CN106446191A - Logistic regression based multi-feature network popular tag prediction method - Google Patents

Abstract

A multi-feature network popular label prediction method based on Logistic regression, comprising the following steps: 1) constructing a network of authorized and undirected network label Tags according to the posting data of a question-and-answer website; 2) extracting popular label sets, non-popular Label collection; 3) extracting the network features of the label, the attribute characteristics of the label proposer, and the attribute change characteristics after the label is proposed; 4) using Logistics multiple regression training and building a label classification model; the present invention considers the correlation between labels, Classifying labels based on multiple features has high accuracy for predicting potential popular labels. It is not only beneficial to guide users to choose reasonable labels, but also beneficial for website builders to provide higher quality labels.

Description

A Multi-feature Network Popular Label Prediction Method Based on Logistic Regression

technical field

The invention relates to the fields of data mining and computer technology, in particular to a Logistic regression-based multi-feature network popular label prediction method.

Background technique

Network tags (Tag) is an organizational form of Internet information content, usually composed of some keywords closely related to content, it can help people describe and classify content conveniently, and also facilitate information retrieval and sharing. Due to the convenience of network tags, tag prediction and tag recommendation have been widely used in many network platforms in recent years, such as StackExchange, a question-and-answer website, Flickr, a photo-sharing website, and Yelp, a restaurant review website. Using the right tags is very important both for the website and for the users. For websites, appropriate tags can help websites make personalized recommendations to users, increasing user stickiness and website click-through rate; for users, tags can help users quickly locate what they need and avoid wasting time browsing useless information. In tag selection, how to select potentially popular tags is a critical step, because popular tags often represent the needs of most users.

At present, the main basis for selecting information tags is the degree of correlation between the information and the text of the tag and the attributes of the information originator. However, there are various disadvantages in this selection, mainly in: 1. Ignoring the potential trend of tags; 2. Ignoring the correlation between tags; 3. Unpopular content leads to unpopular tags, making information not effective Searched; 4. Only a few features are considered, which makes the selection of some labels tend to be one-sided.

Therefore, in order to enable users to better select tags when publishing information content, potential popular tags should be selected as much as possible. The multi-feature network popular label prediction method based on Logistic regression in the present invention solves the following two basic problems: (1) predicts the future popular trend of the label; (2) applies a large number of features to quantitatively describe the popular trend of the label.

Contents of the invention

In order to overcome the shortcomings of the existing label selection system that ignores the potential trend of labels and the correlation between labels, and has a single evaluation feature, the present invention provides a multi-feature network popular label prediction method based on Logistic regression, which not only considers multiple The related features between features and tags can also better predict the popular trend of tags.

The technical solution adopted by the present invention to solve its technical problems is as follows:

A multi-feature network popular label prediction method based on Logistic regression, comprising the steps of:

S1: Data preprocessing: collect the information content and label data of the website, and arrange the information content of the website in ascending order of time, regard the posts with a proportion of the first α% as the transient data before the label network is stable, and delete this part of the transient state Data; select the data of the preset ratio from the remaining data of the website as the training data;

S2: Construct a Tag Network. For Tags that appear in the same information content, make them form an edge between them. Traverse all the information to get the tag network graph G _Tag with the right to undirected network. The weight of the network is two the number of co-occurrences;

S3: Each tag is arranged in descending order according to its frequency of appearance in the post, and the Tag with the first β% ratio is taken as the popular tag set U _PopularTag ;

S4: Find the unpopular tag set U _UnPopularTag , for each popular tag t ∈ U _PopularTag , search for the time when the tag t first appeared, and use this time as the center to search for the closest to the time, the first time , and tags that do not belong to U _PopularTag are used as unpopular tags to form a comparative non-popular tag set U _UnPopularTag ;

S5: For the training sample label set U={U _PopularTag , U _UnPopularTag }, extract the network features of the tags in it, and extract the degree value of the neighbor node connected to the sample label for the first time on the weighted undirected network G _Tag , Neighbor node degree centrality;

S6: For the training sample label set U={U _PopularTag , U _UnPopularTag }, extract the attribute characteristics of the proposer of the Tag in it, specifically including the number of information content published when the Tag proposer proposes the Tag, and the length of the information content ;

S7: For the training sample label set U={U _PopularTag , U _UnPopularTag }, extract the attribute change characteristics of the Tag in it, specifically including the number of replies received by the post corresponding to the Tag within 5 days after the Tag is put forward;

S8: Using Logistic multiple regression, using the features of the tags in the set U={U _PopularTag , U _UnPopularTag } as training data, train and build a tag classifier model.

Further, in the step S1, α% is determined in such a way that when a preset percentage of the total number of Tags on the website appears, it is used as an intercept point of α%. Its purpose is to ensure that the label network is not affected by the staff's debugging of the website label at the beginning of the website's establishment;

Further, in the step S5, formula (1) is used to calculate the node degree value of neighbor i

Among them, g represents the total number of nodes in the network; if nodes i and j have connected edges, then x _ij =1, otherwise x _ij =0;

Use formula (2) to calculate the node degree centrality of neighbor i

The beneficial effects of the present invention are: considering the correlation between tags, classifying tags according to multiple features, and having higher precision for predicting potential popular tags. It is not only beneficial to guide users to choose reasonable labels, but also beneficial for website builders to provide higher quality labels.

Description of drawings

FIG. 1 is a flow chart of a multi-feature label classification method based on Logistic regression according to an embodiment of the present invention.

Fig. 2 is a schematic diagram of tag occurrence frequency according to an embodiment of the present invention.

detailed description

The specific implementation manners of the present invention will be further described in detail below in conjunction with the accompanying drawings.

Referring to Fig. 1 and Fig. 2, a multi-feature network popular label prediction method based on Logistic regression, the present invention uses the data officially released by the question-and-answer website StackExchange sub-site Tex.Stackexchange.com to carry out the modeling analysis of the label classification system, and the original data records Information such as the time when each post appeared, the ID of the poster, and the label of the post. Taking the Tag of this patent research as an example, we extract the time when the tag first appeared, ID of the tag creator, ID of neighbor tags and other information.

In the present embodiment, a kind of multi-feature label classification method based on Logistic regression, its specific steps are:

1) Build a tag Tag network: For the published post data, do the following processing:

1.1) Traverse the post data to get all the Tag label sets T _I , I∈N, where N represents the total number of tags. Take N×20% of the tags as the number of tags required for the stable point of the website tags. The beneficial way is to prevent the site from being built, and the staff’s debugging of the website content will bring noise to the model;

1.2) Arrange the posts in ascending order of time, and traverse the post data again. When the number of different tags is N×20%, record the number of posts traversed at this time as N _{InstablePosts} , and regard the post publishing time at this time as the website label stabilization time;

1.3) OK Where N _Posts is the total number of published posts;

1.4) Build a Tag network: remove the first α% posts, and read the first 80% of the posts in the Q&A website data as training data. Among them, the construction method of the Tag network is as follows: for the Tags appearing in the same post, make them form a connection between them. For all information traversal, the label network graph G _Tag of the authorized undirected network is obtained, and the weight of the network is the number of times the two appear together;

2) Obtain the popular tag set U _PopularTag : For the published post data, do the following processing:

2.1) Traverse the post data to obtain the frequency of each Tag appearing in the post;

2.2) According to the descending order of Tag appearance frequency, take the Tag with the first β% ratio as the popular tag set U _PopularTag , here, we choose β%=5%;

3) Obtain the unpopular tag set U _UnPopularTag , the specific steps are:

3.1) For each label Tag, traverse the post to get the first appearance time of each label;

3.2) For each popular tag t ∈ U _PopularTag , search for all other tags (other tags do not exist in popular tags) The time difference with the label, that is, the time difference ΔT between the rest and the first appearance of the label;

3.3) Arrange the time difference ΔT in ascending order, and take the tag t' with the smallest ΔT as the unpopular tag, thus forming the unpopular tag set U _UnPopularTag

4) Extract the network features of Tag, the specific steps are:

4.1) For each tag t ∈ {U _PopularTag , U _UnPopularTag }, use formula (1) to calculate the node degree value of neighbor i

Among them, g represents the total number of nodes in the network; if nodes i and j have connected edges, then x _ij =1, otherwise x _ij =0;

4.2) Use formula (2) to calculate the node degree centrality of neighbor i

4.3) Normalized neighbor node degree, neighbor node degree centrality, normalized denominator is neighbor node value

5) Extract the attribute characteristics of the sample Tag proposer, the specific steps are:

5.1) For each sample tag t ∈ {U _PopularTag , U _UnPopularTag }, obtain the ID number of the proposer and the time when the tag first appeared when the tag was proposed for the first time;

5.2) Arrange the posts in ascending chronological order, and find out the total number of questions and answers of the presenter ID before the time when the tag first appeared, as the attribute characteristics of the presenter of the Tag;

6) extracting the attribute change feature of the sample Tag, the specific steps are: for the sample tag set U={U _PopularTag , U _UnPopularTag } for training, after the Tag is proposed, the number of answers received by the Tag within 5 days;

7) Logistic multiple regression training classification model: the above sample label set U={U _PopularTag , U _UnPopularTag }, and the corresponding Tag's neighbor node degree value, neighbor node centrality, the number of questions asked by the Tag proposer, and the answer of the Tag proposer The five features of the quantity and the number of answers received within a certain period of time after the Tag is proposed are used as input, and the Logistic multiple regression is used as the classifier to train and build a tag classifier model;

As mentioned above, the label classification embodiment of the present invention in the question-and-answer website StackExchange sub-site Tex.Stackexchange.com is introduced, and the correlation between labels is included in the feature by constructing a network; Features, label time evolution features, etc. increase the feature data of label classification. Through the training model, the judgment of whether the tag is popular can be finally obtained, which provides guidance for the construction of the tag recommendation system of the website.

Claims (3)

1. a kind of multiple features network flow row label Forecasting Methodology based on Logistic recurrence is it is characterised in that methods described bag Include following steps：

S1：Data prediction：Collect the information content and the label data of website, and temporally ascending order is arranged by site information content Row, the model that ratio is front α % is considered as the Temporal Data before label network is stablized, and deletes this part of Temporal Data；From Before choosing in the remaining data in website, the data of preset ratio is as training data；

S2：Build label Tag network, to the Tag occurring in the same information content so as to formation between any two connects side；To institute There is information to travel through, obtain the label network figure G of Undirected networks that has the right_Tag, the weight of network is both common number of times occurring；

S3：The frequency descending that each label occurs in model according to it, the Tag taking front β % ratio is as popular label Set U_PopularTag；

S4：Find non-popular tag set U_UnPopularTag, to each popular label t ∈ U_PopularTag, search for label t first The time of secondary appearance, and centered on this time, search nearest from this time, occur for the first time, be not belonging to simultaneously U_PopularTagLabel as non-streaming row label, the non-streaming row label set U of composition comparison_UnPopularTag；

S5：Sample label set U={ U to training_PopularTag,U_UnPopularTag, extract the network characterization of Tag in it, having Power Undirected networks G_TagOn, extract neighbor node angle value, the neighbor node degree centrality that sample label occurs connecting for the first time；

S6：Sample label set U={ U to training_PopularTag,U_UnPopularTag, the presenter's attribute extracting Tag in it is special Levy, specifically include the quantity that Tag presenter proposes the information content to issue during this Tag, the length of the information content；

S7：Sample label set U={ U to training_PopularTag,U_UnPopularTag, extract the attribute variation feature of Tag in it, After specifically including this Tag proposition, the answer quantity that in 5 days, the corresponding model of this Tag receives；

S8：Using Logistic multiple regression, with set U={ U_PopularTag,U_UnPopularTagIn label feature as training Data, trains and builds label classifier model.

2. a kind of multiple features network flow row label Forecasting Methodology based on Logistic recurrence as claimed in claim 1, it is special Levy and be：In described step S1, the determination mode of α % is, when the preset percentage of website whole Tag number of labels Wait, as the intercept point of α %.

3. a kind of multiple features network flow row label Forecasting Methodology based on Logistic recurrence as claimed in claim 1 or 2, its It is characterised by：The node angle value of neighbours i in described step S5, is calculated using formula (1)

$k_i=\underset{j=1}{\overset{g}{\Σ}}{x}_{ij},(i\≠j)---\left(1\right)$

Wherein, g represents the node total number of network；If node i and j have even side, x_ij=1, otherwise x_ij=0；

Calculate the node degree centrality of neighbours i using formula (2)

$C_D\left(i\right)=\frac{k_i}{g-1}---\left(2\right).$