A method and device for predicting false delivery addresses based on deep learning

Abstract

The present invention proposes a method and device for predicting a false delivery address based on deep learning, the method comprising the following steps: obtaining a user delivery address, and performing word segmentation on the delivery address to obtain common words; encoding the common words to generate a word vector corresponding to the delivery address; padding the word vector according to a uniform length; converting the padded word vector using unique hot encoding, and then converting the word vector into a low-dimensional word vector matrix using word embedding; clustering the word vector matrix according to a clustering algorithm, selecting a delivery address exceeding a threshold in the clustering result, and adding an abnormal label to the delivery address to generate an abnormal delivery address; inputting the abnormal delivery address into a neural network model, and outputting the probability that the abnormal delivery address is a false address. The present invention has a high accuracy rate and an extremely fast running speed, is not interfered by external data, does not require manual data labeling, and can automatically obtain the label of the order address.

Description

Technical Field

[0001] The present invention relates to the field of computer information technology, and in particular to a method and device for predicting a false delivery address based on deep learning. Background Art

[0002] With the development of information technology, e-commerce platforms have become the first choice for many users. However, when e-commerce platforms are doing activities, a small number of abnormal users will register a large number of fake accounts and use different fake order addresses to place a large number of fake orders, which not only prevents real users from enjoying the merchants' discounts, but also damages the interests of the platform.

[0003] According to long-term observation of abnormal accounts, it is found that the delivery locations of many transactions are highly unreasonable, contrary to common sense or inconsistent with normal user consumption habits, or there are coded fields in the delivery addresses. These data show that abnormal users tend to use some special delivery addresses to cover up their true identities. At present, it is time-consuming and laborious to judge the user's order address manually, and the judgment time is too long. There are also some methods for judging false addresses on the market. Most of them are judged by analyzing whether the user account resident range matches the order address, or by matching whether there is a matching address in the map software. These methods generally have the disadvantages of low accuracy and slow operation speed. Summary of the invention

[0004] In view of the above problems, the present invention provides a method and device for predicting false delivery addresses based on deep learning, which has higher accuracy and extremely fast running speed compared with the existing abnormal address detection methods on the market.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for predicting false delivery addresses based on deep learning, comprising the following steps: S101, obtaining a user delivery address, and segmenting the delivery address to obtain common words; S102, encoding the common words to generate a word vector corresponding to the delivery address; S103, padding the word vector according to a uniform length; S104, converting the padded word vector using one-hot encoding, and then converting the word vector into a low-dimensional word vector matrix using word embedding; S105, clustering the word vector matrix according to a clustering algorithm, selecting a delivery address that exceeds a threshold in the clustering results, and adding an abnormal label to the delivery address to generate an abnormal delivery address; S106, inputting the abnormal delivery address into a neural network model, and outputting the probability that the abnormal delivery address is a false address.

[0006] As a preferred solution, the encoding of common words includes: selecting high-frequency words from all common words to create a high-frequency dictionary, and converting the words in the high-frequency dictionary into numbers to generate a word vector corresponding to the delivery address.

[0007] As a preferred solution, the method further includes: S107, regularly obtaining a delivery address, processing the delivery address using steps S101 to S105, and inputting the processing result into a neural network model for training.

[0008] As a preferred solution, the clustering algorithm is the DBSCAN algorithm.

[0009] As a preferred solution, the neural network model is an RNN deep learning model, the recursive cells used in each layer are GRU, and each operation cell has 128 neurons.

[0010] The present invention also provides a false delivery address prediction device based on deep learning, including: a word segmentation module, used to obtain a user delivery address, and segment the delivery address to obtain common words; an encoding module, used to encode the common words and generate a word vector corresponding to the delivery address; a padding module, used to pad the word vector according to a uniform length; a conversion module, used to convert the padded word vector using one-hot encoding, and then convert the word vector into a low-dimensional word vector matrix using word embedding; a clustering module, used to cluster the word vector matrix according to a clustering algorithm, select the delivery address exceeding the threshold in the clustering result, and add an abnormal label to the delivery address to generate an abnormal delivery address; an output module, used to input the abnormal delivery address into a neural network model, and output the probability that the abnormal delivery address is a false address.

[0011] As a preferred solution, it also includes an updating module for periodically acquiring the delivery address, processing the delivery address, and inputting the processing result into the neural network model for training.

[0012] Compared with the prior art, the beneficial effects of the present invention include: segmenting the user's delivery address by using a segmentation algorithm with a custom dictionary added, encoding the phrases using a high-frequency dictionary after segmentation, obtaining the input matrix by unique hot encoding and word embedding, training a deep learning model to classify and predict the delivery address after using the labels obtained based on the DBSCAN clustering algorithm, and regularly annotating the newly acquired data through the clustering algorithm to use the deep learning model to identify the newly generated false addresses. After receiving the user's order, the order address is identified to determine whether it is a false address for brushing orders, which can effectively prevent the loss of merchants and the damage to the rights and interests of normal customers caused by brushing orders. The training data used in the present invention is annotated by a clustering algorithm, and no manual participation is required, thereby saving costs. At the same time, the present invention only needs 2ms for the recognition time of a single address, which significantly improves the efficiency compared with manual work. Compared with the existing abnormal address detection schemes on the market, it has a higher accuracy rate and a very fast running speed. Moreover, the present invention does not rely on external data assistance at all, can be operated in an internal network or in a network-free environment, and has a fast running speed and is not interfered by external data. In addition, the present invention does not require manual data labeling, and automatically obtains the label of the order address through clustering algorithm and user behavior analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the accompanying drawings, the same reference numerals are used to refer to the same components. Among them:

[0014] Figure 1 Schematic diagram of the process of a false delivery address prediction method based on deep learning according to an embodiment of the present invention;

[0015] Figure 2 Schematic diagram of the RNN deep learning model of an embodiment of the present invention;

[0016] Figure 3 Schematic diagram of the structure of a false delivery address prediction device based on deep learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the whole of the present invention or as a limitation or restriction to the technical solution of the present invention.

[0018] The professional terms in the embodiments of the present invention are explained as follows:

[0019] Jieba word segmentation: also known as Jieba word segmentation, is a very popular Chinese open source word segmentation package.

[0020] Custom dictionary: A list of words that are not in the standard dictionary but that the author wants the spell checker to accept as correct words.

[0021] Hidden Markov Model: (Hidden Markov Model, HMM) is a statistical model that is used to describe a Markov process with hidden unknown parameters.

[0022] One-hot encoding: One-hot encoding, also known as one-bit effective encoding, is a method of using an N-bit state register to encode N states. Each state has its own independent register bit, and at any time, only one of them is valid.

[0023] DBSCAN: (Density-Based Spatial Clustering of Applications with Noise) is a representative density-based clustering algorithm. Unlike partitioning and hierarchical clustering methods, it defines a cluster as the largest set of density-connected points. It can divide areas with sufficiently high density into clusters and can find clusters of arbitrary shapes in noisy spatial databases.

[0024] RNN: (Recurrent Neural Network, RNN) is a type of recursive neural network that takes sequence data as input, recursively in the direction of sequence evolution, and all nodes are connected in a chain.

[0025] Sigmoid function: The sigmoid function is also called the Logistic function, which is used for the output of hidden layer neurons. Its value range is (0,1). It can map a real number to the interval (0,1).

[0026] According to one embodiment of the present invention, Figure 1 A method for predicting a false delivery address based on deep learning includes the following steps:

[0027] S101, obtaining a user's delivery address, and segmenting the delivery address into common words.

[0028] Among them, the algorithm for word segmentation of the delivery address is modified based on the open source Jieba word segmentation algorithm, which specifically includes: first loading a pre-built custom dictionary with weights, and then using TrieTree (dictionary tree) to build a word segmentation model, matching the words in the custom dictionary in the delivery address, and identifying whether the words that do not appear are numbers. For the non-numeric part, the Hidden Markov Model (HMM) is used to process it to obtain the word segmentation result, that is, the common vocabulary.

[0029] Since there are special place names in the delivery address, we choose to use a custom dictionary with address-related information added to segment the address. For example, Mr. Zhang's delivery address is "Huai'an City Hongze County Gaoliangjian Town East Seven Road Tianyou Shangdu Hongze Aidong Air Conditioning Maintenance Service Department". After using the word segmentation algorithm, we can get the result ['Huai'an City', 'Hongze County', 'Gaolianjian', 'Zhendong', 'Qidao', 'Tianyou', 'Shangdu', 'Hongze', 'Aidong', 'Air Conditioning', 'Maintenance', 'Service Department']. It can be seen that most of the words are segmented correctly. The main error is that the algorithm segments 'Gaolianjian Town | East Seven Road' into 'Gaolianjian | Zhendong | Qidao'. However, as long as this word segmentation algorithm is fixed, a unified segmentation method is adopted for all addresses, which will not affect the accuracy of the model.

[0030] S102, encode common words to generate word vectors corresponding to the delivery address.

[0031] The encoding of common words includes: selecting high-frequency words from all common words to create a high-frequency dictionary, and converting the words in the high-frequency dictionary into numbers to generate a word vector corresponding to the delivery address.

[0032] In a specific embodiment, 1.2 million addresses are used as training sets, and there are more than 150,000 words in total. However, if all words are used, the model dimension will be too large and the performance will be degraded. Therefore, after repeated tests, it is found that only 20,000 words need to be selected to create a high-frequency dictionary. For phrases in the high-frequency dictionary, they are converted into numbers. For example, "Beijing" is converted to 1, and "building" is converted to 2. For phrases in non-high-frequency dictionaries, we directly ignore the word and do not encode it. The benefit of doing this for model training is that because the frequency of rare words is very low, the deep learning model cannot extract appropriate features, and may interfere with model training and affect the accuracy of the model. Assuming that only 'tianyou' is not selected into the high-frequency dictionary, the common vocabulary ['Huaian City', 'Hongze County', 'Gaolingjian', 'Zhendong', 'Qidao', 'Tianyou', 'Shangdu', 'Hongze', 'Aidong', 'Air Conditioning', 'Maintenance', 'Service Department'] can be converted into the following vector [9393, 15464, 6820, 8029, 13495, 18969, 14060, 11759, 5891, 11015, 12893]. Since 'tianyou' is not in the dictionary, the word vector only has 11 elements.

[0033] S103, padding the word vectors according to a unified length.

[0034] Since the length of each delivery address is different and the number of words contained is also different, but the model needs a uniform length for training, the present invention fixes the word vector of the delivery address to a length of 20 dimensions. For word vectors with a length of more than 20 words, they are truncated and only the first 20 words are used. For word vectors with a length of less than 20 words, 0 is added at the end. After the word vector of Mr. Zhang's delivery address is filled, it will become [9393, 15464, 6820, 8029, 13495, 18969, 14060, 11759, 5891, 11015, 12893, 0, 0, 0, 0, 0, 0, 0].

[0035] S104, using one-hot encoding to convert the padded word vector, and then using word embedding to convert the word vector into a low-dimensional word vector matrix.

[0036] Although we have converted the address into a word vector, these numbers in the word vector itself only record the corresponding position of a word in the dictionary, and there is no corresponding relationship. If it is directly put into the model for training, this will make the model mistakenly believe that there is a relationship between the size order of the two words. Therefore, in the present invention, it is necessary to use the method of unique hot encoding for word vectors to convert the matrix. For example: a word has a value of 9393 in the dictionary. For this word, it will be converted into a 1*20000 matrix in which all numbers are 0, and the 0 in the 9393th position is converted to 1. However, the entire matrix generated by unique hot encoding is too sparse, which is not conducive to model training. Here, the word embedding method is introduced. For example: by pre-building a 20000*128 matrix to compress each word from 20000 dimensions to 128 dimensions, then a single address becomes a 20*128 matrix after conversion. That is, the input 20*20000 matrix is ​​multiplied by the 20000*128 matrix, and the output is a 20*128 matrix.

[0037] S105, cluster the word vector matrix according to the clustering algorithm, select the delivery address that exceeds the threshold in the clustering result, add an abnormal label to the delivery address, and generate an abnormal delivery address. The clustering algorithm is the DBSCAN algorithm, which can cluster similar delivery addresses together, select the addresses with a single class address exceeding the threshold in the clustering result, and after combining the user shopping behavior analysis, label the address whether it is abnormal. The user's shopping behavior is analyzed by calling the existing model, that is, by analyzing the user's personal information and historical shopping records to determine whether the user has abnormal consumption behavior, so as to find out the address used by the abnormal user, mark it as an abnormal address and label it with an abnormal label.

[0038] It should be understood that the above DBSCAN (Density-based spatial clustering of applications with noise) is a density-based clustering algorithm with noise. The algorithm defines a cluster as the maximum set of points in a density necklace. It can divide sufficiently high-density areas into clusters and find clusters in data with a lot of noise. The main process of the DBSCAN algorithm is: for a point x in space, if the ε neighborhood of x contains more than m objects, a cluster based on x as the core point is created. Find and merge objects that are directly density-reachable by the core object, and repeat the above operation of determining new clusters and merging until no clusters that can be merged are found, and return the result.

[0039] S106, inputting the abnormal delivery address into the neural network model, and outputting the probability that the abnormal delivery address is a false address.

[0040] See also Figure 2The above neural network model is a RNN (three-layer recurrent neural network layer) deep learning model, and the recurrent cells used in each layer are GRU (Gated Recurrent Unit), and each computing cell has 128 neurons. The processing result of the RNN deep learning model is sent to the sigmoid function to output the probability of a false address. The delivery address with abnormal labels obtained by the above DBSCAN clustering algorithm can be judged within 2ms using the RNN deep learning model.

[0041] S107, regularly obtain the delivery address, use steps S101 to S105 to process the delivery address, and input the processing results into the neural network model for training, so as to fine-tune the model training set, so that the newly generated false order address also has a high accuracy rate.

[0042] See also Figure 3 The present invention also provides a false delivery address prediction device based on deep learning, comprising:

[0043] The word segmentation module 101 is used to obtain the user's delivery address and segment the delivery address to obtain common words.

[0044] The encoding module 102 is used to encode common words and generate word vectors corresponding to the delivery address.

[0045] The padding module 103 is used to pad the word vectors according to a uniform length.

[0046] The conversion module 104 is used to convert the padded word vectors by using one-hot encoding, and then convert the word vectors into a low-dimensional word vector matrix by using word embedding.

[0047] The clustering module 105 is used to cluster the word vector matrix according to the clustering algorithm, select the delivery addresses exceeding the threshold in the clustering results, and add abnormal labels to the delivery addresses to generate abnormal delivery addresses.

[0048] The output module 106 is used to input the abnormal delivery address into the neural network model and output the probability that the abnormal delivery address is a false address.

[0049] It also includes an updating module 107, which is used to periodically obtain the delivery address, process the delivery address, and input the processing result into the neural network model for training.

[0050] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0051] In summary, the beneficial effects of the present invention include: segmenting the user's delivery address by using a segmentation algorithm with a custom dictionary added, encoding the phrases using a high-frequency dictionary after segmentation, obtaining the input matrix through unique hot encoding and word embedding, using the labels obtained based on the DBSCAN clustering algorithm, training the deep learning model to classify and predict the delivery address, and regularly annotating the newly acquired data through the clustering algorithm to use the deep learning model to identify the newly generated false addresses. After receiving the user's order, the order address is identified to determine whether it is a false address for brushing orders, which can effectively prevent the loss of merchants and the damage to the rights and interests of normal customers caused by brushing orders. The training data used in the present invention is annotated by a clustering algorithm, and no manual participation is required, thereby saving costs. At the same time, the present invention only needs 2ms for the recognition time of a single address, which significantly improves the efficiency compared with manual work. Compared with the existing abnormal address detection schemes on the market, it has a higher accuracy rate and a very fast running speed. Moreover, the present invention does not rely on external data assistance at all, can be operated in an internal network or in a network-free environment, and has a fast running speed and is not interfered by external data. In addition, the present invention does not require manual data labeling, and automatically obtains the label of the order address through clustering algorithm and user behavior analysis.

[0052] It should be understood that if the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0053] The technical scope of the present invention is not limited to the contents in the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A method for predicting false delivery addresses based on deep learning, characterized in that: The steps include: S101, obtaining a user's delivery address, and segmenting the delivery address to obtain common words; S102, encoding the common vocabulary to generate a word vector corresponding to the delivery address; S103, padding the word vector according to a unified length; S104, converting the padded word vectors by one-hot encoding, and then converting the word vectors into a low-dimensional word vector matrix by word embedding; S105, clustering the word vector matrix according to a clustering algorithm, selecting a delivery address exceeding a threshold in the clustering result, and adding an abnormal label to the delivery address to generate an abnormal delivery address; S106, inputting the abnormal delivery address into a neural network model, and outputting the probability that the abnormal delivery address is a false address.

2. The method for predicting false delivery addresses based on deep learning according to claim 1, characterized in that: The encoding of common words includes: selecting high-frequency words from all common words to create a high-frequency dictionary, and converting the words in the high-frequency dictionary into numbers to generate a word vector corresponding to the delivery address.

3. The method for predicting false delivery addresses based on deep learning according to claim 1, characterized in that: Also includes: S107, regularly obtain the delivery address, use steps S101 to S105 to process the delivery address, and input the processing results into the neural network model for training.

4. The method for predicting false delivery addresses based on deep learning according to claim 1, characterized in that: The clustering algorithm is the DBSCAN algorithm.

5. The method for predicting false delivery addresses based on deep learning according to claim 1, characterized in that: The neural network model is an RNN deep learning model, and the recurrent cells used in each layer are GRUs, and each operation cell has 128 neurons.

6. A false delivery address prediction device based on deep learning, characterized in that: include: A word segmentation module is used to obtain a user's delivery address and segment the delivery address to obtain common words; An encoding module, used for encoding the common vocabulary to generate a word vector corresponding to the delivery address; A padding module, used for padding the word vector according to a uniform length; A conversion module, used to convert the padded word vectors by one-hot encoding, and then convert the word vectors into a low-dimensional word vector matrix by word embedding; A clustering module, used to cluster the word vector matrix according to a clustering algorithm, select the delivery addresses exceeding the threshold in the clustering results, and add abnormal labels to the delivery addresses to generate abnormal delivery addresses; The output module is used to input the abnormal delivery address into the neural network model and output the probability that the abnormal delivery address is a false address.

7. The false delivery address prediction device based on deep learning according to claim 6, characterized in that: It also includes an updating module, which is used to periodically obtain the delivery address, process the delivery address, and input the processing result into the neural network model for training.

Images