience, but also makes it easier for the user to understand the information of the store they are interested in, thereby indirectly improving the purchasing power of the customer. How to provide users with the most effective services at the right time and at the right place is a new challenge for intelligent expansion in the era of big data.

SUMMARY OF THE INVENTION

The present invention aims to solve the above problems of the prior art. This paper proposes a big data prediction method for user's store location based on partial mark learning, which enables users to obtain more accurate personalized push services and improves users' shopping experience. The technical scheme of the present invention is as follows:

A big data prediction method for the location of a store where a user is located based on partial mark learning, comprising the following steps:

101. Perform preprocessing operations on the user's location behavior data, including cleaning abnormal samples and filling in missing Wi-Fi information;

102. According to the candidate store set corresponding to each sample, each sample in the data set is a shopping state corresponding to a certain user, and the different shopping states of each user correspond to different candidate store sets. The store set is obtained according to certain rules. For each sample, this rule can be summarized into three steps: 1. Find the 10 closest stores to the user's current shopping status according to the distance; 2. Optimize an innovative convex quadratic programming according to the distance To solve the importance of these 10 stores to the current shopping status of this user; 3. According to the importance, select the stores whose importance is greater than the threshold of 0.4 as the candidate store set, and construct a partial labeling data set;

103. Perform feature extraction on the partial labeled data set, and extract the features of the Wi-Fi distance intensity feature vector to form a feature space. The feature vector is similar to the ONE-HOT feature vector, and each dimension of the feature vector represents each occurrence of the data set. A distance strength value of Wi-Fi in the current shopping state of the user;

104. Build a similarity map according to the feature space, including:

For each sample _xi in the data set, do the same operation repeatedly: 1. Take _xi as a node of the similarity graph; 2. Take _xi as the center point, according to the relationship between _xi and the data set Euclidean distance of the Wi-Fi distance strength feature vector between other samples, select 10 samples with the smallest Euclidean distance for _xi , and then according to _xi and the selected 10 samples, _xi can be regarded as these 10 samples The center sample point of the sample, and its corresponding nodes in the graph are connected by edges in the similarity graph;

105. Carry out probability propagation according to the similarity map; for each sample _xi in the data set, do the same operation repeatedly: 1. Initialization: Calculate the optimal parameters according to the likelihood function (formula (6)), thereby calculating The probability that each candidate store in the candidate store set corresponding to _xi may interact, and this probability distribution is taken as the initialization probability distribution of each candidate store in the candidate store set corresponding to _xi ; 2. For the probability propagation algorithm The t-th iteration: According to the formula based on the similarity graph, the probability distribution of the candidate stores corresponding to the x _i of the t-th iteration is obtained, and the probability propagation of the t-th iteration is realized. The process of calculating this formula is a process of probability propagation. The propagation process can only realize the propagation between the two nodes corresponding to each edge in the similarity graph. Since the process of propagation may cause the interaction probability of shops that are not in the candidate shop set corresponding to x _i to be non-zero, Therefore, it is necessary to disambiguate and normalize the interaction probability of all shops relative to x _i . a. The interaction probability of the shops in the non-candidate shop set is set to 0; b. The interaction probability of the shops in the candidate shop set is the maximum and minimum normalization unify.

106. Using the probability converged by the probability propagation in step 105, predict from the set of candidate stores in the partial labeling data set the stores that the user will have behavioral interaction with in the future.

Further, in the step 101, the specific steps of preprocessing the user's shopping status data are as follows:

1011. Cleaning of abnormal samples: The cleaning of abnormal samples is firstly carried out through the latitude and longitude of the original data set and the Wi-Fi strength information of the current shopping state, according to the formula

τ_i分别为第i个样本所对应用户的经度、纬度和当前状态的Wi-Fi强度，m表示数据集样本数量，若某样本的异常置信度c_i低于0.15或者高于0.85，则将该样本判定为异常样本，并将其从原数据集中过滤出去；Calculate the anomaly confidence for each sample, where λ _i ,

τ _i are the longitude and latitude of the user corresponding to the ith sample and the Wi-Fi strength of the current state respectively, m is the number of samples in the dataset, if the abnormal confidence c _i of a sample is lower than 0.15 or higher than 0.85, then The sample is determined as an abnormal sample, and it is filtered out from the original data set;

1012. Filling of missing Wi-Fi information: First, find the 10 samples with the most similar latitude and longitude to the samples whose Wi-Fi strength information is missing, and the Wi-Fi strength information corresponding to these 10 samples are all known information. The similarity between samples according to the formula

λ_b,

分别为样本a所对应用户的经纬度和样本b所对应用户的经纬度，

分别为经度和纬度在整个数据集中的方差，然后通过这10个样本根据公式is calculated, where λ _a ,

λ _b ,

are the latitude and longitude of the user corresponding to sample a and the latitude and longitude of the user corresponding to sample b, respectively,

are the variances of longitude and latitude in the entire dataset, respectively, and then pass these 10 samples according to the formula

为样本a_i所对应的Wi-Fi强度信息。To fill the missing Wi-Fi strength information of the sample, where sample a is the sample to be filled, a _i (i=1,2,...,10) is the 10 neighbor samples of sample a,

is the Wi-Fi strength information corresponding to sample a _i .

Further, in the step 102, the specific steps of constructing the partial label data set according to the candidate store set corresponding to each sample are as follows:

其中λ_A,

分别表示店铺A的经纬度，λ_a,

分别表示样本a经纬度；(2)根据计算得到的距离d，选取与样本距离最近的10个商铺；(3)根据此样本所对应的距离最近的10个商铺的经纬度，对如下二次规划方程进行优化：For each sample in the original data, repeat the following operations to construct a partial labeled dataset: (1) Calculate the distance between the sample and each store based on the user's latitude and longitude and the store's latitude and longitude in the original data set

where λ _A ,

respectively represent the latitude and longitude of store A, λ _a ,

respectively represent the latitude and longitude of sample a; (2) according to the calculated distance d, select the 10 shops with the closest distance to the sample; (3) according to the latitude and longitude of the 10 shops with the closest distance corresponding to this sample, for the following quadratic programming equation optimize:

分别表示样本a所对应用户经纬度，ω_a,i(i＝1,2,…,10)分别表示距离样本a最近的10个商铺中的商铺i相对于样本a的权重值，

分别表示样本a所对应最近的10个商铺的经纬度，若计算出来的商铺所对应权重大于0.4，则将该商铺添加到该样本的候选商铺集合中。Solve the weight value of the 10 shops corresponding to this sample relative to this sample, where λ _a ,

respectively represent the latitude and longitude of the user corresponding to sample a, ω _a,i (i=1,2,...,10) respectively represent the weight value of store i in the 10 stores closest to sample a relative to sample a,

respectively represent the latitude and longitude of the nearest 10 shops corresponding to sample a. If the weight corresponding to the calculated shop is greater than 0.4, the shop is added to the set of candidate shops of the sample.

Further, the step 103 performs a feature extraction operation on the partial labeled data set, which specifically includes the steps:

Wi-Fi distance strength: First, the Wi-Fi name is discretized into a 1000-dimensional feature vector, and the feature value is the Wi-Fi strength corresponding to the Wi-Fi name, and then according to the conversion formula:

为第i个样本的1000维Wi-Fi距离强度特征向量，

为第i个样本的1000维Wi-Fi名所对应的Wi-Fi强度特征向量，|Y_i|为第i个样本对应的候选商铺集合的大小，

分别表示该样本对应候选商铺A_j的经纬度，λ_a,

分别表示该样本对应用户经纬度。Convert the discretized Wi-Fi strength feature vector to Wi-Fi distance strength feature vector, where

is the 1000-dimensional Wi-Fi distance intensity feature vector of the ith sample,

is the Wi-Fi intensity feature vector corresponding to the 1000-dimensional Wi-Fi name of the ith sample, |Y _i | is the size of the candidate store set corresponding to the ith sample,

respectively represent the latitude and longitude of the candidate shop A _j corresponding to the sample, λ _a ,

respectively represent the latitude and longitude of the user corresponding to the sample.

Further, the specific steps of constructing the similarity map according to the feature space in the step 104 are:

In order to construct the similarity graph <V, E, ω _e > based on the feature space, it is necessary to define the node V of the similar graph, the edge E of the similar graph and the edge weight ω _e of the similar graph respectively;

1041. Definition of the node of similarity graph: treat each sample in the partial labeled dataset as a node in the similarity graph;

1042. Definition of the edge of the similarity graph: For each sample in the partial labeled data set, that is, each node in the similarity graph, select the 10 samples other than itself with the closest Wi-Fi distance strength Euclidean distance as the association The object, that is to connect the corresponding two points in the similar graph, as the edge of the similar graph;

1043. Definition of edge weights of similar graphs: Similar(a, b) in formula (2) is used as the weights of edges (a, b) of similar graphs, where a and b are the two nodes in the similar graphs respectively. The corresponding two samples in the partially labeled dataset.

Further, the step 105 performs probability propagation according to the similarity map, and the specific steps are:

1051. Initialization probability: For each sample, first assume that the probability of a store appearing in its candidate store set is the proportion of the store in the entire data set, that is, the probability of a store appearing in the data set is taken as the probability of a store appearing in the sample candidate store set. , and assume that under the condition that the Wi-Fi distance strength of the ith sample appears, the probability that the store in the candidate set is the real label obeys the logistic distribution, and then according to the existing partial labeled data set, construct the Likelihood function:

即为在此样本的Wi-Fi距离强度特征向量出现的条件下，该商铺y在未来将会被此样本所对应用户的进行交互行为的概率，将作为概率传播的初始化概率；where p(y∈S _i | _xi ,θ) is the probability that the true label exists in the candidate store set of the sample under the condition that the Wi-Fi distance intensity vector of the ith sample appears, and n _y represents the store y The number of occurrences in the entire data set, π _i,y is the probability that store y appears in its candidate store set, p(y|x _i ,θ) is the condition for the occurrence of the Wi-Fi distance intensity vector in the ith sample Next, the store y is the probability of the true label. This likelihood function formalizes the known fact that the true label of each sample in the entire data set exists in the candidate store set, and the parameter value θ can be estimated by maximum likelihood make estimates, which

That is, under the condition that the Wi-Fi distance strength feature vector of this sample appears, the probability that the store y will be interacted by the user corresponding to this sample in the future will be used as the initialization probability of probability propagation;

1052. Probability propagation: in the t-th iteration of probability propagation, according to the probability matrix F _t-1 of the previous iteration and the initialization probability matrix P=[p(y _i =j|x _i ,θ)] _{m× q} , a new round of probability matrix F _t affected by domain sample propagation can be obtained:

Among them, W∈R ^m×m is the similarity matrix between samples, and probability propagation is iterated for 50 rounds in total. In each round of probability propagation, the shop interaction probability corresponding to each sample is based on the similarity between samples. It is propagated to its corresponding neighbor examples, and each sample updates its own interaction probability for this store according to the store interaction probability corresponding to its 10 neighbor samples.

Further, in the partial label learning problem, each iteration needs to perform a disambiguation operation on the updated probability matrix, that is, set the shop interaction probability in each sample non-candidate shop set to 0, and set the shop interaction probability in the candidate shop set to 0. The interaction probability is normalized:

Further, in the step 106, by propagating the converged probability, the specific steps of predicting the shops that the user will interact with in the future from the candidate shop set of the partial marker dataset are as follows:

According to the probability matrix F _t obtained by the propagation and convergence in step 105, the predicted store corresponding to each sample that is most likely to interact with the user can be obtained:

The advantages and beneficial effects of the present invention are as follows:

1. For the store location application itself, the most commonly used prediction method is the basic multi-class machine learning method, and the multi-class method consumes a lot of resources, and the possible labels of each sample should be a subset of all labels, that is, each sample. The true labels of are only possible to appear in a few labels, instead of the multi-classification method treating all labels as possible true labels, which will lead to insufficient accuracy of the multi-classification method. Therefore, this patent innovatively regards the store location application as a partial label learning method for prediction, which can make full use of the label information of only those few stores that each sample may interact with for prediction, which greatly improves the accuracy of the model;

2. In the abnormal sample cleaning step, considering the fact that the samples in the data set are all in the same business circle, this patent innovatively creates a data related to the latitude and longitude of the user corresponding to the sample and the Wi-Fi strength of the current shopping state. The abnormal confidence level is to clean out the samples that deviate from the average confidence level in the data set or whose confidence level is too low.

3. According to the principle that the latitude and longitude of the users corresponding to different samples in the store positioning application are more similar, the shopping states they are in should be more similar. This patent innovatively creates a similarity formula based on this principle to express different Similarity between samples, this similarity has two functions in this patent: (1) According to the 10 samples with the lowest similarity to the samples with missing Wi-Fi information, to fill the missing of the samples with missing Wi-Fi information (2) Similarity can be used as the similarity graph, the edge weights between samples.

4. In the process of constructing the partial labeled dataset, the conventional construction method can only search for the 10 stores closest to the user corresponding to the sample. However, this will bring too much noise to the partial labeled dataset, so we It is also necessary to screen the 10 nearest stores. This patent innovatively creates a quadratic programming equation related to the latitude and longitude of the store and the latitude and longitude of the user corresponding to the sample. The interaction weight is used as the solution variable. According to the optimal solution variable corresponding to the optimized quadratic programming equation, the stores with the closest relative distance (relative to the other 9 stores) of the current shopping status of the user can be selected as much as possible, which can greatly reduce the The noise value caused by the large size of the candidate label set for the partially labeled dataset.

5. In the process of feature extraction, this patent captures the average distance between the user corresponding to each sample in the store location application and each store in the candidate store set, and can compare the candidate stores in each sample with the non-candidate stores. The characteristics of the shops are well differentiated, and at the same time, considering the problem that the average distance cannot well distinguish the shops in the candidate shop set, the Wi-Fi strength corresponding to each sample is combined with the average distance, and innovatively proposed With the Wi-Fi distance strength vector feature, while distinguishing candidate stores from non-candidate stores, the distinction between the stores in the candidate store set is guaranteed.

6. In the process of probability propagation, this patent transforms the classic label propagation algorithm. The classic label propagation algorithm only considers the appearance and non-appearance of candidate stores, but does not consider the potential probability distribution of the candidate store set, so the classic label propagation algorithm cannot achieve satisfactory expressiveness. The structure of the label propagation algorithm, on this basis, the probability propagation algorithm proposed in this patent, according to the maximum likelihood estimation based on the logistic distribution, to mine the probability distribution of the candidate store set corresponding to each sample, and then estimate the Put the probability distribution of the probability distribution into the frame of the label propagation algorithm, and innovatively propose disambiguation normalization (a, the interaction probability for the stores in the non-candidate store set is set to 0; b, the interaction probability for the stores in the candidate store set is set to 0; Perform maximum and minimum normalization) to optimize the problem that the probability of non-candidate shops is not 0 during the propagation process. In essence, the probability propagation algorithm solves the disadvantage that the label propagation algorithm can only perform data mining at the data surface layer, and greatly improves the prediction results of partial label learning.

Description of drawings

FIG. 1 is a flow chart of a method for predicting the location of a store where a user is located based on a big data according to a preferred embodiment of the present invention.

FIG. 2 is a sample similarity graph in a big data prediction method for the location of a store where a user is located based on partial mark learning according to a preferred embodiment of the present invention.

FIG. 3 is an overall framework diagram of the actual application of a partial mark learning model in a big data prediction method for the location of a store where a user is located based on partial mark learning according to a preferred embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be described clearly and in detail below with reference to the accompanying drawings in the embodiments of the present invention. The described embodiments are only some of the embodiments of the invention.

The technical scheme that the present invention solves the above-mentioned technical problems is:

Referring to FIG. 1, FIG. 1 is a flowchart of a method for predicting the location of a store where a user is located based on partial mark learning according to Embodiment 1 of the present invention, which specifically includes:

101. Perform preprocessing operations on the user's shopping status data, as follows: 1011. Abnormal sample cleaning: The abnormal samples are cleaned first through the latitude and longitude of the user corresponding to the sample in the original data set and the Wi-Fi strength information of the current state, according to the formula ( 1) Calculate the abnormal confidence of each sample. If the abnormal confidence c _i of a sample is lower than 0.15 or higher than 0.85, we will judge the sample as an abnormal sample and filter it out from the original data set; 1012. Filling of missing Wi-Fi information: Due to force majeure factors, the Wi-Fi strength information of some samples cannot be accurately obtained. According to the samples with similar latitude and longitude, the Wi-Fi strength information should also be similar. First, find the distance from the Wi-Fi The 10 samples with the most similar longitude and latitude corresponding to the samples with missing strength information, and the Wi-Fi strength information corresponding to these 10 samples are all known information, and the similarity between the two samples is calculated according to formula (2). , and then use the 10 samples to fill in the missing Wi-Fi strength information of the sample according to formula (3).

(其中λ_A,

分别表示店铺A的经纬度，λ_a,

分别表示用户a经纬度)；(2)跟据计算得到的距离d，选取与样本距离最近的10个商铺；(3)根据此样本所对应的距离最近的10个商铺的经纬度，对二次规划方程(公式(4))进行优化，求解该样本所对应的10个商铺相对此样本的权重值，若计算出来的商铺所对应权重大于0.4，则将该商铺添加到该样本的候选商铺集合中。102. Construct a partial label dataset according to the candidate store set corresponding to each user, as follows: For each sample in the original data, repeat the following operations to construct a partial label dataset: (1) According to the user's latitude and longitude in the original dataset And the latitude and longitude of the store, calculate the distance between the sample and each store

(where λ _A ,

respectively represent the latitude and longitude of store A, λ _a ,

(2) According to the calculated distance d, select the 10 shops with the closest distance to the sample; (3) According to the latitude and longitude of the 10 shops with the closest distance corresponding to this sample, for the secondary planning Equation (formula (4)) is optimized to solve the weight value of the 10 shops corresponding to the sample relative to this sample. If the weight corresponding to the calculated shop is greater than 0.4, the shop is added to the set of candidate shops for the sample. .

103. According to the user's latitude and longitude and Wi-Fi strength information, perform feature extraction operation on the partial marked data set, as follows: First, the Wi-Fi name is discretized into a 1000-dimensional feature vector, and the feature value is the corresponding Wi-Fi name. Wi-Fi strength, and then convert the discretized Wi-Fi strength feature vector into a Wi-Fi distance strength feature vector according to the conversion formula (5).

104. Construct the similarity graph according to the feature space, as follows: In order to construct the similarity graph based on the feature space <V, E, ω _e > (see Figure 2), it is necessary to define the node V of the similarity graph and the edge of the similarity graph respectively. E and the edge weights ω _e of similar graphs.

1041. Definition of Nodes in Similarity Graphs: Treat each sample in a partially labeled dataset as a node in a similarity graph.

1042. Definition of the edge of the similarity graph: For each sample in the partial labeled dataset (each node in the similarity graph), select the 10 samples other than itself with the closest Wi-Fi distance strength Euclidean distance ( node) as an association object, that is, to connect two corresponding points in the similar graph as the edge of the similar graph.

1043. Definition of edge weights of similar graphs: Similar(a, b) in formula (2) is used as the weights of edges (a, b) of similar graphs, where a and b are the two nodes in the similar graphs respectively. The corresponding two samples in the partially labeled dataset.

105. Probabilistic propagation according to the similarity map, as follows:

即为在此样本的Wi-Fi距离强度特征向量出现的条件下，该商铺y在未来将会被此样本所对应用户的进行交互行为的概率，将作为概率传播的初始化概率。1051. Initialization probability: For each sample, first assume that the probability of a store appearing in its candidate store set is the proportion of the store in the entire data set, that is, the probability of a store appearing in the data set is taken as the probability of a store appearing in the sample candidate store set. and further assume that under the condition that the Wi-Fi distance strength of the ith sample appears, the probability that the store in the candidate set is the real label obeys the logistic distribution, and then according to the existing partial labeled data set, construct Out of the likelihood function is Equation (6), a likelihood function formalizes the known fact that the true label of each sample in the entire dataset exists in the candidate store set. The parameter value θ can be estimated by maximum likelihood estimation, where

That is, under the condition that the Wi-Fi distance strength feature vector of this sample appears, the probability that the store y will be interacted with by the user corresponding to this sample in the future will be used as the initialization probability of probability propagation.

1052. Probability propagation: in the t-th iteration of probability propagation, according to the probability matrix F _t-1 of the previous iteration and the initialization probability matrix P=[p(y _i =j|x _i ,θ)] _{m× q} , a new round of probability matrix F _t affected by domain sample propagation can be obtained as formula (7). The probability propagation is iterated for 50 rounds in total. In each round of probability propagation, the shop interaction probability corresponding to each sample is propagated to its corresponding neighboring examples according to the similarity between the samples, and each sample is based on the shop interaction probability corresponding to its 10 neighboring samples. Update your probability of interacting with this store. In the partial label learning problem, each iteration needs to perform a disambiguation operation on the updated probability matrix, that is, set the shop interaction probability in each sample non-candidate shop set to 0, and perform the shop interaction probability in the candidate shop set. The normalization operation is shown in Equation (8).

106. By propagating the converged probability, predict from the candidate store set of the partial marker data set the stores that the user will interact with in the future. The details are as follows: According to the probability matrix F _t obtained by the propagation and convergence in step 105, each store can be obtained. The predicted stores corresponding to the samples that are most likely to interact with users are shown in formula (9). The probabilistic propagation method based on partial tagging enables users to obtain more accurate personalized push services, improves the user's shopping experience, and becomes an effective way to solve the problem of difficult tag acquisition today. The overall frame diagram of the practical application of the partial label learning model based on big data in the location of the user's store is shown in Figure 3.

The above embodiments should be understood as only for illustrating the present invention and not for limiting the protection scope of the present invention. After reading the contents of the description of the present invention, the skilled person can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims (4)

1. a method for predicting big data based on the location of the user's shop based on partial mark learning, is characterized in that, comprises the following steps: 101. Perform preprocessing operations on the user's location behavior data, including cleaning abnormal samples and filling in missing Wi-Fi information; 102. According to the candidate store set corresponding to each sample, each sample in the data set is a shopping state corresponding to a certain user, and the different shopping states of each user correspond to different candidate store sets. The store set is obtained according to certain rules. For each sample, this rule is summarized into three steps: 1. Find the 10 closest stores to the user's current shopping status according to the distance; 2. According to the convex quadratic programming problem, to solve this The importance of 10 stores to the current shopping status of the user; 3. According to the importance, select stores with a weight value greater than the threshold of 0.4 as the candidate store set, and construct a partial labeling dataset; 103. Perform feature extraction on the partial labeled dataset, and extract the features of the Wi-Fi distance intensity feature vector to form a feature space. Each dimension of the feature vector represents the current shopping state of the user for each Wi-Fi that appears in the dataset. The distance strength value under ; 104. Build a similarity map according to the feature space, including: For each sample _xi in the data set, do the same operation repeatedly: 1. Take _xi as a node of the similarity graph; 2. Take _xi as the center point, according to the relationship between _xi and the data set The Euclidean distance of the Wi-Fi distance intensity feature vector between other samples, select 10 samples with the smallest Euclidean distance for _xi , and then according to _xi and the selected 10 samples, _xi is regarded as these 10 samples The center sample point of , connect its corresponding nodes in the graph with edges in the similarity graph; 105. Probability propagation according to the similarity map; for each sample _xi in the data set, repeat the same operation: 1. Initialization: Calculate the optimal parameters according to the likelihood function, so as to calculate the candidate corresponding to _xi The probability that each candidate store in the store set may interact, and this probability distribution is taken as the initialization probability distribution of each candidate store in the candidate store set corresponding to x _i ; 2. For the t-th iteration of the probability propagation algorithm: according to Based on the formula of the similarity graph, the probability distribution of the candidate stores corresponding to x _i of the t-th iteration is obtained, and the probability propagation of the t-th iteration is realized. The process of calculating this formula is a process of probability propagation, and this propagation process can only achieve similar For the propagation between the two nodes corresponding to each edge in the graph, since the process of propagation may cause the interaction probability of the shops that are not in the set of candidate shops corresponding to x _i to be non-zero, it is necessary to compare all shops relative to each other. Disambiguation normalization is performed on the interaction probability of x _i , a. The interaction probability of the shops in the non-candidate shop set is set to 0; b, the interaction probability of the shops in the candidate shop set is normalized to the maximum and minimum; 106. Through the probability of probability propagation in step 105, the converged probability is predicted from the candidate store set of the partial marker data set to predict the store that the user will have behavioral interaction in the future; The specific steps of preprocessing the user's shopping status data in step 101 are as follows: 1011. Cleaning of abnormal samples: The cleaning of abnormal samples is firstly carried out through the latitude and longitude of the original data set and the Wi-Fi strength information of the current shopping state, according to the formula

Calculate the anomaly confidence for each sample, where λ _i ,

τ _i are the longitude and latitude of the user corresponding to the ith sample and the Wi-Fi strength of the current state respectively, m is the number of samples in the dataset, if the abnormal confidence c _i of a sample is lower than 0.15 or higher than 0.85, then The sample is determined as an abnormal sample, and it is filtered out from the original data set; 1012. Filling of missing Wi-Fi information: First, find the 10 samples with the most similar latitude and longitude to the samples whose Wi-Fi strength information is missing, and the Wi-Fi strength information corresponding to these 10 samples are all known information. The similarity between samples according to the formula

is calculated, where λ _a ,

λ _b ,

are the latitude and longitude of the user corresponding to sample a and the latitude and longitude of the user corresponding to sample b, respectively,

are the variances of longitude and latitude in the entire dataset, respectively, and then pass these 10 samples according to the formula

To fill the missing Wi-Fi strength information of the sample, where sample a is the sample to be filled, a _i (i=1,2,...,10) is the 10 neighbor samples of sample a,

is the Wi-Fi strength information corresponding to sample a _i ; The specific steps of constructing a partial label data set according to the candidate store set corresponding to each sample in the step 102 are as follows: For each sample in the original data, repeat the following operations to construct a partial labeled dataset: (1) Calculate the distance between the sample and each store based on the user's latitude and longitude and the store's latitude and longitude in the original data set

where λ _A ,

respectively represent the latitude and longitude of store A, λ _a ,

respectively represent the latitude and longitude of sample a; (2) according to the calculated distance d, select the 10 shops with the closest distance to the sample; (3) according to the latitude and longitude of the 10 shops with the closest distance corresponding to this sample, for the following quadratic programming equation optimize:

Solve the weight value of the 10 shops corresponding to this sample relative to this sample, where λ _a ,

respectively represent the latitude and longitude of the user corresponding to sample a, ω _a,i (i=1,2,...,10) respectively represent the weight value of store i in the 10 stores closest to sample a relative to sample a,

Respectively represent the latitude and longitude of the nearest 10 shops corresponding to sample a. If the weight corresponding to the calculated shop is greater than 0.4, the shop will be added to the set of candidate shops of the sample; The step 103 performs a feature extraction operation on the partial labeled data set, which specifically includes the steps: Wi-Fi distance strength: First, the Wi-Fi name is discretized into a 1000-dimensional feature vector, and the feature value is the Wi-Fi strength corresponding to the Wi-Fi name, and then according to the conversion formula:

Convert the discretized Wi-Fi strength feature vector to Wi-Fi distance strength feature vector, where

is the 1000-dimensional Wi-Fi distance intensity feature vector of the ith sample,

is the Wi-Fi intensity feature vector corresponding to the 1000-dimensional Wi-Fi name of the ith sample, |Y _i | is the size of the candidate store set corresponding to the ith sample,

respectively represent the latitude and longitude of the candidate shop A _j corresponding to the sample, λ _a ,

respectively represent the latitude and longitude of the user corresponding to the sample; The step 105 performs probability propagation according to the similarity map, and the specific steps are: 1051. Initialization probability: For each sample, first assume that the probability of a store appearing in its candidate store set is the proportion of the store in the entire data set, that is, the probability of a store appearing in the data set is taken as the probability of a store appearing in the sample candidate store set. , and assume that under the condition that the Wi-Fi distance strength of the ith sample appears, the probability that the store in the candidate set is the real label obeys the logistic distribution, and then according to the existing partial labeled data set, construct the Likelihood function:

where p(y∈S _i | _xi ,θ) is the probability that the true label exists in the candidate store set of the sample under the condition that the Wi-Fi distance intensity vector of the ith sample appears, and n _y represents the store y The number of occurrences in the entire data set, π _i,y is the probability that store y appears in its candidate store set, p(y|x _i ,θ) is the condition for the occurrence of the Wi-Fi distance intensity vector in the ith sample Below, the store y is the probability of the true label. This likelihood function formalizes the known fact that the true label of each sample in the entire dataset exists in the candidate store set, and the parameter value θ is estimated by maximum likelihood. estimated, of which

That is, under the condition that the Wi-Fi distance strength feature vector of this sample appears, the probability that the store y will be interacted by the user corresponding to this sample in the future will be used as the initialization probability of probability propagation; 1052. Probability propagation: in the t-th iteration of probability propagation, according to the probability matrix F _t-1 of the previous iteration and the initialization probability matrix P=[p(y _i =j|x _i ,θ)] _{m× q} , a new round of probability matrix F _t affected by domain sample propagation can be obtained:

Among them, W∈R ^m×m is the similarity matrix between samples, and probability propagation is iterated for 50 rounds in total. In each round of probability propagation, the shop interaction probability corresponding to each sample is based on the similarity between samples. It is propagated to its corresponding neighbor examples, and each sample updates its own interaction probability for this store according to the store interaction probability corresponding to its 10 neighbor samples. 2. The big data prediction method for the location of the store where the user is based on partial mark learning according to claim 1, is characterized in that, the concrete steps of described step 104 constructing similarity degree map according to feature space are: In order to construct the similarity graph <V, E, ω _e > based on the feature space, it is necessary to define the node V of the similar graph, the edge E of the similar graph and the edge weight ω _e of the similar graph respectively; 1041. Definition of the node of similarity graph: treat each sample in the partial labeled dataset as a node in the similarity graph; 1042. Definition of the edge of the similarity graph: For each sample in the partial labeled data set, that is, each node in the similarity graph, select the 10 samples other than itself with the closest Wi-Fi distance strength Euclidean distance as the association The object, that is to connect the corresponding two points in the similar graph, as the edge of the similar graph; 1043. Definition of edge weights of similar graphs: Similar(a, b) in formula (2) is used as the weights of edges (a, b) of similar graphs, where a and b are the two nodes in the similar graphs respectively. The corresponding two samples in the partially labeled dataset. 3. The method for predicting the location of a store where the user is based on partial mark learning according to claim 2, is characterized in that, in the partial mark learning problem, each iteration needs to carry out a disambiguation operation to the updated probability matrix, That is, set the shop interaction probability in each sample non-candidate shop set to 0, and normalize the shop interaction probability in the candidate shop set:

4. The big data prediction method for the location of the shop where the user is based on partial mark learning according to claim 3, is characterized in that, described step 106 predicts from the candidate shop set of partial mark data set by spreading the probability of convergence The specific steps for a store where users have behavioral interaction in the future are: According to the probability matrix F _t obtained by the propagation and convergence in step 105, the predicted store corresponding to each sample that is most likely to interact with the user can be obtained:

Images