CN103209342A - Collaborative filtering recommendation method introducing video popularity and user interest change - Google Patents

Abstract

The invention discloses a collaborative filtering recommendation method introducing video popularity and user interest change. The method comprises the following steps of: acquiring and processing user behavior data, and thus obtaining a user-video binary incidence matrix; acquiring a video popularity weight and a user interest weight on the basis of the matrix, and introducing the video popularity weight and the user interest weight into a user similarity calculation process; searching the first K neighbors maximally similar to a target user, and predicting an interest value of the target user in a video which does not generate an effective behavior according to the magnitude of the similarity between the target user and a neighbor user; and selecting N videos with the maximum interest value to form a recommendation list, and providing a personalized recommendation for the user. In full consideration of the characteristics of difference in the popularity of the videos in the system and the time-dependent change of user interest, the method is in accordance with an objective fact, so that the user similarity can be accurately calculated, the quality of a collaborative filtering recommendation is improved, and a personalized video recommendation in accordance with the user interest is provided for a video user.

Description

A kind of collaborative filtered recommendation method of introducing video popularity and user interest variation

Technical field

The invention belongs to Rich Media's personalized recommendation technical field, more specifically say, relate to a kind of collaborative filtered recommendation method of introducing video popularity and user interest variation.

Background technology

Along with network and application and development thereof, network has welcome the epoch of " information explosion ", be the demand that the search technique of representative has not satisfied the user with Google and Baidu, the one, the user may can not find the information of wanting by search engine, and the 2nd, the user may reach suitable literal to the demand schedule of oneself on one's own initiative allows search engine operate.Thus, commending system arises at the historic moment, and historical behavior and feedback information that it can gather the user find the resource that meets user interest according to these information, make personalized recommendation for the user then.

Recommended technology has been applied to a plurality of fields such as video traffic, ecommerce, personalized reading.Especially in Web TV (IPTV) field, no matter the user watches video frequency program by digital television business or video website, for when nowadays colourful video frequency program rejoices in, feel helpless for being difficult to from the program resource of vastness, to choose the program of oneself really liking again.According to the CNNIC(China Internet Network Information Center) " the China Internet state of development statistical report " of up-to-date issue show that by in by the end of December, 2012, Chinese netizen's scale reaches 5.64 hundred million, Internet video user reaches 3.72 hundred million, than having increased by 4,653 ten thousand people last year.In the face of the video traffic customer group of huge and quick growth like this, be far reaching for it provides personalized video recommendation service, and containing considerable commercial value.

Collaborative filtering is current most widely used personalized recommendation technology, and the commending system that some are well-known such as Amazon, GroupLens and Douban have adopted the method for collaborative filtering.This method is based on a kind of hypothesis of general knowledgeization: if the user is similar to the behavior of some resources, then they are also more similar to the behavior of other resources.The basic thought of collaborative filtering is according to the similarity between user's the historical behavior calculating different user, finds targeted customer's K nearest-neighbors, and the behavior of watching according to this K user generates recommendation list to the targeted customer then.In video system, the user is various to the behavior of video frequency program, comprises duration, number of clicks etc. are marked, watched in the demonstration of program.System extracts these behaviors and they is converted to the binary system association of user-video, and such as scoring, when watching duration and number of clicks greater than setting threshold, user-video is associated as " 1 ", is when not reaching threshold value " 0 ".System uses many similarity calculating methods to comprise space cosine similarity, Pearson's coefficient correlation, Jie Kade similarity factor etc. according to the related similarity of calculating between the user of these binary systems.The most important link of collaborative filtering is K nearest-neighbors of search, and neighbor seaching is according to user's similarity, still, during calculating, present similarity has some problems, the one, system all videos all par treat, ignored the popular degree of program, i.e. popularity.Because the user is different to the meaning of the program generation behavior of different popularities, all watched certain quite high popular film of popularity by Web TV such as two users, this is like all having bought xinhua dictionary as two students, can illustrate that hardly their interest is similar, joint act to the unexpected winner program more can embody similitude on the contrary, is irrational so all programs are treated on an equal basis.The 2nd, ignored user interest over time.Because different user watches the behavior of the video frequency program behavior bigger than time span more can embody user's similarity degree in less time range, is to meet objective fact inadequately so traditional similarity calculating method is not considered the user interest variation.

Summary of the invention

The objective of the invention is to overcome the deficiencies in the prior art, design a kind of collaborative filtering video recommend method of introducing video popularity and user interest variation, calculate similarity between the user according to video popularity weights and user interest decay weight, more tallied with the actual situation and user's similarity more accurately, recommended for the targeted customer makes more accurately.

For achieving the above object, the present invention introduces the collaborative filtered recommendation method of video popularity and user interest variation, it is characterized in that may further comprise the steps:

(1), collects m user to the behavioral data of n video, comprise that the user marks, the user watches duration, the user watches/number of clicks and behavior time of origin, setting mark threshold value, user of corresponding user according to actual conditions watches duration threshold value and user to watch/the number of clicks threshold value, when having a setting threshold more than or equal to correspondence in three data at least, be user's vaild act, set user-video and be associated as " 1 "; Otherwise set user-video and be associated as " 0 "; (m n), has wherein comprised the related information between m user and n the video, the related b of user-video to obtain " user-video " binary system incidence matrices B _UiThe interest value of the video i of expression user u;

(2), (m n) carries out column scan and line scanning respectively, and column scan obtains video i, and the vaild act user of 1≤i≤n gathers U to matrix B _i, line scanning obtains user u, the vaild act video set I of 1≤u≤m _u, note popular _i=| U _i| be the popularity of video i, expression produces user's number of vaild act to video i; Active _u=| I _u| be the liveness of user u, expression user u produces the video number of vaild act; Wherein | U _i| and | I _u| represent U respectively _iAnd I _uMould;

(3), set computing cycle, calculated off-line video popularity weights and user interest decay weight, video popularity weight calculation formula is:

${\mathrm{\ω}}_h\left(i\right)=\frac{1}{\log (1+{\mathrm{popular}}_i)}=\frac{1}{\log (1+|U_i\left|\right)}$

User interest decay weight calculation formula:

${\mathrm{\ω}}_t(u,v,s)=\frac{1}{\log (1+\mathrm{\α}|t_{\mathrm{us}}-t_{\mathrm{vs}}+1\left|\right)},s\∈I_u\∩I_v$

Wherein, s represents user u and user v, and 1≤v≤m, v ≠ u all produce the video of vaild act, t _UsAnd t _VsRepresent the time that the video s of user u and user v produces the last vaild act respectively, α, 0≤α≤1 is interest influence of fading coefficient, according to the actual conditions setting;

(4), usage space cosine similarity is calculated the similarity between targeted customer u and other users except user u:

${\mathrm{sim}}_{\mathrm{uv}}=\frac{\underset{s}{\mathrm{\Σ}}{\mathrm{\ω}}_h\left(s\right){\mathrm{\ω}}_t(u,v,s)}{\sqrt{\left|I_u\right|\left|I_v\right|}},s\∈I_u\∩I_v$

Wherein, | I _u| and | I _v| be the liveness of user u and user v, represent that respectively user u and user v produce the video number of vaild act;

(5), the similarity that obtains is sorted from big to small, K user gathers as the nearest-neighbors of targeted customer u before choosing, be designated as S (u, K), wherein K sets according to actual conditions;

(6), watch behavior prediction targeted customer u to not watching the interest value of video j according to what the nearest-neighbors of targeted customer u was gathered, and sort from big to small by interest value, the top n video is recommended the targeted customer as video recommendation list RecVedio (N), and wherein N sets according to actual conditions.

Wherein, the decision method of the user's vaild act in the step (1) may further comprise the steps:

1.1), judge that user scoring whether more than or equal to user's threshold value of marking, if then the behavior is user's vaild act, judges and finish; Otherwise enter step 1.2);

1.2), judge that the user watches duration whether to watch the duration threshold value more than or equal to the user, if then the behavior is user's vaild act, judge to finish; Otherwise enter step 1.3);

1.3), judge the user watch/whether number of clicks watch/the number of clicks threshold value more than or equal to the user, if then the behavior is user's vaild act, judges and finish; Otherwise the behavior is not user's vaild act, judges and finishes.

Wherein, the computing cycle in the step (3) is one hour.

Wherein, interest influence of fading factor alpha=0.5 in the step (3).

Wherein, the target of prediction user u in the step (6) to the computing formula of the interest value of not watching video j is:

$p(u,j)=\underset{w}{\mathrm{\Σ}}{\mathrm{sim}}_{\mathrm{uw}},w\∈S(u,K)\∩U_j$

Wherein, user w is the user who video j was produced vaild act in the nearest-neighbors set of user u.

Goal of the invention of the present invention is achieved in that

The present invention introduces the collaborative filtered recommendation method of video popularity and user interest variation, handle and obtain user-video binary system incidence matrices by gathering user behavior data, obtain video popularity weights and user interest degree weight based on this matrix, and be incorporated in user's similarity computational process; Find preceding K neighbours with targeted customer's similarity maximum then, by with neighbours user's similarity size target of prediction user to not producing the interest value of vaild act video; N video choosing the interest value maximum at last forms recommendation list and makes personalized recommendation for the user.

The present invention introduces the collaborative filtered recommendation method of video popularity and user interest variation, utilization is representing the behavioral data of individual subscriber interest, change in conjunction with video popularity and user interest, the purpose that provides individualized video to recommend for the user in IPTV or video website system field has been provided, meet network personalized development on the one hand, for the client provides better service, on the other hand also can helping service provider attract and keep more clients here, increase economic efficiency.

Description of drawings

Fig. 1 is a kind of embodiment flow chart that the present invention introduces the collaborative filtered recommendation method of video popularity and user interest variation;

Fig. 2 is the schematic flow sheet that forms user-video binary system incidence matrices among Fig. 1 according to user behavior data.

Embodiment

Below in conjunction with accompanying drawing the specific embodiment of the present invention is described, so that those skilled in the art understands the present invention better.What need point out especially is that in the following description, when perhaps the detailed description of known function and design can desalinate main contents of the present invention, these were described in here and will be left in the basket.

Embodiment

Fig. 1 is a kind of embodiment flow chart that the present invention introduces the collaborative filtered recommendation method of video popularity and user interest variation.As shown in Figure 1, the present invention introduces the collaborative filtered recommendation method that video popularity and user interest change and may further comprise the steps:

Step S101: video system is collected m user to the behavioral data of n video, comprises that the user marks, the user watches duration, the user watches/number of clicks and behavior time of origin.Then these data are stored in the database in high in the clouds.The video i scoring of note user u is score in the present embodiment _Ui, it is length that the video i of user u watches duration _Ui, the number of times that the video i of user u watched or clicked video i is designated as freq _Ui

Step S102: set mark threshold value, user of corresponding user according to actual conditions and watch duration threshold value and user to watch/the number of clicks threshold value, when having a setting threshold more than or equal to correspondence in three data at least, be user's vaild act, set user-video and be associated as " 1 "; Otherwise set user-video and be associated as " 0 "; (m n), has wherein comprised the related information between m user and n the video, the related b of user-video to obtain user-video binary system incidence matrices B _UiThe interest value of the video i of expression user u.

Fig. 2 is the schematic flow sheet that forms user-video binary system incidence matrices among Fig. 1 according to user behavior data.As shown in Figure 2, database flow process that the original user behavioral data is handled is as follows beyond the clouds for video system:

Step S201: user's scoring is judged.In the present embodiment, the points-scoring system of video system adopts the five-grade marking system, and the user can make 1～5 minute scoring for watching video, and setting user's threshold value of marking is 3.Work as score _Ui〉=3 o'clock, then the behavior of the video i of user u was effective, enters step S204; If the user initiatively is not video scoring or score _Ui＜3, then enter step S202.

Step S202: watch duration to judge to the user.In the present embodiment, it is 0.25h that the setting user watches the duration threshold value.Work as length _UiDuring 〉=0.25h, then the behavior of the video i of user u is effective, enters step S204; Otherwise enter step S203.

Step S203: to the user watch/number of clicks judges.In the present embodiment, set the user to watch/the number of clicks threshold value is 3.Work as freq _Ui〉=3 o'clock, then the behavior of the video i of user u was effective, enters step S204; Otherwise enter step S205.

Step S204: set the related b of user-video _Ui=1, enter step S206.

Step S205: set the related b of user-video _Ui=0, enter step S206.

Step S206: according to the user-video relating value of all users to all videos, (m n), has wherein comprised the related information between m user and n the video, the related b of user-video to obtain user-video binary system incidence matrices B _UiThe interest value of the video i of expression user u.

In the present embodiment, use the user-video related information of 4 users and 6 film videos.Table 1 be binary system incidence matrices B (4, F).

?	Film A	Film B	Film C	Film D	Film E	Film F
							The user 1	1	1	0	1	0	1
The user 2	0	1	0	1	1	0
							The user 3	0	0	1	1	1	1
The user 4	1	1	0	1	1	0

Table 1

Step S103: (m n) carries out column scan and line scanning respectively, and column scan gets video i, and the vaild act user of 1≤i≤n gathers U to matrix B _i, line scanning obtains user u, the vaild act video set I of 1≤u≤m _u, note popular _i=| U _i| be the popularity of video i, expression produces user's number of vaild act to video i; Active _u=| I _u| be the liveness of user u, expression user u produces the video number of vaild act; Wherein | U _i| and | I _u| represent U respectively _iAnd I _uMould.

In the present embodiment, matrix is carried out column scan, run into " 1 " back corresponding user is joined this row video i, the vaild act user of A≤i≤F gathers U _iIn, be respectively:

U _A={ user 1, user 4};

U _B={ user 1, and the user 2, user 4};

U _C={ user 3};

U _D={ user 1, and the user 2, and the user 3, user 4};

U _E={ user 2, and the user 3, user 4};

U _F={ user 1, user 3}.

After column scan is finished, the user is gathered U _iAsk mould to obtain | U _i|, | U _i| the popularity of expression film video i is respectively:

|U _A|=2；|U _B|=3；|U _C|=1；|U _D|=4；|U _E|=3；|U _F|=2。

Matrix is carried out line scanning, run into " 1 " back corresponding film video is joined this row user u, the vaild act video set I of 1≤u≤4 _uIn, be respectively:

I ₁={ film A, film B, film D, film F};

I ₂={ film B, film D, film E};

I ₃={ film C, film D, film E, film F};

I ₄={ film A, film B, film D, film E}.

After line scanning is finished, to vaild act video set I _uAsk mould to obtain | I _u|, | I _u| the liveness of expression user u is respectively:

|I ₁|=4；|I ₂|=3；|I ₃|=4；|I ₄|=4。

Step S104: set computing cycle, calculated off-line video popularity weights and user interest decay weight, video popularity weight calculation formula is:

${\mathrm{\ω}}_h\left(i\right)=\frac{1}{\log (1+{\mathrm{popular}}_i)}=\frac{1}{\log (1+|U_i\left|\right)}$

User interest decay weight calculation formula:

${\mathrm{\ω}}_t(u,v,s)=\frac{1}{\log (1+\mathrm{\α}|t_{\mathrm{us}}-t_{\mathrm{vs}}+1\left|\right)},s\∈I_u\∩I_v$

Wherein, s represents user u and user v, and 1≤v≤m, v ≠ u all produce the video of vaild act, t _UsAnd t _VsRepresent the time that the video s of user u and user v produces vaild act respectively, α, 0≤α≤1 is interest influence of fading coefficient, according to the actual conditions setting.

In the present embodiment, the computing cycle of setting is one hour, namely upgrades once every one hour, (when the user uses the peak period, can suitably shorten the cycle), system according to user-video incidence matrices B (4, F) calculated off-line video popularity weights and user interest the decay weight.Interest influence of fading factor alpha=0.5 is set.

A. the popularity of video represents video i is produced user's number of vaild act, and (m is nonzero term sum in the i row in n), and more many users produce vaild act to video i, and then video i is more popular in matrix B.So (m n) calculates the popularity weights value that can obtain each video, the higher video role of popularity when this value has slackened the calculating of user's similarity based on binary system incidence matrices B.

In the present embodiment, according to video popularity weight calculation formula, the video popularity weighted value of 6 films is respectively (all result of calculations all keep 2 decimals in the present embodiment):

${\mathrm{\ω}}_h\left(A\right)=\frac{1}{\log (1+|U_A\left|\right)}=\frac{1}{\log (1+2)}\≈2.10;$

${\mathrm{\ω}}_h\left(B\right)=\frac{1}{\log (1+|U_B\left|\right)}=\frac{1}{\log (1+3)}\≈1.66;$

${\mathrm{\ω}}_h\left(C\right)=\frac{1}{\log (1+|U_C\left|\right)}=\frac{1}{\log (1+1)}\≈3.32;$

${\mathrm{\ω}}_h\left(D\right)=\frac{1}{\log (1+|U_D\left|\right)}=\frac{1}{\log (1+4)}\≈1.43;$

${\mathrm{\ω}}_h\left(E\right)=\frac{1}{\log (1+|U_E\left|\right)}=\frac{1}{\log (1+3)}\≈1.66;$

${\mathrm{\ω}}_h\left(F\right)=\frac{1}{\log (1+|U_F\left|\right)}=\frac{1}{\log (1+2)}\≈2.10.$

B. user interest decay weight, the user that need utilize system acquisition to arrive produces the temporal information of vaild act to video, if by the vaild act of watching number of times to obtain, perhaps a certain video there is repeatedly vaild act, all get the last time that produces vaild act, in the present embodiment, the fate of this time with the distance current time is unit.Because user interest can change in time, watch the method for the time span of same video to weigh interests change to the influence of similarity to calculate between the user.

The fate that table 2 is separated by to the last vaild act of each film video and current time for each user.Wherein "/" expression does not produce vaild act.

?	Film A	Film B	Film C	Film D	Film E	Film F
							The user 1	317	3650	/	270	/	1026
The user 2	/	730	/	150	30	/
							The user 3	/	/	538	169	883	945
The user 4	908	603	/	3200	9102	/

Table 2

In the present embodiment, as the targeted customer, be example with user 1 with user 2, user 2 is film B and film D with the video that user 1 all produces vaild act, according to user interest decay weight calculation formula, obtain in user 2 for user 1 user interest decay weighted value be:

${\mathrm{\ω}}_t(\mathrm{2,1},B)=\frac{1}{\log (1+\mathrm{\α}|t_{2B}-t_{1B}+1\left|\right)}=\frac{1}{\log (1+0.5|730-3650+1\left|\right)}\≈0.32;$

${\mathrm{\ω}}_t(\mathrm{2,1},D)=\frac{1}{\log (1+\mathrm{\α}|t_{2D}-t_{1D}+1\left|\right)}=\frac{1}{\log (1+0.5|150-270+1\left|\right)}\≈0.56.$

In like manner can get user 2 for other users' user interest decay weighted value:

${\mathrm{\ω}}_t(2,3,D)=\frac{1}{\log (1+\mathrm{\α}|t_{2D}-t_{3D}+1\left|\right)}=\frac{1}{\log (1+0.5|150-169+1\left|\right)}\≈1.00;$

${\mathrm{\ω}}_t(2,3,E)=\frac{1}{\log (1+\mathrm{\α}|t_{2E}-t_{3E}+1\left|\right)}=\frac{1}{\log (1+0.5|30-883+1\left|\right)}\≈0.38;$

${\mathrm{\ω}}_t(\mathrm{2,4},B)=\frac{1}{\log (1+\mathrm{\α}|t_{2B}-t_{4B}+1\left|\right)}=\frac{1}{\log (1+0.5|730-603+1\left|\right)}\≈0.55;$

${\mathrm{\ω}}_t(\mathrm{2,4},D)=\frac{1}{\log (1+\mathrm{\α}|t_{2D}-t_{4D}+1\left|\right)}=\frac{1}{\log (1+0.5|150-3200+1\left|\right)}\≈0.31;$

${\mathrm{\ω}}_t(2,4,E)=\frac{1}{\log (1+\mathrm{\α}|t_{2E}-t_{4E}+1\left|\right)}=\frac{1}{\log (1+0.5|30-9102+1\left|\right)}\≈0.27.$

Like this, the vaild act that time span is little between different user is bigger to the influence that similarity produces, and makes that in more little time range that same video is produced the user of vaild act is more similar, more meets objective fact.

Step S105: usage space cosine similarity is calculated the similarity between targeted customer u and other user v except user u:

${\mathrm{sim}}_{\mathrm{uv}}=\frac{\underset{s}{\mathrm{\Σ}}{\mathrm{\ω}}_h\left(s\right){\mathrm{\ω}}_t(u,v,s)}{\sqrt{\left|I_u\right|\left|I_v\right|}},s\∈I_u\∩I_v$

Wherein, | I _u| and | I _v| be the liveness of user u and user v, expression user u and user v produce the video number of vaild act respectively.

In the present embodiment, by above step process, popularity weights value and some user interest decay weighted values of each video correspondence have been stored in the system database.When needs are done recommendation to the targeted customer, according to user's liveness of last update in the database, video popularity, video popularity weights and user interest decay weight calculation targeted customer (user 2) the similarity value with other users, obtain according to space cosine similarity formula:

${\mathrm{sim}}_{21}=\frac{{\mathrm{\ω}}_h\left(B\right){\mathrm{\ω}}_t(\mathrm{2,1},B)+{\mathrm{\ω}}_h\left(D\right){\mathrm{\ω}}_t(\mathrm{2,1},D)}{\sqrt{\left|I_2\right|\left|I_1\right|}}=\frac{1.66\×0.32+1.43\×0.56}{\sqrt{3\×4}}\≈0.385;$

${\mathrm{sim}}_{23}=\frac{{\mathrm{\ω}}_h\left(D\right){\mathrm{\ω}}_t(\mathrm{2,3},D)+{\mathrm{\ω}}_h\left(E\right){\mathrm{\ω}}_t(\mathrm{2,3},E)}{\sqrt{\left|I_2\right|\left|I_3\right|}}=\frac{1.43\×1.00+1.66\×0.38}{\sqrt{3\×4}}\≈0.595;$

${\mathrm{sim}}_{24}=\frac{{\mathrm{\ω}}_h\left(B\right){\mathrm{\ω}}_t(\mathrm{2,4},B)+{\mathrm{\ω}}_h\left(D\right){\mathrm{\ω}}_t(\mathrm{2,4},D)+{\mathrm{\ω}}_h\left(E\right){\mathrm{\ω}}_t(\mathrm{2,4},E)}{\sqrt{\left|I_2\right|\left|I_4\right|}}.$

$=\frac{1.66\×0.55+1.43\×0.31+1.66\×0.27}{\sqrt{3\×4}}\≈0.521$

Adopt the similarity of introducing video popularity and user interest variation to calculate, the vaild act that produces at same user in unexpected winner video and the little time span is bigger to the contribution of similarity, more meets objective fact.For example, if do not consider the influence of video popularity and user interest degree weight, by incidence matrices as can be seen, user 2 and user's 4 space cosine similarity value can be greater than user 2 and user's 3 space cosine similarity value.But above result of calculation shows, sim ₂₃Greater than sim ₂₄, this mainly is that user interest decay weight has produced effective lifting to their similarity, has obtained more preferably user's analog result because the time of user 3 and 2 couples of film D of user and film E generation vaild act is very approaching.

Step S106: the similarity that obtains is sorted from big to small, and K user gathers as the nearest-neighbors of targeted customer u before choosing, be designated as S (u, K); Wherein K sets according to actual conditions.

The user 2 that obtains and the similarity size between other users sorted to get sim ₂₃＞sim ₂₄＞sim ₂₁, set K=3 in the present embodiment, therefore choose these 3 users and gather as user 2 nearest-neighbors: S (user 2,3)={ user 3, and the user 4, user 1}.

Step S107: watch behavior prediction targeted customer u to not watching the interest value of video j according to what the nearest-neighbors of targeted customer u was gathered, and sort from big to small by interest value, the top n video is recommended the targeted customer as video recommendation list RecVedio (N), and wherein N sets according to actual conditions.

The computing formula of specifically carrying out the interest prediction is:

$p(u,j)=\underset{w}{\mathrm{\Σ}}{\mathrm{sim}}_{\mathrm{uw}}{b}_{\mathrm{wj}},w\∈S(u,K)\∩U_j$

Wherein, user w is the user who video j was produced vaild act in the nearest-neighbors set of user u, b _WjSo the interest value of the video j of expression user w is all b in this computing formula _Wj=1.Therefore the final computing formula of interest prediction is:

$p(u,j)=\underset{w}{\mathrm{\Σ}}{\mathrm{sim}}_{\mathrm{uw}},w\∈S(u,K)\∩U_j$

In the present embodiment, with the big or small predictive user 2 of the similarity of its nearest-neighbors collection it is not produced the interest level of vaild act film according to user 2.Scan matrix B (4, F), obtain the video set that user 2 do not produce vaild act and be: H _u={ film A, film C, film F}.

Then to each film j ∈ H _u, the interest value size of the video j of predictive user u is respectively:

P (user 2, film A)=sim ₂₄+ sim ₂₁=0.521+0.385=0.906;

P (user 2, film C)=sim ₂₃=0.595;

P (user 2, film F)=sim ₂₃+ sim ₂₁=0.595+0.385=0.980.

The interest value that calculates obtains film F＞film A＞film C by descending.In the present embodiment, N=2 is set, forms recommendation list: RecVedio (N)={ film F, film A}.This tabulation is pushed to the targeted customer, recommends thereby form.

Although above the illustrative embodiment of the present invention is described; so that those skilled in the art understand the present invention; but should be clear; the invention is not restricted to the scope of embodiment; to those skilled in the art; as long as various variations appended claim limit and the spirit and scope of the present invention determined in, these variations are apparent, all utilize innovation and creation that the present invention conceives all at the row of protection.

Claims (5)

1. introduce the collaborative filtered recommendation method that video popularity and user interest change for one kind, it is characterized in that may further comprise the steps:

(1), collects m user to the behavioral data of n video, comprise that the user marks, the user watches duration, user click frequency and behavior time of origin, setting mark threshold value, user of corresponding user according to actual conditions watches duration threshold value and user to watch/the number of clicks threshold value, when having a setting threshold more than or equal to correspondence in three data at least, be user's vaild act, set user-video and be associated as " 1 "; Otherwise set user-video and be associated as " 0 "; (m n), has wherein comprised the related information between m user and n the video, the related b of user-video to obtain user-video binary system incidence matrices B _UiThe interest value of the video i of expression user u;

(2), (m n) carries out column scan and line scanning respectively, and column scan obtains video i, and the vaild act user of 1≤i≤n gathers U to matrix B _i, line scanning obtains user u, the vaild act video set I of 1≤u≤m _u, note popular _i=| U _i| be the popularity of video i, expression produces user's number of vaild act to video i; Active _u=| I _u| be the liveness of user u, expression user u produces the video number of vaild act;

(3), set computing cycle, calculated off-line video popularity weights and user interest decay weight, video popularity weight calculation formula is:

${\mathrm{\ω}}_h\left(i\right)=\frac{1}{\log (1+{\mathrm{popular}}_i)}=\frac{1}{\log (1+|U_i\left|\right)}$

User interest decay weight calculation formula:

${\mathrm{\ω}}_t(u,v,s)=\frac{1}{\log (1+\mathrm{\α}|t_{\mathrm{us}}-t_{\mathrm{vs}}+1\left|\right)},s\∈I_u\∩I_v$

Wherein, s represents user u and user v, and 1≤v≤m, v ≠ u all produce the video of vaild act, t _UsAnd t _VsRepresent the time that the video s of user u and user v produces the last vaild act respectively, α is interest influence of fading coefficient, according to the actual conditions setting;

(4), usage space cosine similarity is calculated the similarity between targeted customer u and other users except user u:

${\mathrm{sim}}_{\mathrm{uv}}=\frac{\underset{s}{\mathrm{\Σ}}{\mathrm{\ω}}_h\left(s\right){\mathrm{\ω}}_t(u,v,s)}{\sqrt{\left|I_u\right|\left|I_v\right|}},s\∈I_u\∩I_v$

Wherein, | I _u| and | I _v| be the liveness of user u and user v;

(5), the similarity that obtains is sorted from big to small, K user gathers as the nearest-neighbors of targeted customer u before choosing, be designated as S (u, K), wherein K sets according to actual conditions;

(6), according to the set of the nearest-neighbors of targeted customer u watch behavior prediction targeted customer u to not watching the interest value of video j, and sort from big to small by interest value, the top n video is recommended the targeted customer as video recommendation list RecVedio (N).

2. collaborative filtered recommendation method according to claim 1 is characterized in that, the decision method of the user's vaild act in the described step (1) may further comprise the steps:

1.1), judge that user scoring whether more than or equal to user's threshold value of marking, if then the behavior is user's vaild act, judges and finish; Otherwise enter step 1.2);

1.2), judge that the user watches duration whether to watch the duration threshold value more than or equal to the user, if then the behavior is user's vaild act, judge to finish; Otherwise enter step 1.3);

1.3), judge the user watch/whether number of clicks watch/the number of clicks threshold value more than or equal to the user, if then the behavior is user's vaild act, judges and finish; Otherwise the behavior is not user's vaild act, judges and finishes.

3. according to claim 1 and 2 arbitrary described collaborative filtering recommending methods, it is characterized in that the target of prediction user u in the described step (6) to the computing formula of the interest value of not watching video j is:

$p(u,j)=\underset{w}{\mathrm{\Σ}}{\mathrm{sim}}_{\mathrm{uw}},w\∈S(u,K)\∩U_j$

Wherein, user w is the user who video j was produced vaild act in the nearest-neighbors set of user u.

4. collaborative filtered recommendation method according to claim 1 is characterized in that, the computing cycle in the described step (3) is one hour.

5. collaborative filtered recommendation method according to claim 1 is characterized in that, interest influence of fading factor alpha=0.5 in the described step (3).

Images