CN105005576A - System and method for searching similar users of video website - Google Patents

Abstract

The invention provides a system and a method for searching similar users of a video website based on video contents according to the conditions of seed users. According to the system and the method, the descriptors of videos are taken as the descriptions of the video contents, the video watching behaviors of users are extracted as that the uses watch the video contents, and therefore, a user group having the similar watching behaviors can be found out by virtue of indexes to the video contents. According to the system and the method, more similar video users can be found out, the similar video website user group can be expanded, and then the promotion effect of products in advertisements can be enhanced.

Description

A kind of video website similar users search system and method

Technical field

The present invention relates to a kind of video website similar users search system and method.

Background technology

Video ads in current video website is thrown in usually can run into such problem: in a certain advertisement putting, certain user has gone out interest to this advertisement performance, there occurs click or consumer behavior.Similar video user, the interest for advertised product has similarity to a certain degree, such as: two users that have viewed body building Yoga video equally, may all can be interested in diet products, cosmetics etc.; And two users that have viewed cutter tower game video equally, may become interested to some web games.

Similar user watches behavior, can for inferring that its hobby is given a clue.In advertisement release process, some users create positive feedback (such as: click, purchase etc.) to the advertisement of throwing in, using these users as seed user, carry out the expanded search of similar crowd, in the advertisement putting in future, just can lock this class crowd with a definite target in view, thus make advertisement putting produce larger benefit.

Therefore, how to utilize the viewing behavior of seed user, effective similar users search is carried out to video website, just become a very important problem.

Summary of the invention

The present invention, according to the situation of seed user, proposes a kind of video website similar users search system based on video content and method.The descriptor of video is considered as the description of video content by the present invention, refines the video-see behavior of user for user is to the viewing of video content, thus utilizes the index to video content, search for the customer group with similar viewing behavior.

System and method of the present invention can search out more similar video user, expand similar video website customer group, thus the promotion effect of product in lifting advertisement, make the input of this advertisement produce larger benefit, have important using value in advertisement putting field.

Accompanying drawing explanation

The present invention with reference to accompanying drawingfurther describe, wherein:

fig. 1it is video website similar users searching method flow process of the present invention figure;

fig. 2it is video website similar users search system of the present invention structural representation figure.

Embodiment

Although with reference to containing preferred embodiment of the present invention accompanying drawingabundant description the present invention, but should be appreciated that before this describes, those of ordinary skill in the art can revise invention described herein, obtains technique effect of the present invention simultaneously.Therefore, Yan Weiyi discloses widely to those of ordinary skill in the art must to understand following description, and its content does not lie in restriction exemplary embodiment described in the invention.

Reference fig. 1shown in, video website similar users searching method of the present invention comprises:

Step 1, carries out statistical study to user's view content, and in statistics a period of time, the user video of (such as a week) watches record, obtains the viewing number of times of each user on each video content and frequency in conjunction with video content description word.Wherein, video content description word describes mainly through video tab, keyword and video title participle, video tab, keyword and video title participle have all carried out brief and abstract description to video content, more efficiently can portray the content information of video, different videos containing similar content, may show that they may have identical label or keyword.Utilize the viewing record of user, in conjunction with video content description word, the viewing frequency of counting user on different content, can reflect the interest preference of user effectively.

Wherein, step 1 comprises further:

Step 1.1, utilizes the viewing record of video user, the video-see number of times in counting user a period of time, obtains the video-see list of " user ID---video labeling---viewing number of times ";

Step 1.2, for video information, extracts video information list " video labeling---label 1, label 2 ..., label i ", in conjunction with video-see list generating content viewing list " user ID---label i---watches number of times ";

Step 1.3, merges the content viewing record with same subscriber mark, utilizes the content viewing number of times of label i to calculate the viewing frequency of label i, the viewing frequency namely on inherent label i of each user's a period of time, and computing method are:

${\mathrm{tf}}_i=\frac{C_i}{{\mathrm{\Σ}}_{j\∈T}{C}_j}$

Wherein, tf _ifor the frequency of label i, C _i, C _jfor user watches the number of times of label i, label j, T is the set of all labels of this user viewing.

By step 1, the video content of viewing and the viewing frequency of often kind of video content in each user nearest a period of time can be obtained.

Step 2, set up the inverted index of user, according to the viewing record that statistical study in step 1 obtains, the inverted index of user is set up based on video content description word, this index form using the descriptor of video content as index key, using watch this descriptor all user ID and viewing frequency as index value.

Wherein, step 2 comprises further:

Step 2.1 take label as index key, adds up all users of this label viewed and the viewing frequency of each user, calculates the total number of users of this label viewed;

Step 2.2, utilizes hash method, carries out Hash calculation to label, carries out piecemeal to index file;

Step 2.3, is stored to the piecemeal place corresponding to cryptographic hash by the viewing information of label.

Step 3, carries out similar users search and calculates similarity, utilizes the video-see record of seed user, with video content description word for search key, indexed file carries out the search of similar users, calculate the similarity of relative users simultaneously, obtain preliminary Search Results.

Wherein, step 3 comprises further:

Step 3.1, analyzes the viewing record of seed user, searches for each label of seed user, obtains all total numbers of users of this label viewed, user ID and watches frequency accordingly;

Step 3.2, search for each the user returned and calculate similarity, its computing method are as follows:

$S_{\mathrm{ui}}={\mathrm{tf}}_{\mathrm{ui}}*\log \left(\frac{D}{P_i}\right)$

Wherein, S _uirepresent the similarity of user u on label i, tf _uirepresent that user u watches the frequency of label i, D represents all total numbers of users, P _irepresent the total number of users of viewing label i;

Step 3.4, searches for the viewing label of all seed user the result returned and comprehensively analyzes, and calculate each comprehensive similarity returning user, computing method are:

${\mathrm{Score}}_u=\underset{i}{\mathrm{\Σ}}{S}_{\mathrm{ui}}$

Wherein, Score _urepresent the comprehensive similarity of user u, S _uirepresent the similarity of user u on label i.

Step 4, carries out search results ranking, utilizes similarity to carry out descending sequence to initial search result, obtains final similar crowd's Search Results through filtration treatment.For the comprehensive similarity of searching for customer group and each user returned, carry out descending sequence, suitable similarity threshold can be adopted to carry out result filtration according to similarity, the result after sequence being filtered exports.

Reference fig. 2, the present invention also provides a kind of video website similar users search system, comprising:

Statistical study device, statistical study is carried out to user's view content, user video viewing record in statistics a period of time, each user is obtained to the viewing number of times of each video content and frequency in conjunction with video content description word, wherein, above-mentioned video content description word is described by video tab, keyword and video title participle.

Wherein, statistical study device utilizes the viewing record of video user, the video-see number of times in counting user a period of time, obtains " user ID---video labeling---viewing number of times " video-see list; For video information, extract video information list " video labeling---label 1, label 2 ..., label i ", in conjunction with video-see list generating content viewing list " user ID---label i---watches number of times "; Merge the content viewing record with same subscriber mark, utilize the viewing number of times of label i to calculate the viewing frequency of label i, the viewing frequency namely on inherent label i of each user's a period of time, computing method are:

${\mathrm{tf}}_i=\frac{C_i}{{\mathrm{\Σ}}_{j\∈T}{C}_j}$

Wherein, tf _ifor the frequency of label i, C _i, C _jfor user watches the number of times of label i, label j, T is the set of all labels of this user viewing.

Indexing unit, set up the inverted index of user, according to the viewing record that statistical study in the first step obtains, the inverted index of user is set up based on video content description word, this index form using video content description word as index key, using watch this descriptor all user ID and viewing frequency as index value.

Wherein, indexing unit is index key with label, adds up all users of this label viewed and the viewing frequency of each user, calculates the total number of users of this label viewed; Utilize hash method, Hash calculation is carried out to label, piecemeal is carried out to index file; The viewing information of label is stored to the piecemeal place corresponding to cryptographic hash.

Calculation element, carries out similar users search and calculates similarity, utilizes the video-see record of seed user, with video content description word for search key, indexed file carries out the search of similar users, calculate the similarity of relative users simultaneously, obtain preliminary Search Results.

Wherein, the viewing record of calculation element to seed user is analyzed, and searches for each label of seed user, obtains all total numbers of users of this label viewed, user ID and watches frequency accordingly;

Search for each the user returned and calculate similarity, its computing method are as follows:

$S_{\mathrm{ui}}={\mathrm{tf}}_{\mathrm{ui}}*\log \left(\frac{D}{P_i}\right)$

Wherein, S _uirepresent the similarity of user u on label i, tf _uirepresent that user u watches the frequency of label i, D represents all total numbers of users, P _irepresent the total number of users of viewing label i;

Search for the viewing label of all seed user the result returned comprehensively to analyze, calculate each comprehensive similarity returning user, computing method are:

${\mathrm{Score}}_u=\underset{i}{\mathrm{\Σ}}{S}_{\mathrm{ui}}$

Wherein, Score _urepresent the comprehensive similarity of user u, S _uirepresent the similarity of user u on label i.

Collator, carries out search results ranking, utilizes similarity to carry out descending sequence to initial search result, obtains final similar crowd's Search Results through filtration treatment.

Wherein, collator, for the comprehensive similarity of searching for customer group and each user returned, carries out descending sequence according to similarity, suitable similarity threshold can be adopted to carry out result filtration, and the result after sequence being filtered exports.

Below, system and method for the present invention is further described by two examples.

Example one: the similar people's group hunting of certain video website.

There is video set S={V certain website ₁..., V _n, each video packets, containing one group of content descriptor (i.e. label), may also have identical descriptor between different video.Watching record of user R={U simultaneously in this website records nearest a week ₁---V _x---C _1x..., U _n---V _y---C _ny.

Step 1, using label as the description of video content, according to the label information of each video, adds up viewing number of times on each tab in each user one week, obtains the viewing record of shape as " user ID---label---viewing number of times "; Viewing record for same subscriber mark carries out joint account, and obtain all labels of each user viewing, and calculate the viewing frequency of each label, computing method are:

${\mathrm{tf}}_i=\frac{C_i}{{\mathrm{\Σ}}_{j\∈T}{C}_j}$

Wherein, tf _ifor the frequency of label i, C _i, C _jfor user watches the number of times of label i, label j, T is the set of all labels of this user viewing.Like this, just obtain each user viewing frequency on each tab, part viewing record example is as follows:

table 1user's view content record example

Step 2, using label as index key, sets up inverted index to viewing information.Hash is carried out to label, obtains cryptographic hash; Suitable piecemeal is carried out to inverted index file, cryptographic hash and file block is set up and maps; Viewing information (comprising: the total number of users of this label viewing, watch all user ID of this label and the viewing frequency of each user) corresponding to each label is stored to file block place corresponding to this label cryptographic hash.

Step 3, for given seed user viewing record, utilizes viewing label information at the enterprising line search of inverted index file.For each label of seed user viewing, identical hash function is adopted to calculate cryptographic hash, thus find corresponding inverted index blocks of files, read viewing information wherein, obtain the total number of users of this label viewed, all user ID and viewing frequency, calculate the similarity of each user of this label viewed, method is as follows:

$S_{\mathrm{ui}}={\mathrm{tf}}_{\mathrm{ui}}*\log \left(\frac{D}{P_i}\right)$

Wherein, S _uirepresent the similarity of user u on label i, tf _uirepresent that user u watches the frequency of label i, D represents all total numbers of users, P _irepresent the total number of users of viewing label i.

The result returned for all seed user viewing tag search is comprehensively analyzed, and calculate each comprehensive similarity returning user, its computing method are:

${\mathrm{Score}}_i=\underset{i}{\mathrm{\Σ}}{S}_{\mathrm{ui}}$

Wherein, Score _irepresent the comprehensive similarity of user, S _uirepresent the similarity of user u on label i.

Step 4, according to the sequence that comprehensive similarity is carried out from big to small, through certain filtering screening, exports the result after sequence.

The seed file finally obtained and search result examples as follows:

table 2seed user

User ID	Comprehensive similarity
		1414805406362bou	7.457061423316192
1411422657876HQS	7.457061423316192
		1414897033491tst	6.188232499062661
1414225525441rHY	5.067268407706754
		1376735750584cE7	4.97137438163828
1413197549819YYw	4.97137438163828
		1414929307620uum	4.125488415218207
1401986230544u2n	4.125488415218207
		1396228567787C4I	4.125488415218207
1413550544110F75	4.125488415218207
		1414835997319Vst	4.125488415218207
14148266200180f2	4.125488415218207
		1413333051347w4D	4.125488415218207
1403043694606LSF	4.125488415218207

table 3part searches returns results

table 4the content viewing record of part similar users

Example two: certain product summary crowd expands

A certain product has locked a small amount of target group U={U1 ..., Um}, is desirably in certain video website and carries out product promotion, require to promote audient be with the target group U locked in there is the customer group of similar interests.Watching record of user R={U simultaneously in this website records nearest a week ₁---V _x---C _1x..., U _n---V _y---C _ny.

Step 1, utilizes the viewing record of website, searches the video-see record of user in target group U, in conjunction with video information, obtains the viewing record of target group based on video tab.In conjunction with the information of this product, screen the viewing label of target group, filtering has nothing to do label.Using the seed of the viewing record after filtration as search.Afterwards, for all viewing records in nearest a week, using label as the description of video content, according to the label information of each video, add up viewing number of times on each tab in each user one week, obtain the viewing record of shape as " user ID---label---viewing number of times ".

Viewing record for same subscriber mark carries out joint account, obtains all labels of each user viewing, and calculates the viewing frequency of each label.Computing method are:

${\mathrm{tf}}_i=\frac{C_i}{{\mathrm{\Σ}}_{j\∈T}{C}_j}$

Wherein, tf _ifor the frequency of label i, C _i, C _jfor user watches the number of times of label i, label j, T is the set of all labels of this user viewing.We just obtain each user viewing frequency on each tab like this.

Step 2, using label as index key, sets up inverted index to viewing information.Hash is carried out to label, obtains cryptographic hash; Suitable piecemeal is carried out to inverted index file, cryptographic hash and file block is set up and maps; Viewing information (comprising: the total number of users of this label viewing, watch all user ID of this label and the viewing frequency of each user) corresponding to each label is stored to file block place corresponding to this label cryptographic hash.

Step 3, for given seed user viewing record, utilizes viewing label information at the enterprising line search of inverted index file.For each label of seed user viewing, identical hash function is adopted to calculate cryptographic hash, thus find corresponding inverted index blocks of files, read viewing information wherein, obtain the total number of users of this label viewed, all user ID and viewing frequency, calculate the similarity of each user of this label viewed, computing method are as follows:

$S_{\mathrm{ui}}={\mathrm{tf}}_{\mathrm{ui}}*\log \left(\frac{D}{P_i}\right)$

Wherein, S _uirepresent the similarity of user u on label i, tf _uirepresent that user u watches the frequency of label i, D represents all total numbers of users, P _irepresent the total number of users of viewing label i.

The result that all seed user viewing tag search return comprehensively is analyzed, calculates each comprehensive similarity returning user.Its computing method are:

${\mathrm{Score}}_u=\underset{i}{\mathrm{\Σ}}{S}_{\mathrm{ui}}$

Wherein, Score _urepresent the comprehensive similarity of user u, S _uirepresent the similarity of user u on label i.

Step 4, according to the sequence that comprehensive similarity is carried out from big to small, through certain filtering screening, exports the result after sequence.

The view content example of this product summary user is as follows:

Targeted customer	The video tab of target customer's viewing	The product information that target customer pays close attention to
			Client one	Huawei honor 3c	Honor 3c UNICOM version
Client two	I Phone 5s	Iphone5s hand shell
			Client three	Meizu .mx3	mx3
Client four	Xplay news conference	Bubukao xplay3

table 5client's view content example

The partial results example searched is as follows:

User ID	Comprehensive similarity
		1403339995050JHU	6.344749147553775
1414920358781u4V	6.344749147553775
		14046215115455ID	6.344749147553775
1414887141725RG9	6.344749147553775
		1403888781082S88	6.344749147553775
1408775203633njo	6.344749147553775
		1400511822703RkC	6.344749147553775
1414852321322EFa	6.007725367960162
		1414934708013Eut	6.007725367960162
141126880285943L	6.007725367960162
		1414923557154foW	6.007725367960162
1414856887921mCx	6.007725367960162

table 6part searches returns results

table 7the viewing record of part similar users

After detailed description preferred embodiment of the present invention; those of ordinary skill in the art can clearly understand; various change and change can be carried out under the protection domain not departing from claim of enclosing and spirit, and the present invention is not also limited to the embodiment of examples cited embodiment in instructions.

Claims (9)

1. a video website similar users searching method, comprising:

Step 1, statistical study is carried out to user's view content, user video viewing record in statistics a period of time, each user is obtained to the viewing number of times of each video content and frequency in conjunction with video content description word, wherein, above-mentioned video content description word is described by video tab, keyword and video title participle;

Step 2, set up the inverted index of user, according to the viewing record that statistical study in above-mentioned steps 1 obtains, the inverted index of user is set up based on video content description word, this index form using video content description word as index key, using watch this descriptor all user ID and viewing frequency as index value;

Step 3, carries out similar users search and calculates similarity, utilizes the video-see record of seed user, with video content description word for search key, indexed file carries out the search of similar users, calculate the similarity of relative users simultaneously, obtain preliminary Search Results;

Step 4, carries out search results ranking, utilizes similarity to carry out descending sequence to initial search result, obtains final similar crowd's Search Results through filtration treatment.

2. method according to claim 1, wherein, step 1 comprises further:

Step 1.1, utilizes the viewing record of video user, the video-see number of times in counting user a period of time, obtains the video-see list of " user ID---video labeling---viewing number of times ";

Step 1.2, for video information, extracts video information list " video labeling---label 1, label 2 ..., label i ", in conjunction with video-see list generating content viewing list " user ID---label i---watches number of times ";

Step 1.3, merge the content viewing record with same subscriber mark, utilize the viewing number of times of label i to calculate the viewing frequency of label i, be i.e. viewing frequency on to that tag in each user's a period of time, computing method are:

${\mathrm{tf}}_i=\frac{C_i}{{\mathrm{\Σ}}_{j\∈T}{C}_j}$

Wherein, tf _ifor the frequency of label i, C _i, C _jfor user watches the number of times of label i, label j, T is the set of all labels of this user viewing.

3. method according to claim 1, wherein, step 2 comprises further:

Step 2.1 take label as index key, adds up the frequency that all users of this label viewed and each user watch this label, calculates the total number of users of this label viewed;

Step 2.2, utilizes hash method, carries out Hash calculation to label, carries out piecemeal to index file;

Step 2.3, is stored to the position corresponding to cryptographic hash by the viewing information of label.

4. method according to claim 1, wherein, step 3 comprises further:

Step 3.1, analyzes the viewing record of seed user, searches for each label of seed user, obtains all total numbers of users of this label viewed, user ID and watches frequency accordingly;

Step 3.2, search for each the user returned and calculate similarity, its computing method are as follows:

$S_{\mathrm{ui}}={\mathrm{tf}}_{\mathrm{ui}}*\log \left(\frac{D}{P_i}\right)$

Wherein, S _uirepresent the similarity of user u on label i, tf _uirepresent that user u watches the frequency of label i, D represents all total numbers of users, P _irepresent the total number of users of viewing label i;

Step 3.4, searches for the viewing label of all seed user the result returned and comprehensively analyzes, and calculate each comprehensive similarity returning user, computing method are:

${\mathrm{Score}}_u=\underset{i}{\mathrm{\Σ}}{S}_{\mathrm{ui}}$

Wherein, Score _urepresent the comprehensive similarity of user u, S _uirepresent the similarity of user u on label i.

5. method according to claim 1, wherein, step 4 comprises further: for the comprehensive similarity of searching for customer group and each user returned, descending sequence is carried out according to similarity, suitable similarity threshold can be adopted to carry out result filtration, and the result after sequence being filtered exports.

6. a video website similar users search system, comprising:

Statistical study device, statistical study is carried out to user's view content, user video viewing record in statistics a period of time, each user is obtained to the viewing number of times of each video content and frequency in conjunction with video content description word, wherein, above-mentioned video content description word is described by video tab, keyword and video title participle;

Indexing unit, set up the inverted index of user, according to the viewing record that statistical study in statistical study device obtains, the inverted index of user is set up based on video content description word, this index form using video content description word as index key, using watch this descriptor all user ID and viewing frequency as index value;

Calculation element, carries out similar users search and calculates similarity, utilizes the video-see record of seed user, with video content description word for search key, indexed file carries out the search of similar users, calculate the similarity of relative users simultaneously, obtain preliminary Search Results;

Collator, carries out search results ranking, utilizes similarity to carry out descending sequence to initial search result, obtains final similar crowd's Search Results through filtration treatment.

7. system according to claim 6, wherein, statistical study device utilizes the viewing record of video user, the video-see number of times in counting user a period of time, obtains " user ID---video labeling---viewing number of times " video-see list; For video information, extract video information list " video labeling---label 1, label 2 ..., label i ", in conjunction with video-see list generating content viewing list " user ID---label i---watches number of times "; Merge the content viewing record with same subscriber mark, utilize the viewing number of times of label to calculate the viewing frequency of label, be i.e. viewing frequency on to that tag in each user's a period of time, computing method are:

${\mathrm{tf}}_i=\frac{C_i}{{\mathrm{\Σ}}_{j\∈T}{C}_j}$

Wherein, tf _ifor the frequency of label i, C _i, C _jfor user watches the number of times of label i, label j, T is the set of all labels of this user viewing.

8. system according to claim 6, wherein, indexing unit is index key with label, adds up the frequency that all users of this label viewed and each user watch this label, calculates the total number of users of this label viewed; Utilize hash method, Hash calculation is carried out to label, piecemeal is carried out to index file; The viewing information of label is stored to the position corresponding to cryptographic hash.

9. system according to claim 6, wherein, the viewing record of calculation element to seed user is analyzed, and searches for each label of seed user, obtains all total numbers of users of this label viewed, user ID and watches frequency accordingly;

Search for each the user returned and calculate similarity, its computing method are as follows:

$S_{\mathrm{ui}}={\mathrm{tf}}_{\mathrm{ui}}*\log \left(\frac{D}{P_i}\right)$

Wherein, S _uirepresent the similarity of user u on label i, tf _uirepresent that user u watches the frequency of label i, D represents all total numbers of users, P _irepresent the total number of users of viewing label i;

Search for the viewing label of all seed user the result returned comprehensively to analyze, calculate each comprehensive similarity returning user, computing method are:

${\mathrm{Score}}_u=\underset{i}{\mathrm{\Σ}}{S}_{\mathrm{ui}}$

Wherein, Score _urepresent the comprehensive similarity of user u, S _uirepresent the similarity of user u on label i.