CN108846692A - A kind of consumer spending behavior prediction method based on multifactor Recognition with Recurrent Neural Network - Google Patents

Abstract

The present invention relates to a kind of consumer behavior prediction method based on multi-factor cyclic neural network, comprising: 1) data preprocessing obtains training data, carries out standardization processing to four sequences S, H, C, T, obtains training data X= [SHCT]; 2) Generate a recurrent neural network model, including input layer X, hidden layer Z, and output layer Y; 3) Use BPTT algorithm and training data to train the model, and use BPTT algorithm to update iteratively according to the gradient descent method weight values (U, V, W) and two bias values (b _z , b _y ); 4) Use the trained neural network to calculate the output sequence Y and restore it to the final prediction result s _T+1 . The beneficial effects of the present invention are: the present invention considers some natural factors that will affect consumption behavior, and can accurately and efficiently predict the offline consumption behavior of different users; in addition to the three natural factors proposed by the present invention, this method also Can expand and continue to introduce other factors that may affect users' offline consumption behavior.

Description

A Method for Predicting Consumer Behavior Based on Multi-factor Recurrent Neural Network

technical field

The invention relates to a method for predicting offline consumption behavior of consumers based on a multi-factor cyclic neural network, which mainly involves processing historical consumption data of consumers and predicting the most likely future consumption of the user through a multi-factor cyclic neural network. choose.

Background technique

With the continuous development of e-commerce technology, users will generate a large amount of consumption data when they consume, and these data contain a lot of useful information. By analyzing this information, the consumption habits of each consumer can be analyzed to predict the possible future consumption behavior of consumers, which is very important for providing better personalized recommendation services for users and optimizing business and management strategies for businesses.

In daily life, there are some natural factors that will affect people's consumption behavior. For example, people tend to consume at restaurants and stores at work on weekdays, but are more likely to consume at stores near home during holidays; people like to consume iced drinks on hot days, but prefer to consume hot drinks on cold days.

How to make better use of historical consumption data and combine some external factors that can affect user behavior, so as to effectively and quickly predict the future consumption behavior of consumers, is an urgent problem to be solved by those skilled in the art.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies in the prior art and provide a method for predicting consumer behavior based on multi-factor recurrent neural networks.

This consumer consumption behavior prediction method based on the multi-factor cyclic neural network includes the following steps:

Step 1. Data preprocessing to obtain training data:

Extract the consumer's historical merchant consumption data {shop ₁ , shop ₂ , ..., shop _T } with a length of T, and number the offline merchants of the consumer's historical consumption, and then convert the consumption data according to the merchant number according to the time sequence Form a merchant sequence S(s ₁ s ₂ ...s _T );

According to the local holiday arrangement, use 1 to represent the holiday, 0 to represent the working day, and then generate the holiday sequence H(h ₁ h ₂ ... h _T ) corresponding to the time according to the consumer's business sequence S;

According to common weather conditions and their severity, the weather is divided into the following different situations and given different labels {Sunny: 0, Light Rain: -0.5, Heavy Rain: -1, Light Snow: -1.5, Heavy Snow: -2} ;

According to the consumer's business sequence S, extract the weather data corresponding to the time and place, and generate the weather sequence C(c ₁ c ₂ ... c _T );

According to the consumer's business sequence S, extract the temperature data corresponding to the time and place, and generate the temperature sequence T(t ₁ t ₂ ...t _T );

Then standardize the four sequences S, H, C, T to obtain the training data X=[SHCT];

Step 2. Generate a recurrent neural network model:

Generate a three-layer cyclic neural network model, including input layer X, hidden layer Z and output layer Y;

Wherein the input data of the input layer X is a four-dimensional vector X=[SHCT];

In the hidden layer Z, the state z _t at the current moment t not only depends on the input data x _t at the current moment, but also depends on the hidden layer state z _t-1 at the previous moment: z _t =f(Ux _t +Wz _{t -1} +b _z ), where U is the weight between the input layer and the hidden layer, W is the weight between the hidden layer and the hidden layer, b _z is the bias value, and f is the activation function;

The output layer Y is a fully connected layer, each node of which is connected to each node of the hidden layer: y _t =g(Vz _t +b _y ), where y _t is the value of the output layer at time t, V is the hidden layer Contains the weight of the layer and the output layer, b _y is the deviation value, and g is the activation function;

Step 3. Use the BPTT algorithm and training data to train the model:

Use the BPTT algorithm to iteratively update three weight values (U, V, W) and two bias values (b _z , b _y ) according to the gradient descent method;

Step 4: Use the trained neural network to calculate the output sequence Y and restore it to the final prediction result s _T+1 .

As a preference: the specific calculation method of the weight value (U, V, W) and the deviation value (b _z , b _y ) in step 3 is:

Define e _t as the error of each step of the neural network, then the error of the whole function is E=∑ _t e _t ;

The gradient of the weight V is Among them, y′ _t is the supervised value at time t, that is, the actual value;

Define two operators and

The gradient of the weight U is

The gradient of the weight W is

The gradient of the bias value b _z is

The gradient of the bias value b _y is

The parameters are updated iteratively according to the stochastic gradient descent method.

The beneficial effects of the present invention are: the present invention considers some natural factors that will affect consumption behavior, and can accurately and efficiently predict the offline consumption behavior of different users; in addition to the three natural factors proposed by the present invention, this method also Can expand and continue to introduce other factors that may affect users' offline consumption behavior.

Description of drawings

Fig. 1 is the frame diagram of multi-factor recurrent neural network model;

Figure 2 is a comparison chart of the forecasting effect of different forecasting methods;

Figure 3 is a comparison chart of considering different natural factors to improve the prediction effect.

Detailed ways

The present invention will be further described below in conjunction with the examples. The description of the following examples is provided only to aid the understanding of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, some improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

As shown in Figure 1, the described consumer consumption behavior prediction method based on multi-factor recurrent neural network, its steps are as follows:

1. Data preprocessing to obtain training data:

Extract the consumer's historical merchant consumption data {shop ₁ , shop ₂ , ..., shop _T } with a length of T, and number the offline merchants of the consumer's historical consumption, and then convert the consumption data according to the merchant number according to the time sequence Form a merchant sequence S(s ₁ s ₂ ...s _T );

According to the local holiday arrangement, use 1 to indicate a holiday and 0 to indicate a working day. Then according to the consumer's merchant sequence S, generate the holiday sequence H(h ₁ h ₂ ... h _T ) corresponding to the time;

According to common weather conditions and their severity, the weather is divided into the following different situations and given different labels {Sunny: 0, Light Rain: -0.5, Heavy Rain: -1, Light Snow: -1.5, Heavy Snow: -2} .

According to the consumer's business sequence S, extract the weather data corresponding to the time and place, and generate the weather sequence C(c ₁ c ₂ ... c _T );

According to the consumer's business sequence S, extract the temperature data corresponding to the time and place, and generate the temperature sequence T(t ₁ t ₂ ...t _T );

Then standardize the four sequences S, H, C, T to obtain the training data X=[SHCT].

2. Generate a recurrent neural network model:

Generate a three-layer cyclic neural network model, including input layer X, hidden layer Z, and output layer Y;

Wherein the input data of the input layer X is a four-dimensional vector X=[SHCT];

In the hidden layer Z, the state z _t at the current moment t not only depends on the input data x _t at the current moment, but also depends on the hidden layer state z _t-1 at the previous moment: z _t =f(Ux _t +Wz _{t -1} +b _z ), where U is the weight between the input layer and the hidden layer, W is the weight between the hidden layer and the hidden layer, b _z is the bias value, and f is the activation function;

The output layer Y is a fully connected layer, each node of which is connected to each node of the hidden layer: y _t =g(Vz _t +b _y ), where y _t is the value of the output layer at time t, V is the hidden layer Contains the weight of the layer and the output layer, b _y is the deviation value, and g is the activation function;

3. Use the BPTT algorithm and training data to train the model:

e _t is the error of each step of the neural network, then the error of the whole function is E=∑ _t e _t ;

The gradient of the weight V is Among them, y′ _t is the supervised value at time t, that is, the actual value;

Define two operators and

The gradient of the weight U is

The gradient of the weight W is

The gradient of the bias value b _z is

The gradient of the bias value b _y is

Iteratively update three weight values and two bias values according to the stochastic gradient descent method;

4. Use the trained neural network to calculate the output value Y and restore it to the final prediction result s _T+1 .

Experiments and results: In order to verify the prediction effect of the method, we conducted experiments on a real data set. The dataset contains 1057 different users and 2000 merchants, and each user has more than 120 consumption records. We first calculated the information entropy of each user's consumption sequence, and then divided the 1057 users into three groups according to the size of the information entropy: the first group contained 448 users, and the information entropy of these user consumption sequences was 0-0.5; the second group It contains 296 users whose consumption sequence information entropy is 0.5-1.0; the third group contains 313 users whose consumption sequence information entropy is greater than 1. The higher the information entropy, the greater the amount of information contained in the user's consumption sequence, the more difficult it is to predict, so the third group of users among the three groups of users is the most difficult to predict. We compare the multi-factor recurrent neural network prediction method (MF-RNN for short) with the most frequent item prediction method (most frequent, MF for short), hidden Markov prediction method (HMM for short), and long-short memory network (LSTM for short). ,as shown in picture 2. Among all three groups of users, MF-RNN achieved the best prediction accuracy, and in the third group, which is the most difficult to predict, the prediction accuracy of MF-RNN method was 83.45%; the prediction accuracy of MF method Only 57.51%; the prediction accuracy of the HMM method is 59.42%; the prediction accuracy of the LSTM method is 76.74%. It can be seen that compared with the other three prediction methods, our method can achieve better prediction results. And to verify the influence of different natural factors on the prediction results, we combined different natural factors to conduct experiments on the third group of user data, as shown in Figure 3. Experiments show that the prediction accuracy rate without introducing natural factors is 81%, indicating that considering the impact of natural factors on offline consumer behavior can improve the accuracy of offline consumer behavior prediction; the introduction of three natural factors can improve the prediction effect Function; the more factors are introduced, the higher the prediction accuracy will be.

Claims (2)

1. a kind of consumer spending behavior prediction method based on multifactor Recognition with Recurrent Neural Network, which is characterized in that including following Step：

Step 1: data prediction obtains training data：

The length for extracting consumer is history businessman's consumption data { shop of T₁, shop₂..., shoop_TAnd to consumer's History consumption offline businesses are numbered, and consumption data is then converted into businessman's sequence according to merchant number according to time sequencing S(s₁s₂...s_T)；

It is arranged according to local festivals or holidays, indicates festivals or holidays using 1,0 indicates working day, then according to businessman's sequence of consumer S generates the festivals or holidays sequence H (h of corresponding time₁h₂...h_T)；

According to common weather conditions and its severity, weather is divided into following several different situations, and give respectively different Label { fine day：0, light rain：- 0.5, heavy rain：- 1, slight snow：- 1.5, heavy snow：-2}；

According to the businessman sequence S of consumer, the weather data of corresponding when and where is extracted, weather sequence C is generated (c₁c₂...c_T)；

According to the businessman sequence S of consumer, the temperature data of corresponding when and where is extracted, temperature sequence T is generated (t₁t₂...t_T)；

Then four sequences S, H, C, T are standardized, obtain training data X=[SHCT]；

Step 2: generating a Recognition with Recurrent Neural Network model：

The Recognition with Recurrent Neural Network model an of three-decker is generated, includes input layer X, hidden layer Z and output layer Y；

Wherein the input data of input layer X is exactly a four dimensional vector X=[SHCT]；

In hidden layer Z, the state z of current time t_tDepend not only on the input data x at current time_t, when also depending on previous The implicit layer state z carved_t-1：z_t=f (Ux_t+Wz_t-1+b_z), wherein U is the weight between input layer and hidden layer, and W is hidden layer Weight between hidden layer, b_zIt is deviation, f is activation primitive；

Output layer Y is a full articulamentum, its each node is connected with each node of hidden layer：y_t=g (Vz_t+b_y), Middle y_tIt is the value of t moment output layer, V is the weight of hidden layer and output layer, b_yIt is deviation, g is activation primitive；

Step 3: being trained using BPTT algorithm and training data to model：

Three weighted values (U, V, W) and two deviation (b are updated according to gradient descent method iteration using BPTT algorithm_z,b_y)；

Step 4: calculating output sequence Y, and be reduced into final prediction result s using trained neural network_T+1。

2. the consumer spending behavior prediction method according to claim 1 based on multifactor Recognition with Recurrent Neural Network, special Sign is, weighted value (U, V, W) and deviation (b in step 3_z,b_y) circular is：

Define e_tFor the error of each step of neural network, then the error of entire function is E=∑_te_t；

The gradient of weight V isWherein y '_tFor supervision value, that is, actual value of t moment；

Define two operatorsWith

The gradient of weight U is

The gradient of weight W is

Deviation b_zGradient be

Deviation b_yGradient be

According to stochastic gradient descent method iteration undated parameter.