CN103309972A - Recommend method and system based on link prediction - Google Patents

Abstract

The invention discloses a recommend method and system based on link prediction, in particular a recommend method based on link prediction. The recommend method comprises the steps as follows: traversing all users in a current network to obtain the historical behavior data of all the users; creating a user-article bigraph according to the historical behavior data of all the users; predicting the preference degree of the users on unselected articles by a link prediction method; and screening the prediction results of all the users to generate recommended articles for all the users.

Description

Recommendation method and system based on link prediction

Technical Field

The invention relates to the technical field of network recommendation, in particular to a recommendation method and a recommendation system based on link prediction.

Background

The vigorous development of the Web2.0 technology and interactive social media has improved people's participation, making people from past information retrievers information providers. However, any one of the technologies is a double-edged sword, and although the free interaction between the user and the internet makes the user feel the existence and brings convenience to people, the large amount of information generated by frequent interaction also brings people into the era of information explosion, so that people are difficult to find valuable contents from the colorful network world, and the large amount of fresh and known information becomes 'dark information' in the network.

The recommendation system is considered to be a very potential information filtering technology in the 21 st century to help solve the problem of internet information overload. In 1997, Resnick and Varian gave a definition of the recommendation system: the intelligent system provides commodity information and suggestions to a customer by using an electronic commerce website, helps the user decide what product should be purchased, and simulates salesmen to help the customer complete a purchasing process. In recent years, recommendation systems have received high attention from both academic and industrial fields, and have been widely applied to many fields and succeeded, such as movie recommendation, music recommendation, friend recommendation, book recommendation, travel recommendation, restaurant recommendation, and the like.

Different from a traditional information filtering technology search engine, the recommendation system does not need a user to provide searched keywords, and automatically completes the processes of user interest and hobby mining and content recommendation by analyzing the historical behaviors of the user. Collaborative filtering is currently the most widely used recommendation system. As the name suggests, collaborative filtering means that users can collaborate with each other, and through continuous interaction with websites, the recommendation list of the users can continuously filter out contents which are not interesting to the users, so that the needs of the users are more and more met. At the core of the collaborative filtering system, a collaborative filtering algorithm designed based on user behavior data assumes that a user will be interested in items liked by users with similar interests and preferences, and will also tend to like items with similar past liked items. The advantages of easy deployment, interpretability, better recommendation precision and the like of the collaborative filtering algorithm are favored by the academic and industrial circles, so that the collaborative filtering algorithm can stand out from numerous recommendation algorithms. However, as the scale of users and goods is increased, the user behavior data is more prominent than the sparsity of a large-scale system, and the collaborative filtering algorithm faces more serious challenges of data sparsity and computational complexity. In addition, the users selecting fewer articles and the articles selected less times cannot get recommendations or are recommended because proper similarity neighbors cannot be found, and on the contrary, active users and popular articles can always be compared with other users or articles in similarity, so that more recommendations and recommended opportunities are obtained, which indicates that the collaborative filtering algorithm is prone to recommend popular articles, and the recommendation result of the collaborative filtering algorithm lacks diversity.

In the traditional collaborative filtering method, under the condition of sparse data, similar neighbors cannot be found and inaccurate recommendations are given, so that a recommendation method and a recommendation system based on link prediction are necessary to be provided so as to effectively improve the accuracy and diversity of recommendations of a classical link prediction method.

Disclosure of Invention

The technical problem to be solved by the invention is to provide a recommendation method and a recommendation system based on link prediction, which expand the search range of neighbors into the whole system, thereby effectively avoiding the problem of lack of neighbors and improving the precision and robustness of recommendation.

To this end, according to one aspect of the present invention, there is provided a link prediction-based recommendation method including: traversing all users in the current network to obtain historical behavior data of all the users; establishing a user-item bipartite graph according to the historical behavior data of all the users; predicting the preference degree of the user for the unselected articles by using a link prediction method according to the user-article bipartite graph; and screening the prediction result of each user to generate a recommended item for each user.

In the above recommendation method, the creating a user-item bipartite graph according to the historical behavior data of all the users includes: obtaining a relation vector describing past preference of the user to articles according to the historical behavior data of all the users; constructing a user-item relationship matrix from the relationship vectors; and obtaining the user-item bipartite graph according to the user-item relation matrix.

In the above recommendation method, the predicting, by using the link prediction method, the user's preference for the unselected item includes: respectively calculating the sum of all path weights with any length between two nodes in the user-article bipartite graph according to the user-article bipartite graph; and weighting the sum of all path weights with any length between the two nodes by utilizing a decreasing weight factor to obtain a predicted value of the preference degree of the user on the unselected goods.

In the above recommendation method, the filtering the prediction result of each user to generate the recommended item for each user includes: sorting the unselected items according to the predicted value from high to low; and recommending one or more items ranked most front to the user as candidate items.

In the recommendation method, the creating of the user-item bipartite graph also takes into account the liveness of the user and the popularity of the items.

In the above recommendation method, the weight of the user-item preference degree side is represented by a complex number, and the weight of the user-user, item-item similarity degree side is represented by a real number.

According to another aspect of the present invention, there is provided a recommendation system based on link prediction, including: the acquisition module is used for traversing all users in the current network to acquire historical behavior data of all the users; the establishing module is used for establishing a user-article bipartite graph according to the historical behavior data of all the users; the link prediction module is used for predicting the preference degree of the user on the unselected articles by using a link prediction method according to the user-article bipartite graph; and the recommending module is used for screening the prediction result of each user to generate recommended articles for each user.

In the above recommendation system, the establishing module includes: the relation vector generation module is used for obtaining a relation vector which describes the past preference of the user to the articles according to the historical behavior data of all the users; the relation matrix generation module is used for constructing a user-article relation matrix according to the relation vector; and a bipartite graph generation module for obtaining the user-item bipartite graph according to the user-item relation matrix.

In the above recommendation system, the link prediction module includes: the weight calculation module is used for calculating the sum of all path weights with any length between two nodes in the user-article bipartite graph according to the user-article bipartite graph; and the predicted value generation module is used for weighting the sum of all path weights with any length between the two nodes by utilizing a descending weight factor to obtain the predicted value of the preference degree of the user on the unselected goods.

In the above recommendation system, the recommendation module includes: the sorting module sorts the unselected articles from high to low according to the predicted value; and the candidate item generation module is used for recommending one or more items which are ranked most front to the user as candidate items.

Drawings

The features and advantages of the present invention will be more clearly understood by reference to the accompanying drawings, which are illustrative and not to be construed as limiting the invention in any way, and in which:

FIG. 1 shows a flow diagram of a recommendation method based on link prediction according to an embodiment of the invention;

FIG. 2 illustrates an example of a user-item bipartite graph according to an embodiment of the invention; and

fig. 3 shows a schematic diagram of a recommendation system based on link prediction according to an embodiment of the invention.

Detailed Description

Embodiments of the present invention will now be described in detail with reference to the drawings, and it should be noted that the embodiments are illustrative and not restrictive.

The link prediction-based recommendation method provided by the invention mainly comprises two stages. The first stage is to use the historical behavior data of the user to obtain the past preference relationship of the user to the article and to use a plurality of numbers to represent, namely the user-article bipartite graph establishing stage. The second stage is to respectively calculate the sum of the path weights with any length between two nodes according to the user-article bipartite graph; then weighting the sum of the path weights with any length by utilizing a descending weight factor to obtain a predicted value, and if the predicted value is a real number, indicating that the two nodes are both user nodes or both article nodes; otherwise, if the predicted value is a plurality, the two node types are different; and finally, for each user, sorting the corresponding complex predicted values of the users from high to low, selecting the articles which are not selected by the first N users and recommending the articles to the user, namely, the preference degree prediction and article recommendation stages of the articles which are not selected by the users.

The link prediction based recommendation method will be described in detail below with reference to fig. 1. A flow chart of a recommendation method based on link prediction according to an embodiment of the present invention is shown in fig. 1. As shown in fig. 1, the establishment phase of the user-item bipartite graph includes the following steps:

step S101: traversing all users in the current network system to obtain historical behavior data of all users;

step S102: obtaining a relation vector describing the past preference of the user to the articles according to the historical behavior data of all the users;

step S103: constructing a user-item relationship matrix according to the relationship vector;

step S104: and obtaining a user-item bipartite graph according to the user-item relation matrix.

By now the user-item bipartite graph set-up phase has been completed and the results are saved in the form of a adjacency matrix. As shown in fig. 1, the preference prediction and item recommendation stage for the items not selected by the user includes the following steps:

step S105: predicting the preference degree of the user on the unselected goods by using a link prediction method;

step S106: and screening the obtained prediction results, and generating recommended articles for each user.

The following describes the above recommendation method in further detail, and first describes the user-item bipartite graph creation phase in detail.

Firstly, traversing all articles for a user u in user historical behavior data, judging the favorite relationship of the user to an article by utilizing the past behaviors (such as purchase, grading, click, comment and the like) of the user u to the article i, and obtaining a past favorite relationship vector of the user u

As in equation (1):

wherein,

representing a relationship vector

I (n is the number of items in the system), i.e. the preference relationship of the user u to the item i.

Then, a user-item relationship matrix B is constructed according to the obtained past preference relationship vectors of all users, which is specifically described as formula (2).

Where B is a matrix of dimensions m x n, m being the number of users in the system and n being the number of items in the system.

Then, a user-item bipartite graph is built according to the obtained user-item relation matrix B, and the result is stored in the adjacency matrix a, specifically calculated as shown in formula (3).

Where a is a matrix of (m + n) × (m + n) dimensions, and when the rows and columns of the matrix represent two user nodes or two commodity nodes simultaneously, the corresponding element values in the adjacent matrix are 0. In general, we can express the adjacency matrix a as equation (4).

$A=\left[\begin{array}{cc}0& \mathrm{jB}\\ -{\mathrm{jB}}^T& 0\end{array}\right]---\left(4\right)$

Next, the preference degree prediction and item recommendation stages of items that have not been selected by the user will be described in detail.

First, the adjacency matrix A describing the user-item bipartite graph is used to calculate the sum A of all path weights of any length l (l ≧ 3) between two nodes in the user-item bipartite graph^l. Equation (4) may be further expressed as equation (5).

Wherein the sum of all path weights with length l from node x to node y in the figure can be represented as A^l(x,y)。

The previously obtained a may then be weighted with decreasing weighting factors^l(x, y) to obtain a prediction value matrix P of the user's preference for unselected items. The weighting factors may be, for example, equal ratios, and in the case where the weighting factors are equal ratios, the prediction value matrix P is as shown in equation (6).

$P=\underset{l}{\mathrm{\Σ}}{\mathrm{\α}}^l\·A^l---\left(6\right)$

The geometric factor alpha satisfies 0< alpha <1, the selection is determined according to specific conditions, and the smaller the alpha value is, the smaller the influence of the weight of the long path on the final predicted value is; conversely, the greater the effect. In addition, if P (x, y) is a real number, the fact that x and y are both user nodes or both article nodes is indicated, the size of the value represents the similarity degree between x and y, and the larger the value is, the more similar the value is; if P (x, y) is a complex number, it indicates that the node types of x and y are different, further assuming that x is a user node and y is an item node, the value thereof is characterized by the preference degree of the user x for the item y, the sign is positive to indicate "like" and the sign is negative to indicate "dislike".

Of course, the weighting factors are not necessarily equal but follow a principle that the weighting factor corresponding to each power of the adjacent matrix is necessarily smaller than the reciprocal of the maximum absolute value in the power matrix. Such as A³If the maximum absolute value of the sum is 7, the sum A in the calculation of the prediction result is³The corresponding multiplied weight factor must be less than 1/7 and greater than 0.

Then, for user u, the predicted value matrix P is selected to select the non-selected items of user uPrediction value

Vector and sort the items from high to low according to the size of the predicted value.

And finally, recommending the N items which are ranked most at the top as candidate items to the user.

Through the steps, the recommendation method based on the link prediction is completed.

In order to make the description clearer, the above recommendation method is illustrated by a simple example. As shown in Table 1, this example includes 3 users (u)₁～u₃) For 6 items (i)₁～i₆) Each score takes an integer value of 1-5, with the size indicating the user's preference for the item and a null indicating that the user has not selected the item. N (N = 1) most likely favorite items need to be recommended for each user.

TABLE 1

If the historical rating value of the user for the item is greater than 2, it indicates that the user likes the item, and the real number 1 indicates the relationship, whereas the real number-1 indicates a relation of dislike, and if the user has not scored the item, it is indicated by the real number 0. After mapping the user rating data in table 1 to like or dislike, a user-item relationship matrix B is obtained as shown in equation (7).

$B=\left[\begin{array}{cccccc}1& -1& 1& 0& 0& 0\\ 0& 1& -1& 1& 0& 0\\ 1& 0& 1& 0& 1& -1\end{array}\right]---\left(7\right)$

The relationship matrix B is used to build a user-item bipartite graph, as shown in fig. 2. An example of a user-item bipartite graph according to an embodiment of the invention is shown in fig. 2. Since there is no known relationship between the user and the user (user-user), and between the item and the item (item-item) in the initial condition, it is represented in the user-item bipartite graph that there is no edge between two users and two items. Of course, in practice, sometimes, similar or dissimilar relationships between some users and between some items can be obtained through some demographic information of the users or content information of the items, and then the real number "1" is used to represent the weight of the edge in the user-item bipartite graph between users and between items with known similar relationships, and the real number "-1" is used to represent the weight of the edge in the user-item bipartite graph between users and between items with known dissimilar relationships. Similarly, a user-item of known relationship is represented by the plural number "j" or "-jThe weight of the edge in the user-item bipartite graph. Specifically, the user u is in the user-item relationship matrix obtained above₁Like item i₁Then there is a slave u in the user-item bipartite graph₁Point of direction i₁Has a weight of j, and a slave i exists₁Point u₁The weight is-j; in addition, user u in the user-item relationship matrix₁Disliked article i₂Then there is a slave u in the user-item bipartite graph₁Point of direction i₂Has a weight of-j, and a slave i exists₂Point u₁The weight is j. Since fig. 2 has limited space, all edges cannot be weighted, but the process of constructing a user-item bipartite graph according to an embodiment of the invention has been described more vividly.

The user-item bipartite graph in fig. 2 is represented by a adjacency matrix, resulting in equation (8). In equation (8), the user-item adjacency matrix is obtained by using complex numbers to represent the weights of the user-item preference degree edges. In the initial condition, since the user-user and item-item similarity degree edges are not considered, the upper left corner and the lower right corner of the adjacency matrix are both zero matrices.

The object of the invention is to predict the user's preference for non-selected items, e.g. to predict the user u in the above system₁To article i₆The degree of preference. Abstracted into a user-item bipartite graph, as shown in FIG. 2, from user u₁Node to item i₆Nodes having more paths reachable, e.g. from u₁The node may pass through i₁The node then passes through u₃Node finally to i₆A node, the path length being 3; of course, can also be selected from u₁The node may pass through i₂Node is passed through u₂The node then passes through i₃Node is passed through u₃Node is the mostBack to i₆Node, the path length is 5. The shorter the length of these paths, the user u₁Like item i₆The greater the likelihood. Mathematically, the statistics of such path numbers can be computed by multiplication of a adjacency matrix.

For simplicity, only the sum of all path weights for length between two nodes, l =3, is considered here, and thus no weight factor is needed, resulting in a predictor matrix P as shown in equation (9). In the predictor matrix P, the value-2 j in the ninth column of the first row indicates the user u₁To article i₆Wherein the minus sign indicates dislike, and modulo-2 indicates the degree of dislike.

$P=A^3=\left[\begin{array}{ccccccccc}0& 0& 0& 5j& -5j& 7j& -2j& 2j& -2j\\ 0& 0& 0& -3j& 5j& -6j& 3j& -j& j\\ 0& 0& 0& 6j& -3j& 7j& -j& 4j& -4j\\ -5j& 3j& -6j& 0& 0& 0& 0& 0& 0\\ 5j& -5j& 3j& 0& 0& 0& 0& 0& 0\\ -7j& 6j& -7j& 0& 0& 0& 0& 0& 0\\ 2j& -3j& j& 0& 0& 0& 0& 0& 0\\ -2j& j& -4j& 0& 0& 0& 0& 0& 0\\ 2j& -j& 4j& 0& 0& 0& 0& 0& 0\end{array}\right]---\left(9\right)$

And finally, for each user, sorting the items according to the predicted value of the preference degree of the user on the unselected items from high to low, and selecting the top N =1 items from the sorted list to recommend the items to the user. The ranking and recommendation results for this example are shown in table 2.

User' s	Results of the sorting	Recommendation (N = 1)
			u1	i5 i4 i6	i5
u2	i6 i5 i1	i6
			u3	i4 i2	i4

TABLE 2

Next, a recommendation system based on link prediction is described with reference to fig. 3. A schematic diagram of a recommendation system based on link prediction according to an embodiment of the present invention is shown in fig. 3. As shown in fig. 3, the recommendation system includes an acquisition module 301, an establishment module 302, a link prediction module 303, and a recommendation module 304.

Specifically, the obtaining module 301 traverses all users in the current network to obtain historical behavior data of all users; the building module 302 builds a user-item bipartite graph from historical behavior data of all users; the link prediction module 303 predicts the user's preference for the unselected item by using a link prediction method according to the user-item bipartite graph; the recommendation module 304 filters the prediction results for each user to generate recommended items for each user.

As shown in fig. 3, the establishing module 302 includes: the relation vector generation module 302a obtains a relation vector describing the past preference of the user for the article according to the historical behavior data of all the users; a relationship matrix generation module 302b, which constructs a user-item relationship matrix according to the relationship vector; and a bipartite graph generation module 302c for obtaining a user-item bipartite graph according to the user-item relationship matrix.

The link prediction module 303 includes: the weight calculation module 303a is used for calculating the sum of the weights of all paths with any length between two nodes in the graph according to the user-article bipartite graph; and a predicted value generation module 303b for weighting the sum of all path weights of any length between two nodes by using a decreasing weight factor to obtain a predicted value of the user's preference degree for the unselected item.

The recommendation module 304 includes: a sorting module 304a for sorting the unselected items according to the predicted value from high to low; and a candidate item generation module 304b recommending one or more items ranked the top as candidate items to the user.

Therefore, the link prediction-based recommendation system is realized.

In the invention, the complex number and the real number are respectively used for representing the weight of the user-item preference degree edge and the weight of the user-user and item-item similarity degree edge, so that two different types of nodes (user nodes and item nodes) in the user-item bipartite graph are effectively distinguished, the classic link prediction method is conveniently deployed in the recommendation system, and the accuracy and diversity of the classic link prediction method for recommendation are effectively improved without being influenced by the two different types of nodes in the user-item bipartite graph.

Preferably, the liveness of the user and the popularity of the item can be considered in the process of establishing the user-item bipartite graph, so that better recommendation diversity is obtained.

Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations fall within the scope defined by the appended claims.

Claims (10)

1. A method for link prediction based recommendation, comprising:

traversing all users in the current network to obtain historical behavior data of all the users;

establishing a user-item bipartite graph according to the historical behavior data of all the users;

predicting the preference degree of the user for the unselected articles by using a link prediction method according to the user-article bipartite graph; and

and screening the prediction result of each user to generate a recommended item for each user.

2. The method of claim 1, wherein said building a user-item bipartite graph from the historical behavior data of all users comprises:

obtaining a relation vector describing past preference of the user to articles according to the historical behavior data of all the users;

constructing a user-item relationship matrix from the relationship vectors; and

and obtaining the user-item bipartite graph according to the user-item relation matrix.

3. The method of claim 1, wherein said predicting user preference for unselected items using a link prediction method comprises:

respectively calculating the sum of all path weights with any length between two nodes in the user-article bipartite graph according to the user-article bipartite graph; and

and weighting the sum of all path weights with any length between the two nodes by utilizing a descending weight factor to obtain a predicted value of the preference degree of the user on the unselected goods.

4. The method of claim 3, wherein the filtering the predicted outcome for each user to generate recommended items for each user comprises:

sorting the unselected items according to the predicted value from high to low; and

recommending one or more items ranked most forward to the user as candidate items.

5. The method of claim 1, wherein the creating a user-item bipartite graph further takes into account liveness of the user and popularity of items.

6. The method according to any of claims 1-5, wherein the weight of the user-item like degree edge is represented by a complex number and the weight of the user-user, item-item like degree edge is represented by a real number.

7. A link prediction based recommendation system comprising:

the acquisition module is used for traversing all users in the current network to acquire historical behavior data of all the users;

the establishing module is used for establishing a user-article bipartite graph according to the historical behavior data of all the users;

the link prediction module is used for predicting the preference degree of the user on the unselected articles by using a link prediction method according to the user-article bipartite graph; and

and the recommending module is used for screening the prediction result of each user to generate recommended articles for each user.

8. The recommendation system of claim 7, wherein the establishment module comprises:

the relation vector generation module is used for obtaining a relation vector which describes the past preference of the user to the articles according to the historical behavior data of all the users;

the relation matrix generation module is used for constructing a user-article relation matrix according to the relation vector; and

and the bipartite graph generation module is used for obtaining the user-item bipartite graph according to the user-item relation matrix.

9. The recommendation system of claim 7, wherein the link prediction module comprises:

the weight calculation module is used for calculating the sum of all path weights with any length between two nodes in the user-article bipartite graph according to the user-article bipartite graph; and

and the predicted value generation module is used for weighting the sum of all path weights with any length between the two nodes by utilizing a descending weight factor to obtain the predicted value of the preference degree of the user on the unselected goods.

10. The recommendation system of claim 9, wherein the recommendation module comprises:

the sorting module sorts the unselected articles from high to low according to the predicted value; and

and the candidate item generation module is used for recommending one or more items which are ranked most front to the user as candidate items.

Images