TWI771284B - Method and device for predicting user problems based on data-driven - Google Patents

Abstract

The present invention discloses a method and device for predicting user problems based on data driving. When receiving a question raised by a user, the method collects user behavior data and preprocesses it, and obtains the preprocessed user behavior data from the user behavior data. Intercept the candidate behavior data that contribute to the questions raised by the user, filter the candidate behavior data through the set target behavior data set, and select the candidate behavior data contained in the target behavior data set from the candidate behavior data. The filtered candidate behavior data is input into the trained classifier model to predict the category of the question raised by the user. The device of the present invention includes a preprocessing module, an intercepting module and a prediction module. The method and device of the present invention can significantly improve the prediction effect.

Description

Method and device for predicting user problems based on data-driven

The present invention relates to the technical field of data processing, and in particular, to a method and device for predicting user problems based on data-driven.

Users often encounter problems that they cannot handle when using products or services, and then seek customer service help. Usually, the customer service personnel need to go through multiple rounds of dialogue with the user to determine what problem the user encounters, which requires a lot of labor costs. If you can predict the problems that users may encounter in advance, you can intelligently push relevant answers or help customer service personnel to locate user problems more effectively.

Predicting the problems that users may encounter in advance is a typical multi-classification problem, which usually consists of two parts: feature selection and model modeling. In the existing methods, when the feature selection end extracts features, some rules are usually set by humans, and these rules are empirically considered to be related to the questions that the user may ask, such as whether the user has opened a certain service, in the past Whether there has been a consumption record within a few days, etc. By matching with these rules, the characteristics describing the user's state before questioning can be obtained. These features are then modeled using logistic regression techniques to obtain classifiers that are used for new features Prediction.

In the prior art, some rules that are empirically considered to be related to the problems the user may encounter are manually set, and the features describing the state of the user before the question are obtained by matching with these rules. There are two problems in this: 1. It is not driven by data, but requires strong human intervention, and requires the intervener to fully understand and be familiar with the corresponding product or business, which will introduce a lot of inconvenience when the product changes frequently or the business coverage expands. Sex is not strong. 2. Failure to consider the relationship between the user's behavior and the user's problems within a short period of time before seeking help from the customer service staff. Usually, the user will have a series of These behaviors include but are not limited to mobile phone, tablet user-side clicks, PC web browsing, and other operations performed by the user, which include the user's behavioral trajectory information before asking questions. In theory, these behavioral trajectories are related to There is a strong correlation between users' subsequent help-seeking.

The object of the present invention is to provide a data-driven method and device for predicting user problems, which is used to predict the user encounters as accurately as possible before the user describes the problem when some user information and operations are known. It can avoid the influence of human intervention in the prior art and improve the accuracy of classification prediction.

In order to achieve the above object, the technical solution of the present invention is as follows: a method for predicting user problems based on data-driven, the method for predicting user problems includes: When receiving questions from users, collect user behavior data and preprocess it; intercept candidate behavior data that contribute to the questions raised by users from the preprocessed user behavior data; The data set is used to filter the behavior data to be selected, the selected behavior data contained in the target behavior data set is screened from the candidate behavior data, and the selected behavior data is input into the trained classifier model to predict the behavior proposed by the user. The category to which the question belongs.

Further, for the trained classifier model, the training process includes the following steps: collecting the questions fed back by users and their corresponding behavior data, and preprocessing the collected user behavior data; The behavior data that contributes to the user's feedback is intercepted from the data as the candidate behavior data; according to the questions fed back by all users and their corresponding candidate behavior data, a data-driven method is used to correspond to each user's feedback problem. The selected behavior data is scored, and the target behavior data that meets the set conditions are screened out, and the target behavior data corresponding to the questions fed back by all users is combined to form the selected target behavior data set; The problem and target behavior data sets are trained to obtain a classifier model.

Further, the preprocessing includes: removing interference behavior data whose frequency is lower than a set frequency threshold.

The preprocessing also includes: Digital identification of user behavior data. Digital identification of behavioral data is for the convenience of processing directly with the digitalized identification in subsequent steps, so that it is not necessary to process according to the specific data of behavioral data, such as URL or API name, and other long-string data, which is simpler to process. .

Further, the present invention intercepts the candidate behavior data that contributes to the problem raised by the user from the preprocessed user behavior data and adopts the method of windowing and truncation, and the windowing and truncation includes: intercepting the most recent section before the problem occurs. User behavior data over time.

Further, after the target behavior data set corresponding to the questions fed back by all users is formed into the selected target behavior data set, the method further includes: re-digitizing the target behavior data in the target behavior data set.

Further, before the classifier model is obtained from the training, the method further includes the step of: performing vectorization processing on the target behavior data in the target behavior data set.

Further, before inputting the selected behavior data to be selected into the trained classifier model, the method further includes: performing vectorization processing on the behavior data to be selected. The vectorized user behavior data can directly train the classifier model and be used for actual prediction, and the calculation is simpler.

The present invention also proposes a data-driven prediction user problem based The device for predicting user problems includes: a preprocessing module, used for collecting user behavior data and preprocessing when a question raised by a user is received; The candidate behavior data that contributes to the questions raised by the user is intercepted from the user behavior data; the prediction module is used to filter the candidate behavior data through the set target behavior data set, and select the target behavior from the candidate behavior data. The candidate behavior data contained in the data set is input into the trained classifier model to predict the category of the question raised by the user.

Further, the device also includes a model training module for training the classifier model, and the model training model performs the following operations when training the classifier model: collecting the questions and corresponding behavior data fed back by the user, The collected user behavior data is preprocessed; the behavior data that contributes to the user feedback is intercepted from the preprocessed user behavior data as the candidate behavior data; Select behavior data, use a data-driven method to score the candidate behavior data corresponding to each user feedback question, and filter out the target behavior data that meets the set conditions, and select the target behavior data corresponding to all user feedback questions. The union constitutes a set of filtered target behavior data; according to the set of questions and target behavior data fed back by each user, a classifier model is obtained by training.

When preprocessing the collected user behavior data, the preprocessing module of the present invention performs the following steps: removing the interference behavior data whose frequency is lower than the set frequency threshold.

The preprocessing module is also used to digitally identify the user behavior data.

Further, when the interception module intercepts the candidate behavior data that contributes to the problem raised by the user from the preprocessed user behavior data, a method of windowing and truncation is adopted, and the windowing and truncation includes: intercepting User behavior data for the most recent period prior to the issue.

Further, the model training module is also used to re-digitize the target behavior data in the target behavior data set after taking the union set of the target behavior data corresponding to the questions fed back by all users to form the selected target behavior data set. identification.

Further, the model training module is further configured to perform vectorization processing on the target behavior data in the target behavior data set before training to obtain the classifier model.

Further, before inputting the selected behavior data to be selected into the trained classifier model, the prediction module is also used to perform vectorization processing on the behavior data to be selected.

A method and device for predicting user problems based on data-driven proposed by the present invention utilizes short-term behavioral trajectory information of users to classify and predict user problems to improve the classification accuracy and significantly improve the prediction effect of models that do not contain such information. .

F1: Process

F2: Process

F3: Process

F4: Process

S1: Step

S2: Step

S3: Step

FIG. 1 is a flow chart of training a classifier model according to the present invention; FIG. 2 is a flow chart of a method for predicting user problems based on data driving according to the present invention; FIG.

The technical solutions of the present invention will be further described in detail below with reference to the drawings and examples, and the following examples do not constitute a limitation of the present invention.

The general idea of the present invention is to use the training data to train a classifier model, and to analyze the user behavior data according to the trained classifier model to predict the problems encountered by the user.

As shown in Figure 1, the process of using the training data to train the classifier model in this embodiment is as follows:

F1. Collect user feedback problems and corresponding behavior data, and preprocess the collected user behavior data. The preprocessing includes removing interfering behavior data, and digitally marking the behavior data.

For any user feedback problem, the user's behavior data are collected, so as to obtain a large amount of behavior data. Behavioral data are some user operations, including mobile phone, tablet user-side clicks, PC web browsing, and other operations performed by the user. These operations are represented by URLs or API names, preceded by unix timestamps. For example a user X in the past A period of behavior can be represented as: 1438661879:alipay.mappprod.shop.queryPage

1438661885: alipay.client.mobileapp.checkResult

1438661889: alipay.commerce.category.queryByCategoryId

1438661899: alipay.siteprobe.sync.queryWifis

1438661909: alipay.charity.mobile.donate.deduct.unsign

.....

.....

1438661999: https://couriercore.alipay.com/errorRepeatSubmit.htm

1438662999: https://cshall.alipay.com/lab/question.htm

In order to be more accurate and facilitate subsequent processing, the preprocessing in this embodiment includes removing interfering behavior data, and digitally identifying the behavior data.

The removal of interfering behavior data refers to behavior data with extremely low frequency, such as lower than a set frequency threshold. These behavioral data with extremely low frequency may cause a relatively low possibility of user feedback, which is not considered in this embodiment, so as to eliminate the interference caused by behavioral data with extremely low frequency.

The digital identification of the behavioral data is for the convenience of directly processing the digitalized identification in the subsequent steps, so that it is not necessary to process according to the specific data of the behavioral data, such as URL or API name and other long-string data, which makes it easier to process. Simple.

To digitally identify the behavioral data, the website or API of the above behavioral data can be digitally identified according to the mapping table prepared in advance; The number serves as a digitized identifier for the behavioral data; Or according to the specific content of the behavior data, the corresponding digitized identification can be obtained through HASH calculation. The behavior data after digitizing the logo becomes: 1438661879:2

1438661885:65

1438661889:11

1438661899:6

1438661909:18

.....

.....

1438661999:108

1438662999:111

The digitized identification is directly used for screening and processing in subsequent steps.

F2. Intercept from the preprocessed user behavior data, the candidate behavior data that contributes to the user's feedback.

For a large amount of user behavior data, what really affects the problems reported by users is often the behavior data of users in the recent period before the problem occurs. That is, the behavior data that contributes to the user's feedback is the user's recent behavior data, and the influence of historical behavior data can be ignored. Therefore, in this embodiment, the user behavior data needs to be intercepted, and the behavior data of the user in the most recent time is selected as the candidate behavior data.

Specifically, the interception is performed by adding a window, and a fixed window length or a variable window length can be selected. For example, the fixed window length is 30-120 behavior data, that is, 30-120 behavior data are selected from the current behavior data forward; the variable window length is from the current behavior data. The previous behavior data is selected from the behavior data of a certain period of time, for example, the behavior data within 0.5 hours to 2 hours before the current time.

For example, for the above behavior data, when windowing is truncated from the last behavior data, that is, 1438662999:111, backtracking, the length is fixed window length (30-120 data) or variable window length (0.5 hours-2 hours, through unix time stamp ok). Suppose that the data becomes: 1438661885:65 after windowing and truncation

1438661889:11

1438661899:6

1438661909:18

.....

.....

1438661999:108

1438662999:111

Thereby, the candidate behavior data that contributes to the question fed back by the user is obtained, the behavior data corresponding to each user feedback question is traversed, and the candidate behavior data corresponding to each user feedback question is obtained.

F3. According to the questions fed back by all users and the corresponding candidate behavior data, use a data-driven method to score the candidate behavior data corresponding to each user feedback question, and filter out the target behavior data that meets the set conditions. , the target behavior data set corresponding to the questions fed back by all users is taken to form the selected target behavior data set.

In this embodiment, the questions fed back by all users are regarded as a file set, and the questions fed back by each user are regarded as a document. data-driven The method is the TF-IDF method. The TF-IDF method is a common weighted statistical method used for information retrieval and data acquisition to evaluate the importance of a word to a document set or a document in a corpus. The importance of a word increases proportionally to the number of times it appears in the document, but decreases inversely to the frequency it appears in the corpus. For this embodiment, the words are equivalent to the behavior data to be selected, all the questions fed back by the users are regarded as a file set, the questions fed back by each user are regarded as a document, and the questions fed back by each user are corresponding to the questions fed back by each user through TF-IDF. The candidate behavior data are scored.

The main idea of TF-IDF is: if a word or phrase appears frequently in one article TF, and rarely appears in other articles, it is considered that this word or phrase has a good ability to distinguish between categories, suitable for classification. TFIDF is actually: TFXIDF, TF Term Frequency, IDF Inverse Document Frequency. TF stands for the term, the frequency of occurrence in document d. The main idea of IDF is: if there are fewer documents containing term t, that is, the smaller n is, and the larger the IDF is, it means that term t has a good ability to distinguish between categories. If a class C. The number of files containing entry t is m, and the total number of files containing t in other classes is k. Obviously, the number of files containing t is n=m+k. When m is large, n is also large, according to the IDF formula. The value of IDF will be small, which means that the classification ability of the entry t is not strong.

The detailed algorithm is as follows:

The first step is to calculate the word frequency.

The second step is to calculate the reverse file frequency.

The third step is to calculate TF-IDF.

TF-IDF = term frequency (TF) × inverse document frequency (IDF)

User behavior data can be regarded as words to some extent, and the importance of words to one of the documents in a document set or a corpus. The TF-IDF technology is borrowed and applied to the screening of behavior data. Screen out the behavior data suitable for classification, and call these screened out behavior data as the target behavior data. In this embodiment, the top N (50-200) with the highest score or higher than the questions fed back by each user are selected. The behavior data of a certain threshold is used as the target behavior data. A set of target behavior data corresponding to the questions fed back by all users is obtained to form a set of filtered target behavior data, and the amount of behavior data contained in the set is much smaller than that in all training data.

For example, the behavior data corresponding to question A is (represented by a digitized identifier): 6 18 1 9 77 98 69 ...................... .................................88 189 87

Perform TF-IDF scores on all digitized identifiers, and take the most important top N (such as 50) for problem A from high to low, and you can get the most important behavior data for problem A: A: 11 18.... ......108......

The union of the digital identification sets corresponding to all the questions constitutes the target behavior data set. It can be seen that when the above-mentioned user feedback questions are identified In the case of known problems, the above target behavior data collection contains target behavior data for all known problems.

Further, the target behavior data in the target behavior data set is re-digitized and identified, which makes the set simpler and facilitates subsequent processing.

F4. According to the set of questions and target behavior data fed back by each user, a classifier model is obtained by training.

A classifier model is trained by using known problems and their corresponding target behavior data. Through the classifier model, when a user feedbacks a problem, the user's behavior data can be analyzed to predict the possibility of the problem. Which is a known problem, so that the customer service can answer the user's question and provide a solution.

Classifier models include, but are not limited to, logistic regression models, deep neural network models, support vector machine models, recurrent neural network models, and the like. In view of the fact that there are many methods for obtaining a model by training according to the training data in the prior art, details are not repeated here.

As shown in FIG. 2 , a method for predicting user problems based on data driving in this embodiment includes:

Step S1, when a question raised by a user is received, user behavior data is collected and preprocessed.

Step S2, intercepting behavior data that contributes to the question raised by the user from the preprocessed user behavior data as candidate behavior data.

After receiving the question raised by the user, the customer service can capture the user behavior data for preprocessing. The specific methods of preprocessing and how to proceed. The row windowing truncation has been described in the training of the classifier model above, and will not be repeated here.

Step S3, screening the to-be-selected behavior data through the set target behavior data set, screen out the target behavior data set from the candidate behavior data, and input the selected behavior data into the trained classifier A model that predicts the category of questions asked by users.

For example, the behavior data of user X becomes: 1438661889:11 after filtering through the target behavior data set

1438661909:18

.....

.....

1438661999:108

Assuming that 1438661885:65, 1438661899:6, and 1438662999:111 are not included in the target behavior data set, these three pieces of data will be removed because they are not included in the target behavior data set.

Because the target behavior data set has been obtained through screening, and the classifier model has been trained. Therefore, when a user submits a question to the customer service, the customer service can submit the user's candidate behavior data to the trained classifier model, and the classifier model calculates the specific type of problem raised by the user, and outputs the corresponding According to the probability of different questions, the question with the highest probability is selected as the category of the question raised by the user.

Further, in order to facilitate the training of the classifier model, as well as the subsequent pre- In this embodiment, a data-driven method for predicting user problems further performs vectorization processing on the target behavior data in the target behavior data set, and vectorization processing on the behavior data to be selected. vectorize

The vectorization process is divided into binarization and quantization. Binarization refers to setting 1 at the corresponding vector position when it occurs, and setting 0 when it does not. Quantization refers to the number of times the behavior occurs at the corresponding vector position. The vectorized user behavior data can directly train the classifier model and be used for actual prediction, or it can be combined with the original features to train the classifier model and used for actual prediction.

As shown in FIG. 3 , corresponding to the above method, the present invention also proposes a device for predicting user problems based on data driving. The device includes: a preprocessing module, which is used for collecting data when receiving a problem raised by a user. User behavior data and preprocessing; interception module, used to intercept candidate behavior data that contributes to the questions raised by users from the preprocessed user behavior data; prediction module, used to pass the set goal The behavior data set filters the behavior data to be selected, and selects the behavior data contained in the target behavior data set from the behavior data to be selected, and inputs the selected behavior data into the trained classifier model, and predicts that the user proposes The category the question belongs to.

The device for predicting user problems in this embodiment further includes a model training module for training the classifier model. When training the classifier model, the model training model performs the following operations: collecting user feedback problems and their corresponding behavior data, pair collection preprocess the user behavior data; intercept the behavior data that contributes to the user feedback from the preprocessed user behavior data as the candidate behavior data; based on all the user feedback questions and their corresponding candidates Behavior data, using a data-driven method to score the candidate behavior data corresponding to each user feedback question, and filter out the target behavior data that meets the set conditions, and connect the target behavior data corresponding to all user feedback questions. The set constitutes a set of selected target behavior data; according to the set of questions and target behavior data fed back by each user, a classifier model is obtained by training.

When the preprocessing module in this embodiment preprocesses the collected user behavior data, the following steps are performed: removing the interference behavior data whose frequency is lower than the set frequency threshold.

Further, the preprocessing module is also used to digitally identify the user behavior data.

In this embodiment, when the interception module intercepts the candidate behavior data that contributes to the question raised by the user from the preprocessed user behavior data, a method of windowing and truncation is adopted, and the windowing and truncation includes: User behavior data for the most recent period prior to the issue.

Corresponding to the above method, the model training module of this embodiment is also used to re-analyze the target behavior data set in the target behavior data set after forming the selected target behavior data set from the target behavior data set corresponding to the questions fed back by all users. Behavioral data is digitally identified.

The model training module of this embodiment is further configured to perform vectorization processing on the target behavior data in the target behavior data set before training to obtain the classifier model.

Corresponding to the above method, the prediction module of this embodiment is also used to perform vectorization processing on the to-be-selected behavior data before inputting the screened out to-be-selected behavior data into the trained classifier model.

The above embodiments are only used to illustrate the technical solutions of the present invention but not to limit them. Those with ordinary knowledge in the technical field can make various corresponding changes according to the present invention without departing from the spirit and essence of the present invention. and deformation, but these corresponding changes and deformations should all belong to the protection scope of the appended patent application scope of the present invention.

Claims (12)

A method for predicting user problems based on data-driven, the method for predicting user problems includes: collecting user behavior data and performing preprocessing when receiving a question raised by a user, wherein the preprocessing includes: removing Interfering behavior data whose frequency is lower than the set frequency threshold; adopting the method of windowing and truncation to intercept the candidate behavior data that contributes to the problem raised by the user from the preprocessed user behavior data, and the windowing and truncation includes: Intercept the user behavior data in the most recent period before the problem occurs; filter the to-be-selected behavior data through the set target behavior data set, and select the candidate behavior data contained in the target behavior data set from the candidate behavior data, and then filter The selected behavior data is input into the classifier model, and the category of the question raised by the user is predicted. According to the data-driven method for predicting user problems according to item 1 of the scope of the application, wherein, the training process of the classifier model includes the following steps: collecting user feedback problems and their corresponding behavior data; The preprocessing is performed on the user behavior data; the behavior data that contributes to the user's feedback is intercepted from the preprocessed user behavior data as the candidate behavior data; according to all the user feedback questions and their corresponding Select behavior data, use a data-driven method to score the candidate behavior data corresponding to each user feedback question, and filter out the goals that meet the set conditions Behavior data, the target behavior data set corresponding to the questions fed back by all users is combined to form the selected target behavior data set; the classifier model is obtained by training according to the set of questions and target behavior data fed back by each user. According to the data-driven method for predicting user problems according to item 1 of the scope of the patent application, the preprocessing further comprises: digitally identifying the user behavior data. According to the data-driven method for predicting user problems according to item 3 of the scope of the patent application, after the target behavior data set corresponding to the questions fed back by all users is formed into the selected target behavior data set, the Including: digitally marking the target behavior data in the target behavior data collection again. According to the data-driven method for predicting user problems according to item 2 of the scope of the patent application, before the classifier model is obtained from the training, the method further includes the step of: vectorizing the target behavior data in the target behavior data set. According to the data-driven method for predicting user problems according to item 5 of the scope of the patent application, before inputting the screened candidate behavior data into the classifier model, the method further includes: vectorizing the to-be-selected behavior data. A data-driven device for predicting user problems, the device for predicting user problems includes: The preprocessing module is used to collect and preprocess the user behavior data when receiving the questions raised by the user, wherein the preprocessing module performs the following steps when preprocessing the collected user behavior data Steps: remove the interference behavior data whose frequency is lower than the set frequency threshold; the interception module is used for intercepting the candidate behaviors that contribute to the problem raised by the user from the preprocessed user behavior data by adopting the method of windowing and truncation The windowing and truncation includes: intercepting the user behavior data in the most recent period before the problem occurs; the prediction module is used to filter the to-be-selected behavior data through the set target behavior data set, and from the to-be-selected behavior data The candidate behavior data included in the target behavior data set is filtered out, and the filtered candidate behavior data is input into the classifier model to predict the category of the question raised by the user. The device for predicting user problems based on data-driven claims according to claim 7, wherein the device further includes a model training module for training a classifier model, and the model training model is used when training the classifier model. , and perform the following operations: collect the questions fed back by the users and their corresponding behavior data, and perform the preprocessing on the collected user behavior data; The behavior data of each user is used as the candidate behavior data; according to the questions fed back by all users and the corresponding candidate behavior data, the data-driven method is used to score the candidate behavior data corresponding to each user feedback question, and screen out The target behavior data that meets the set conditions, and the target behavior data corresponding to the questions fed back by all users The union set constitutes a set of filtered target behavior data; a classifier model is obtained by training according to the set of questions and target behavior data fed back by each user. According to the device for predicting user problems based on data-driven claimed in claim 7, wherein the preprocessing module is further used to digitally identify the user behavior data. According to the data-driven device for predicting user problems according to item 9 of the scope of the patent application, wherein the model training module takes a union set of target behavior data corresponding to the questions fed back by all users to form the filtered target behavior data After the collection, it is also used to digitally identify the target behavior data in the target behavior data collection. According to the device for predicting user problems based on data driven according to item 8 of the scope of the patent application, the model training module is further configured to perform a training operation on the target behavior data in the target behavior data set before training to obtain the classifier model. vectorized processing. The device for predicting user problems based on data-driven claims according to claim 11, wherein the prediction module is further configured to perform a data-to-be-selected behavior analysis before inputting the selected behavior data into the classifier model. vectorized processing.

Images