WO2023024411A1

WO2023024411A1 - Association rule assessment method and apparatus based on machine learning

Info

Publication number: WO2023024411A1
Application number: PCT/CN2022/071425
Authority: WO
Inventors: 蒋雪涵
Original assignee: 平安科技（深圳）有限公司
Priority date: 2021-08-25
Filing date: 2022-01-11
Publication date: 2023-03-02
Also published as: CN113656558A; CN113656558B

Abstract

The present application relates to the technical field of artificial intelligence. Disclosed are an association rule assessment method and apparatus based on machine learning, and a computer device and a readable storage medium. The method comprises: mining association rules from an item set by using an item co-occurrence condition, wherein each association rule comprises an antecedent and a consequent, and the item co-occurrence condition is that an item occurs in both the antecedent and the consequent; performing feature extraction on collected item text information by using a pre-trained text information encoder and an antecedent prediction machine, so as to obtain a code vector representation of the item text information, wherein the text information encoder is used for predicting whether the consequent occurs in the association rule, and the antecedent prediction machine is used for predicting whether the antecedent occurs in the association rule; and in response to an assessment instruction for the association rules, assessing each association rule according to the code vector representation of the item text information, so as to obtain an assessment result, which reflects a causal relationship between the antecedent and the consequent in the association rule. By means of the present application, a causal relationship assessment can be performed on an association rule, thereby improving the interpretability of the association rule.

Description

Method and device for evaluating association rules based on machine learning

This application claims the priority of the Chinese patent application submitted to the China Patent Office on August 25, 2021, with the application number 202110980623.X and the application name "Method and device for evaluating association rules based on machine learning", the entire content of which Incorporated in the application by reference.

technical field

The present application relates to the technical field of artificial intelligence, in particular to a method, a device, a computer device and a readable storage medium for evaluating association rules based on machine learning.

Background technique

Association analysis is a commonly used mining algorithm, which is used to mine the internal associations between data, and can be applied to many application scenarios in life. For example, in shopping scenarios, association rules are used to discover the internal commonality of group buying habits and guide supermarket products. Placement, in medical scenarios, uses association rules to mine the possibility of patients consuming medical items to guide doctors in case diagnosis.

Usually, association rules can be proposed by domain experts, and candidate sets that meet certain measurement values, such as confidence, support, and promotion, can also be obtained through data mining, and then the rationality can be confirmed by experts. However, the inventor realized that the items in the association rules are determined by different factors, and the combined effect of these factors has biased the evaluation of the relationship between items. For example, the association rule is "oral anesthesia → root canal" where "oral anesthesia" may It is caused by the patient's "tooth extraction" or "root canal treatment", and the "root canal" is only caused by the patient's "root canal treatment", which makes the "oral anesthesia" deduce that the "root canal" has a certain deviation However, the mining process of the above-mentioned association rules has the following two deficiencies. One is that there are a large number of false positives in the mining association rules, and the rules are too complex, which will lead to weak interpretability of the association rules; The mining of association rules depends on expert experience, and the opinions of different experts may differ, resulting in the subjectivity of association rules.

Contents of the invention

In view of this, the present application provides a method, device, computer equipment and readable storage medium for evaluating association rules based on machine learning, the main purpose of which is to solve the subjectivity and interpretability of association rules mined in the prior art. problem of weakness.

According to one aspect of the present application, a method for evaluating association rules based on machine learning is provided, the method comprising:

Use item co-occurrence condition to mine association rules from item collection, described association rule comprises antecedent and subsequent item, and described item co-occurrence condition is that item occurs simultaneously in antecedent and subsequent item;

Use the pre-trained text information encoder and antecedent predictor to extract the features of the collected item text information, and obtain the coded vector representation of the item text information, and the text information encoder is used to determine whether the consequent appears in the association rules Forecasting, the antecedent predictor is used to predict whether the antecedent appears in the association rule;

In response to the evaluation instruction of the association rules, each association rule is evaluated according to the coded vector representation of the item text information, and an evaluation result reflecting the causal relationship between the antecedent and the subsequent in the association rule is obtained.

According to another aspect of the present application, a device for evaluating association rules based on machine learning is provided, the device comprising:

A mining unit, configured to use item co-occurrence conditions to mine association rules from the item collection, the association rules include antecedents and subsequent items, and the item co-occurrence conditions are that items in the antecedent and subsequent items appear simultaneously;

The extraction unit is used to perform feature extraction on the collected item text information by using the pre-trained text information encoder and antecedent predictor to obtain the coded vector representation of the item text information, and the text information encoder is used for the association rule Predict whether the consequent appears, and the antecedent predictor is used to predict whether the antecedent appears in the association rules;

The evaluation unit is configured to respond to the evaluation instruction of the association rules, evaluate each association rule according to the coded vector representation of the item text information, and obtain an evaluation result reflecting the causal relationship between the antecedent and the consequent in the association rule.

According to still another aspect of the present application, a computer device is provided, including a memory and a processor, the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, it implements association rules based on machine learning. Steps in the method of evaluation.

According to still another aspect of the present application, a computer storage medium is provided, on which computer readable instructions are stored, and when the computer readable instructions are executed by a processor, the steps of the method for evaluating association rules based on machine learning are realized.

When evaluating association rules, this application introduces causal correction to evaluate the causal relationship of association rules obtained by mining, removes the features that are only related to the antecedent or the latter of the association rules, and obtains the causal explanation of the latter for the former. Increase the interpretability of association rules, thereby reducing false positives in association rules and avoiding the influence of subjective factors on association rule screening.

Description of drawings

Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiment. The drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the application. Also throughout the drawings, the same reference numerals are used to designate the same parts. In the attached picture:

FIG. 1 shows a schematic flowchart of a method for evaluating association rules based on machine learning provided by an embodiment of the present application;

FIG. 2 shows a schematic flowchart of another method for evaluating association rules based on machine learning provided by an embodiment of the present application;

FIG. 3 shows a schematic structural diagram of a device for evaluating association rules based on machine learning provided by an embodiment of the present application;

FIG. 4 shows a schematic structural diagram of another apparatus for evaluating association rules based on machine learning provided by an embodiment of the present application.

Detailed ways

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided for more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

The embodiments of the present application may acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. .

Artificial intelligence basic technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation/interaction systems, and mechatronics. Artificial intelligence software technology mainly includes computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning/deep learning.

The embodiment of this application provides a method for evaluating association rules based on machine learning. By using the coded vector representation of project text information to evaluate the causal relationship of each association rule, the causal screening of association rules is realized, and the number of association rules is increased. Interpretability, as shown in Figure 1, the method includes:

101. Mining association rules from item collections using item co-occurrence conditions.

Among them, the association rule has the form of condition A → condition B, which means that condition B can be obtained when condition A is satisfied. Here, condition A and condition B are the antecedents and postconditions contained in the association rule respectively. In the association rule , the item on the left of the arrow is the antecedent of the association rule, and the item on the right of the arrow is the aftermath of the association rule. Here, the antecedent and the aftermath can be one item or multiple items, and the set of items can involve different fields. For example, for drug items and inspection items in the medical field, and for payment items and evaluation items in the online shopping field, a large amount of user text information can be obtained through a pre-set interface channel, and a large amount of user text information can be aggregated to form an item collection . The item co-occurrence condition is that the antecedent and the subsequent item in the association rules appear at the same time. For example, in the medical consultation text information, the patient purchased item A and item B at the same time. The premise of the item co-occurrence condition is that the item in the antecedent and the latter item Items in the file appear at the same time.

Specifically, in the process of using item co-occurrence conditions to mine candidate association rules from the item set, the first is the generation of the item set, which can be generated using the PF growth algorithm, and then filter out the association rules that meet the preset conditions from the full arrangement of the item set , where the preset condition is that the support and confidence are greater than a given threshold at the same time, the support is defined as the co-occurrence frequency of the antecedent and the consequent, and the confidence is defined as the ratio of the co-occurrence frequency of the antecedent and the latter to the probability of the antecedent , the antecedent probability is the co-occurrence frequency of all items in the antecedent.

For the frequency of co-occurrence of the antecedent and the posterior, there are 1,000 inspection items in the item collection, among which, there are 800 medical visit text information for both blood routine examination and urine routine examination, so the co-occurrence frequency of blood routine and urine routine items is 0.8. For the probability of the antecedent, the antecedent is a collection of items. There can be one item or multiple items. If it is one item, the antecedent probability is the occurrence probability of the item. If it is multiple items, the antecedent probability is multiple The frequency with which items co-occur. The given threshold here can be set according to the actual project requirements. If the actual project requirement is quality inspection, it will be judged as a violation sample if it violates the rules. It is necessary to set the preset conditions of high confidence and low support.

In this embodiment of the application, the execution subject can be a device that evaluates association rules based on machine learning, and is specifically applied on the server side. Here, the item co-occurrence condition is used to mine the association rules from the item collection to meet the preset conditions, which can be used as association rules Preliminary screening of , which can distinguish the association relationship existing in the project collection.

The above servers can be independent servers, or provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery network (Content Delivery Network) , CDN), and cloud servers for basic cloud computing services such as big data and artificial intelligence platforms.

102. Use the pre-trained text information encoder and antecedent predictor to perform feature extraction on the collected item text information, and obtain the encoded vector representation of the item text information.

Among them, the text information encoder is used to predict whether the consequent appears in the association rules. Natural language models such as TextCNN and BERT can be used. The input parameter is the item text information, and the output parameter is the encoded vector representation of the item text information. Further, the item text The coded vector representation of information is used to classify, and it can also output the predicted value of whether the subsequent part in the association rule appears in the item text information. The antecedent predictor is used to predict whether the antecedent appears in the association rule, and a deep neural network model can be used , that is, a multi-layer perceptron, the input of the Lth layer is the output of the (L-1)th layer, and its calculation formula is, z ^l = ReLU(w ^l z ^l-1 + b ^l ), where w ^l and b ^l is the model parameter of the L-th layer, RELU is the activation function, its calculation formula is max(0,x), the input parameter is the coded vector representation of the item text information and the predicted value of whether the consequent appears in the item text information in the association rule , the output parameter is the predicted value of whether the antecedent in the association rule appears in the item text information.

The item text information here can be medical text data, and the medical text data can be electronic healthcare records (Electronic Healthcare Record), electronic personal health records, including medical records, electrocardiograms, medical images, and a series of electronic records that are valuable for future reference. .

Here, the text information encoder and the antecedent prediction machine can use the machine algorithm of artificial intelligence to combine whether the antecedent and the aftermath in the association rules appear in the item text information as label data to train the network model to vectorize the item text information It is expressed as an encoded vector representation, and the text information encoder and the antecedent predictor perform adversarial learning during the training process, that is, the optimization goals of the two are opposite. Through adversarial learning, only The information related to the predecessor in the association rule, retains the information related to the former and the latter.

103. In response to an association rule evaluation instruction, evaluate each association rule according to the coded vector representation of the item text information, and obtain an evaluation result reflecting the causal relationship between the antecedent and the consequent in the association rule.

It is understandable that, for each association rule, in the confrontational learning process between the text information encoder and the antecedent predictor, the coded vector of the item text information is used to represent the aftermath of the predicted association rule, and then the latter and the antecedent of the association rule are used. The coding vector of the item text information represents the antecedent of the predicted association rules, which can remove the information irrelevant to the subsequent item in the item text information, and realize the causal correction in the item text information. Further, through the corrected text item information, similar text items The information evaluates the causal contribution of the antecedent to the consequent in association rules.

Specifically for each association rule, the project text information contains a large amount of text, and the project text information that occurred in the previous item can be selected as the text sample set. Here, each association rule will filter out multiple text samples, and then target each text in the text sample set Sample, traversing the coded vector representation of the item text information, querying the K items of text information that are most similar to the coded vector representation of the text sample, which can be obtained by calculating the distance between the coded vectors, and further calculating K for each text sample The probability value of the subsequent occurrence of the association rule in the project text information, and calculate the average value of the occurrence probability of the subsequent occurrence calculated by all text samples, as the evaluation value reflecting the causal relationship between the antecedent and the subsequent in the association rule, the The evaluation value is the representation of the causal relationship between the antecedent and the subsequent in association rules.

The embodiment of the present application provides a method for evaluating association rules based on machine learning, using item co-occurrence conditions to mine association rules from item collections, the association rules include antecedents and subsequent items, and the item co-occurrence conditions are antecedents and The items in the subsequent items appear at the same time, and the pre-trained text information encoder and the antecedent predictor are used to extract the features of the collected item text information, and the encoded vector representation of the item text information is obtained. The text information encoder is used for the association rules. Predict whether the consequent appears. The antecedent predictor is used to predict whether the antecedent appears in the association rules. In response to the evaluation instruction of the association rules, each association rule is evaluated according to the coded vector representation of the item text information, and the obtained It reflects the evaluation result of the causal relationship between the antecedent and the consequent in the association rule. Compared with the method of association rules obtained through data mining in the prior art, this application introduces the method of causal correction to evaluate the causal relationship of the association rules obtained by mining, and removes the features only related to the antecedent or the latter of the association rules, and obtains The causal interpretation of the latter to the antecedent increases the interpretability of association rules, thereby reducing false positives in association rules and avoiding the influence of subjective factors on association rule screening.

The embodiment of this application provides another method for evaluating association rules based on machine learning. By using the coded vector representation of item text information to evaluate the causal relationship of each association rule, the causal screening of association rules can be realized, and association rules can be added. Interpretability, as shown in Figure 2, the method includes:

201. Perform a full arrangement on the subset of frequent items included in the item set.

Wherein, the item set is equivalent to a set of different items, and each item is an item in the item set, and the item can be a customer consumption item, for example, milk, biscuits, medical payment items, for example, blood routine, urine test. Since the association between items in the item collection can guide consumption or auxiliary medical reimbursement to a certain extent, for example, customers will purchase item C while purchasing item A and item B, and patients will pay for medical item C while paying Medical Item D and Medical Item E. In order to reflect the relationship between items, the frequent item subset is to contain at least one item in the item set, and the number of times the contained items appear in a record at the same time is greater than or equal to the minimum support. Specifically, in the process of full arrangement, according to the item set The number of items in the medium determines the frequent item subsets containing different item numbers, and lists all the item subsets according to the item number, and further filters out the frequent item subsets whose support degree is greater than the preset threshold from the item subset. Specifically, the process of screening frequent item subsets can follow the following two principles. If an item subset is a frequent item subset, then the item subset is a frequent item subset. If an item subset is an infrequent item Subset, then the superset of the item subset is the infrequent item subset, and this process can save the generation time of the frequent item subset.

For example, the collection of items is {A,B,C,D}, first list the subset of items containing one item as follows: {A}, {B}, {C}, {D}, then list the subset of items containing two items The subset of items in is as follows: {A,B}, {A,C}, {A,D}, {B,C}, {B,D}, {C,D}, and then list the three items The subset of items is as follows: {A,B,C}, {A,B,D}, {A,C,D}, {B,C,D}, and the subset of frequent items with support greater than 3/5{ A}, {B}, {A,B}, {B,C}, {A,C}, {A,B,C}.

202. Generate candidate association rules for the frequent item subset, and filter the candidate association rules by using preset parameter indicators to obtain candidate rules that meet preset conditions.

Among them, the parameter index includes support and confidence at least, the support is the co-occurrence frequency of the antecedent and the consequent, and the confidence is the ratio of the support to the probability of the antecedent. The candidate association rules generated for the frequent item subset are equivalent to the derivation relationship of the items in the frequent item subset. For example, the frequent item subset is {A, B, C}, and the derivation relationship of the items may include: A, B=>C, A, C => B, B, C => A, A => B, C, B => A, C, C => A, B.

Furthermore, in order to generate effective association rules between items, it is necessary to calculate whether the candidate association rules formed between items meet the parameter indicators. The correlation between them is weak and not referential. Here, the threshold value of the parameter index threshold is set as a preset condition to filter out candidate association rules with weak correlation and improve the reliability of the association rules.

Specifically, the items in each frequent item subset can form multiple candidate association rules, and the confidence and support can be calculated separately for each candidate association rule. If the confidence and support meet the preset conditions, that is, both are greater than the set Confidence threshold and support threshold of , indicating that the candidate association rule has strong relevance and can be retained; otherwise, the candidate association rule is filtered.

203. For each association rule, use a pre-determined whether the antecedent and the consequent appear in the item text information as tag data.

In this application, each item text information will contain at least one item, specifically for each item association rule, the appearance of the preceding item in the item text information is equivalent to the occurrence of all items in the item text information in the item, for example , the antecedent is the blood routine and urine routine of the project, and if the project text information contains the blood routine and urine routine, it is considered that the antecedent appears in the project text information, and similarly, whether the latter appears in the project text information is Items in the aftermath appear in the item text information.

204. Input the item text information carrying the label data into the first network model for training, and construct a text information encoder.

Specifically, during the training process, the text information encoder can obtain the coded vector representation of the item text information, and use the coded vector representation to predict whether the consequence of the association rule appears in the item text information. The optimization goal of the text information coder is Maximize the prediction of whether the consequents in the association rules appear in the item text information. That is to say, for each association rule, the label data that appears in the item text information of the subsequent items in the association rule will be used for training, and the multi-label loss function will be combined during the training process, where each label corresponds to a cross-entropy loss function , multiple labels are added for multiple cross-entropy loss functions, and the specific loss function is publicly expressed as:

Among them, y is whether the consequent appears in the item text information in the association rule,

is the predicted value of whether the consequent appears in the item text information in the encoder output association rule, x is whether the antecedent appears in the item text information in the association rule,

is the predicted value of whether the antecedent appears in the project text information in the association rules output by the antecedent predictor.

205. Input the encoded vector representation of the item text information output by the first network model and the prediction value of whether the consequent appears in the item text information in the association rules to the second network model for training, and build an antecedent predictor.

Specifically, during the training process, the antecedent predictor can predict whether the antecedent in the association rule appears in the item text information by using the coded vector representation and the predicted value of the latter in the user text information in the association rule. The optimization goal of the antecedent predictor is to maximize the prediction of whether the antecedent appears in the project text information in the association rules. That is to say, for each association rule, the label data that appears in the item text information of the previous item in the association rule will be used for training, and the multi-label loss function will be combined during the training process. The loss function is also the loss of the multi-label problem function, the formula is expressed as:

206. Use the pre-trained text information encoder and antecedent predictor to perform feature extraction on the collected item text information, and obtain an encoded vector representation of the item text information.

It should be noted that here, the text information encoder and the antecedent predictor perform confrontational learning during the training process, so that the information related to the antecedent in the association rules is removed from the item text information, and the antecedent and the aftermath in the association rules are retained. Related information.

207. In response to the evaluation instruction of the association rule, for each association rule, calculate an evaluation value reflecting the causal relationship between the antecedent and the consequent in the association rule according to the coded vector representation of the item text information.

Among them, the item text information contains multiple texts. Specifically, for each association rule, the text that appears in the preceding item in the association rule can be selected from the item text information as the sample text, and then the encoded vector representation of the item text information is traversed, and the query and each The coded vectors of each sample text represent the texts that meet the similarity condition, as the similar target text of each sample text, for each sample text similar target text, calculate the evaluation reflecting the causal relationship between the antecedent and the consequent in the association rule value.

Specifically, in the process of calculating the evaluation value reflecting the causal relationship between the antecedent and the consequent in the association rules, we can calculate the probability value of the latter in the association rules in the target similar texts for the similar target texts of each sample text, and obtain each The probability value of each sample text meeting the evaluation condition is obtained by weighting the probability value of each sample text meeting the evaluation condition to obtain the evaluation value reflecting the causal relationship between the antecedent and the latter in the association rules.

For example, the project text information contains 100 texts. For each association rule, the project text information that occurs before the selected item contains 10 sample texts, that is, sample text 1-10. For sample text 1, traverse the encoding vectors of 100 texts means, find 5 target texts that are similar to the coded vector representation of the sample text, and further calculate the probability value of the occurrence of the consequent of the association rule in the 5 target texts, if the probability value is 0.8, it means that in the 5 target texts There are 4 postconditions of association rules. Similarly, for the sample texts 2-10, the qualified probability values a2, a3, a4, and a5 can be calculated, and the probability value is further weighted to obtain the average value (a1+a2+a3 +a4+a5)/5, to obtain the evaluation value reflecting the causal relationship between the antecedent and the consequent in the association rule.

208. If the evaluation value is greater than the preset threshold, it is determined that there is a causal relationship between the antecedent and the consequent in the association rule.

It can be understood that the evaluation value here can represent the mutual causal interpretation between the antecedent and the subsequent in the association rules, which can more intuitively reflect whether there is a causal relationship between the antecedents and the latter, and increase the interpretability of the association rules. If the evaluation value is greater than the preset threshold, it means that the causal explanatory power between the antecedent and the subsequent in the association rule is strong, indicating that there is a causal relationship between the antecedent and the latter; otherwise, the explanatory power of the association rule is weak, namely Although the association rules have been mined, the antecedents and consequents in the association rules are less rational.

In practical application scenarios, the evaluation of association rules can be used to filter or explain the mined association rules to achieve data collocation and data prediction, for example, for clothing collocation in shopping scenarios, epidemic situation judgment for livestock breeding scenarios, Business push for page access scenarios, etc.

In this application, for each association rule, the pre-trained text information encoder and antecedent predictor are used to extract the features of the pre-collected item text information, and the vector encoding representation of the extracted item text information is used to extract the antecedents in the association rules. Whether there is a causal relationship between the item and the subsequent item can be evaluated, and the relevant features of the antecedent or subsequent item in the association rules can be removed, and the causal explanation of the latter item for the anterior item can be obtained, thereby reducing false positives of potential rules and reducing the time spent in mining association rules. Subjectivity, using information encoders and antecedent predictors, enables fast and stable iterations, improving the interpretability of association rules.

Further, as a specific implementation of the method described in FIG. 1 , an embodiment of the present application provides a device for evaluating association rules based on machine learning. As shown in FIG. 3 , the device includes: a mining unit 31, an extraction unit 32 , evaluation unit 33 .

The mining unit 31 can be used to mine association rules from the item collection using item co-occurrence conditions, the association rules include antecedents and subsequent items, and the item co-occurrence conditions are that items in the antecedent and subsequent items appear simultaneously;

The extraction unit 32 can be used to perform feature extraction on the collected item text information by using a pre-trained text information encoder and antecedent predictor to obtain a coded vector representation of the item text information, and the text information encoder is used for the associated Predict whether the latter appears in the rule, and the antecedent predictor is used to predict whether the antecedent appears in the association rule;

The evaluation unit 33 may be configured to respond to the evaluation instruction of the association rules, evaluate each association rule according to the coded vector representation of the item text information, and obtain an evaluation result reflecting the causal relationship between the antecedent and the subsequent in the association rule .

The embodiment of the present application provides a device for evaluating association rules based on machine learning, using item co-occurrence conditions to mine association rules from item collections, the association rules include antecedents and subsequent items, and the item co-occurrence conditions are The items in the subsequent items appear at the same time, and the pre-trained text information encoder and the antecedent predictor are used to extract the features of the collected item text information, and the encoded vector representation of the item text information is obtained. The text information encoder is used for the association rules. Predict whether the consequent appears. The antecedent predictor is used to predict whether the antecedent appears in the association rules. In response to the evaluation instruction of the association rules, each association rule is evaluated according to the coded vector representation of the item text information, and the obtained It reflects the evaluation result of the causal relationship between the antecedent and the consequent in the association rule. Compared with the method of association rules obtained through data mining in the prior art, this application introduces the method of causal correction to evaluate the causal relationship of the association rules obtained by mining, and removes the features only related to the antecedent or the latter of the association rules, and obtains The causal interpretation of the latter to the antecedent increases the interpretability of association rules, thereby reducing false positives in association rules and avoiding the influence of subjective factors on association rule screening.

As a further description of the device for evaluating association rules based on machine learning shown in FIG. 3 , FIG. 4 is a schematic structural diagram of another device for evaluating association rules based on machine learning according to an embodiment of the present application, as shown in FIG. 4 , the item co-occurrence condition is that the preceding item and the subsequent item appear simultaneously in the association rule, and the mining unit 31 includes:

The arrangement module 311 can be used to perform full arrangement on the subset of frequent items included in the item set;

The selection module 312 can be used to generate candidate association rules for the subset of frequent items, and use preset parameter indicators to filter the candidate association rules to obtain candidate rules that meet the preset conditions, and the parameter indicators include at least supporting degree and confidence, the support degree is the co-occurrence frequency of the antecedent and the consequent, and the confidence degree is the ratio of the support degree to the probability of the antecedent.

In a specific application scenario, as shown in Figure 4, the device further includes:

The generation unit 34 can be used to perform feature extraction on the collected item text information using the pre-trained text information encoder and antecedent predictor, and obtain the coded vector representation of the item text information. For each association rule, use Predetermining whether said antecedent and said consequent appear in item text information as tag data;

The first construction unit 35 can be used to input the item text information carrying the label data into the first network model for training, and construct a text information encoder whose optimization goal is to maximize the prediction in the association rules. Whether the aftermath appears in the project text information;

The second construction unit 36 can be used to input the encoded vector representation of the first network model output item text information and the predicted value of whether the consequent appears in the item text information in the association rules to the second network model for training, and construct An antecedent predictor, the optimization objective of the antecedent predictor is to maximize the prediction of whether the antecedent in the association rule appears in the project text information.

In a specific application scenario, the text information encoder and the antecedent predictor perform adversarial learning during the training process, so that the information related to the antecedents in the association rules is removed from the item text information, and the antecedents in the association rules are retained. Information related to the item and the subsequent item.

In a specific application scenario, as shown in FIG. 4, the evaluation unit 33 includes:

The calculation module 331 can be used to calculate, for each association rule, an evaluation value that reflects the causal relationship between the antecedent and the consequent in the association rule according to the coded vector representation of the item text information;

The determination module 332 may be configured to determine that there is a causal relationship between the antecedent and the consequent in the association rule if the evaluation value is greater than a preset threshold.

In a specific application scenario, as shown in Figure 4, the item text information includes a plurality of texts, and the calculation module 331 includes:

The selection sub-module 3311 can be used to select the text that appears in the preceding item in the association rule from the item text information as the sample text for each association rule;

The query sub-module 3312 can be used to traverse the coded vector representation of the item text information, and query the text that meets the similarity condition with the coded vector representation of each sample text, as the similar target text of each sample text;

The calculation sub-module 3313 can be used to calculate the evaluation value reflecting the causal relationship between the antecedent and the consequent in the association rule for the similar target text of each sample text.

In a specific application scenario, as shown in FIG. 4 , the calculation submodule 3313 can specifically be used to calculate the probability value of the occurrence of the consequent in the association rule in the similar target text of each sample text, Obtain the probability value of each sample text meeting the evaluation conditions;

The calculation sub-module 3313 can also be used to obtain the evaluation value reflecting the causal relationship between the antecedent and the consequent in the association rule by weighting the average probability value of each sample text meeting the evaluation condition.

It should be noted that for other corresponding descriptions of the functional units involved in the apparatus for evaluating association rules based on machine learning provided in this embodiment, reference may be made to the corresponding descriptions in FIG. 1 and FIG. 2 , and details are not repeated here.

Based on the method shown in Figure 1 and Figure 2 above, correspondingly, this embodiment also provides a readable storage medium, the readable storage medium may be non-volatile or volatile, and Computer-readable instructions are stored on it, and when the computer-readable instructions are executed by the processor, the above-mentioned method for evaluating association rules based on machine learning as shown in FIG. 1 and FIG. 2 is realized.

Based on this understanding, the technical solution of the present application can be embodied in the form of software products, which can be stored in a non-volatile storage medium (which can be CD-ROM, U disk, mobile hard disk, etc.), including several The instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute the methods described in various implementation scenarios of the present application.

Based on the method shown in Figure 1 and Figure 2 above, and the virtual device embodiment shown in Figure 3 and Figure 4, in order to achieve the above purpose, the embodiment of this application also provides a computer device, which can be a personal computer, Servers, network devices, etc., the physical device includes a readable storage medium and a processor; the readable storage medium is used to store computer-readable instructions; the processor is used to execute computer-readable instructions to achieve the above as shown in Figure 1 and Figure 2 The method for evaluating association rules based on machine learning shown in

Optionally, the computer device may also include a user interface, a network interface, a camera, a radio frequency (Radio Frequency, RF) circuit, a sensor, an audio circuit, a WI-FI module, and the like. The user interface may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the like, and optional user interfaces may also include a USB interface, a card reader interface, and the like. Optionally, the network interface may include a standard wired interface, a wireless interface (such as a Bluetooth interface, a WI-FI interface) and the like.

Those skilled in the art can understand that the physical device structure of the device for evaluating association rules based on machine learning provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine some components, or different component arrangements.

The readable storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the above-mentioned computer equipment, and supports the operation of information processing programs and other software and/or programs. The network communication module is used to realize the communication among various components inside the readable storage medium, and communicate with other hardware and software in the physical device.

Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be realized by means of software plus a necessary general-purpose hardware platform, or by hardware. By applying the technical solution of this application, compared with the current prior art, this application introduces the method of causal correction to evaluate the causal relationship of the association rules obtained by mining, and removes the features only related to the antecedent or the latter of the association rules, and obtains The causal interpretation of the latter to the antecedent increases the interpretability of association rules, thereby reducing false positives in association rules and avoiding the influence of subjective factors on association rule screening.

Those skilled in the art can understand that the accompanying drawing is only a schematic diagram of a preferred implementation scenario, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the present application. Those skilled in the art can understand that the modules in the devices in the implementation scenario can be distributed among the devices in the implementation scenario according to the description of the implementation scenario, or can be located in one or more devices different from the implementation scenario according to corresponding changes. The modules of the above implementation scenarios can be combined into one module, or can be further split into multiple sub-modules.

The serial numbers of the above application are for description only, and do not represent the pros and cons of the implementation scenarios. The above disclosures are only a few specific implementation scenarios of the present application, but the present application is not limited thereto, and any changes conceivable by those skilled in the art shall fall within the protection scope of the present application.

Claims

A method for evaluating association rules based on machine learning, wherein the method includes:

Using item co-occurrence conditions to mine association rules from the item collection, the association rules include antecedents and subsequent items, and the item co-occurrence conditions are that the items in the antecedent and subsequent items appear simultaneously;

Use the pre-trained text information encoder and antecedent predictor to extract the features of the collected item text information, and obtain the coded vector representation of the item text information, and the text information encoder is used to determine whether the consequent appears in the association rules Forecasting, the antecedent predictor is used to predict whether the antecedent appears in the association rule;

In response to the evaluation instruction of the association rules, each association rule is evaluated according to the coded vector representation of the item text information, and an evaluation result reflecting the causal relationship between the antecedent and the subsequent in the association rule is obtained.
The method according to claim 1, wherein the item co-occurrence condition is that the antecedent and the subsequent item in the association rule appear simultaneously, and the use of the item co-occurrence condition to mine the association rule from the item collection specifically includes:

Perform a full permutation of the frequent item subsets contained in the item set;

Generate candidate association rules for the subset of frequent items, and use preset parameter indicators to filter the candidate association rules to obtain candidate rules that meet preset conditions, the parameter indicators include at least support and confidence, and the The support is the co-occurrence frequency of the antecedent and the consequent, and the confidence is the ratio of the support to the probability of the antecedent.
The method according to claim 1, wherein, before said utilizing the pre-trained text information encoder and antecedent predictor to perform feature extraction on the collected item text information, and obtain the coded vector representation of the item text information, said method Also includes:

For each association rule, predetermining whether the antecedent and the consequent appear in item text information is used as tag data;

Input the item text information carrying the labeled data into the first network model for training, and build a text information encoder whose optimization goal is to maximize the prediction of whether the consequent appears in the item text information in the association rules ;

The coded vector representation of the output item text information of the first network model and the predicted value of whether the consequent appears in the item text information in the association rules are input to the second network model for training, and the antecedent predictor is constructed, and the antecedent The optimization goal of the predictor is to maximize whether the antecedents in the association rules appear in the project text information.
The method according to claim 3, wherein the text information encoder and the antecedent predictor perform confrontational learning during the training process, so that information related to the antecedent in the association rule is removed from the item text information, and Preserve the information related to the former and the latter in the association rules.
The method according to any one of claims 1-4, wherein, each association rule is evaluated according to the encoded vector representation of the item text information, and the causality between the antecedent and the subsequent in the association rule is obtained The results of the assessment of the relationship, including:

For each association rule, according to the coding vector representation of the item text information, the evaluation value reflecting the causal relationship between the antecedent and the subsequent part in the association rule is calculated;

If the evaluation value is greater than the preset threshold, it is determined that there is a causal relationship between the antecedent and the consequent in the association rule.
The method according to claim 5, wherein the item text information contains a plurality of texts, and for each association rule, according to the encoding vector representation of the item text information, the antecedent and the subsequent item in the association rule are reflected The estimated value of the causal relationship between, specifically includes:

For each association rule, select the text that appears in the preceding item in the association rule from the item text information as the sample text;

Traverse the coded vector representation of the item text information, query the text that meets the similarity condition with the coded vector representation of each sample text, and use it as the similar target text of each sample text;

For the similar target text of each sample text, an evaluation value reflecting the causal relationship between the antecedent and the consequent in the association rule is calculated.
The method according to claim 6, wherein, for the similar target text of each sample text, calculating an evaluation value reflecting the causal relationship between the antecedent and the consequent in the association rule specifically includes:

For the similar target text of each sample text, calculate the probability value of the consequent in the association rule in the target similar text, and obtain the probability value that each sample text meets the evaluation condition;

By weighting and averaging the probability values of each sample text meeting the evaluation conditions, the evaluation value reflecting the causal relationship between the antecedent and the subsequent in association rules is obtained.
A device for evaluating association rules based on machine learning, wherein the device includes:

A mining unit, configured to use item co-occurrence conditions to mine association rules from the item collection, the association rules include antecedents and subsequent items, and the item co-occurrence conditions are that items in the antecedent and subsequent items appear simultaneously;

The extraction unit is used to perform feature extraction on the collected item text information by using the pre-trained text information encoder and antecedent predictor to obtain the coded vector representation of the item text information, and the text information encoder is used for the association rule Predict whether the consequent appears, and the antecedent predictor is used to predict whether the antecedent appears in the association rules;

The evaluation unit is configured to respond to the evaluation instruction of the association rules, evaluate each association rule according to the coded vector representation of the item text information, and obtain an evaluation result reflecting the causal relationship between the antecedent and the consequent in the association rule.
A computer device, comprising a memory and a processor, wherein the memory stores computer-readable instructions, wherein, when the processor executes the computer-readable instructions, the steps of a method for evaluating association rules based on machine learning are implemented, including :

Using item co-occurrence conditions to mine association rules from the item collection, the association rules include antecedents and subsequent items, and the item co-occurrence conditions are that the items in the antecedent and subsequent items appear simultaneously;

Use the pre-trained text information encoder and antecedent predictor to extract the features of the collected item text information, and obtain the coded vector representation of the item text information, and the text information encoder is used to determine whether the consequent appears in the association rules Forecasting, the antecedent predictor is used to predict whether the antecedent appears in the association rule;

In response to the evaluation instruction of the association rules, each association rule is evaluated according to the coded vector representation of the item text information, and an evaluation result reflecting the causal relationship between the antecedent and the subsequent in the association rule is obtained.
The computer device according to claim 9, wherein the item co-occurrence condition is that the antecedent and the subsequent item in the association rule appear at the same time, and the use of the item co-occurrence condition to mine the association rule from the item collection specifically includes:

Perform a full permutation of the frequent item subsets contained in the item set;

Generate candidate association rules for the subset of frequent items, and use preset parameter indicators to filter the candidate association rules to obtain candidate rules that meet preset conditions, the parameter indicators include at least support and confidence, and the The support is the co-occurrence frequency of the antecedent and the consequent, and the confidence is the ratio of the support to the probability of the antecedent.
The computer device according to claim 9, wherein, before the feature extraction of the collected item text information by using the pre-trained text information encoder and antecedent predictor to obtain the coded vector representation of the item text information, the Methods also include:

For each association rule, predetermining whether the antecedent and the consequent appear in item text information is used as tag data;

Input the item text information carrying the labeled data into the first network model for training, and build a text information encoder whose optimization goal is to maximize the prediction of whether the consequent appears in the item text information in the association rules ;

The coded vector representation of the output item text information of the first network model and the predicted value of whether the consequent appears in the item text information in the association rules are input to the second network model for training, and the antecedent predictor is constructed, and the antecedent The optimization goal of the predictor is to maximize whether the antecedents in the association rules appear in the project text information.
The computer device according to claim 11, wherein the text information encoder and the antecedent predictor perform adversarial learning during the training process, so that information related to the antecedent in the association rule is removed from the item text information, And retain the information related to the former and the latter in the association rules.
The computer device according to any one of claims 9-12, wherein, each association rule is evaluated according to the coded vector representation of the item text information, and the relationship between the antecedent and the consequent in the association rule is obtained The results of the assessment of causality, including:

For each association rule, calculate an evaluation value reflecting the causal relationship between the antecedent and the subsequent in the association rule according to the coded vector representation of the item text information;

If the evaluation value is greater than the preset threshold, it is determined that there is a causal relationship between the antecedent and the consequent in the association rule.
The computer device according to claim 13, wherein the item text information contains a plurality of texts, and for each association rule, according to the coded vector representation of the item text information, the calculation reflects the antecedent and the subsequent item in the association rule Estimates of the causal relationship between events, including:

For each association rule, select the text that appears in the preceding item in the association rule from the item text information as the sample text;

Traverse the coded vector representation of the item text information, query the text that meets the similarity condition with the coded vector representation of each sample text, and use it as the similar target text of each sample text;

For the similar target text of each sample text, the evaluation value reflecting the causal relationship between the antecedent and the consequent in the association rule is calculated.
A readable storage medium, on which computer-readable instructions are stored, wherein, when the computer-readable instructions are executed by a processor, the steps of the method for evaluating association rules based on machine learning are implemented, including:

Using item co-occurrence conditions to mine association rules from the item collection, the association rules include antecedents and subsequent items, and the item co-occurrence conditions are that the items in the antecedent and subsequent items appear simultaneously;

Use the pre-trained text information encoder and antecedent predictor to extract the features of the collected item text information, and obtain the coded vector representation of the item text information, and the text information encoder is used to determine whether the consequent appears in the association rules Forecasting, the antecedent predictor is used to predict whether the antecedent appears in the association rule;

In response to the evaluation instruction of the association rules, each association rule is evaluated according to the coded vector representation of the item text information, and an evaluation result reflecting the causal relationship between the antecedent and the subsequent in the association rule is obtained.
The readable storage medium according to claim 15, wherein the item co-occurrence condition is that the antecedent and the subsequent item in the association rule appear at the same time, and the use of the item co-occurrence condition to mine the association rule from the item set specifically includes:

Perform a full permutation of the frequent item subsets contained in the item set;

Generate candidate association rules for the subset of frequent items, and use preset parameter indicators to filter the candidate association rules to obtain candidate rules that meet preset conditions, the parameter indicators include at least support and confidence, and the The support is the co-occurrence frequency of the antecedent and the consequent, and the confidence is the ratio of the support to the probability of the antecedent.
The readable storage medium according to claim 15, wherein, before performing feature extraction on the collected item text information by using the pre-trained text information encoder and antecedent predictor to obtain the coded vector representation of the item text information, The method also includes:

For each association rule, predetermining whether the antecedent and the consequent appear in item text information is used as tag data;

Input the item text information carrying the labeled data into the first network model for training, and build a text information encoder whose optimization goal is to maximize the prediction of whether the consequent appears in the item text information in the association rules ;

The coded vector representation of the output item text information of the first network model and the predicted value of whether the consequent appears in the item text information in the association rules are input to the second network model for training, and the antecedent predictor is constructed, and the antecedent The optimization goal of the predictor is to maximize whether the antecedents in the association rules appear in the project text information.
The readable storage medium according to claim 17, wherein the text information encoder and the antecedent predictor perform confrontational learning during the training process, so that items related to the antecedent in the association rule are removed from the item text information. information, and retain the information related to the former and the latter in the association rules.
The readable storage medium according to any one of claims 15-18, wherein, each association rule is evaluated according to the encoded vector representation of the item text information, and an antecedent and a consequent in the association rule are obtained The results of the assessment of the causal relationship between, specifically include:

For each association rule, calculate an evaluation value reflecting the causal relationship between the antecedent and the subsequent in the association rule according to the coded vector representation of the item text information;

If the evaluation value is greater than the preset threshold, it is determined that there is a causal relationship between the antecedent and the consequent in the association rule.
The readable storage medium according to claim 19, wherein the item text information contains a plurality of texts, and for each association rule, according to the encoding vector representation of the item text information, the calculation reflects the antecedent in the association rule The estimated value of the causal relationship between the event and the consequent, including:

For each association rule, select the text that appears in the preceding item in the association rule from the item text information as the sample text;

Traverse the coded vector representation of the item text information, query the text that meets the similarity condition with the coded vector representation of each sample text, and use it as the similar target text of each sample text;

For the similar target text of each sample text, an evaluation value reflecting the causal relationship between the antecedent and the consequent in the association rule is calculated.