CN104462215A - Scientific and technical literature quoting number predicting method based on time sequence - Google Patents

Abstract

The invention relates to a scientific and technical literature quoting number predicting method based on a time sequence. The predicting method includes the steps that statistics is carried out on scientific and technical literature quoting numbers, and then average literature quoting numbers of all the months are calculated; in combination with the average literature quoting numbers of all the months, the quoting numbers of the corresponding months are processed in a normalization mode to obtain the quoting time sequence; cluster analysis is carried out according to the time sequence, and a quoting number model with the optimal predicting performance is obtained through dividing training sets and verifying sets, building a regression model and performing error analysis; according to similarity analysis of literature to be predicted and time sequences of various kinds of literature, the class with the highest similarity is obtained, and the quoting number, in the next month, of the literature to be predicted is obtained through the model with the optimal predicting performance. The quoting conditions of each piece of published literature can be automatically analyzed, the average literature quoting numbers of all the months are obtained, different quoting modes of the literature are excavated through clustering, and then the future quoting number is predicted according to the existing time sequence of the literature to be predicted.

Description

A kind of to be cited number Forecasting Methodology based on seasonal effect in time series scientific and technical literature

Technical field

The invention belongs to field of computer technology, relate to and a kind of to be cited number Forecasting Methodology based on seasonal effect in time series scientific and technical literature.

Background technology

The number that is cited refers to that in time period of specifying at one period, scientific and technical literature is by the number of times of other reference citations, is the important method of assessment scientific and technical literature influence power and quality.But quantity to be quoted object statistics is vulnerable to the restriction of current point in time, be difficult to obtain the situation that is cited in future time section, and then impact is to the assessment of scientific and technical literature in scientific and technological contribution.Urgently provide a kind of to be cited number Forecasting Methodology based on seasonal effect in time series scientific and technical literature, identify potential document faster, the propagation of advance science research and new knowledge.

Summary of the invention

The object of the present invention is to provide and a kind of to be cited number Forecasting Methodology based on seasonal effect in time series scientific and technical literature, by obtaining and analyzing the time series that is cited of scientific and technical literature, the number that is cited in prediction following a period of time, help the scientific and technical potential of assessment document, the suggestion of reading is rapidly and efficiently provided.

Realize the object of the invention technical scheme:

Step 1: collect each Literature publication days and index, adds up the number be cited by each document each moon after publication.

Step 2: in units of month, calculate the sum that monthly all documents that will analyze are cited and the document sum be cited, be divided by obtain the number avecitecount (month) that is on average cited in this month;

Step 3: to each document, from publication this month, calculated the difference of be after this cited number and avecitecount (month) monthly, obtained the time series that is cited of the document;

Step 4: according to the time series similarity that is cited to literature collection cluster, set up multiple regression model to the time series in every class, utilizes error analysis to select best performance model;

Step 5: utilize vector similarity to calculate document to be measured and all kinds of document seasonal effect in time series similarity, calculates the document to be measured number that is cited of following month with the regression model of the highest class of similarity.

In step 1, utilize the index of each document of database retrieval, according to the label of document each in database and publish days, the concrete time that statistical literature is cited and number of times, the number that is cited in each month after obtaining each Literature publication.

In step 4, first screen the document participating in cluster according to the time series that is cited, the foundation of screening is seasonal effect in time series length.To the time series of length more than N, to divide surplus length portion and block.To the time series of length lower than N, give up.N value is set by the user.

In step 4, when carrying out cluster analysis, first calculate the seasonal effect in time series distance that is respectively cited, distance calculates and adopts Euclidean distance, then uses unweighted average Furthest Neighbor to generate clustering tree.

Be cited time series X _i=(X _i1, X _i2... X _i8): the time series vector value that is cited representing document i;

Be cited time series X _j=(X _j1, X _j2... X _j8): the time series vector value that is cited representing document j;

Distance d (X _i, X _j): the seasonal effect in time series Euclidean distance that is cited representing document i and j;

Distance computing formula is as follows:

$d(X_i,X_j)={\left[\underset{k=1}{\overset{8}{\mathrm{\Σ}}}{(X_{\mathrm{ik}}-X_{\mathrm{jk}})}^2\right]}^{1/2}$

By calculating the distance be cited between time series, obtain a distance matrix.According to Spectral Clustering, unweighted average Furthest Neighbor is used to generate clustering tree.

Between class distance D _pq: represent classification G _p, G _qbetween distance.Wherein G _pelement number be n _p, G _qelement number be n _q.

Element spacing d _ij: represent the distance between time series i, j.

Between class distance computing formula is as follows:

$D_{\mathrm{pq}}=\frac{1}{n_p{n}_q}\underset{i\∈G_p}{\mathrm{\Σ}}\underset{j\∈G_q}{\mathrm{\Σ}}{d}_{\mathrm{ij}}$

By cluster analysis, each document in set is divided into different classes.

In step 4, when building regression model to time series in class, first divide training set and checking collection, choose a time point in time series, using the data before this time point as training set, the later data of this time point are as checking collection.Modling model on training set, assessment models accuracy on checking collection.Finally training set and checking collection data are merged into a data set, and on this data set, operate in the optimum prediction model that training set obtains.

In step 5, for two document p and document p _j, use (X respectively _i1, X _i2... X _i8) and (X _j1, X _j2... X _j8) represent corresponding time series vector value, then time series similarity Similarity (p, p between document _j) computing formula as follows:

$\mathrm{Similarity}(p,p_j)=\cos \mathrm{\θ}=\frac{{\mathrm{\Σ}}_k{X}_{\mathrm{ik}}\×X_{\mathrm{jk}}}{\sqrt{\left({\mathrm{\Σ}}_k{X_{\mathrm{ik}}}^2\right)\left({\mathrm{\Σ}}_k{X_{\mathrm{jk}}}^2\right)}}$

And then can be calculated by time series similarity between document and survey document and all kinds of document seasonal effect in time series similarity.

The formula calculating document to be measured and all kinds of document seasonal effect in time series similarity is as follows:

$\mathrm{Similarity}(p,C_i)=\frac{1}{n}\×\left[\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Similarity}(p,p_j)\right]$

Similarity (p, C _i) represent document p and C to be measured _iclass document seasonal effect in time series Similarity value;

Similarity (p, p _j) represent document p to be measured and document p _jseasonal effect in time series Similarity value, tried to achieve by cosine angle function.Document p _j∈ C _iclass, j=1,2 ..., (n represents C to n _itotal number of class Literature).

The beneficial effect that the present invention has:

The present invention utilizes statistics of database scientific and technical literature publication time and publishes the number that is cited in rear each month; At data preprocessing phase, calculate the number sum that all documents in each month are cited and the document sum be cited, be divided by and obtain the number that is on average cited in this month; For each document, from publication this month, the be cited number of number to this month that be on average cited in conjunction with every month does normalized, obtains the time series that is cited of the document; According to being cited, the seasonal effect in time series degree of correlation carries out cluster analysis to literature collection, in each class, by dividing training set and checking collection, building regression model, carrying out error analysis, obtains the number estimated performance optimization model that is cited; Finally according to document to be measured and all kinds of document seasonal effect in time series similarity analysis, obtain the class that similarity is the highest, calculate with such prediction optimization model, obtain the document to be measured number that is cited of following month.The present invention not only can the situation that is cited after each Literature publication of automatic analysis, obtain the number that is on average cited in each month, also gone out the different reference patterns of document by cluster result, and then go out following number that is cited according to the existing time series forecasting of document to be measured.

The present invention calculates the number that is on average cited in each month in data preprocessing phase and step 2, build each document be cited time series time, use number and the average quantity to be quoted object difference actual value as this month that is cited in corresponding month, can effectively cut down because of the academic liveness difference of seasonality to predicting the error caused, raising predictablity rate.In step 4 by the foundation of be cited Time Series Clustering analysis and regression model, the difference fully can excavating document is cited pattern, after error analysis obtains optimization model, training set and checking collection are merged once again and rerun optimization model, in prediction, fully can be applied to latest data, effectively improve the degree of accuracy of forecast model.

Embodiment:

Step 1: collect each Literature publication days and index, adds up the number be cited by each document each moon after publication.

Utilize the index of each document of database retrieval, according to the label of document each in database and publish days, the concrete time that statistical literature is cited and number of times, the number that is cited in each month after obtaining each Literature publication.

Each document in traversal set, reads the quoted passage label (refid in publication time (time) and index ₁, refid ₂..., refid _n).To each quoted passage label refid _i, statistics quotes refid in every month after publishing _idocument number be the number that is cited in this month.

Step 2: in units of month, calculate the sum that monthly all documents that will analyze are cited and the document sum be cited, be divided by obtain the number avecitecount that is on average cited in this month;

On average be cited number Avecitecount (month): represent the number value that is on average cited within the month month.

Be cited the moon number Citecount (P _i, month) and (P _i∈ N) (N represents the literature collection be cited in the month month): represent document P _iin the number value that is cited of the month month.

The monthly average number computing formula that is cited is as follows:

$\mathrm{Avecitecount}\left(\mathrm{month}\right)=\frac{1}{n}\×\left[\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Citecount}(P_i,\mathrm{mounth})\right]$

The number that is on average cited in corresponding month can be obtained by the monthly average number computing formula that is cited, when building the time series of each document, using number and the average quantity to be quoted object difference actual value as this month that is cited in corresponding month, can effectively cutting down because of the academic liveness difference of seasonality predicting the error caused.

Step 3: to each document, from publication this month, calculated the difference of be after this cited number and avecitecount (month) monthly, obtained the time series that is cited of the document;

Step 4: according to the time series similarity that is cited to literature collection cluster, set up multiple regression model to the time series in every class, utilizes error analysis to select best performance model;

First, according to being cited, time series is screened the document participating in cluster, and the foundation of screening is seasonal effect in time series length.To the time series of length more than N, to divide surplus length portion and block.To the time series of length lower than N, give up.N value is set by the user.N=8 in this experiment.

When carrying out cluster analysis, first calculate the seasonal effect in time series distance that is respectively cited, distance calculates and adopts Euclidean distance, then uses unweighted average Furthest Neighbor to generate clustering tree.

Be cited time series X _i=(X _i1, X _i2... X _i8): the time series vector value that is cited representing document i;

Be cited time series X _j=(X _j1, X _j2... X _j8): the time series vector value that is cited representing document j;

Distance d (X _i, X _j): the seasonal effect in time series Euclidean distance that is cited representing document i and j;

Distance computing formula is as follows:

$d(X_i,X_j)={\left[\underset{k=1}{\overset{8}{\mathrm{\Σ}}}{(X_{\mathrm{ik}}-X_{\mathrm{jk}})}^2\right]}^{1/2}$

By calculating the distance be cited between time series, obtain a distance matrix.According to Spectral Clustering, unweighted average Furthest Neighbor is used to generate clustering tree.

Between class distance D _pq: represent classification G _p, G _qbetween distance.Wherein G _pelement number be n _p, G _qelement number be n _q.

Element spacing d _ij: represent the distance between time series i, j.

Between class distance computing formula is as follows:

$D_{\mathrm{pq}}=\frac{1}{n_p{n}_q}\underset{i\∈G_p}{\mathrm{\Σ}}\underset{j\∈G_q}{\mathrm{\Σ}}{d}_{\mathrm{ij}}$

By cluster analysis, each document in set is divided into different classes.Then multiple regression model is built to the time series in class.Linear trend model is constructed, exponential trend model and polynomial trend mode in this experiment.

If certain moon is cited, number is output variable Y _t, predictive variable be month t (t=1,2,3 ...), then linear trend model is:

Y _t＝β ₀+β ₁×t+ε

Wherein Y _tthe number that is cited in month t, β ₀, β ₁, ε is corresponding seasonal effect in time series level, trend and noise respectively.

Exponential trend model is:

log Y _t＝β ₀+β ₁×t+ε

Quadratic polynomial trend model is:

Y _t＝β ₀+β ₁×t+β ₂×t ²+ε

When building regression model to the time series in class, first dividing training set and checking collection, choosing a time point t in time series, using the data before this time point as training set, the later data of this time point are as checking collection.Modling model on training set, assessment models accuracy on checking collection.Use root-mean-square error RMSE as evaluation index during assessment.

Root-mean-square error computing formula is:

$\mathrm{RMSE}=\sqrt{\frac{1}{v}\underset{t=1}{\overset{v}{\mathrm{\Σ}}}{e_t}^2}$

Wherein, e _tthe actual value of expression time t and the difference of predicted value, v represents the time period number of checking collection.

Finally training set and checking collection data are merged into a data set, and on this data set, operate in the optimum prediction model that training set obtains.

Step 5: utilize vector similarity to calculate document to be measured and all kinds of document seasonal effect in time series similarity, calculates the document to be measured number that is cited of following month with the regression model of the highest class of similarity.

For two document p and document p _j, use (X respectively _i1, X _i2... X _i8) and (X _j1, X _j2... X _j8) represent corresponding time series vector value, then time series similarity Similarity (p, p between document _j) computing formula as follows:

$\mathrm{Similarity}(p,p_j)=\cos \mathrm{\θ}=\frac{{\mathrm{\Σ}}_k{X}_{\mathrm{ik}}\×X_{\mathrm{jk}}}{\sqrt{\left({\mathrm{\Σ}}_k{X_{\mathrm{ik}}}^2\right)\left({\mathrm{\Σ}}_k{X_{\mathrm{jk}}}^2\right)}}$

And then can be calculated by time series similarity between document and survey document and all kinds of document seasonal effect in time series similarity.

The formula calculating document to be measured and all kinds of document seasonal effect in time series similarity is as follows:

$\mathrm{Similarity}(p,C_i)=\frac{1}{n}\×\left[\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Similarity}(p,p_j)\right]$

Similarity (p, C _i) represent document p and C to be measured _iclass document seasonal effect in time series Similarity value;

Similarity (p, p _j) represent document p to be measured and document p _jseasonal effect in time series Similarity value.Document p _j∈ C _iclass, j=1,2 ..., (n represents C to n _itotal number of class Literature).

After filtering out the highest class of similarity, using existing for document to be measured time series as input variable, use such regression model can dope the number that is cited in document future.

Claims (2)

1. to be cited a number Forecasting Methodology based on seasonal effect in time series scientific and technical literature, to it is characterized in that:

Step 1: collect each Literature publication days and index, adds up the number be cited by each document each moon after publication.

Step 2: in units of month, calculate the sum that monthly all documents that will analyze are cited and the document sum be cited, be divided by obtain the number avecitecount (month) that is on average cited in this month;

Step 3: to each document, from publication this month, calculated the difference of be after this cited number and avecitecount (month) monthly, obtained the time series that is cited of the document;

Step 4: time series is screened the document participating in cluster according to being cited, the foundation of screening is that seasonal effect in time series is long; To the time series of length more than N, to divide surplus length portion and block; To the time series of length lower than N, give up; N value is set by the user;

When carrying out cluster, first calculate the seasonal effect in time series distance that is respectively cited, distance calculates and adopts Euclidean distance, then uses unweighted average Furthest Neighbor to generate clustering tree;

Be cited time series X _i=(X _i1, X _i2... X _i8): the time series vector value that is cited representing document i;

Be cited time series X _j=(X _j1, X _j2... X _j8): the time series vector value that is cited representing document j;

Distance d (X _i, X _j): the seasonal effect in time series Euclidean distance that is cited representing document i and j;

Distance computing formula is as follows:

$d(X_i,X_j)={\left[\underset{k=1}{\overset{8}{\mathrm{\Σ}}}{(X_{\mathrm{ik}}-X_{\mathrm{jk}})}^2\right]}^{1/2}$

By calculating the distance be cited between time series, obtain a distance matrix.According to Spectral Clustering, unweighted average Furthest Neighbor is used to generate clustering tree.

Between class distance D _pq: represent classification G _p, G _qbetween distance.Wherein G _pelement number be n _p, G _qelement number be n _q.

Element spacing d _ij: represent the distance between time series i, j.

Between class distance computing formula is as follows:

$D_{\mathrm{pq}}=\frac{1}{n_p{n}_q}\underset{i\∈G_p}{\mathrm{\Σ}}\underset{j\∈G_q}{\mathrm{\Σ}}{d}_{\mathrm{ij}}$

Pass through cluster analysis, each document in set is divided into different classes, when regression model is built to time series in class, first training set and checking collection is divided, choose a time point in time series, using the data before this time point as training set, the later data of this time point are as checking collection; Modling model on training set, assessment models accuracy on checking collection.Finally training set and checking collection data are merged into a data set, and on this data set, operate in the optimum prediction model that training set obtains;

Step 5: utilize vector similarity to calculate document to be measured and all kinds of document seasonal effect in time series similarity, calculates the document to be measured number that is cited of following month with the regression model of the highest class of similarity;

For two document p and document p _j, use (X respectively _i1, X _i2... X _i8) and (X _j1, X _j2... X _j8) represent corresponding time series vector value, then time series similarity Similarity (p, p between document _j) computing formula as follows:

$\mathrm{Similarity}(p,p_j)=\cos \mathrm{\θ}=\frac{{\mathrm{\Σ}}_k{X}_{\mathrm{ik}}\×X_{\mathrm{jk}}}{\sqrt{\left({\mathrm{\Σ}}_k{X_{\mathrm{ik}}}^2\right)\left({\mathrm{\Σ}}_k{X_{\mathrm{jk}}}^2\right)}}$

And then can be calculated by time series similarity between document and survey document and all kinds of document seasonal effect in time series similarity.

The formula calculating document to be measured and all kinds of document seasonal effect in time series similarity is as follows:

$\mathrm{Similarity}(p,C_i)=\frac{1}{n}\×\left[\underset{j=1}{\overset{n}{\mathrm{\Σ}}}\mathrm{Similarity}(p,p_j)\right]$

Similarity (p, C _i) represent document p and C to be measured _iclass document seasonal effect in time series Similarity value;

Similarity (p, p _j) represent document p to be measured and document p _jseasonal effect in time series Similarity value, document p _j∈ C _iclass, j=1,2 ..., (n represents C to n _itotal number of class Literature).

2. to be according to claim 1ly cited number Forecasting Methodology based on seasonal effect in time series scientific and technical literature, it is characterized in that: in step 1, utilize the index of each document of database retrieval, according to the label of document each in database and publish days, the concrete time that statistical literature is cited and number of times, the sum that is cited in each month after obtaining each Literature publication.