CN106991199A - The commending system score in predicting of probability is inclined to recommending method based on user behavior - Google Patents

Abstract

The invention provides a kind of commending system score in predicting and recommendation method that probability is inclined to based on user behavior, score calculation is carried out for the relation between scoring the article in current commending system, the label of article and the tendentiousness of user, and the recommendation service of similar article is provided when targeted customer searches for article from current commending system to the targeted customer, it is characterised in that it includes following steps：Step one, item database is built；Step 2, builds user's scoring storehouse；Step 3, calculates user u_rTo label T_qTendentiousness probabilistic forecasting scoring P (u_r, T_q)；Step 4, calculates user u_rTo article I_iTendentiousness scoring R (u_r, I_i)；Step 5, in targeted customer u^*Article I is searched for from current commending system_iWhen, calculate article I_iWith article I_jSimilarity S (I_i, I_j)；Step 6, calculates targeted customer u^*To article I_iPrediction score value r (u_r, I_i)；Step 7, according to prediction score value r (u_r, I_i) neighbours' article is ranked up, the neighbours' article for coming predetermined figure is recommended into targeted customer u^*。

Description

The commending system score in predicting of probability is inclined to recommending method based on user behavior

Technical field

The present invention relates to a kind of recommendation method, and in particular to a kind of commending system scoring that probability is inclined to based on user behavior Prediction is with recommending method.

Background technology

Commending system can be recommended to feel based on the similitude between article as a content of personalized service to user Interest or suitable film, the commodity and other items of its own so that user, which can obtain rapidly meeting, to be liked and select tendency Item Information.Meanwhile, commending system can save system resource and the band that user is consumed during article is largely browsed The costs such as width so that commending system is trusted by a large number of users, favors and used.

In the prior art, commending system mainly uses collaborative filtering recommending method, and this method utilizes the user having similar tastes and interests The common hobby of colony is recommended article to targeted customer.

But, above-mentioned collaborative filtering recommending method deposits problem both ways.On the one hand it is that conventional recommendation method is relied solely on Similarity between user chooses the neighbours of targeted customer, causes to recommend the diversity for the article that precision is relatively low, recommend poor； On the other hand it is that conventional recommendation method depends on user unduly to the scoring of article to be predicted, it is impossible to avoid commenting for user completely Divide prejudice and malice score data, cause the article and user's article interested recommended to have deviation.Accordingly, it would be desirable to which one kind is more Effectively and with appropriate multifarious commending system, convenient and efficient recommendation service is provided the user.

The content of the invention

The present invention is carried out to solve the above problems, it is therefore intended that provide a kind of based on user behavior tendency probability Commending system score in predicting with recommend method.

The invention provides a kind of commending system score in predicting for being inclined to probability based on user behavior with recommending method, it is used for Relation between scoring the article in current commending system, the label of article and the tendentiousness of user carries out score calculation, and The recommendation service of similar article, its feature are provided when targeted customer searches for article from current commending system to the targeted customer It is, comprises the following steps：

Step one, item database is set up, the item database includes N number of article I_iWith for describing article I_iThe L of property Individual label T_q, article I_iSet be set toLabel T_qSet be set to TG={ T_q, q= 1 ..., L }, each article I_iL (I_i) individual label t_pSet be set to T (I_i)={ t_p|t_p∈ TG, the t as p ≠ q_p≠t_q, its Middle p, q=1 ..., L (I_i), each article I_iWith each label T_qRelation be δ (t_p, T_q)；

Step 2, sets up user's scoring storehouse, and user scoring storehouse includes M user u_rWith each user u_rTo K (u_r) it is individual Evaluate article I_iScoring, user u_rSet be set to UR={ u_r, r=1 ..., M }, evaluate article I_iCollection be combined into I (u_r) ={ I_i|I_i∈ IM, i=1 ..., K (u_r), score as r ' (u_r∈ UR, I_i∈IM)；

Step 3, user u is calculated according to pre-defined rule_rTo label T_qTendentiousness probabilistic forecasting scoring P (u_r, T_q)；

Step 4, user u is calculated according to following formula_rTo article I_iTendentiousness scoring R (u_r, I_i)：

Step 5, in targeted customer u^*Article I is searched for from current commending system_iWhen, article I is calculated according to following formula_iWith thing Product I_jSimilarity S (I_i, I_j)：

In above formula, article I_iDegree be d (I_i), targeted customer u^*Degree be d (u_r), a (u_r, I_i) represent targeted customer u^*It is It is no to article I_iScoring was carried out, as targeted customer u^*To article I_iWhen commenting undue, a (u_r, I_i)=1, as targeted customer u^*It is not right Article I_iWhen commenting undue, a (u_r, I_i)=0；

Step 6, calculates targeted customer u^*To article I_iPrediction score value, article I_iWith article I_jKnown average score It is r respectively_i' and r_j', user u_rTo article I_jScore value be r ' (u_r, I_j), setting article I_iNeighbours' article number be n, The prediction score value for calculating each neighbours' article according to following formula is r (u_r, I_i)：

Step 7, according to prediction score value r (u_r, I_i) neighbours' article is ranked up from high to low, and pre-determined bit will be come Several neighbours' articles recommend targeted customer u^*。

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior that the present invention is provided, can also have There is such feature, wherein, in step one, article I_i∈IM。

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior that the present invention is provided, can also have There is such feature, wherein, L (I_i) individual label t_pBetween different, and L (I_i)≤L。

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior that the present invention is provided, can also have There is such feature, wherein, in step one,

Work as t_p=T_qWhen, δ (t_p, T_q)=1,

Work as t_p≠T_qWhen, δ (t_p, T_q)=0.

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior that the present invention is provided, can also have There is such feature, including：In step 2, r ' (u_r∈ UR, I_i∈ IM) value be 1~5, as user u_rTo article I_iDo not comment Valency is out-of-date, r ' (u_r∈ UR, I_i∈ IM)=0.

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior that the present invention is provided, can also have There is such feature, wherein, in step 3, the formula of pre-defined rule is as follows：

As user u_rEvaluated article I_iWhen,

As user u_rArticle I was not evaluated_iWhen, P (u_r, T_q)=0.

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior that the present invention is provided, can also have There is such feature, wherein, in step 5 and step 6, article I_j∈IM。

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior that the present invention is provided, can also have There is such feature, wherein, in step 6, r ' (u_r, I_j) value be 1~5, as user u_rTo article I_jIt is out-of-date not evaluate, r ' (u_r, I_j)=0.

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior that the present invention is provided, can also have There is such feature, wherein, in step 6, n ∈ { 10,20,30,40 }.

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior that the present invention is provided, can also have There is such feature, wherein, in step 7, predetermined figure is 5~10.

The effect of invention and effect

According to the commending system score in predicting involved in the present invention based on user behavior tendency probability with recommending method, by Relation between the scoring of the article in commending system, label and the tendentiousness of user carries out score calculation, and is calculating mesh Mark user to before the prediction score value of article to article I_iWith article I_jSimilarity Measure is carried out so that commending system is to mesh Mark the article more diversification that user recommends.In addition, because the commending system score in predicting of probability should be inclined to based on user behavior User's scoring is not depended on unduly with recommendation method so that the article that commending system is recommended to user meets the interest of user, so that Improve the accuracy of recommendation.

Brief description of the drawings

Fig. 1 is the commending system score in predicting and recommendation method for being inclined to probability in embodiments of the invention based on user behavior Flow chart.

Embodiment

In order that the technological means realized of the present invention is easy to understand with effect, with reference to embodiments and accompanying drawing to Invention is specifically addressed.

<Embodiment>

Fig. 1 is the commending system score in predicting and recommendation method for being inclined to probability in embodiments of the invention based on user behavior Flow chart.

As shown in figure 1, being inclined to the commending system score in predicting of probability with recommending method to include following step based on user behavior Suddenly：

Step one, item database is set up, the item database includes N number of article I_iWith for describing article I_iThe L of property Individual label T_q, article I_iSet be set toLabel T_qSet be set to TG={ T_q, q= 1 ..., L }, each article I_iL (I_i) individual mutually different label t_pSet be set to T (I_i)={ t_p|t_p∈ TG, as p ≠ q When t_p≠t_q, wherein p, q=1 ..., L (I_i), each article I_iWith each label T_qRelation be δ (t_p, T_q), work as t_p=T_q When, δ (t_p, T_q)=1, works as t_p≠T_qWhen, δ (t_p, T_q)=0, i.e.,

In the present embodiment, the item database is the item database related to film, including 6 article I_iIt is (i.e. electric Shadow 1, film 2, film 3, film 4, film 5, film 6), and for describing article I_i6 label T_q(i.e. love, terrified, acute Feelings, science fiction, history, action).Wherein, each article I_iThere are multiple mutually different label t_p, each article I_iWith each label T_qRelation be δ (t_p, T_q)。

6 film I_iRespectively with 6 label T_qRelation δ (t_p∈T(I_i∈ IM), T_q∈ TG) as shown in table 1.For example, electric Shadow 1 has 3 label t_p(i.e. terrified, science fiction, action), so, film 1 and three label T_qThe pass of (i.e. terrified, science fiction, action) System is 1, film 1 and other 3 label T_qThe relation of (i.e. love, the story of a play or opera, history) is 0.

Table 1

Step 2, sets up user's scoring storehouse, and user scoring storehouse is related to film user's scoring storehouse, including M is used Family u_rWith each user u_rTo K (u_r) individual evaluate article I_iScoring, user u_rSet be set to UR={ u_r, r=1 ..., M }, Article I is evaluated_iCollection be combined into I (u_r)={ I_i|I_i∈ IM, i=1 ..., K (u_r), score as r ' (u_r∈ UR, I_i∈ IM), should r′(u_r∈ UR, I_i∈ IM) value be 1~5, as user u_rTo article I_iIt is out-of-date not evaluate, r ' (u_r∈ UR, I_i∈ IM)=0.

In the present embodiment, user u_rTo each film I_iScore data r ' (u_r∈ UR, I_i∈ IM) as shown in table 2.With Family scoring storehouse includes 6 user u_r(u1, u2, u3, u4, u5, targeted customer u^*), and each user u_rElectricity has been evaluated to multiple Shadow I_iScoring r ' (u_r∈ UR, I_i∈ IM), such as scorings of the user u3 to film 2 is 5, and scorings of the user u3 to film 4 is 0.

Table 2

Step 3, user u is calculated according to pre-defined rule_rTo label T_qTendentiousness probabilistic forecasting scoring P (u_r, T_q), make a reservation for The formula of rule is as follows：As user u_rEvaluated article I_iWhen,As user u_r Article I was not evaluated_iWhen, P (u_r, T_q)=0, i.e.,

In the present embodiment, it can be calculated by above formula and obtain user u_rTo each label T_qTendentiousness probability P (u_r∈ UR, T_q∈ TG), as shown in table 3, such as user u3, which is evaluated, to flash back past events 2 and film 3, the label of film 2 for the story of a play or opera, history and action with And the label of film 3 is love, science fiction and action, user u3 did not evaluated the film for including label 2 (terror), so, user U3 is that the tendentiousness probability of films of the user u3 to belonging to love class is 1/6 to label 1, and user u3 is u3 couples of user to label 2 The tendentiousness probability for belonging to the film of terrified class is 0.

Table 3

Step 4, user u is calculated according to following formula_rTo article I_iTendentiousness scoring R (u_r, I_i)：

In the present embodiment, it can be calculated by above formula and obtain user u_rTo film I_iTendentiousness scoring R (u_r∈ UR, I_i∈ IM), as shown in table 4, for example understand that tendentiousness scorings of the user u3 to film 2 is 0.66 by calculating.

Table 4

Step 5, in targeted customer u^*Article I is searched for from current commending system_iWhen, article I is calculated according to following formula_iWith thing Product I_jSimilarity S (I_i, I_j)：

In above formula, I_j∈ IM, article I_iDegree be d (I_i), targeted customer u^*Degree be d (u_r), a (u_r, I_i) represent target User u^*Whether to article I_iScoring was carried out, as targeted customer u^*To article I_iWhen commenting undue, a (u_r, I_i)=1, when target is used Family u^*Not to article I_iWhen commenting undue, a (u_r, I_i)=0.

In the present embodiment, in targeted customer u^*Film I is searched for from current commending system_iWhen, electricity can be calculated by above formula Shadow I_iWith film I_jSimilitude S (I_i∈ IM, I_j∈ IM), as shown in table 5.The similitude of such as film 1 and film 1 is 1, electricity The similitude of shadow 1 and film 2 is 0.008333.

Table 5

Step 6, calculates targeted customer u^*To article I_iPrediction score value, article I_iWith article I_jKnown average score It is r respectively_i' and r_j', user u_rTo article I_jScore value be r ' (u_r, I_j), r ' (u_r, I_j) value be 1~5, as user u_rIt is right Article I_jIt is out-of-date not evaluate, r ' (u_r, I_j)=0, setting article I_iNeighbours' article number be n, n ∈ { 10,20,30,40 }, The prediction score value for calculating each neighbours' article according to following formula is r (u_r, I_i),

In the present embodiment, neighbours' article number is set as 10, and targeted customer u is calculated by above formula^*To film I_iPrediction Score value r (u^*, I_i(i=2,3,4)), as shown in table 6, such as targeted customer u^*Calculated value to film 2 is 2.415, through four houses Five enter after must predict score value be 2, wherein, targeted customer u^*Evaluation is flashed back past events 1 and film 6, so targeted customer u^*To film 1 Prediction score value with film 6 is that existing score value is respectively 5 and 3.

Table 6

Step 7, according to prediction score value r (u_r, I_i) neighbours' article is ranked up from high to low, and pre-determined bit will be come Several neighbours' articles recommend targeted customer u^*, predetermined figure is 5~10.

In the present embodiment, the prediction score value of film 2, film 3, film 4 and film 5 in table 6, which sorts, is respectively Film 4, film 5, film 3, film 2.Due to recommending predetermined figure to be 5, and digit is currently recommended to be less than 5, therefore by film 4th, film 5, film 3 and film 2 recommend targeted customer u in order^*。

The effect of embodiment and effect

Commending system score in predicting and recommendation method that probability is inclined to based on user behavior according to involved by the present embodiment, Relation between due to scoring the article in commending system, label and the tendentiousness of user carries out score calculation, and is calculating Targeted customer to before the prediction score value of article to article I_iWith article I_jCarried out Similarity Measure so that commending system to The article more diversification that targeted customer recommends.In addition, because the commending system scoring that probability should be inclined to based on user behavior is pre- Survey and do not depend on user's scoring unduly with recommendation method so that the article that commending system is recommended to user meets the interest of user, from And improve the accuracy of recommendation.

Claims (10)

1. a kind of commending system score in predicting and proposed algorithm that probability is inclined to based on user behavior, for current commending system In article, the label of article and the scoring of the tendentiousness of user between relation carry out score calculation, and in targeted customer from institute State the recommendation service that similar article is provided when article is searched in current commending system to the targeted customer, it is characterised in that including Following steps：

Step one, item database is set up, the item database includes N number of article I_iWith for describing the article I_iProperty The L label T of matter_q, the article I_iSet be set toThe label T_qSet set For TG={ T_q, q=1 ..., L }, each article I_iL (I_i) individual label t_pSet be set to T (I_i)={ t_p|t_p∈ TG, The t as p ≠ q_p≠t_q, wherein p, q=1 ..., L (I_i), each article I_iWith each label T_qRelation be δ (t_p, T_q)；

Step 2, sets up user's scoring storehouse, and user scoring storehouse includes the M user u_rWith each user u_rTo K (u_r) It is individual to have evaluated article I_iScoring, the user u_rSet be set to UR={ u_r, r=1 ..., M }, it is described to have evaluated article I_i Collection be combined into I (u_r)={ I_i|I_i∈ IM, i=1 ..., K (u_r), the scoring is r ' (u_r∈ UR, I_i∈IM)；

Step 3, the user u is calculated according to pre-defined rule_rTo the label T_qTendentiousness probabilistic forecasting scoring P (u_r, T_q)；

Step 4, the user u is calculated according to following formula_rTo the article I_iTendentiousness scoring R (u_r, I_i)：

$R(u_r,I_i)=\underset{T_q\&Element;TG}{\&Sigma;}\underset{t_p\&Element;T\left(I_i\right)}{\&Sigma;}\mathrm{\&delta;}(t_p,T_q)P(u_r,T_q);$

Step 5, in the targeted customer u^*The article I is searched for from the current commending system_iWhen, institute is calculated according to following formula State article I_iWith article I_jSimilarity S (I_i, I_j)：

$S(I_i,I_j)=\frac{1}{d\left(I_i\right)}{\mathrm{\&Sigma;}}_{r=1}^M\frac{a(u_r,I_i)a(u_r,I_j)(1-|R\left(u_r,I_i\right)-R\left(u_r,I_j\right)\left|\right)}{d\left(u_r\right)}$

In above formula, the article I_iDegree be d (I_i), the targeted customer u^*Degree be d (u_r), a (u_r, I_i) represent the target User u^*Whether to the article I_iScoring was carried out, as the targeted customer u^*To the article I_iWhen commenting undue, a (u_r, I_i) =1, as the targeted customer u^*Not to the article I_iWhen commenting undue, a (u_r,I_i)=0；

Step 6, calculates the targeted customer u^*To the article I_iPrediction score value, the article I_iWith the article I_j's Known average score is r respectively_i' and r_j', the user u_rTo the article I_jScore value be r ' (u_r, I_j), set the thing Product I_iThe number of neighbours' article be n, the prediction score value for calculating each neighbours' article according to following formula is r (u_r, I_i)：

$r(u_r,I_i)={r_i}^{\&prime;}+\frac{{\mathrm{\&Sigma;}}_{I_j\&Element;IM}\&lsqb;r^{\&prime;}(u_r,I_j)-{r_j}^{\&prime;}\&rsqb;}{{\mathrm{\&Sigma;}}_{I_j\&Element;IM}S(I_j,I_i)};$

Step 7, according to the prediction score value r (u_r, I_i) neighbours' article is ranked up from high to low, and will come pre- Neighbours' article of positioning number recommends the targeted customer u^*。

2. the commending system score in predicting and recommendation method according to claim 1 that probability is inclined to based on user behavior, its It is characterised by：

Wherein, in the step one, the article I_i∈IM。

3. the commending system score in predicting and recommendation method according to claim 1 that probability is inclined to based on user behavior, its It is characterised by：

Wherein, the L (I_i) the individual label t_pBetween it is different,

L (the I_i)≤L。

4. the commending system score in predicting and recommendation method according to claim 1 that probability is inclined to based on user behavior, its It is characterised by：

Wherein, in the step one, t is worked as_p=T_qWhen, the δ (t_p, T_q)=1,

Work as t_p≠T_qWhen, the δ (t_p, T_q)=0.

5. the commending system score in predicting and recommendation method according to claim 1 that probability is inclined to based on user behavior, its It is characterised by：

Wherein, in the step 2, the r ' (u_r∈ UR, I_i∈ IM) value be 1~5, as the user u_rTo the article I_iIt is out-of-date not evaluate, the r ' (u_r∈ UR, I_i∈ IM)=0.

6. the commending system score in predicting and recommendation method according to claim 1 that probability is inclined to based on user behavior, its It is characterised by：

Wherein, in the step 3, the formula of the pre-defined rule is as follows：

As the user u_rEvaluated the article I_iWhen, it is described

As the user u_rThe article I was not evaluated_iWhen, the P (u_r, T_q)=0.

7. the commending system score in predicting and recommendation method according to claim 1 that probability is inclined to based on user behavior, its It is characterised by：

Wherein, in the step 5 and the step 6, the article I_j∈IM。

8. the commending system score in predicting and recommendation method according to claim 1 that probability is inclined to based on user behavior, its It is characterised by：

Wherein, in the step 6, the r ' (u_r, I_j) value be 1~5, as the user u_rTo the article I_jDo not evaluate It is out-of-date, the r ' (u_r, I_j)=0.

9. the commending system score in predicting and recommendation method according to claim 1 that probability is inclined to based on user behavior, its It is characterised by：

Wherein, in the step 6, the n ∈ [10,20,30,40 }.

10. the commending system score in predicting and recommendation method according to claim 1 that probability is inclined to based on user behavior, its It is characterised by：

Wherein, in the step 7, the predetermined figure is 5~10.