WO2019165944A1

WO2019165944A1 - Transition probability network based merchant recommendation method and system thereof

Info

Publication number: WO2019165944A1
Application number: PCT/CN2019/076049
Authority: WO
Inventors: 李旭瑞; 郑建宾; 赵金涛
Original assignee: 中国银联股份有限公司
Priority date: 2018-02-28
Filing date: 2019-02-25
Publication date: 2019-09-06
Also published as: CN108573399A; CN108573399B

Abstract

A transition probability network based merchant recommendation method and a system thereof. The method comprises: carrying out OneHot coding on a number N of merchants, wherein each merchant is mapped to an N-dimensional sparse vector; for a prescribed account, recording and sorting merchant codes corresponding to relevant transactions so as to construct a consumption/merchant sequence, wherein the merchant codes are represented by vectors after OneHot coding; constructing a neural network, wherein the vector after OneHot coding of each merchant is taken as an input layer, and the occurrence probability distribution of merchants that may appear after the current merchant is taken as an output layer; and recommending merchants to a user on the basis of the merchant occurrence probability output by the transition probability network construction steps. The described method proposes an architecture of a transition-probability three-layer neural network; according to the consumption sequence of massive amounts of users, the relationships between merchant sequences can be quickly analyzed and more accurate recommendation thus can be made.

Description

Merchant recommendation method and system based on transition probability network

Technical field

The present invention relates to data processing, and in particular to a merchant recommendation method and system thereof for calculating a transition probability by constructing a neural network.

Background technique

When users are spending, they are accustomed to first obtaining the merchant information from the network, and then selecting the interested merchants to consume, even if the online consumption scene is the same. Some Internet websites also frequently recommend merchants to users, thereby saving users' shopping time, improving efficiency, and better improving the user service experience.

However, many merchant recommendation functions are to recommend merchants by popularity. This leads to a large number of users receiving similar recommendation information, and there is no recommendation for the characteristics of the user and the merchant. In this way, users will still have a lot of confusion and find it difficult to find a business that suits them. Through the search, some companies have also used some machine learning methods to improve the efficiency of merchant recommendation, but most of these solutions are based on the accurate acquisition of various user privacy data.

The information disclosed in the Background of the Invention is only intended to provide an understanding of the general background of the invention, and should not be construed as an admission or in any way.

Summary of the invention

In view of the above problems, the present invention aims to provide a merchant recommendation method based on transition probability network and a merchant recommendation based on transition probability network, which can quickly analyze the association between merchant sequences according to the consumption sequence of a large number of users for more accurate recommendation. system.

The merchant recommendation method based on the transition probability network of the present invention is characterized in that it comprises:

The OneHot encoding step performs OneHot encoding on N merchants, wherein each merchant is mapped to an N-dimensional sparse vector;

a consumer merchant sequence construction step of constructing a consumer merchant sequence for a specified account, recording a business code corresponding to the related transaction, and sorting, wherein the merchant code is represented by a vector after OneHot encoding by the OneHot encoding step;

The transition probability network construction step constructs a neural network in which the vector after the OneHot encoding of each merchant is used as an input layer, and the probability distribution of the merchants of the merchants that may appear next after the merchant is taken as an output layer.

Optionally, after the transition probability network construction step, further comprising:

The merchant recommendation step is to recommend the merchant to the user based on the probability of occurrence of the merchant outputted by the transition probability network construction step.

Optionally, the transition probability network construction step comprises the following sub-steps:

Constructing a neural network model sub-step to construct a three-layer neural network including an input layer, a hidden layer, and an output layer;

Generating a training sample sub-step to generate a training sample based on the consumer merchant sequence;

The neural model substep is trained to train the neural network based on a prescribed algorithm.

Optionally, in the training neural model sub-step, the dimension of the input layer is N, the dimension of the hidden layer is m, and the weight matrix of the input layer to the hidden layer is recorded as W _1, then the weight matrix W _{1 is} written as N rows * m columns, and when the kth dimension of the input vector is 1, only the row vector of the kth row in W ₁ is extracted as an output.

Optionally, in the training neural model sub-step, the positive samples are weight-updated, and a part of the negative samples are randomly selected for weight update.

Optionally, in the consumer merchant sequence construction step, an account whose consumption times are greater than a preset threshold in a predetermined time period is counted, and for the statistical account, a merchant code corresponding to each transaction is recorded, and each of the accounts is The merchant codes of the merchants that are consumed in the specified period are arranged in order.

Optionally, in the generating the training sample sub-step, for the consumer merchant sequence, selecting a nearby merchant of the merchant as the input layer according to the specified location parameter to obtain a training sample or selecting a vicinity of the merchant as the input layer according to the specified time parameter Merchants get training samples.

The merchant recommendation system based on the transition probability network of the present invention is characterized in that it has:

OneHot encoding module, which performs OneHot encoding on N merchants, wherein each merchant is mapped to an N-dimensional sparse vector;

a consumer merchant sequence building module, for a specified account, recording a merchant code corresponding to the related transaction and sorting to construct a consumer merchant sequence, wherein the merchant code is represented by a vector after OneHot encoding by the OneHot encoding step;

The transition probability network construction module constructs a neural network in which the vector after the OneHot encoding of each merchant is used as an input layer, and the probability distribution of the merchants of the merchants that may appear next after the merchant is used as an output layer.

Optionally, further comprising:

The merchant recommendation module recommends the merchant to the user based on the merchant appearance probability output by the transition probability network construction device.

Optionally, the transition probability network building module has:

Constructing a neural network model sub-module, and constructing a three-layer neural network including an input layer, a hidden layer, and an output layer;

Generating a training sample sub-module, generating a training sample using the consumer merchant sequence;

The neural model sub-module is trained to train the neural network based on a prescribed algorithm.

Optionally, in the training neural model sub-module, the dimension of the input layer is N, the dimension of the hidden layer is m, and the weight matrix of the input layer to the hidden layer is W1 _{, and} the weight matrix is W _{1 is} denoted as N rows * m columns, and when the kth dimension of the input vector is 1, only the row vector of the kth row in W ₁ is extracted as an output.

Optionally, in the training neural model sub-module, the positive samples are weight-updated, and a part of the negative samples are randomly selected for weight update.

Optionally, the consumer merchant sequence building module counts an account whose consumption times are greater than a preset threshold within a specified time period, and for a statistical account, records a merchant code corresponding to each transaction, and specifies a time for each segment of the account. The merchant codes of the merchants in which they are consumed are arranged in order.

Optionally, in the generating the training sample sub-module, selecting, according to the specified location parameter, a merchant corresponding to the merchant location as the input layer according to the specified location parameter, obtaining the training sample, or selecting the parameter according to the specified time parameter as the input layer Train samples are obtained for merchant-related merchants.

The computer readable medium of the present invention has stored thereon a computer program, characterized in that the computer program is executed by a processor to implement the above-described transition probability network based merchant recommendation method.

The computer device of the present invention includes a memory, a processor, and a computer program stored on the memory and operable on the processor, wherein the processor executes the computer program to implement the above-described transition probability network based merchant Recommended method.

According to the merchant recommendation method based on the transition probability network and the merchant recommendation system based on the transition probability network, the architecture of the transition probability three-layer neural network can greatly make the calculation within the controllable range. Moreover, in the process of calculating the weight of the neural network, the calculation can be accelerated by using the "table lookup method" and the "negative sample sampling" method. According to the present invention, a three-layer neural network architecture with transition probability is proposed, which can quickly analyze the relationship between merchant series and make more accurate recommendations according to the consumption sequence of a large number of users.

Other features and advantages of the method and apparatus of the present invention will become more apparent or <RTIgt; clarified</RTI> by the accompanying drawings and the accompanying drawings. .

DRAWINGS

1 is a schematic diagram showing a merchant recommendation method based on a transition probability network of the present invention.

Fig. 2 is a flow chart showing the first embodiment of the merchant recommendation method based on the transition probability network of the present invention.

Figure 3 is a diagram showing the architecture of a merchant transition probability network model.

The business sequence of 7 merchants from merchant 1 to merchant 7 is shown in FIG.

Fig. 6 is a block diagram showing the construction of a merchant recommendation system based on a transition probability network of the present invention.

Detailed ways

The following are some of the various embodiments of the invention, which are intended to provide a basic understanding of the invention. It is not intended to identify key or critical elements of the invention or the scope of the invention.

First, the merchant recommendation method based on the transition probability network of the present invention will be described.

As shown in FIG. 1, the merchant recommendation method based on the transition probability network of the present invention includes the following steps:

OneHot encoding step S100: Perform OneHot encoding on N merchants, wherein each merchant is mapped to an N-dimensional sparse vector, where N is a natural number;

a consumer merchant sequence construction step S200: for a prescribed account, recording a merchant code corresponding to the related transaction and sorting to construct a consumer merchant sequence, wherein the merchant code is represented by a vector after OneHot encoding by the OneHot encoding step S100;

Transition probability network construction step S300: constructing a neural network, wherein a vector after the OneHot encoding of each merchant is used as an input layer, and a probability distribution of a merchant occurrence of a merchant that may appear next after the merchant is taken as an output layer;

The merchant recommendation step S400 is to recommend the merchant to the user based on the merchant appearance probability output by the transition probability network construction step S300.

Next, the OneHot encoding step S100, the consumer merchant sequence construction step S200, the transition probability network construction step S300, and the merchant recommendation step S400 will be specifically described.

First, in the OneHot encoding step S100, OneHot encoding is performed for N merchants. Each merchant is mapped to an N-dimensional sparse vector, where only the corresponding location of the merchant is 1, and all other bits are zero.

The so-called OneHot code (one-hot code), that is, the unique heat code, is intuitively speaking, how many bits are there, and only one bit is 1, and all other codes are 0. Generally, in the communication network protocol stack, an eight-digit or six-bit status single-heat code is used, and the system occupies one of the status codes, and the rest can be used by the user.

For example, there are six states of the unique heat code status code: 000001,000010,000100, 001000, 010000, 100000. For another example, the one-hot state code of sixteen states should be: 0000000000000001,0000000000000010,0000000000000100, 0000000000001000, 0000000000010000, 0000000000100000, ..., 10000000000000000.

But usually we simplify the writing to simplify the binary to a hexadecimal representation (every four digits from right to left are represented by a hexadecimal number), then the above sixteen state of the heat code can be expressed as 0x0001, 0x0002, 0x0004, 0x0008, 0x0010, 0x0020, ..., 0x8000 (where 0x is a hexadecimal prefix representation, and there are other representations in programs such as PLC).

Next, in the consumer merchant sequence construction step S200, the account whose number of consumptions in the time T is greater than the threshold θ is counted, and the threshold may be preset or may be changed as needed. For these accounts, record the merchant code for each transaction. The merchant codes of each account in the period of time are arranged in order. Note that the merchants here are all vectors after OneHot encoding.

Again, in the transition probability network construction step S300, a three-layer neural network such as "input-hidden layer-output" is constructed. The input is a merchant's OneHot code, with a fixed dimension hidden layer in the middle, and the output is the probability that the next merchant will appear. The training sample is generated by using the consumer merchant sequence, and the neural network is quickly trained by using the "lookup table method" and the "negative sample sampling". After the training is completed, the probability distribution of the merchants that may appear next for each merchant can be obtained.

Specifically, the transition probability network construction step S300 may mainly include the following specific steps:

(1) Constructing a neural network model substep

Construct a three-layer neural network such as “input-hidden layer-output”. The input is the sparse OneHot vector corresponding to the current merchant. The dimension of the output layer is the same as the input layer, except that each neuron on the layer outputs the probability corresponding to each merchant.

(2) Generate training sample substeps based on the consumer merchant sequence

A training sample is generated using the constructed consumer merchant sequence above. The constructed training sample is in the form of a key-value pair: (input merchant → export merchant). Optionally select a merchant in a consumer merchant sequence as the input merchant, and then select another merchant in the sequence that appears near the "input merchant" as the "export merchant."

(3) Training neural network model substeps

Based on these training data, the above-mentioned three-layer neural network is trained. The purpose of training a neural network is actually to obtain the parameters of the optimized network weight matrix. Here, by using the "lookup table method" and the "negative sample sampling", the amount of calculation can be greatly reduced when optimizing the "input layer→hidden layer" and "hidden layer→output layer" weight matrix, respectively. The specific method is as follows:

(I) Use the "look-up table method" to achieve fast calculation of the input layer to the hidden layer

First, regarding the weight matrix W ₁ of the "input layer → hidden layer", if the input layer dimension is N and the hidden layer dimension is m, the weight matrix W ₁ should be N rows * m columns. Considering the weight matrix of N*m as a table, when the kth dimension of the input vector is 1, only the row vector of the kth row of W ₁ needs to be extracted is the calculation result we need.

(II) Using the "negative sample sampling" method to achieve fast calculation of the hidden layer to the output layer

Secondly, the positive samples are weighted and then a small number of negative samples are randomly selected to update the corresponding weights, thereby reducing the amount of computation in the gradient descent process. A negative sample here refers to a sample that does not exist in the training data. At the time of sampling, we use the “random tap method” to make the probability that a popular merchant is selected as a negative sample is greater than the probability that an unpopular merchant is selected as a negative sample.

(4) Smart merchant recommendation substeps

After training the above neural network, each merchant's OneHot code is taken as input, and then the output layer is calculated, given an input merchant, and the probability value of each merchant in the merchant dictionary is output. That is to say, for a user, the merchant of the last transaction is used as the input of the neural network, then the merchant with the higher output probability should be more likely to appear in the context of the input merchant. In this way, when an account has a consumption, the neural network can be used to infer the ranking of the merchant that is most likely to be consumed next, and recommend it to the user from high to low.

Next, a first embodiment of the merchant recommendation method based on the transition probability network of the present invention will be described.

2 is a flow chart showing a first embodiment of a merchant recommendation method based on a transition probability network of the present invention.

As shown in FIG. 2, in step S1, merchant OneHot encoding is performed. First, create a business dictionary. For example, if there are a total of N merchants, then an initial business dictionary is built: {1: merchant 1, 2: merchant 2, ..., N: merchant N}. In this way, each merchant can first correspond to a number. If N is 100,000, then OneHot can be coded for these 100,000 merchants. In this way, each merchant is mapped to a 100,000-dimensional sparse vector, where only the corresponding location of the merchant is 1, and all other bits are zero.

For example, the vector after the merchant 1 is encoded by OneHot is [1, 0, 0, 0... 0], the vector after the merchant 2 is encoded by OneHot is [0, 1, 0, 0... 0], and the merchant 3 passes the OneHot. The vector after encoding is [0,0,1,0...0].

Next, in step S2, an account-level consumer merchant sequence is constructed. Specifically, an account that counts the number of consumptions within a period of time T greater than the threshold θ is counted. For these accounts, record the merchant code for each transaction. The merchant codes of each account in the period of time are arranged in order: {business 1, merchant 2, ..., merchant n}. Among them, the merchant code here is a vector after OneHot encoding.

Next, in step S3, a neural network model is constructed. Specifically, a three-layer neural network such as "input layer-hidden layer-output layer" is constructed. The input layer is the sparse OneHot vector corresponding to the current merchant, and the hidden layer is a compression vector of a specified length (the length generally does not need to be too large, the default dimension is 100 is enough), the dimension of the output layer is the same as the input layer, but the layer Each neuron on the output is the probability that each merchant corresponds.

In Fig. 3, the left side represents the input layer, the middle represents the hidden layer, and the right side represents the output layer. The input layer inputs the merchant OneHot code, the hidden layer represents the merchant scenario compression coding, and the output layer outputs the merchant probability distribution. Among them, the weight matrix from the "input layer → hidden layer" is denoted as W ₁ , and the weight matrix from the "hidden layer → output layer" is denoted as W ₂ . In addition, the output layer can be constructed with a softmax regression classifier. As shown in Figure 3, the dimensions of the input and output layers are equal to the total number of all merchants.

Next, in step S4, a training sample is generated based on the consumer merchant sequence. Specifically, the training samples are generated using the constructed consumer merchant sequence above. The constructed training sample is in the form of a key-value pair: (input merchant → export merchant). The overall scheme is to arbitrarily select one of the merchants in the sequence of a certain consumer merchant as the input merchant, and then select another merchant in the sequence that appears in the vicinity of the "input merchant" as the "export merchant".

In addition, you need to define a parameter called win_size, which represents the number of merchants selected from the left or right of the current "enter merchant". The other parameter is n_pick, which represents how many different merchants we choose from the entire window as our output.

Optionally, a time threshold max_T can also be defined. If the key-value pairs we select are: (input merchant → export merchant), the corresponding consumption time interval is greater than max_T, we can think that there is no scenario between the two merchants. Dependent, so remove it from the training samples.

Here, the vendor sequence of FIG. 4 is taken as an example to illustrate the parameters of win_size and the n_pick parameters.

The business sequence of 7 merchants from merchant 1 to merchant 7 is shown in FIG. Thus, for a consumer merchant sequence {business 1, merchant 2, ..., merchant 7}, if we select merchant 4 as the "input merchant". If win_size=2 is set, it means that 2 and 2 merchants on the left side of “Input Merchant” are selected to enter our window, that is, select: Merchant 2, Merchant 3, Merchant 5, and Merchant 6.

When n_pick=2, two sets of training samples will be randomly obtained, namely (Business 4 → Merchant 2), (Business 4 → Merchant 5).

And if we select merchant 1 as the "input merchant" and the win_size and n_pick parameters are also set to 2, then the final training sample must be (business 1 → merchant 2), (business 1 → merchant 3). Among them, the above input and output merchants are actually their corresponding OneHot encoding.

Next, in step S5, the neural network model is trained. Through the above steps, a large number of training samples have been generated from the consumer merchant sequence, and the above-mentioned three-layer neural network is trained based on these training data.

The purpose of training a neural network is actually to obtain the parameters of the optimized network weight matrix. However, considering the large number of merchants, we do not intend to use traditional methods to train the model, otherwise it will be very computationally intensive. Here, we use two technical means to greatly reduce the amount of calculation when optimizing the "input layer → hidden layer" and "hidden layer → output layer" weight matrix. The specific method is as follows:

First, a description will be given of the fast calculation of the input layer to the hidden layer using the "lookup table method".

As already mentioned above, the weight matrix of "Input Layer → Hidden Layer" is set to W ₁ . If the input layer dimension is N and the hidden layer dimension is m, then the weight matrix W ₁ should be N rows * m columns. For example, if the size of the merchant dictionary is 100,000 and you want to use 100 features to represent a merchant, then the weight matrix W ₁ should be 100,000 rows * 100 columns. In the calculation, it is necessary to multiply the input matrix of 1*N and the weight matrix of N*m. If the number of merchants is large, a large N will result in a lot of calculations.

By observing, our input vector is N-dimensional, but in fact, only one bit of each input vector is 1, and the other bits are all 0. In this class, we don't actually need to calculate the whole amount. We only need to treat the weight matrix of N*m as a table. When the kth dimension of the input vector is 1, we only need to extract the row vector of the kth row of W1. The calculation results we need. For example, the matrix of 1*4 is multiplied by the matrix of 4*3, and only the third digit of the input matrix is 1, then the calculation result is exactly the third row in the weight matrix. As shown below:

Secondly, the fast calculation of the hidden layer to the output layer is described using the "negative sample sampling" method.

The output layer is a softmax regression classifier in which each neuron will output a probability value between 0 and 1, and the sum of the probabilities of all output layer neuron nodes is one. The function formula of Softmax is as follows:

Where a _j represents the output of the jth neuron of the last layer; z _j represents the input of the jth neuron of the last layer, and e is a natural constant.

Represents the summation of the inputs to all neurons in the last layer. Thus, the log likelihood cost function is as follows:

C = -Σ _{_{_k}} y _k loga _k

Where a _k represents the output value of the kth neuron of the last layer; y _k corresponds to the true value of the kth neuron, and the value is 0 or 1. Our goal is to minimize this cost function. Notice that there is a ∑ _k summation in the cost function, which means that if the number of merchants is large, the optimization calculation process of the function will be very time consuming.

Therefore, here, we use the method of negative sample sampling, which is different from the original weight of each training sample, but each time a training sample is updated with only a small part of the weight, which will reduce the gradient drop process. Calculated amount. However, under this strategy, our optimization goal needs to be: maximize the probability of positive samples while minimizing the probability of negative samples. Negative samples are samples that we construct that do not exist in the training data.

For example, for merchant a, we first extract a few positive samples by context, and the rest of the words that are not in the context are called negative samples. For example, in the above example, for merchant 4, because (businesser 4 → merchant 2 ) is a positive sample, and for the input merchant 4, we will construct several samples that did not exist in the training data, such as (Business 4 → Merchant 25), (Business 4 → Merchant 342), (Business 4 → Merchant) 1253) These are called negative samples.

In this way, we first update the weight of our positive samples, and then randomly select a small part of the negative samples to update the corresponding weights. For large-scale data sets, the number of negative samples is generally about 5 or so.

Of course, at the time of sampling, we hope that the probability that a popular merchant will be selected as a negative sample is greater than the probability that an unpopular merchant will be selected as a negative sample. To achieve this function, the "random dot method" is used in the present invention for selection. That is, for each merchant k, we map it to a line segment of length expressed by the following formula:

Cnt(k)/∑ _i cnt(i)

Among them, cnt(k) indicates the number of times the merchant k appears in total, and the denominator is the sum of the statistics of the number of times of all merchants. We put these line segments together to form a complete line segment of length 1, and we get the line segment as shown in Figure 5. If you randomly click on this line segment, the point falls on that line segment interval, and the result is that the line segment corresponds to Merchants. As shown in Figure 5, it can be imagined that the line segment corresponding to the popular merchant is relatively long, and the probability of being hit is large.

Finally, in step S7, the merchant is recommended based on the seriality of the output layer probability analysis. Specifically, after training the above neural network, each merchant's OneHot code is taken as an input, and then the output layer is calculated, given an input merchant, and the probability value of each merchant in the merchant dictionary is output. That is to say, for a user, the merchant of the last transaction is used as the input of the neural network, then the merchant with the higher output probability should be more likely to appear in the context of the input merchant. In this way, when an account has a consumption, we can use the neural network to infer the ranking of the merchants that are most likely to consume next, and recommend it to the user from high to low.

The above describes the merchant recommendation based on the transition probability network of the present invention, and then briefly introduces the merchant recommendation system based on the transition probability network of the present invention.

As shown in FIG. 6, the merchant recommendation system based on the transition probability network of the present invention has:

The OneHot encoding module 100 performs OneHot encoding on N merchants, wherein each merchant is mapped to an N-dimensional sparse vector;

The consumer merchant sequence building module 200 constructs a consumer merchant sequence for the specified account, records the merchant codes corresponding to the related transactions, and sorts the merchant code, wherein the merchant code is represented by a vector after the OneHot encoding by the OneHot encoding step;

The transition probability network building module 300 constructs a neural network, wherein a vector after the OneHot encoding of each merchant is used as an input layer, and a probability distribution of a merchant occurrence of a merchant that may appear next after the merchant is taken as an output layer;

The merchant recommendation module 400 recommends the merchant to the user based on the merchant appearance probability output by the transition probability network construction device.

The transition probability network construction module 300 has:

Constructing a neural network model sub-module 310, constructing a three-layer neural network including an input layer, a hidden layer, and an output layer;

Generating a training sample sub-module 320 for generating training samples using the consumer merchant sequence;

The neural model sub-module 330 is trained to train the neural network based on a prescribed algorithm.

In a preferred manner, in the training neural model sub-module 330, the dimension of the input layer is N, the dimension of the hidden layer is m, and the weight matrix of the input layer to the hidden layer is recorded as W _1, then the weight matrix W _{1 is} denoted as N rows * m columns, and when the kth dimension of the input vector is 1, only the row vector of the kth row in W ₁ is extracted as an output.

Further, in the training neural model sub-module 330, the positive samples are weight-updated, and a part of the negative samples are randomly selected for weight update.

In a preferred manner, the consumer merchant sequence building module 200 counts the accounts whose consumption times in the time T is greater than the threshold θ, and for the account that is counted, records the merchant code corresponding to each transaction, and the segment of each of the accounts is The merchant codes of the merchants that consume the specified time are arranged in order.

Specifically, the generated training sample sub-module 320 selects a merchant that is related to the merchant location of the input layer according to the specified location parameter, obtains the training sample according to the specified location parameter, or selects the merchant as the merchant time associated with the input layer according to the specified time parameter. Get training samples.

Further, the present invention further provides a computer readable medium having stored thereon a computer program, characterized in that the computer program is executed by a processor to implement the above-described transfer probability network-based merchant recommendation method of the present invention.

Further, the present invention also provides a computer device comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, wherein the processor implements the above-described invention when the computer program is executed A merchant recommendation method based on a transfer probability network.

The merchant recommendation method based on the transition probability network and the merchant recommendation system based on the transition probability network discard the idea of the transfer matrix, and propose a framework of the transition probability three-layer neural network. The network input is a merchant's unique heat code, and the output is the most likely merchant to appear next. Because a hidden layer with a fixed dimension is added in the middle (generally not too large, such as 100 dimensions), it is possible to avoid direct calculation of two columns between mass merchants (such as 100,000 * 100,000), instead of "input → The hidden layer "100,000 * 100, and the "hidden layer → output" of 100 * 100, so that the calculation is within the controllable range.

Moreover, in the process of calculating the weight of the neural network, the "check table method" and the "negative sample sampling" method are further used to accelerate the calculation. Among them, the "check table method" can further avoid the "input → hidden layer" of the matrix operation of 100,000 * 100, only need to index, "negative sample sampling" accelerates the "hidden layer → output" to some extent The calculation of the weight.

The merchant recommendation method based on the transition probability network and the merchant recommendation system based on the transition probability network do not depend on the “user-business score matrix”, but extract effective merchant-pair training samples according to the user's consumption sequence, avoiding hyperscale The "user-business" matrix.

The above examples mainly illustrate the merchant recommendation method based on the transition probability network of the present invention and the merchant recommendation system based on the transition probability network. Although only a few of the specific embodiments of the present invention have been described, it is understood that the invention may be embodied in many other forms without departing from the spirit and scope of the invention. Accordingly, the present invention is to be construed as illustrative and not restrictive, and the invention may cover various modifications without departing from the spirit and scope of the invention as defined by the appended claims With replacement.

Claims

A merchant recommendation method based on a transition probability network, which is characterized in that it comprises:

The OneHot encoding step performs OneHot encoding on N merchants, wherein each merchant is mapped to an N-dimensional sparse vector, where N is a natural number;

a consumer merchant sequence construction step of constructing a consumer merchant sequence for a specified account, recording a business code corresponding to the related transaction, and sorting, wherein the merchant code is represented by a vector after OneHot encoding by the OneHot encoding step;

The transition probability network construction step constructs a neural network in which a vector after OneHot encoding of each merchant is used as an input layer, and an occurrence probability distribution of a merchant that may appear next after the merchant is used as an output layer.
The transition recommendation network-based merchant recommendation method according to claim 1, further comprising: after the transition probability network construction step:

The merchant recommendation step recommends the merchant to the user based on the probability of occurrence of the output of the transition probability network construction step.
The transition recommendation network-based merchant recommendation method according to claim 1 or 2, wherein the transition probability network construction step comprises the following sub-steps:

Constructing a neural network model sub-step to construct a three-layer neural network including an input layer, a hidden layer, and an output layer;

Generating a training sample sub-step to generate a training sample based on the consumer merchant sequence;

The neural model substep is trained to train the neural network based on a prescribed algorithm.
A merchant recommendation method based on a transition probability network according to claim 3, wherein

In the training neural model sub-step, the dimension of the input layer is N, the dimension of the hidden layer is m, the weight matrix of the input layer to the hidden layer is recorded as W1, and the weight matrix W1 is recorded as N rows* In the m column, when the kth dimension of the input vector is 1, only the row vector of the kth row in W1 is extracted as an output.
A merchant recommendation method based on a transition probability network according to claim 4, wherein

In the training neural model substep, the positive samples are weighted and the negative samples are randomly selected for weight update.
A merchant recommendation method based on a transition probability network according to claim 3, wherein

In the consumer merchant sequence construction step, an account whose consumption times are greater than a preset threshold in a predetermined time period is counted, and for a statistical account, a merchant code corresponding to each transaction is recorded, and each of the accounts is within a specified time period of the segment. The merchant code of the merchant where the consumer is located is arranged in order.
A merchant recommendation method based on a transition probability network according to claim 3, wherein

In the step of generating a training sample, for the consumer merchant sequence, selecting a merchant related to a merchant location as an input layer according to a specified location parameter to obtain a training sample, or selecting a merchant time as an input layer according to a predetermined time parameter Get training samples from relevant merchants.
A merchant recommendation system based on a transition probability network, characterized in that it has:

The OneHot encoding module performs OneHot encoding on N merchants, wherein each merchant is mapped to an N-dimensional sparse vector, where N is a natural number;

a consumer merchant sequence building module, for a specified account, recording a merchant code corresponding to the related transaction and sorting to construct a consumer merchant sequence, wherein the merchant code is represented by a vector after OneHot encoding by the OneHot encoding step;

The transition probability network construction module constructs a neural network in which a vector after OneHot encoding of each merchant is used as an input layer, and an occurrence probability distribution of a merchant that may appear next after the merchant is used as an output layer.
The transition recommendation network-based merchant recommendation system according to claim 8, further comprising:

The merchant recommendation module recommends the merchant to the user based on the appearance probability output by the transition probability network construction device.
The transition recommendation network-based merchant recommendation system according to claim 6 or 7, wherein the transition probability network construction module has:

Constructing a neural network model sub-module, and constructing a three-layer neural network including an input layer, a hidden layer, and an output layer;

Generating a training sample sub-module, generating a training sample using the consumer merchant sequence;

The neural model sub-module is trained to train the neural network based on a prescribed algorithm.
A transition probability network-based merchant recommendation system according to claim 8, wherein

In the training neural model sub-module, the dimension of the input layer is N, the dimension of the hidden layer is m, the weight matrix of the input layer to the hidden layer is recorded as W1, and the weight matrix W1 is recorded as N rows. *m column, when the kth dimension of the input vector is 1, only the row vector of the kth row in W1 is extracted as an output.
A merchant recommendation system based on a transition probability network according to claim 11, wherein

In the training neural model sub-module, the positive samples are weighted and the negative samples are randomly selected for weight update.
A transition probability network-based merchant recommendation system according to claim 8, wherein

The consumer merchant sequence building module counts accounts whose consumption times are greater than a preset threshold within a specified time period, and for the statistical accounts, records the merchant codes corresponding to each transaction, and stores the merchants in each of the accounts within the specified time period. The business code is arranged in order.
A transition probability network-based merchant recommendation system according to claim 10, wherein

And generating, by the generated training sample sub-module, the training sample according to the specified location parameter, selecting a merchant related to the merchant location as the input layer, or selecting the predetermined time parameter as the merchant time associated with the input layer. Trainers get training samples.
A computer readable medium having stored thereon a computer program, wherein the computer program is executed by a processor to implement the transition probability network based merchant recommendation method according to any one of claims 1-7.
A computer device comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, wherein the processor executes the computer program to implement any one of claims 1-7 The merchant recommendation method based on the transition probability network.