CN102254028A - A personalized product recommendation method and system integrating attribute and structure similarity - Google Patents

Abstract

The invention discloses a personalized commodity recommendation method integrating attributes and structural similarity. The method integrates attribute and structural similarity information, maps users and commodities as nodes with characteristic information to a network, establishes an information network graph according to the purchasing relation between customers and commodities, measures interest preference between user node pairs by utilizing the integrated attribute and the structural similarity in the information network graph, and selects nearest neighbors according to the interest preference so as to improve recommendation accuracy. On the basis of the recommendation method, the invention also discloses a personalized commodity recommendation system integrating the attribute and the structural similarity measurement. The system accurately measures the interest preference of the user by using a calculation method with similar integration attributes and similar node structure backgrounds in an information network graph, and improves the generation efficiency of nearest neighbors by using a clustering technology in the recommendation process. The method and the system can be used in electronic commerce application and provide personalized commodity recommendation for users.

Description

A personalized product recommendation method and system integrating attribute and structure similarity

technical field

The present invention relates to the field of computer Internet technology, in particular to the field of e-commerce, in particular to a personalized product recommendation method and system integrating attributes and structural similarities.

Background technique

With the continuous development of e-commerce, the number and types of goods are increasing rapidly. In order to find the goods they need as soon as possible, users want a function similar to a shopping guide to help them choose the right goods or services. Personalized recommendation systems should be used. And born. Based on massive data analysis and mining technology, personalized recommendation recommends products and information of interest to users according to their behavior habits and interest characteristics. At present, almost all large-scale e-commerce websites integrate personalized recommendation technology into the system to varying degrees, among which collaborative filtering technology is widely used.

The recommendation based on collaborative filtering first collects information representing user interests, then selects neighboring users, and finally predicts the interests of target users to generate recommendation results. Its core problem is to find a group of users whose interests are similar to the target user. The similarity between this group of users and the target user is obtained by collecting and comparing the behavior selection vectors representing the interests of two users. At present, the similarity calculation methods for comparing behavior selection vectors mainly include Poisson correlation coefficient and cosine similarity, which are generally based on user-item rating matrix, and are characterized by simple calculation and easy understanding. However, this method based on scoring matrix makes it difficult for the system to find users with the same interest because the matrix is sparse when measuring the similarity of users or items. In addition, modeling user interests through scoring cannot fully and truly describe user interests, and some information will be lost. In order to retain more information that potentially reflects user interests in the data, some methods map users and items to the network as nodes, and use node structure similarity for collaborative recommendation. In addition, some content-based filtering methods use the attribute description information of users and items to measure the similarity between users or items, and then recommend products. However, the above two calculation methods based on structural similarity and attribute similarity will lose some information when making recommendations. According to the similarity enhancement hypothesis: the similarity between two objects not only depends on their own attributes but also depends on their attributes. Similarity between related other objects. Therefore, how to combine attribute description information and node structure information when recommending, propose a comprehensive and effective similarity measurement method to describe the interests of users, and select the nearest neighbor accordingly to improve the accuracy of recommendation is a problem that needs to be solved. The problem.

Contents of the invention

In order to solve the above problems, the present invention proposes a personalized product recommendation method that integrates attributes and structural similarities. Perform preprocessing to obtain the basic characteristics, statistical characteristics and behavioral characteristics of users; map users and products to the network as nodes with attribute information, and establish an information network graph based on the purchase relationship between users and products; use Combine the measurement method of attribute and structure similarity to measure the interest preference between user node pairs; use the similarity between node pairs as input to cluster users, so as to narrow the selection range of nearest neighbors and improve the recommendation speed; from the activity In the cluster where the user is located, select the M nearest neighbor users as their nearest neighbor users and generate a neighbor set; predict and score the products that the active users have not purchased in the candidate product database; make Top-N recommendations for the active users, so as to Returned as a set of recommended candidate items.

As an embodiment of the present invention, the collection of data includes the basic information of the user, the basic information of the commodity, and the transaction data of the commodity purchased by the user, that is, the comment data. The user's basic information refers to the registration information submitted by the user when shopping online, including the user's name when registering, registration time, current level, and region from. The basic information of the product refers to the descriptive information of the product when it is put on the shelf, such as the product number, product name, product brand, product field, product type, and product launch time. The transaction data of the product purchased by the user includes the product number, user name, purchase time, review time, review advantages, review disadvantages, review title, review main content, and rating.

As an embodiment of the present invention, data preprocessing includes removing noise data, filling vacant numerical items, data regularization, and user feature extraction. Among them, user feature extraction refers to the acquisition of statistical features and behavioral features after statistics on the user's historical purchase transaction data. These features include the user's historical purchase times, the number of brands he is keen on, the average time interval between purchase time and comment time, and average consumption amount. Time interval between registration time and purchase of new products, time interval between product listing time and purchase time, useful and useless times of comments, proportion of negative reviews, and the category, type, and field of the currently purchased product. ratio, the average length of advantages and disadvantages in the comments, the average length of the entire comment, and the proportion of deficiencies in the comments that are the default comments.

As an embodiment of the present invention, the construction of the user commodity information network map is to map data to the network. According to the preprocessed data and information, users and products are represented as nodes with attributes in the network graph. If a user purchases a certain product, there is a directed connection edge between the user and the product node, and the direction of the edge is from Items point to users.

As an embodiment of the present invention, the similarity measurement method between users is based on attribute information and structure information of nodes in the graph. If the user's attribute values are more similar and the user's historical purchase records are more similar, the distance between users is closer, which in turn indicates that the interests and preferences of users are more similar. The formula for calculating the similarity between users in this recommendation method is:

$\left\{\begin{array}{c}{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; ASimRank</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; attribute</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}<span\; class="notranslate">\; *</span>{\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; link</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; \&NotEqual;</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; )</span>\\ {\mathrm{<span\; class="notranslate">\; S</span>}}_{\mathrm{<span\; class="notranslate">\; ASimRank</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; b</span>}<span\; class="notranslate">\; )</span>\end{array}\right.$

in, $S_{\mathrm{attribute}}(a,b)=e^{-\underset{r=1}{\overset{p}{\mathrm{\&Sigma;}}}|x_{\mathrm{ij}.}{A}_r-x_{\mathrm{ik}.}{A}_r|+\mathrm{\&mu;}\underset{r=p+1}{\overset{N}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(x_{\mathrm{ij}.}{A}_r,x_{\mathrm{ik}.}{A}_r)},$ $S_{\mathrm{link}}(a,b)=\frac{C}{\left|I\left(a\right)I\left(b\right)\right|}\underset{i=1}{\overset{\left|I\left(a\right)\right|}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{\left|I\left(b\right)\right|}{\mathrm{\&Sigma;}}}{S}_{\mathrm{link}}(I_i\left(a\right),I_j\left(b\right))$

In the calculation formula of S _ASimRank (a, b), S _attribute (a, b) represents the attribute similarity score of nodes a and b, and S _link (a, b) is based on the connection between a and b in the information network graph Structural similarity scores computed from relationships. In the calculation of S _attribute , x _ij .A _r and x _ik .A _r represent the rth attribute value of objects x _ij and x _ik ; N is the total number of attributes, and the first to pth attributes are numerical attributes or have ordinal category attributes, and the p+1th attribute to the Nth attribute are unordered category attributes. The function δ(x _ij .A _r , x _ik .A _r ) is an attribute value difference function, and the function value is 0 when the two attribute values are the same, otherwise it is 1. The parameter μ in the formula is used to adjust the importance of the disordered category attribute relative to the value attribute in the calculation of the entire attribute similarity, and its value range is [0,1]. In the example of the present invention, μ takes a value of 0.5 . Among them, the parameter C in the formula of S _link (a, b) is a constant between [0, 1], which represents the decay rate when calculating the similarity. In the embodiment of the present invention, the value of C is 0.8. I(a) and I(b) represent the in-degree neighbor node sets of node a and node b respectively. I _i (a) and I _j (b) respectively represent the i-th in-degree neighbor node of a and the j-th in-degree neighbor node of b. |I(a)| and |I(b)| represent the in-degree of node a and node b, respectively. If the in-degree of a or b is 0, the value of S _link (a, b) is 0, and if a and b represent the same node object, the value of S _link (a, b) is 1. From the calculation formula of S _link (a, b), it can be seen that, generally speaking, the similarity between a and b is the average similarity between the in-degree neighbor nodes of a and the in-degree neighbor nodes of b. The similarity scores between node pairs calculated by S _link (a, b) are symmetrical, that is, S(a, b)=S(b, a). Using the S _link (a, b) formula, the structural similarity score between any user node pair a and b in the graph can be calculated after multiple iterative operations until convergence. The final similarity score S _ASimRank (a, b) of nodes a and b is jointly determined by the above two similarity scores. The relative importance of the S _attribute and the S _link is adjusted by the parameter λ in the formula, and the value of λ is [0, 1], and the value of λ is 0.5 in the example of the present invention.

As an embodiment of the present invention, in order to improve the recommendation speed, all users are clustered by clustering technology according to the similarity among users. Narrow down the active user nearest neighbor search range from the global user to a certain cluster after clustering.

As an embodiment of the present invention, the clustering result is used to generate the nearest neighbors of the active users, and the M users most similar to the active users are selected from the cluster where the active users are located as their nearest neighbors.

As an embodiment of the present invention, predicting scoring refers to predicting and scoring the commodities that active users have not purchased in the candidate commodity database, and using the weighted addition method to weight the scores of all users in the neighbor set to the target commodity. The rating of the target item.

As an embodiment of the present invention, Top-N recommendation refers to sorting the predicted scores of all target commodities of an active user, recommending N commodities with the highest scores to the active user, and returning them as recommended candidate commodity sets.

On the basis of the above recommendation method, the present invention also provides a personalized commodity recommendation system integrating attributes and structural similarities. The system at least includes the following components and modules: user terminal, shared information server, user and commodity information collector, user basic information database, commodity basic information database, user history transaction database, user preference model processor, data mapping converter, user Preference Metric, User Match Database, Recommendation Accelerator and Personalized Recommendation Processor.

The user terminal is used to submit a user recommendation request, and the return terminal displays a list of products recommended by the system. The shared information server is used to store the information shared by the system. The user and commodity information collectors are used to collect data, and the collected data are respectively stored in the user basic information database, commodity basic information database, and user historical transaction database. The user preference model processor refers to processing data and establishing a user preference model. The data mapping converter maps users and products to the network graph and represents them as nodes with attribute information, and constructs an information network graph based on the purchase relationship between users and products. The user preference measurer measures the preferences among users in the information network graph by using the measure method of integrated attribute and structure similarity proposed in the present invention, and the result of the measure is stored in the user matching degree database. The recommendation accelerator takes the similarity between user node pairs as input, and uses clustering technology to narrow down the search range of active users' nearest neighbors to improve recommendation efficiency. Based on the set of nearest neighbors, a personalized recommendation processor is used to predict the predicted ratings of target items by active users.

The present invention adopts various characteristics and attributes of users to establish a user preference model; uses the similarity calculation method of similarity in integrated attributes and similarity in node structure background in the information network graph to accurately measure user interest preferences; uses clustering technology to improve With the generation efficiency of the nearest neighbor, the system can finally respond to the user's recommendation request in real time and quickly, and provide the client with personalized product recommendation.

The advantages and features of the present invention will be given emphatically in the detailed analysis below, or be reflected through system practice. Combining attributes and structural similarities, the present invention can deeply dig out user interests and preferences, and more accurately construct user interest models, thereby generating accurate recommendation content.

Description of drawings

Fig. 1 is a flow chart of measuring user interests in the recommendation method of the present invention.

Fig. 2 is an example diagram of user product relationship information.

Fig. 3 is a schematic diagram of the construction of the user product information network graph.

Fig. 4 is an overall flow chart of the personalized commodity recommendation method in the present invention.

Fig. 5 is a flowchart of the recommended accelerator in the embodiment of the present invention.

Fig. 6 is a structural diagram of a personalized commodity recommendation system integrating attributes and structural similarities in the present invention.

Detailed ways

In order to illustrate the details of the technical solutions of the present invention more clearly, the present invention will be explained and described in detail below through embodiments in conjunction with the accompanying drawings.

The main feature of the invention is to propose an effective recommendation method and design an efficient and practical personalized recommendation system. The important feature of the method of the present invention is to accurately measure the user's interest and preference in the information network graph in combination with the attribute and structure similarity, and use the clustering technology to narrow the search range of the nearest neighbor. The system responds to the user's recommendation request in real time and promptly returns the list of products that the user is really interested in to the client.

Figure 1 shows the process of calculating the similarity between users by combining the measurement method of attribute and structural similarity. The recommendation method disclosed in the present invention takes the user interest preference model and basic commodity information as input s101; the user and commodity data are mapped to the network and represented as nodes, the nodes have attribute information, and the directed connection edge represents the purchase between the user and the commodity relationship, if the user purchases a product, the direction of the connecting edge is from the product to the user, and the information network graph s102 is formed accordingly; the similarity calculation method combining attribute similarity and structural background similarity is used to measure the distance between any user node pair in the network graph similarity of interests and preferences s103; return the similarity between all user node pairs, and store them in the user matching degree database s104.

As an embodiment of the present invention, the information network graph is constructed with user purchased product data in the recommendation system, and a small example is used to describe in detail the calculation process of the measurement method combining attribute and structure similarity.

The similarity calculation in the present invention is based on the directed graph model in the information network, such as G=(V, E), wherein V represents the set of nodes in the figure, and E represents the set of edges <u, v> in the figure, <u, v>∈E, (u, v∈V). In the user and commodity information network graph G, users and commodities correspond to a certain node v in G, the present invention uses I(v) to represent the in-degree neighbor node set of node v, and I _i (v) represents v The i-th in-degree node of , where 1≤i≤|I(v)|, where each node object in G has multiple attribute characteristics. The generation of connection edges in the user product information network graph relies on the purchase relationship between users and products, which is reflected in the comment table, as shown in Figure 2. If a user has commented on a product, there is an edge connecting the user and the product. At the same time, user and product nodes have their own attribute lists.

As shown in Figure 3, we take the two types of objects in the recommendation system as an example, users and products, where P represents a product, C1, C2 and C3 represent different users, and the legend shows that product P is simultaneously used by users C1, C2 and C3 buys. In order to have a clear understanding of the calculation of the similarity between user node pairs in the network graph, this example is used to describe the calculation process of the similarity between users. For the sake of calculation convenience, only four attributes are considered here: the user level, the region where the user is located, the time interval between the time when the user posts a comment and the time when the product is purchased, and the user's rating of the product. The attribute values of the 3 users are shown in the table below.

user level area Interval (days) score C1 D SC 10 5 C2 T EC 16 4 C3 T NC 15 4

From the example in Figure 3, if only the structural similarity between nodes is considered, the similarity between {C1, C2}, {C1, C3} and {C2, C3} node pairs cannot be distinguished, because they are all controlled by the same Commodity P refers to, C1, C2 and C3 have the same in-degree neighbor nodes as commodity P. According to the formula $S_{\mathrm{link}}(a,b)=\frac{C}{\left|I\left(a\right)I\left(b\right)\right|}\underset{i=1}{\overset{\left|I\left(a\right)\right|}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{\left|I\left(b\right)\right|}{\mathrm{\&Sigma;}}}{S}_{\mathrm{link}}(I_i\left(a\right),I_j\left(b\right)),$ Then when C is 0.8, the structural similarity values between {C1, C2}, {C1, C3} and {C2, C3} node pairs are all 0.8. However, considering the attribute features in the similarity calculation, the similarity of {C1, C2}, {C1, C3} and {C2, C3} node pairs is better distinguished. The calculation process of the similarity score between them will be introduced in detail below. Before using the formula to calculate the attribute similarity, the data needs to be processed first. If the user's current level is quantified by an ordered numerical value, the user level is divided into 6 levels, namely T (Diamond Member), A (Gold Member), B (Silver Member), C (Bronze Member), D (Iron Member) Member), E (registered member), the quantitative values corresponding to these levels are 5, 4, 3, 2, 1, 0. The area attribute belongs to the disorder category attribute value only needs to judge its similarity and difference. According to the formula $S_{\mathrm{attribute}}(a,b)=e^{-(\underset{r=1}{\overset{p}{\mathrm{\&Sigma;}}}|x_{\mathrm{ij}.}{A}_r-x_{\mathrm{ik}.}{A}_r|+\mathrm{\&mu;}\underset{r=p+1}{\overset{N}{\mathrm{\&Sigma;}}}\mathrm{\&delta;}(x_{\mathrm{ij}.}{A}_r,x_{\mathrm{ik}.}{A}_r))},$ The attribute similarity value between user node pairs can be calculated as follows:

S _attribute (C1, C2)＝e ^{-(|1-5|+|10-16|+|5-4|+0.5*1)} ＝1.013* ^10-5 ,

S _attribute (C1, C3)＝e ^{-(|1-5|+|10-15|+|5-4|+0.5*1)} ＝2.754* ^10-5 ,

S _attribute (C2, C3) = e ^{- (|5-5|+|16-15|+|4-4|+0.5*1)} = 0.2231.

Using the formula S _ASimRank (a, b)=(1-λ)*S _attribute (a, b)+λ*S _link (a, b), when λ=0.5, the following values are calculated:

S _ASimRank (C1, C2)=(1-0.5)*1.013*10 ⁻⁵ +0.5*0.8=0.4+5.065*10 ⁻⁶ ,

S _ASimRank (C1, C3)=(1-0.5)*2.754*10 ⁻⁵ +0.5*0.8=0.4+1.377*10 ⁻⁵ ,

S _ASimRank (C2, C3)=(1−0.5)*0.2231+0.5*0.8=0.51155.

It can be seen from the final similarity score that although C1, C2 and C3 are all referred to by the same P and have the same in-degree neighbor nodes, but the attribute values of {C1, C2} are quite different, making the similarity score of the two The lowest is 0.400005065, while {C2, C3} has the highest similarity score of 0.51155 because the attribute values of the two are basically the same. From this example, it can be shown that the combination of attribute and structural similarity can retain more real information, so that the obtained similarity score can better distinguish the interest preference matching between users.

Fig. 4 is a general flow chart of the personalized commodity recommendation method in the present invention. The whole process is divided into 8 major steps, and the brief description of each step is as follows: collecting data s401, data preprocessing and establishing user model s402, data conversion and establishing information network graph s403, user node interest similarity measurement s404, generating nearest neighbors s405, selecting Nearest neighbor s406, predicted score s407, recommended Top-N products s408.

In more detail, in s401, collecting data refers to organizing user information registered on the e-commerce platform, product information on the shelves, and user purchase transaction records on the website. All data are stored in a unified format. For example, the format of the basic information of the product is: product [ID=143076, name=Kingston DDR3 1333 2G, brand=Kingston, field=computer product, type=core accessories, launch time=2008-12-11 17:18:45]; user The basic information format is: user [user name=lihui581203, region from=Shanghai, current level=silver member, registration time=2009-11-13]; the user’s purchase transaction is reflected in the comment information, and its format is: comment[commodity ID = 143076, user name = lihui581203, purchase time = 2010-05-12, comment time = 2010-07-02 14:11, advantage = compatibility, good quality and guaranteed, disadvantage = no disadvantage found yet! , title = genuine goods at a fair price, main content = have been using this product with peace of mind, score = 5].

In s402, data preprocessing refers to filling some vacant data items with default values or average values, and removing some noise data. Store user, product, and comment data into the corresponding database. In the database, SQL statements are used to perform statistics on the data and join operations between tables to obtain the final user attribute characteristics and establish a user model.

In s403, data conversion and establishment of information network graph Utilize the user model and processed data established in s402 to map products and users to the network graph as nodes, and the generation of connection edges between nodes is based on the purchase relationship between users and products . Since the in-degree information of nodes is needed when calculating the measure of structural similarity, the direction of the connection edge is from the product to the user. According to this method, the information relationship network graph of products and users can be constructed. The nodes in the graph have attribute information, which provides input for the calculation of node similarity in the next step in the network graph.

In s404, the recommendation method in the present invention is based on user interests, so the measurement of user node interest similarity is a key step in the flow of the method. In the present invention, the combination of attribute and structure similarity measurement method can comprehensively and accurately describe the interest preferences among users. This similarity measurement method is based on the assumption of similarity enhancement: the similarity between two data objects not only depends on their own attributes but also depends on the similarity between other objects related to them. The similarity measurement method proposed in the present invention not only considers the structural background information of the node in the data conversion to the network graph, but also considers the attribute information of the node, thereby greatly retaining the potential interest information of the user. Using the similarity calculation formula disclosed in the present invention, the similarity between any pair of user nodes in the network graph can be finally obtained by iteratively calculating the similarity scores between nodes in the information network graph.

In s405, in order to improve the recommendation speed and greatly reduce the search range of the nearest neighbors, this method adopts the K-modies clustering technology to cluster the users. Locate the search scope of the nearest neighbor from all global users to a certain user cluster, so as to improve the recommendation speed.

The execution of step s405 is described in detail in conjunction with FIG. 5 . First, randomly select K users from all user nodes as clustering centers, use the user similarity results saved in s404 to assign other user objects to the cluster most similar to a certain clustering center, and then reselect the clustering center. In each cluster, a user object is sequentially selected, and the consumption and cost E of replacing the original cluster center with the selected object are calculated. Select the user object with the smallest E to replace the original cluster center as the new cluster center. This cycle is repeated until the convergence condition is met and the centers of K clusters do not change. Next, according to the clustering results, the most similar M users are selected as the nearest neighbors from the cluster where the users needing to be recommended are located, thereby generating the nearest neighbor set of active users.

In s406, the selected M neighbor users are returned using the nearest neighbor set generated in step s405.

In s407, the weighted sum of the scores of all users in the neighbor set for the target product is used as the score of the active user for the target product by using the weighted addition method in predicting the score. Assuming that the neighbor set U={u ₁ , u ₂ ,..., u _n } based on the active user u _a , then the evaluation of the unrated product t _i by the user u _a is defined as the evaluation of the product t _i by all users in the neighbor set U The weighted sum of the score values, the formula is as follows:

$\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; \&lambda;</span>}\underset{\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; the\; s</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; t</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; R</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; )</span>$

为u_k对所有已评价商品的平均评分；

为当前活动用户u_a先验平均评分；λ为规范化系数。Among them, s(u _a , u _k ) is the similarity between active user u _a and neighbor u _k ; R(u _k , t _i ) is u _k ’s rating on product t _i ;

is the average rating of u _k on all evaluated products;

is the prior average score of the current active user u _a ; λ is the normalization coefficient.

In s408, use the predicted scores calculated in s407 to sort the predicted scores of all target commodities of the active user, and recommend N commodities with the highest scores to the active user, ie Top-N recommendation.

Fig. 6 is a structural diagram of a personalized commodity recommendation system designed according to the recommendation method disclosed in the present invention. The system is mainly composed of the following components: user terminal, shared information server, user and commodity information collector 600, user basic information database 610, commodity basic information database 620, user history transaction database 630, user preference model processor 640, data mapping A converter 650 , a user preference measurer 660 , a user matching degree database 670 , a recommendation accelerator 680 , and a personalized recommendation processor 690 .

As an embodiment of the present invention, the terminal mainly refers to a PC, but is not limited thereto, and may also refer to any electronic device with a network communication function such as a mobile handheld device.

As an embodiment of the present invention, the shared information server is used to store all public information and knowledge of the e-commerce platform.

As an embodiment of the present invention, the user and product information collector 600 is used to collect the user's original registration information on the e-commerce platform, process the registration information and store it in the user basic information database 610 . The collector also collects the basic information of all the commodities on the website, and preprocesses the basic information of the commodities and stores them in the basic commodity information database 620 . In addition, user comments on e-commerce sites are also included to reflect their purchase records and transaction information. The information collector 600 monitors user comments and processes the data and stores the data in the user history transaction database 630 .

As an embodiment of the present invention, the user basic information database 610 is used to store the user basic information output by the user and product information collector 600, as one of the inputs to the user preference model processor 640, and the stored content is Refers to the registration information submitted by the user when shopping online, including the user level, the region the user comes from, the user name when the user registers, and the registration time.

As an embodiment of the present invention, the basic commodity information database 620 is used to store the basic commodity information output by the user and the commodity information collector 600 . The stored content refers to the descriptive information of the product on the e-commerce website when it is put on the shelf, such as the product number, product name, product launch time, product price, product category, product type, product field, and product brand.

As an embodiment of the present invention, the user historical transaction database 630 is used to store the comment information published by the user after purchasing the product, which includes the user name, product name, purchase time, comment time and comment content, etc. These information contain The user's purchasing behavior characteristics.

As an embodiment of the present invention, the user preference model processor 640 uses the data stored in the user basic information database 610 and the user history transaction database 630 to establish a model for the user. The processor includes steps such as data preprocessing, user feature extraction and integration during execution. Finally, all the characteristics of the user are used to build a model for the user to accurately describe a user. The establishment of the user preference model is carried out in real time when the user makes a product recommendation request. If the user browses a certain type of product, the establishment of the user model will be triggered. The user’s basic information and user historical purchase transaction records are used to establish a model for the user. The establishment of the user model is not only In addition to basic user information, it also includes statistics and behavioral information, including the number of times users have purchased products in history, the number of brands they are passionate about, the average time interval between purchase time and comments, the average amount of consumption per time, the time interval between registration time and new product purchases , the time interval between the time when the product was put on the shelf and the time when the product was purchased, the number of useful and useless comments posted by users, the proportion of negative reviews, the proportion of purchased products in the category, type, and field of the currently purchased product, and the advantages and disadvantages of the comments The average length of the entire comment, the average length of the entire comment, and the proportion of the default comment in the comment.

As an embodiment of the present invention, the data mapping converter 650 refers to using the established user preference model to map users and commodities with attribute information to the network as nodes, and map users to the network according to the purchase relationship between users and commodities. Nodes and product nodes are connected with directed edges to construct an information network graph. Based on this, the commodity user information is converted into the network, and then expressed as nodes and connection edges to form a network graph.

As an embodiment of the present invention, the user preference measurer 660 measures the interest preference of the user node pairs in the information network graph by using a method combining attributes and structural similarities. The measurement value depends on the attribute description information and structural background information of the node. The structural background information here refers to the in-degree neighbor nodes of the user node pair, and the measurement result will be used as the input of the user matching degree database 670 . Specifically, user preference measurement includes the following steps:

According to the attribute information of user nodes, the attribute similarity between user node pairs is obtained by using the attribute similarity calculation method;

According to the structural background information of user nodes, the structural similarity between user node pairs is obtained by using the structural similarity calculation method;

For user node pairs with similar attribute and structure, combine the previous two similarity measures and adjust the weight factor to measure the preference similarity between end user pairs.

As an embodiment of the present invention, the user matching degree database 670 is used to store similarity values between user node pairs, and use this as a basis for active users to select nearest neighbors.

As an embodiment of the present invention, the recommendation accelerator 680 uses the data in the user matching database 670 as input, and uses clustering technology to greatly narrow the user's nearest neighbor search range, thereby improving the real-time performance of the recommendation. Specifically, clustering the user nodes in the network graph mainly includes the following steps.

First, randomly select M cluster center users;

Second, assign non-cluster center users to the nearest cluster;

Again, adjust the cluster center until the clustering result no longer changes;

Finally, user clusters are formed, and each user belongs to its own cluster.

As an embodiment of the present invention, based on the user preference model, the personalized recommendation processor 690 uses the weighted addition method to score the commodities that the active user has not purchased in the candidate commodity database, and the scoring of the active user depends on the recommendation accelerator According to the ratings of the nearest neighbor users on the target products generated in 680, the predicted ratings of the products are sorted, and N products with the highest ratings are recommended to active users. Feedback the product recommendation request made by the user in real time, and return the recommendation result to the client.

Claims (6)

1. A personalized product recommendation method integrating attributes and structural similarities, characterized in that it comprises steps: Step A. Collect basic user information, basic product information, and user historical purchase record information on the e-commerce platform; Step B. After collecting the data, preprocess the data and extract the user's attribute characteristics. Data preprocessing refers to processing the incomplete, noisy, and inconsistent data in the data using corresponding technologies. The user's attribute characteristics Including basic attributes, behavioral characteristics and statistical characteristics, and building a model for users after obtaining all characteristics of users; Step C, using the attribute characteristics of users and products, and the purchase relationship between users and products, mapping users and products to network forming nodes, and constructing user and product information network diagrams; Step D, the integrated attribute and structural similarity measurement method performs interest preference measurement on the user model constructed in step C; Step E, taking the similarity between pairs of user nodes as input, using clustering technology to cluster user nodes, thereby narrowing the search range of nearest neighbors, thereby improving the recommendation speed; Step F, according to the clustering result, generate the nearest neighbor set, sort the similarity between the node pairs in descending order, and return the top M neighbor users; Step G, using the weighted addition method and the ratings of the M neighbor users returned to the target product in step F to predict the rating of the active user on the unevaluated product; Step H, sort the predicted scores of all target commodities, and recommend N commodities with the highest scores to the active user. 2. The method according to claim 1, wherein said step D specifically comprises the following steps: Step D1, inputting the user commodity information network diagram constructed in step C, the network diagram includes user nodes with attribute feature information, commodity nodes and the connection relationship between them; Step D2, calculating the attribute similarity between user node pairs in the network graph, and selecting different attribute value matching methods according to different attribute types; Step D3, calculating the structural similarity between the user node pairs in the network graph, the calculation of the structural similarity depends on the average value of the similarity between the in-degree neighbor nodes of the user node pair; Step D4, selecting an appropriate trade-off factor, using the results of steps D2 and D3 to weight and sum the attribute similarity and structural similarity to obtain the final similarity score between user node pairs in the network graph; 3. The method according to claim 1, wherein said step E specifically comprises the following steps: Step E1, input the similarity scores between all user node pairs in the user and product information network graph; Step E2, randomly selecting K user nodes in the information network diagram as clustering centers; Step E3, based on the similarity of user node pairs in step E1, assign other non-cluster center user objects to the cluster most similar to a certain cluster center; Step E4. Reselect the cluster center. The method of selection is to sequentially select a user object in each cluster, calculate the consumption and cost E after replacing the original cluster center with the selected object, and select the user object with the smallest E to replace the original The cluster center is used as the new cluster center, and this step is repeated until all user cluster centers no longer change; Step E5, taking the clustering result in step E4 as input, sorting the similarity between all users in the user cluster and the current active user in descending order, and selecting the top M users as the nearest neighbor user set. 4. The method according to claim 1, wherein said step F specifically comprises the following steps: Step F1, input the neighbor set of the active user generated in step E and the user's rating on the product; Step F2, using the weight addition method to weight the ratings of all users in the neighbor set on the unrated items of the active users and predict the ratings of the active users on the target items; Step F3, returning the active user's predicted scores for all target commodities. 5. A personalized commodity recommendation system integrating attributes and structural similarities, characterized in that the components and modules include: user terminals, shared information servers, user and commodity information collectors, user basic information databases, commodity basic information Database, user history transaction database, user preference model processor, data mapping converter, user preference measurer, user matching degree database, recommendation accelerator and personalized recommendation processor, characterized in that: The terminal mainly refers to a PC, and can also be any electronic device with network communication functions such as a mobile handheld device; The shared information server refers to a computer that stores all public information and knowledge of the e-commerce platform; The user and product information collector collects the original registration information of the user on the e-commerce platform, processes the registration information and stores it in the user basic information database, and the collector also collects the basic information of all the products on the website, and the basic information of the products The information is pre-processed and stored in the basic information database of the product. In addition, it also includes the user's comment behavior on the e-commerce website. The information collector monitors the user's comment behavior and stores the data in the user's historical transaction database after processing; The basic user information database is used to store basic user information output by the user and commodity information collector, such as user name, region from, current level, and registration time, as one of the inputs of the user preference model; The commodity basic information database is used to store basic commodity information output by users and commodity information collectors, such as commodity numbers, commodity names, commodity brands, commodity fields, commodity types, and commodity on-shelf time; The user history transaction database refers to the transaction data used to store the user and the commodity information collector output by the user to purchase the commodity, which contains information such as the behavior characteristics of the user, the purchase relationship between the user and the commodity, and its information includes : product number, user name, purchase time, review time, review advantages, review disadvantages, review title, review main content, rating, if a user has commented on a product, it means that the user has purchased the product, the user The purchase relationship between the user and the product and the behavioral characteristics of the user's purchase of the product will be used as the input of the data mapping converter and the user preference model processor respectively; The user preference model processor refers to the use of the user basic information database, commodity basic information database and user history transaction database to extract various types of user characteristics such as basic attribute characteristics, statistical characteristics and behavioral characteristics, etc. Building a model, using this model to describe a user, is called building a user model; The data mapping converter refers to using the established user preference model to map users and commodities with attribute information to the network as nodes, and use directed edges to connect user nodes and commodity nodes according to the purchase relationship between users and commodities. Connect to build an information network graph, thereby converting and mapping raw data to nodes and connection edges in the network graph; The user preference measurer refers to the use of the method of integrating attributes and structural similarities to measure the interest preference of user nodes in the information network graph, and the measurement value depends on the attribute description information and structural background information of the nodes. Here, Structural background information refers to the difference between the in-degree neighbor nodes of the user node, and the measurement result will be used as the input of the user matching database; The user matching degree database refers to the similarity value used to store user node pairs. This data will reflect the similarity of interest preferences between users and the data in the user matching degree database will also be selected as the nearest neighbor for active users. basis; The recommendation accelerator refers to using the data in the user matching database as input, using clustering technology to greatly narrow the search range of the user's nearest neighbors, thereby improving the recommendation efficiency; The personalized recommendation processor is based on the user preference model and the data in the user matching degree database, and uses the weight addition method to score the commodities that the active user has not purchased in the candidate commodity database, and the scoring of the active user depends on The recommendation accelerator generates the ratings of the nearest neighbor users on the target items, and then sorts the predicted ratings of the products, and recommends the top N products with the highest ratings to the active users. Return the client. 6. The personalized commodity recommendation system according to claim 5, wherein the user preference similarity calculation method in the user preference measurer comprises the following steps: Step 1, for the attribute information of the user node, utilize the attribute similarity calculation method to obtain the attribute similarity between the user node pairs; Step J, according to the structural background information of the user node, using the structural similarity calculation method to obtain the structural similarity between the user node pairs; Step K: For the user node pairs with similar attributes and similar structures, combine the previous two similarity measures and adjust the weight factor to calculate the preference similarity between the end user pairs.

Images