CN103209342B - A kind of introduce video popularity and the collaborative filtered recommendation method of user's interests change - Google Patents

Abstract

The present invention discloses and a kind of introduces video popularity and the collaborative filtered recommendation method of user's interests change, carry out process by collection user behavior data and obtain user-video scale-of-two incidence matrix, obtain video popularity weights and user interest degree weight based on this matrix, and it is incorporated in user's Similarity Measure process; Then front K the neighbours maximum with target user's similarity are found, by the similarity size target of prediction user with neighbor user to the interest value not producing effective behavior video; Finally choose the maximum N number of video of interest value to form recommendation list for user and make personalized recommendation. The present invention fully considers the characteristic that the popularity degree of video in system is different and user's interest changes in time, more meet objective fact, make user's Similarity Measure more accurate, it is to increase the quality of collaborative filtered recommendation, it is that video user provides the individualized video adapting to user's interest to recommend.

Description

A kind of introduce video popularity and the collaborative filtered recommendation method of user's interests change

Technical field

The invention belongs to Rich Media's personalized recommendation technical field, more specifically say, it relates to a kind of introduce video popularity and the collaborative filtered recommendation method of user's interests change.

Background technology

Along with the development of network and application thereof, network has welcome the epoch of " information explosion ", the demand of user has been can not meet by the search technique of representative of Google and Baidu, one is that user may can not find the information wanted by search engine, and two is that user possibly cannot reach suitable word the demand schedule of oneself on one's own initiative search engine operates to allow. Thus, commending system arises at the historic moment, and it can gather history behavior and the feedback information of user, finds the resource meeting user's interest according to these information, then for user makes personalized recommendation.

Recommended technology has been applied to multiple fields such as video traffic, ecommerce, personalized reading. Especially in Web TV (IPTV) field, no matter user is by digital television business or video website viewing video frequency program, for nowadays colourful video frequency program feel joyful while, choose the program oneself really liked and feel helpless for being difficult to from the program resource of vastness again. According to the CNNIC(China Internet Network Information Center) " the China Internet Status of development statistical report " of up-to-date issue display, end in by the end of December, 2012, Chinese netizen's scale reaches 5.64 hundred million, and Internet video user reaches 3.72 hundred million, relatively adds 4,653 ten thousand people last year. In the face of video traffic customer group that is so huge and fast growth, the video recommendations service providing personalized for it is significant, and contains considerable commercial value.

Collaborative filtering is current most widely used personalized recommendation technology, and some well-known commending systems such as Amazon, GroupLens and Douban have employed the method for collaborative filtering. This kind of method is the hypothesis based on a kind of general knowledge: if 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 that history behavior according to user calculates the similarity between different user, finds K the nearest-neighbors of target user, and then target user is generated recommendation list by viewing behavior according to this K user. In video system, the behavior of video frequency program is various by user, comprises that the display to program is marked, length, number of clicks etc. when watching. System is extracted these behaviors and they is converted to the scale-of-two association of user-video, and when when such as scoring, viewing, length and number of clicks are greater than setting threshold value, user-video is associated as " 1 ", is " 0 " when not reaching threshold value. System calculates the similarity between user according to the association of these scale-of-two, it may also be useful to many similarity calculating methods comprise space cosine similarity, Pearson correlation coefficients, Jie Kade similarity factor etc. The most important link of collaborative filtering is search K nearest-neighbors, and neighbor seaching follows user's similarity, but, current Similarity Measure exists some problems, one is that system treats all equal for all videos status, have ignored the popular degree of program, i.e. popularity. Because the meaning that the program of different popularity is produced behavior by user is different, such as two users have viewed certain quite high popular film of popularity by Web TV, this all have purchased xinhua dictionary like such as two students, almost can not illustrate that their interest is similar, on the contrary the common behavior of unexpected winner program more can be embodied similarity, so it is irrational that all programs are treated on an equal basis. Two is have ignored user's interest over time. Because different user watches the similarity degree that the behavior of the video frequency program behavior bigger than time span more can embody user in less time range, tradition similarity calculating method does not consider that user's interests change meets objective fact not.

Summary of the invention

It is an object of the invention to overcome the deficiencies in the prior art, design a kind of collaborative filtering video recommendation method introducing video popularity and user's interests change, the similarity between user is calculated according to video popularity weights and user's interest attenuation weight, obtain more realistic situation and user's similarity more accurately, recommend more accurately for target user makes.

For achieving the above object, the present invention introduces video popularity and the collaborative filtered recommendation method of user's interests change, it is characterised in that comprise the following steps:

(1) m user, is collected to the behavioral data of n video, comprise user's scoring, user length, user's viewing/number of clicks and behavior time of origin when watching, the user corresponding according to practical situation setting marks threshold value, user long threshold value and user's viewing/number of clicks threshold value when watching, when three data having at least one be more than or equal to corresponding setting threshold value, being the effective behavior of user, setting user-video is associated as " 1 "; Otherwise setting user-video is associated as " 0 "; Obtain " user-video " scale-of-two incidence matrix B (m, n), wherein contain the related information between m user and n video, user-video association b_uiRepresent that user u is to the interest value of video i;

(2), to matrix B (m, n) carrying out column scan and line scanning respectively, column scan obtains video i, and effective behavior user of 1��i��n gathers U_i, line scanning obtains user u, effective behavior video collection I of 1��u��m_u, note popular_i=| U_i| it is the popularity of video i, represents user's number that video i is produced effective behavior; Active_u=| I_u| it is the active degree of user u, represents that user u produces the video number of effective behavior; Wherein | U_i| with | I_u| represent U respectively_iAnd I_uMould;

(3), the setup algorithm cycle, off-line calculation video popularity weights and user's interest attenuation 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's interest attenuation 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 produces the video of effective behavior, t_usAnd t_vsRepresenting that video s is produced the time of the last effectively behavior by user u and user v respectively, ��, 0�ܦ���1 is interest attenuation influence coefficient, arranges according to practical situation;

(4), usage space cosine similarity calculates the similarity between target user 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| with | I_v| it is the active degree of user u and user v, represents that user u and user v produces the video number of effective behavior respectively;

(5), to the similarity obtained sorting from big to small, choose the nearest-neighbors set of front K user as target user u, be designated as S (u, K), wherein K sets according to practical situation;

(6), according to the viewing behavior prediction target user u of the nearest-neighbors set of target user u to the interest value not watching video j, and sort from big to small by interest value, as video recommendations list RecVedio (N), front N number of video is recommended target user, and wherein N sets according to practical situation.

Wherein, the decision method of the effective behavior of user in step (1) comprises the following steps:

1.1), judge whether user's scoring is more than or equal to user and marks threshold value, if it does, then the behavior effective behavior that is user, judge to terminate; Otherwise enter step 1.2);

1.2), when judging that user watches whether length is more than or equal to long threshold value when user watches, if it does, then the behavior effective behavior that is user, judges to terminate; Otherwise enter step 1.3);

1.3), judge whether user's viewing/number of clicks is more than or equal to user's viewing/number of clicks threshold value, if it does, then the behavior effective behavior that is user, judge to terminate; Otherwise the behavior is not the effective behavior of user, judges to terminate.

Wherein, the computation period in step (3) is one hour.

Wherein, interest attenuation influence coefficient ��=0.5 in step (3).

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

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

Wherein, user w is the user that video j produced effective behavior in the nearest-neighbors set of user u.

The goal of the invention of the present invention is achieved in that

The present invention introduces video popularity and the collaborative filtered recommendation method of user's interests change, carry out process by collection user behavior data and obtain user-video scale-of-two incidence matrix, obtain video popularity weights and user interest degree weight based on this matrix, and it is incorporated in user's Similarity Measure process; Then front K the neighbours maximum with target user's similarity are found, by the similarity size target of prediction user with neighbor user to the interest value not producing effective behavior video; Finally choose the maximum N number of video of interest value to form recommendation list for user and make personalized recommendation.

The present invention introduces video popularity and the collaborative filtered recommendation method of user's interests change, utilize the behavioral data that represent individual subscriber interest, in conjunction with video popularity and user's interests change, achieve the object providing individualized video to recommend for user at IPTV or video website system regions, meet network personalized development on the one hand, for client provides better service, service provider also can be helped on the other hand to attract and keep more clients here, it is to increase economic benefit.

Accompanying drawing explanation

Fig. 1 is a kind of embodiment schema that the present invention introduces the collaborative filtered recommendation method of video popularity and user's interests change;

Fig. 2 is the schematic flow sheet forming user-video scale-of-two incidence matrix in 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 the technician of this area understands the present invention better. Requiring particular attention is that, in the following description, when perhaps the detailed description of known function and design can desalinate the main contents of the present invention, these descriptions will be ignored here.

Embodiment

Fig. 1 is a kind of embodiment schema that the present invention introduces the collaborative filtered recommendation method of video popularity and user's interests change. As shown in Figure 1, the collaborative filtered recommendation method that the present invention introduces video popularity and user's interests change comprises the following steps:

Step S101: video systematic collection m user, to the behavioral data of n video, comprises user's scoring, user length, user's viewing/number of clicks and behavior time of origin when watching. Then these data are stored in the database in high in the clouds. The present embodiment being remembered, video i is marked as score by user u_ui, when video i is watched by user u, length is length_ui, the number of times that video i was watched or clicked video i by user u is designated as freq_ui��

Step S102: the user corresponding according to practical situation setting marks threshold value, user long threshold value and user's viewing/number of clicks threshold value when watching, when three data having at least one be more than or equal to corresponding setting threshold value, being the effective behavior of user, setting user-video is associated as " 1 "; Otherwise setting user-video is associated as " 0 "; Obtain user-video scale-of-two incidence matrix B (m, n), wherein contain the related information between m user and n video, user-video association b_uiRepresent that user u is to the interest value of video i.

Fig. 2 is the schematic flow sheet forming user-video scale-of-two incidence matrix in Fig. 1 according to user behavior data. As shown in Figure 2, the flow process that original user behavioral data is processed by video system in cloud database is as follows:

Step S201: user's scoring is judged. In the present embodiment, the points-scoring system of video system adopts the five-grade marking system, user can for viewing video make the scoring of 1��5 point, setting user's threshold value of marking is 3. Work as score_uiWhen >=3, then user u is effective to the behavior of video i, enters step S204; If user is not initiatively video scoring or score_ui< 3, then enter step S202.

Step S202: when user being watched, progress row judges. In the present embodiment, when setting user watches, long threshold value is 0.25h. Work as length_uiDuring >=0.25h, then user u is effective to the behavior of video i, enters step S204; Otherwise enter step S203.

Step S203: user's viewing/number of clicks is judged. In the present embodiment, setting user's viewing/number of clicks threshold value is 3. Work as freq_uiWhen >=3, then user u is effective to the behavior of video i, enters step S204; Otherwise enter step S205.

Step S204: setting user-video association b_ui=1, enter step S206.

Step S205: setting user-video association b_ui=0, enter step S206.

Step S206: according to all users to the user-video relating value of all videos, obtain user-video scale-of-two incidence matrix B (m, n), wherein contain the related information between m user and n video, user-video association b_uiRepresent that user u is to the interest value of video i.

In the present embodiment, it may also be useful to the user of 4 users and 6 film videos-video related information. Table 1 is scale-of-two incidence matrix B (4, F).

	Film A	Film B	Film C	Film D	Film E	Film F
							User 1	1	1	0	1	0	1
User 2	0	1	0	1	1	0
							User 3	0	0	1	1	1	1
User 4	1	1	0	1	1	0

Table 1

Effective behavior user of step S103: matrix B (m, n) is carried out column scan and line scanning respectively, and column scan obtains video i, 1��i��n gathers U_i, line scanning obtains user u, effective behavior video collection I of 1��u��m_u, note popular_i=| U_i| it is the popularity of video i, represents user's number that video i is produced effective behavior; Active_u=| I_u| it is the active degree of user u, represents that user u produces the video number of effective behavior; Wherein | U_i| with | I_u| represent U respectively_iAnd I_uMould.

In the present embodiment, matrix being carried out column scan, after running into " 1 ", the user of correspondence is joined these row video i, effective behavior user of A��i��F gathers U_iIn, it is respectively:

U_A={ user 1, user 4};

U_B={ user 1, and user 2, user 4};

U_C={ user 3};

U_D={ user 1, and user 2, and user 3, user 4};

U_E={ user 2, and user 3, user 4};

U_F={ user 1, user 3}.

After column scan completes, user is gathered U_iMould is asked to obtain | U_i|, | U_i| represent the popularity of film video i, it be 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, after running into " 1 ", the film video of correspondence is joined this row user u, effective behavior video collection I of 1��u��4_uIn, it is respectively:

I₁={ film A, film B, film D, film F};

I₂={ film B, film D, film E};

I₃={ film C, film D, film E, film F};

I₄={ film A, film B, film D, film E}.

After line scanning completes, to effective behavior video collection I_uMould is asked to obtain | I_u|, | I_u| represent the active degree of user u, it be respectively:

|I₁|=4; | I₂|=3; | I₃|=4; | I₄|=4��

Step S104: setup algorithm cycle, off-line calculation video popularity weights and user's interest attenuation weight, video popularity weight calculation formula is:

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

User's interest attenuation weight calculation formula:

${\mathrm{\&omega;}}_t(u,v,s)=\frac{1}{\log (1+\mathrm{\&alpha;}|t_{\mathrm{us}}-t_{\mathrm{vs}}+1\left|\right)},s\&Element;I_u\&cap;I_v$

Wherein, s represents user u and user v, and 1��v��m, v �� u all produces the video of effective behavior, t_usAnd t_vsRepresenting that video s is produced the time of effective behavior, �� by user u and user v respectively, 0�ܦ���1 is interest attenuation influence coefficient, arranges according to practical situation.

In the present embodiment, the computation period of setting is one hour, namely upgrades once every one hour, (when user uses peak period, can suitably shorten the cycle), system is according to user-video incidence matrix B (4, F) off-line calculation video popularity weights and user's interest attenuation weight. Interest attenuation influence coefficient ��=0.5 is set.

A. the popularity of video represents user's number that video i produces effective behavior, is nonzero term sum in the i-th row in matrix B (m, n), and video i is produced effective behavior by more many users, then video i is more popular. So calculating the popularity weights value that can obtain each video based on scale-of-two incidence matrix B (m, n), the video institute role that when this value has slackened user's Similarity Measure, popularity is higher.

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

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

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

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

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

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

${\mathrm{\&omega;}}_h\left(F\right)=\frac{1}{\log (1+|U_F\left|\right)}=\frac{1}{\log (1+2)}\&ap;2.10.$

B. user's interest attenuation weight, need to utilize system acquisition to user video produced the time information of effective behavior, if the effective behavior obtained by viewing number of times, or a certain video is had repeatedly effectively behavior, all get the last time producing effective behavior, in the present embodiment, this time is in units of the number of days of distance current time. Because user's interest can change in time, with calculate the time span watching same video between user method to weigh interests change to the impact of similarity.

Table 2 is for each user is to the last number of days that effectively behavior and current time are separated by of each film video. Wherein "/" represents and does not produce effective behavior.

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

Table 2

In the present embodiment, using user 2 as target user, for user 1, it is film B and film D that user 2 and user 1 all produce the video of effective behavior, according to user's interest attenuation weight calculation formula, obtain in user 2 for user's interest attenuation weighted value of user 1 be:

${\mathrm{\&omega;}}_t(\mathrm{2,1},B)=\frac{1}{\log (1+\mathrm{\&alpha;}|t_{2B}-t_{1B}+1\left|\right)}=\frac{1}{\log (1+0.5|730-3650+1\left|\right)}\&ap;0.32;$

${\mathrm{\&omega;}}_t(\mathrm{2,1},D)=\frac{1}{\log (1+\mathrm{\&alpha;}|t_{2D}-t_{1D}+1\left|\right)}=\frac{1}{\log (1+0.5|150-270+1\left|\right)}\&ap;0.56.$

The user interest attenuation weighted value of user 2 for other users can be obtained with reason:

${\mathrm{\&omega;}}_t(2,3,D)=\frac{1}{\log (1+\mathrm{\&alpha;}|t_{2D}-t_{3D}+1\left|\right)}=\frac{1}{\log (1+0.5|150-169+1\left|\right)}\&ap;1.00;$

${\mathrm{\&omega;}}_t(2,3,E)=\frac{1}{\log (1+\mathrm{\&alpha;}|t_{2E}-t_{3E}+1\left|\right)}=\frac{1}{\log (1+0.5|30-883+1\left|\right)}\&ap;0.38;$

${\mathrm{\&omega;}}_t(\mathrm{2,4},B)=\frac{1}{\log (1+\mathrm{\&alpha;}|t_{2B}-t_{4B}+1\left|\right)}=\frac{1}{\log (1+0.5|730-603+1\left|\right)}\&ap;0.55;$

${\mathrm{\&omega;}}_t(\mathrm{2,4},D)=\frac{1}{\log (1+\mathrm{\&alpha;}|t_{2D}-t_{4D}+1\left|\right)}=\frac{1}{\log (1+0.5|150-3200+1\left|\right)}\&ap;0.31;$

${\mathrm{\&omega;}}_t(2,4,E)=\frac{1}{\log (1+\mathrm{\&alpha;}|t_{2E}-t_{4E}+1\left|\right)}=\frac{1}{\log (1+0.5|30-9102+1\left|\right)}\&ap;0.27.$

Like this, the impact that similarity is produced by effective behavior that between different user, time span is little is bigger so that the user that same video produces effective behavior in more little time range is more similar, more meets objective fact.

Step S105: usage space cosine similarity calculates the similarity between target user u and other user v except user u:

${\mathrm{sim}}_{\mathrm{uv}}=\frac{\underset{s}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_h\left(s\right){\mathrm{\&omega;}}_t(u,v,s)}{\sqrt{\left|I_u\right|\left|I_v\right|}},s\&Element;I_u\&cap;I_v$

Wherein, | I_u| with | I_v| it is the active degree of user u and user v, represents that user u and user v produces the video number of effective behavior respectively.

In the present embodiment, by above step process, system database stores popularity weights value corresponding to each video and some user's interest attenuation weighted values. When needs do to target user recommend time, according to the active degree of the user upgraded the last in database, video popularity, video popularity weights and user's interest attenuation weight calculation target user (user 2) and the Similarity value of other users, obtain according to space cosine similarity formula:

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

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

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

$=\frac{1.66\&times;0.55+1.43\&times;0.31+1.66\&times;0.27}{\sqrt{3\&times;4}}\&ap;0.521$

Adopting the Similarity Measure introducing video popularity and user's interests change, the effective behavior produced for same user in unexpected winner video and little time span is relatively big to the contribution of similarity, more meets objective fact. Such as, if not considering video popularity and the impact of user interest degree weight, by incidence matrix it may be seen that the space cosine similarity value of user 2 and user 4 can be greater than the space cosine similarity value of user 2 and user 3. But above calculation result shows, sim₂₃It is greater than sim₂₄, this time mainly because of user 3 and user 2, film D and film E being produced effective behavior is very close, and their similarity is created effective lifting by user's interest attenuation weight, obtains more preferably user's analog result.

Step S106: the similarity obtained sorted from big to small, chooses the nearest-neighbors set of front K user as target user u, is designated as S (u, K); Wherein K sets according to practical situation.

Similarity size between the user 2 obtained and other users is carried out sequence and can obtain sim₂₃> sim₂₄> sim₂₁, the present embodiment sets K=3, therefore chooses the nearest-neighbors set of these 3 users as user 2: S (user 2,3)={ user 3, and user 4, user 1}.

Step S107: according to the viewing behavior prediction target user u of the nearest-neighbors set of target user u to the interest value not watching video j, and sort from big to small by interest value, as video recommendations list RecVedio (N), front N number of video is recommended target user, and wherein N sets according to practical situation.

The calculation formula specifically carrying out interest prediction is:

$p(u,j)=\underset{w}{\mathrm{\&Sigma;}}{\mathrm{sim}}_{\mathrm{uw}}{b}_{\mathrm{wj}},w\&Element;S(u,K)\&cap;U_j$

Wherein, user w is the user that video j produced effective behavior in the nearest-neighbors set of user u, b_wjRepresent that user w is to the interest value of video j, so all b in this calculation formula_wj=1. Therefore the final calculation formula of interest prediction is:

$p(u,j)=\underset{w}{\mathrm{\&Sigma;}}{\mathrm{sim}}_{\mathrm{uw}},w\&Element;S(u,K)\&cap;U_j$

In the present embodiment, according to the similarity size prediction user 2 of user 2 and its nearest-neighbors collection, it is not produced the interest level of effective behavior film. Scan matrix B (4, F), obtaining the video collection that user 2 do not produce effective behavior is: H_u={ film A, film C, film F}.

Then to each film j �� H_u, user u is to the interest value size of video j in prediction, is respectively:

P (user 2, film A)=sim₂₄+sim₂₁=0.521+0.385=0.906;

P (user 2, film C)=sim₂₃=0.595;

P (user 2, film F)=sim₂₃+sim₂₁=0.595+0.385=0.980.

The interest value calculated, by descending sort, obtains film F > film A > film C. In the present embodiment, N=2 is set, forms recommendation list: RecVedio (N)={ film F, film A}. This list is pushed to target user, thus forms recommendation.

Although above the embodiment of the present invention's explanation property being 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 change is in appended scope and the spirit and scope of the present invention determined, these changes are apparent, and all utilize the innovation and creation of present inventive concept all at the row of protection.

Claims (5)

1. introduce video popularity and the collaborative filtered recommendation method of user's interests change for one kind, it is characterised in that comprise the following steps:

(1) m user, is collected to the behavioral data of n video, comprise user's scoring, user length, user's viewing/number of clicks and behavior time of origin when watching, the user corresponding according to practical situation setting marks threshold value, user long threshold value and user's viewing/number of clicks threshold value when watching, when three data having at least one be more than or equal to corresponding setting threshold value, being the effective behavior of user, setting user-video is associated as " 1 "; Otherwise setting user-video is associated as " 0 "; Obtain user-video scale-of-two incidence matrix B (m, n), wherein contain the related information between m user and n video, user-video association b_uiRepresent that user u is to the interest value of video i;

(2), to matrix B (m, n) carrying out column scan and line scanning respectively, column scan obtains video i, and effective behavior user of 1��i��n gathers U_i, line scanning obtains user u, effective behavior video collection I of 1��u��m_u, note popular_i=| U_i| it is the popularity of video i, represents user's number that video i is produced effective behavior; Active_u=| I_u| it is the active degree of user u, represents that user u produces the video number of effective behavior;

(3), the setup algorithm cycle, off-line calculation video popularity weights and user's interest attenuation weight, video popularity weight calculation formula is:

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

User's interest attenuation weight calculation formula:

${\mathrm{\&omega;}}_t(u,v,s)=\frac{1}{log(1+\mathrm{\&alpha;}|t_{us}-t_{vs}+1\left|\right)},s\&Element;I_u\&cap;I_v$

Wherein, s represents user u and user v, and 1��v��m, v �� u all produces the video of effective behavior, t_usAnd t_vsRepresenting that video s is produced the time of the last effectively behavior by user u and user v respectively, �� is interest attenuation influence coefficient, arranges according to practical situation;

(4), usage space cosine similarity calculates the similarity between target user u and other users except user u:

${\mathrm{sim}}_{uv}=\frac{\underset{s}{\&Sigma;}{\mathrm{\&omega;}}_h\left(s\right){\mathrm{\&omega;}}_t(u,v,s)}{\sqrt{\left|I_u\right|\left|I_v\right|}},s\&Element;I_u\&cap;I_v$

Wherein, | I_u| with | I_v| it is the active degree of user u and user v;

(5), to the similarity obtained sorting from big to small, choose the nearest-neighbors set of front K user as target user u, be designated as S (u, K), wherein K sets according to practical situation;

(6), according to the viewing behavior prediction target user u of the nearest-neighbors set of target user u to the interest value not watching video j, and sort from big to small by interest value, front N number of video is recommended target user as video recommendations list RecVedio (N).

2. collaborative filtered recommendation method according to claim 1, it is characterised in that, the decision method of the described effective behavior of user in step (1) comprises the following steps:

1.1), judge whether user's scoring is more than or equal to user and marks threshold value, if it does, then the behavior effective behavior that is user, judge to terminate; Otherwise enter step 1.2);

1.2), when judging that user watches whether length is more than or equal to long threshold value when user watches, if it does, then the behavior effective behavior that is user, judges to terminate; Otherwise enter step 1.3);

1.3), judge whether user's viewing/number of clicks is more than or equal to user's viewing/number of clicks threshold value, if it does, then the behavior effective behavior that is user, judge to terminate; Otherwise the behavior is not the effective behavior of user, judges to terminate.

3. according to the arbitrary described collaborative filtered recommendation method of claim 1 and 2, it is characterised in that, the calculation formula of the interest value not watching video j is by the described target of prediction user u in step (6):

$p(u,j)=\underset{w}{\&Sigma;}{\mathrm{sim}}_{uw},w\&Element;S(u,K)\&cap;U_j$

Wherein, user w is the user that video j produced effective behavior in the nearest-neighbors set of user u.

4. collaborative filtered recommendation method according to claim 1, it is characterised in that, the described computation period in step (3) is one hour.

5. collaborative filtered recommendation method according to claim 1, it is characterised in that, described interest attenuation influence coefficient ��=0.5 in step (3).