TWM573493U - System for predicting conversion probability by visitors' browsing paths - Google Patents

Abstract

This invention provides a system for predicting conversion probability by visitors' browsing paths comprising a retrieval unit which collects visitors' identities and access paths of browsing pages, a path database which records the visitors' identities and page orders composed of the access paths, a tag unit which labels the page orders have been converted, a processor which loads the page orders into the memory and performs an algorithm module to calculate the page orders as vectors, a training module which inputs a predetermined proportion of the vectors into a recurrent neural network (RNN) to establish a RNN model, and a prediction module which inputs updated vectors into the RNN model to obtain visitors with high conversion probability.

Description

System for predicting conversion probability by visitor browsing path

This creation is about a prediction system, especially a system that uses a visitor browsing path to predict conversion probability.

The existing advertisement is to learn the user browsing path, including the indirect or directly linked webpage, to collect the content of the browsing page, analyze the content of the browsing page, and find the meaningful vocabulary according to the vocabulary according to the vocabulary, and according to the vocabulary The frequency of occurrence gives weight.

At the same time, according to the previous browsing page and weight, the conditions for searching for the advertisement are determined, and the conditions for the advertisement search can also be selected according to the channel selected by the user, the title of the article, the name of the linked website, the keyword input, and the like. Conditional ads and place the ad in a prominent position on the page.

However, existing webpage advertisements are information about behaviors of users clicking through advertisements, and by using keywords of webpages as weights to provide other advertisements for users to approximate content, but failing to truly analyze user behavior and browsing pages. The model does not really align the user's needs and create user needs, let alone predict the user's behavior patterns.

Therefore, the existing advertising technology has its shortcomings and needs to be improved.

For the foregoing purposes, the present invention provides a system for predicting conversion probability by a visitor browsing path, including: a retrieval unit for collecting access identities of multiple visitors and access paths of the plurality of visitors; a path database, Connecting the capturing unit to record the access status of the plurality of visitors and recording the page order formed by the access paths of each of the visitors; the marking unit is connected to the path database to record the path database The converted page order is marked in the page order; the processor includes a memory and an algorithm module for receiving the visitor identity recorded from the path database and the page order corresponding thereto Entering in the memory, the page order is calculated as a corresponding vector through the algorithm module; the training module receives the vector calculated by the algorithm module, and presets the number of the vector Entering a recursive neural network to establish a recursive neural network model, and let the recursive neural network model learn the marked page order And a prediction module, wherein the page order of the update access path collected by the capture unit and recorded by the path database is calculated as a corresponding vector through the algorithm module to input the regressive neural network The network model derives a predictive conversion page order that corresponds to the vector pattern of the marked page order.

The foregoing system for predicting a conversion probability of a visitor path includes: an evaluation module, the preset proportional number of the vector of the algorithm module The remaining vectors other than the vector are input to the recursive neural network model to evaluate the accuracy of the prediction transformation.

The foregoing guest browsing path predicts a conversion probability system, wherein the converted page order marked by the marking unit is marked as a value 1 in a consumption behavior in the access path, and the marking unit is the path database The page order in which the conversion does not appear in the page order of the record is marked as a value of zero.

The foregoing guest browsing path predicts a conversion probability system, wherein the training module randomly selects a percentage of the vector of the 80% input to the recursive neural network to establish a recursive neural network model.

The foregoing system for predicting a conversion probability of a visitor path, wherein the recursive neural network system comprises an input layer and an output layer, wherein the input layer multiplies an output value of the previous output layer by a first weight value plus The current vector provided by the algorithm module is used as an input value, and the output layer multiplies the input value of the input layer by a second weight value plus an offset value as the input value.

The foregoing method for predicting conversion probability of a visitor browsing path, wherein the recursive neural network further comprises a normalized index layer, which converts the output value of the output layer into a numerical range between 0 and 1, As a chance to predict conversion.

The aforementioned guest browsing path predicts a conversion probability system, wherein the algorithm module uses a word2vec algorithm.

The foregoing system for predicting a conversion probability of a visitor path, wherein the path database records the access path group of each visitor in chronological order The order of the pages.

The foregoing system for predicting a conversion probability of a visitor path, wherein the path database is to update the page access order by adding the update access path collected by the capture unit to an existing access path based on the same guest identity.

10‧‧‧System

11‧‧‧Capture unit

12‧‧‧Path database

13‧‧‧Marking unit

14‧‧‧ Processor

141‧‧‧ memory

142‧‧‧ algorithm module

15‧‧‧ training module

16‧‧‧Return-type neural network

161‧‧‧Input layer

162‧‧‧Output layer

163‧‧‧ normalized index layer

17‧‧‧Evaluation module

18‧‧‧ Forecasting Module

Figure 1 is a system architecture diagram of the system for predicting conversion probability by the guest browsing path.

Figure 2 is a schematic diagram of the recursive neural network architecture of the system for predicting conversion probability by the guest browsing path.

The embodiments of the present invention are described in the following specific embodiments, and those skilled in the art can easily understand other advantages and functions of the present invention by the contents disclosed in the present specification, and can also be implemented by other different embodiments. Or application.

Referring to FIG. 1, the present invention provides a system 10 for predicting conversion probability by a visitor browsing path, which includes: a retrieval unit 11, a path database 12, a marking unit 13, a processor 14, a training module 15, and a delivery The back-type neural network 16, the evaluation module 17, and the prediction module 18.

The retrieval unit 11 can be a collection engine for performing cookies generated by, for example, Google Analytics, server log files, user registration forms, or online advertisements, and multiple visitor identifications (eg, IP location, machine code, visitor use). The browser, or the registered account), and the collection of data such as the access path (eg, IP address) of the plurality of guest browsing pages.

The retrieval unit 11 stores the collected data in the path database 12, and the path database 12 mainly records the access status of the plurality of visitors (for example, IP location, machine code, browser used by the visitor, or registered account). And the page order formed by the access paths of the respective guest browsing pages, the path database 12 is sorted in chronological order to record the plurality of access paths of the guest browsing page as the page order, and the page order of the related guest identity and the access path may be Refer to Table 1 below.

The conversion in Table 1 shows the consumer in the browsing process. The behavior that the consumer behavior can be used to place the item into the shopping cart, order, checkout, credit card payment, or other online payment can be determined by whether to enter a particular web page or to make a specific instruction. The marking unit 13 is the path information. In the order of the pages recorded by the library 12, the converted page order is marked. In the present embodiment, the flag 1 indicates that there is a consumer behavior, and conversely, the consumer behavior is marked as 0.

The processor 14 includes a memory 141 and an algorithm module 142. The processor 14 receives the visitor identity recorded from the path repository 12 and the page order corresponding to the visitor identity (including the page order of the markup) And temporarily stored in the memory 141 to calculate the page order temporarily stored in the memory 141 by the algorithm module 142 as a corresponding vector, wherein the algorithm module 142 uses the word2vec algorithm, the word2vec algorithm Generally, it is used to process natural language. In the present creation, the processor 14 uses the word2vec algorithm to calculate the page order as a vector through the algorithm module 142, for example, to operate into a vector having eight dimensions.

The training module 15 receives the vector calculated by the algorithm module 142 and inputs a preset proportion of the vector into the Recurrent Neural Network (RNN) to train the recurrent neural network. a model, wherein the preset proportional number of vectors is randomly selected by the training module 15 by a vector of 80% of the existing vector data, and is input to the regressive neural network 16 for training purposes. Back-type neural network model By learning the memory function, vector patterns with approximate page order can be learned from the aforementioned marked page order.

Please also refer to Figure 2, the recursive neural network 16 series includes The input layer 161 and the output layer 162 multiply the output value of the previous output layer 162 by the first weight value plus the current vector provided by the algorithm module 142 as an input value, and the output layer 162 is The input value of the input layer 161 is multiplied by the second weight value plus the offset value as an output value, wherein the recursive neural network 16 further includes a normalized index layer (or softmax layer) 163, which The output value of the output layer 162 is scaled to a value range between 0 and 1 as a probability of predictive conversion.

The evaluation module 17 inputs the preset vector of the predetermined proportion of the vector of the algorithm module 142 to the recursive neural network model to verify and evaluate the recursive neural network model. In the present embodiment, the evaluation module 17 uses the remaining twenty percent of the vector of the existing vector data as the verification and evaluation of the accuracy of the recursive neural network model. If it is known that the trained recursive neural network model cannot correctly predict that the page order will be converted, then the data source of the retrieval unit 11 is re-examined, or the data source is replaced, and then recalculated as The vector is input into the recursive neural network 16 to establish a new recursive neural network model; or more visitors are added to the page order to make the recursive neural network model under the training of a large amount of data. Gradually accurate; or, re-adjust the number of nodes and layers of the recursive neural network to re-establish the recursive neural network model.

The prediction module 18 collects the update access path collected by the capture unit 11, and forms a page order through the path database 12 and operates as a corresponding vector through the algorithm module 152, and then inputs the recursive neural network model to obtain a predictive conversion. Page order; take the above Table 1 as an example, due to the recursive class The recursive neural network model generated by the neural network 16 is learned to learn that the "page A→page C→page D" in the page order of the visitor path is converted and marked as 1, when the capture unit 11 collects multiple Approximating the page order of "Page A → Page C → Page D", and the recursive neural network model will calculate that the visitor is browsing the page containing the predicted conversion of "Page A → Page C → Page D" In the order, the normalized index layer 163 calculates the probability that the visitor has a value close to 1 (very high probability) to perform the consumption behavior; further, the system 10 of the present creation can be set to, when the recursion The system-like neural network model calculates that when the visitor has browsed the page order including "page A→page C→page D", the system 10 automatically delivers the same type of product advertisement as the page D, or passes through the recursive nerve. The page order predicted by the network model is calculated, and these page order with high probability conversion is counted to the product supplier as a reference for adjusting the product page, item sorting, or page level.

In addition, the access path collected by the capturing unit 11 may evolve over time, and a new access path may be added at any time. Therefore, the capturing unit 11 collects the updated access path at least daily and stores it in the path database 12. The newly browsed page of the same guest identity may be added to the updated page order after the existing access path, and the updated page order is calculated into the corresponding vector by the algorithm module 142, and the recursive class is input. Neural network models to predict conversion probabilities.

The detailed description above is a detailed description of one of the possible embodiments of the present invention, and is not intended to limit the scope of the present invention, and equivalent implementations or modifications that are not departing from the spirit of the present invention should be included in The scope of the patent application for this creation.

Claims (9)

A system for predicting a conversion probability by a visitor browsing path, comprising: a capturing unit for collecting access identities of a plurality of visitors and access paths of the plurality of visitors; a path database connecting the capturing units to record The access status of the plurality of visitors and the page order formed by the visit paths of the visitor; the mark unit is connected to the path database to convert the page order in the page order recorded by the path database The processor includes a memory and an algorithm module for receiving the visitor identity recorded from the path database and the page order corresponding thereto and loading the same in the memory to transmit The algorithm module calculates the page order as a corresponding vector; the training module receives the vector calculated by the algorithm module, and inputs the preset proportional quantity of the vector into the recursive neural network to Establishing a recursive neural network model, and letting the recursive neural network model learn a vector pattern of marked page order; and a prediction module, The page order of the update access path collected by the capture unit and recorded in the path database is calculated as a corresponding vector through the algorithm module, and the recursive neural network model is input to obtain a predicted conversion page order. Wherein, the corresponding vector of the predicted conversion page order is close to the vector mode of the marked page order. For example, the system described in claim 1 includes: The evaluation module inputs the remaining vector of the predetermined proportion of the vector of the algorithm module to the recursive neural network model to evaluate the accuracy of the prediction conversion. The system of claim 1, wherein the converted page order marked by the marking unit is marked as a value 1 in a consumption behavior in the access path, and the marking unit is the path database The recorded page order in which the conversion does not appear in the order of the page is marked as a value of zero. The system of claim 1, wherein the training module randomly selects a percentage of the vector of the eighty percent input to the recursive neural network to establish a recursive neural network model. The system of claim 1, wherein the recursive neural network comprises an input layer and an output layer, the input layer multiplying the output value of the previous output layer by the first weight value plus The current vector provided by the algorithm module is used as an input value, and the output layer multiplies the input value of the input layer by a second weight value plus an offset value as an output value. The system of claim 5, wherein the recursive neural network further comprises a normalized index layer that converts the output value of the output layer to a value range between 0 and 1. To take the chance of predicting conversion. The system of claim 1, wherein the algorithm module uses a word2vec algorithm. The system of claim 1, wherein the path information The library records the page order of the visit paths of each of the visitors in chronological order. The system of claim 1, wherein the path database is to update the page order after the updated access path collected by the capturing unit is added to an existing access path based on the same guest identity.