CN108876536A - Collaborative filtering recommending method based on arest neighbors information - Google Patents

Abstract

The present invention relates to a collaborative filtering recommendation method based on nearest neighbor information, comprising the following steps: Step 1: Input the rating data set in the user-item matrix to determine the target user u; Step 2: Through the rating data and Pearson similarity, Calculate the availability between the target user and other users, and select k `users with higher availability to generate the nearest neighbor candidate set of the target user _u ; Step 3: Calculate each The trust degree between a user v and the target user u; step 4: filter out users with high availability but low trust, select K users with high trust as the nearest neighbor users, and generate the nearest neighbor set ; Step 5: Using the rating information of Nu, calculate the predicted ratings of target user _u for all unrated items; Step 6: Select items with high predicted ratings to generate a list of recommended items.

Description

Collaborative filtering recommendation method based on nearest neighbor information

technical field

The invention belongs to the technical field of recommendation based on collaborative filtering, and in particular relates to a collaborative filtering recommendation method based on nearest neighbor information.

Background technique

With the rapid development of the Internet today, the network has long been integrated into human daily life. At the same time, network information has also become complicated and rapidly increasing, which makes the Internet face problems such as "information overload" when sharing information. In the field of e-commerce, the phenomenon of "information overload" is quite serious. In this context, personalized recommendation technology has gradually developed.

Compared with traditional search methods, the personalized recommendation system provides users with unique services. It can collect users' historical behavior data, analyze users' interests and potential interests, provide users with products they are interested in, and reduce users' search for products. time and energy. On the other hand, the personalized recommendation system attracts more consumers for e-commerce websites, recommends products to consumers, increases website sales, and obtains more profits. To achieve a win-win situation for users and suppliers. The core technology in the personalized recommendation system is the personalized recommendation technology. There is often a scoring system in the personalized recommendation system, and the user's preference can be estimated by using the system's historical scoring records. In addition to the rating system, some platforms only have an evaluation system. Using existing technologies such as sentiment analysis, users' comments such as text and expressions can be converted into numerical ratings, which can then be converted into a rating system.

As mentioned above, personalized recommendations exist on major platforms such as Taobao, Amazon, JD.com, and Toutiao. However, as a user, it is not difficult to find that there are still some problems. For example, a user bought a laptop on an e-commerce platform. Recommend the laptop to this user for the first time. At the same time, it is not difficult to find that for some platforms that only have ratings or reviews, the reason why they can achieve high personalized recommendations is because they have a large number of users and data before starting personalized recommendations, while for some small and medium-sized For a company that does not have a large number of users, even if it has a large number of users, it is useless to have more historical user behavior records, resulting in the existing data being quite sparse, making it impossible to achieve good personalized recommendations.

In order to solve the problems of low accuracy of nearest neighbor selection and low recommendation accuracy in the case of sparse data, the present invention establishes a nearest neighbor optimization selection method on the basis of traditional collaborative filtering algorithm to provide better technical support for personalized recommendation.

Contents of the invention

In order to solve the problems of low accuracy of nearest neighbor selection and low recommendation accuracy under the condition of sparse data, the present invention provides a more accurate personalized recommendation method. On the basis of traditional collaborative filtering algorithm, a collaborative filtering based on nearest neighbor information is established. recommended method. To achieve the above object, the present invention takes the following technical solutions:

A collaborative filtering recommendation method based on nearest neighbor information, comprising the following steps:

Step 1: Input the scoring data set in the user-item matrix to determine the target user u;

Step 2: Calculate the availability between the target user and other users through the scoring data and Pearson similarity, and select k` users with high availability to generate the nearest neighbor candidate set N <sub>u of the target user u</sub> `;

Step 3: Calculate the trust degree tru(u,v) between each user v in the same time window θN u` and the target user _u :

a) According to whether user v’s preference for an item i in the same time window θ is close to or far from the target user u’s preference, it is considered that user u has consistent preferences and inconsistent preferences for user v in window θ, which are recorded as and

b) Since user v may evaluate multiple items for target user u in window θ, the value of consistent or inconsistent preference in window θ will be calculated as the sum of the user’s consistent or inconsistent preference on all items, denoted as g ^θ (u,v) and b ^θ (u,v)

c) Assign the same forgetting factor f to all ratings falling in the same time window θ, and gradually forget the user’s previous preferences: For the time window θ, assign the consistent preference forgetting factor f _g and the inconsistent preference forgetting factor f _b , respectively, consistent and The total value of inconsistent preferences is denoted as G ^θ (u,v) and B ^θ (u,v);

d) Finally, calculate the trust degree between user v and target user u according to the total value of user v’s consistent and inconsistent preference for user u. The higher the total value of consistent preference or the lower the total value of inconsistent preference, the higher the trust degree. high;

Step 4: Filter out users with high availability but low trust, select K users with high trust as nearest neighbor users, and generate nearest neighbor set N _u ;

Step 5: Using the rating information of Nu, calculate the predicted ratings of target user _u for all unrated items;

Step 6: Select items with high predicted scores to generate a list of recommended items.

The present invention has the following advantages due to the adoption of the above technical scheme:

(1) Most existing user similarity calculation mechanisms, such as Pearson similarity, calculate the similarity between a pair of users as a symmetric value, but in actual situations, the ability of two users to recommend each other is not the same. Based on the traditional similarity calculation method, the present invention considers the asymmetry of user similarity and recommendation availability, and ensures the recommendation ability of the nearest neighbor;

(2) Most current neighbor selection methods do not consider the consistency of users' preferences for different items, and ignore the dynamic changes of users' preferences over time. The present invention considers the user's preference consistency for different items, increases the time window and the forgetting factor, and further ensures that the preferences of the nearest neighbor and the target user are consistent in different time periods.

Description of drawings

Figure 1 is the process flow of the collaborative filtering recommendation method based on the nearest neighbor optimization selection method.

Detailed ways

The present invention establishes a collaborative filtering recommendation method based on the nearest neighbor information on the basis of the traditional collaborative filtering algorithm, which includes two key models: an availability calculation model and a dynamic trust calculation model.

To achieve the above object, the present invention takes the following technical solutions:

(1) Availability calculation model: The traditional similarity calculation method mainly relies on the common scoring item set, so when the number of common scoring items of two users is small, the deviation between the obtained similarity and the actual situation is large. At the same time, the traditional similarity calculation method considers that the recommendation ability between users is symmetric, but in the actual situation, the recommendation ability of two users to each other is not the same. In view of the above problems, the present invention considers the ratio of the number of common scoring items between two users to the number of scoring items of the target user, and combines the traditional similarity calculation method to propose a user availability model.

(2) Calculation model of dynamic trust degree: In the past traditional recommendation algorithm, the data mainly relied on in the method of calculating similarity is the value of common scoring items between users. One of the problems is the data sparsity problem. Therefore, in the case of less common scoring data, the calculated similarity cannot accurately represent the similarity between users, which negatively affects the subsequent recommendation results. In addition, there are many users with false or malicious ratings in the recommendation system, which makes the collaborative filtering recommendation system vulnerable to attacks. The untrue evaluation of the product or the evaluation content with other purposes make it mislead other users' perception of the product. . The invention proposes a trust degree model, which can improve the problem of data sparseness on the one hand, and on the other hand, by calculating the trust degree among users, select those trustworthy users and exclude those malicious users, thereby further improving the accuracy of recommendation.

In the collaborative filtering recommendation method based on the nearest neighbor optimization selection method of the present invention, in the process of calculating the dynamic trust model, the traditional trust model is improved according to the actual situation, the time window and the forgetting factor are proposed, and the user is calculated from the perspective of the target user. The trust value of , and consider the time factor to capture the user's preference changes over time, and select neighbors with higher trust. The present invention will greatly improve the accuracy of the system to make recommendations to target users. Figure 1 shows the flow of the present invention. The specific implementation steps are as follows:

(1) Input the rating data set in the user-item matrix to determine the target user u.

(2) Calculate the usability between the target user and other users by scoring data, the formula is as follows:

Among them, ava(u, v) represents the availability of user v to user u, I _v and I _u represent the scoring items of users v and u respectively, and sim(u, v) represents the similarity between users v and u. It can be seen that when That is, when all rating items of user v have been evaluated by user u, user v has no recommendation ability for user u, so ava(u, v)=0 at this time. When sim(u, v)<0, it means that the similarity between users u and v is too low, therefore, ava(u, v)=0 also exists at this time. |I _u ∩I _v | represents the number of common scoring items of user u and v, when user v has unrated items of user u, by considering the ratio of the number of common scoring items to the number of scoring items of user u and combining the traditional The similarity calculation method offsets the excessive reliance of the traditional similarity calculation method on common scoring items. When the sim(u,v) is the same, the larger |I _u ∩I _v |/|I _u |, indicating that when calculating the traditional similarity The more rating data used, and thus the higher the availability of user v to user u.

(3) Select K' users with high availability, and generate the nearest neighbor candidate set N'(u) of the target user u.

K'=[ε×K]

Where ε is the nearest neighbor selection coefficient, ε∈{ε∈R|ε≥1}, [x] is the rounding function, that is, the largest integer that does not exceed the real number x. The users in the N'(u) set generally have higher availability, that is, they have higher similarity with target users and have higher recommendation ability.

(4) Calculate the trust degree tru(u,w) between all users w in N′(u) and the target user u.

a) The present invention gradually forgets the user's previous preferences by introducing a time window and a forgetting factor. Given a particular score provided at time _tc , the index of the window into which that score falls is denoted θ,

θ=[(t _c -t _s )/t _w ]+1

where t _s and t _w represent the training start time and window length, respectively. In the method proposed by the present invention, all scores falling in the same time window θ will be assigned the same forgetting factor for processing.

b) According to whether v’s preference for a certain item is close to or far from the target user u’s preference, the present invention considers that user v has “good behavior” (that is, consistent preference) or “bad behavior” (that is, inconsistent preference) for user u , respectively recorded as g _i (uv) and _bi (uv), each of which can be quantized as a continuous value in the range of (0, 1). The present invention can quantify the scoring behavior of each user v in the window θ (denoted as or

where R _max and R _min represent the maximum and minimum rating values in the recommendation system, respectively, Indicates the rating value of user v on item i in time window θ, Indicates the rating value of target user u on item i in time window θ.

c) Since user v may rate multiple items for target user u in window θ, the total value of good or bad behavior in window θ will be calculated as the sum of the user’s good or bad behavior values on all items.

d) In order to allow the more recent behavior to obtain a higher weight in calculating the user's credibility, the present invention uses the forgetting factor to gradually forget the user's previous behavior. Specifically, for a time window θ, the total value of "good behavior" and "bad behavior" (denoted as G ^θ (u,v) and B ^θ (u,v)) can be calculated as:

G ^θ (u,v)=G ^θ-1 (u,v)×f _g +g ^θ (u,v)

B ^θ (u,v)=B ^θ-1 (u,v)×f _b +b ^θ (u,v)

Among them, f _g and f _b are forgetting factors representing "good behavior" and "bad behavior", and the value range is (0,1).

e) Finally, the trust degree tru(u, v) of target user u to user v is calculated as follows, it can be seen that the better the behavior of user v, the higher the trust value the user will obtain.

(5) Filter out users with high availability but low trust, select K users with high trust as the nearest neighbor users, and generate the nearest neighbor set N(u), thus obtaining the availability and trust The users with a higher value are the nearest neighbors of the target user, and these users have similar preferences with the user, and the similarity persistence is relatively high.

(6) Using the rating information of the nearest neighbor user set, calculate the predicted ratings of the target user u for all unrated items (for example, the predicted rating P _ui for the target item i). The following calculation method is mainly used.

(7) The higher the score data in the predicted score list, it means that the present invention believes that the target user has a higher preference for the item corresponding to the score. Therefore, items with high predicted scores are selected to generate a list of recommended items.

Claims (1)

1. a kind of collaborative filtering recommending method based on arest neighbors information, includes the following steps：

Step 1：The score data collection in user-project matrix is inputted, determines target user u；

Step 2：By score data and Pearson's similarity, the availability between target user and other users is counted It calculates, and selects the higher k` user of availability, generate the nearest-neighbors Candidate Set N of target user u<sub>u</sub>`；

Step 3：Calculate same time window θ N_uEach user v in the ` and degree of belief tru (u, v) between target user u：

A) according to user v in same time window θ to the preference of some project i whether close to or away from target user u preference, Think that user u produces consistent preference and inconsistent preference to user v in window θ, is denoted as respectivelyWith

B) since user v may be the multiple projects of target user u evaluation in window θ, consistent or inconsistent preference in window θ Value will be calculated as that the user on all items is consistent or the summation of inconsistent preference, be denoted as g respectively^θ(u, v) and b^θ(u,v)

C) identical forgetting factor f is distributed to all scorings fallen in same time window θ, it is pervious partially gradually forgets user It is good：For time window θ, consistent preference forgetting factor f is assigned respectively_gWith inconsistent preference forgetting factor f_b, consistent and inconsistent The total value of preference is denoted as G respectively^θ(u, v) and B^θ(u,v)；

D) finally, according to user v to user u unanimously and the height of inconsistent preference total value, calculate user v and target user u it Between degree of belief, consistency preference total value is higher or inconsistent preference total value is lower, and degree of belief is higher；

Step 4：It filters out that availability is higher but the lower user of degree of belief, the higher K user of degree of belief is chosen, as most Nearly neighbor user, and generate nearest-neighbors collection N_u；

Step 5：Utilize N_uScore information, calculate target user u and score the predictions of all non-scoring items；

Step 6：Choose the project generation recommended project list that prediction is scored high.