CN103824213A - Individualized recommendation method based on user preferences and commodity properties - Google Patents

Abstract

The invention relates to an individualized recommendation method based on user preferences and commodity properties, belonging to the field of machine learning. The individualized recommendation method comprises the following steps of adding attribute information of a project and doing recommendation by commodity property information when preference information does not exist. Meanwhile, the recall ratio of a recommendation system is improved by the recommendation method. With the adoption of the individualized recommendation method based on the user preferences and the commodity properties, the problem of cold starting based on a new project is solved.

Description

A kind of personalized recommendation method based on user preference and item property

Technical field

The present invention relates to a kind of personalized recommendation method based on user preference and item property, belong to machine learning field.

Background technology

Along with further developing of intelligent computation and e-commerce technology, Collaborative Filtering Recommendation System research has become the hot subject in e-commerce field.Collaborative filtering is an important technology in ecommerce, about the research of collaborative filtering recommending technology, be mainly at present under the Internet model,, cluster analysis integrated from user gradation, project scoring, information filtering, language ambience information, Association Rule Analysis equal angles are inquired into.Basic thought can be summarized as in the situation that guaranteeing that score data is abundant, finds the nearest-neighbors user who has same or similar user behavior with targeted customer, selects the highest commodity of user behavior similarity to feed back to targeted customer as recommendation list collection.

A major issue of recommended engine is cold start-up.The reason that cold start-up produces is cannot make reliable recommendation owing to lacking initial ordering of optimization preference.We have three kinds at current known cold start-up problem: new communities, new projects and new user.The cold start-up of new user type is the already present middle sixty-four dollar question of commending system.New communities' problem refers to that the very difficult data (preference) of simultaneously obtaining mass efficient are made recommendation in the time of startup commending system.New projects' problem refers in the time having new project to join in commending system, because this project does not have ordering of optimization preference, therefore this project can not recommended to user.Conversely, not recommended article will be ignored by user, and therefore these article will can not given preference value by user always, like this, will be absorbed in a circulation, cause these uncared-for article always outside preference/recommendation process.New projects' problem can represent that the problem (as film) of new projects wants large many to the impact of some problems (ecommerce, microblogging, picture, video etc.) by other approach than those.One of method that solves new projects is in commending system, to exist some to make specially the user of sequence for new projects.New customer problem is a very large problem in commending system.Because new user in ordering of optimization preference without any information, therefore, in the collaborative filtering method based on memory, make any recommendation cannot to new user.When new user submits to after their ordering of optimization preference, they expect that commending system can make recommendation to them, but the data (ordering of optimization preference) that they submit to may be very little to such an extent as to cannot be made effectively reliably and recommending, therefore, new user may think that commending system does not reach their expectation, thereby abandons using commending system.Conventionally the approach that solves new customer problem is to add additional information to ordering of optimization preference, thus according to each user can with information make recommendation.

The present invention, in order to solve the cold start-up problem based on new projects, has proposed a kind of personalized recommendation method based on user preference and item attribute.

Summary of the invention

A kind of personalized recommendation method based on user preference and item attribute involved in the present invention, technical scheme is specially:

Step 1, definite similar matrix based on item attribute;

Defined item object proper vector: item _i=(p ₁, p ₂..., p _m); The attribute number that wherein m is project, p _i(1≤i≤m) represented value of i proper vector of this project.Then each project can be converted to a vectorial item _i=(w ₁, w ₂..., w _m) represent, wherein vectorial dimension is m, i.e. the attributive character number of project.Then by the distance A between reckoner aspect object vector _ijrepresent item _iand item _jbetween similarity, thereby form similar matrix

Under the computing method of distance between vector between project i and project j comprise: distance that Pearson came is relevant, Euclidean distance, cosine distance, Spearman distance and based on paddy relevant distance.

Step 2, definite co-occurrence matrix based on user preference;

The list of preferences of definition user to project: prefs=(user_id, item_id, pref), the wherein scoring of pref representative of consumer to project, the scoring composition scoring list prefs of all users to project.Appear at the number of times B in same user's list of preferences by calculating every a pair of project simultaneously _ij(represent item _iand item _jappear at the number of times in identical user's list of preferences simultaneously) form co-occurrence matrix

Step 3, determine final similar matrix;

Final similar matrix is defined as

wherein

with β be self-defining weight.

Step 4, compute user preferences vector:

User's preference is regarded as to a vector, and in the data model of n project, user's preference is a n dimensional vector; User's preference value is corresponding to the element in vector, and the preference not providing is by 0 expression; There is a user preference vector P with respect to each user _{(u, j)}, 1<j<n, represents the preference of user u to project j, the preference vector of the final user u generating is

The recommendation that step 5, calculated candidate are recommended article;

The user preference vector that the similar matrix obtaining according to step 3 and step 4 obtain, calculated candidate is recommended the recommendation of article;

For user u, the recommendation scores of project i is:

$P_{(u,j)}=\underset{j<n}{\mathrm{\&Sigma;}}{\mathrm{sim}}_{(j,i)}{P}_{(u,j)}$

Wherein, sim _{(j, i)}the similarity of expression project j and project i; P _{(u, j)}represent the preference value of user u to project j;

For the recommendation of user u all items be

Step 6, to user u, calculate the recommendation scores of all items, and the project that is zero using preference value recommends article as candidate, and recommends the recommendation of article to sort according to order from big to small to candidate;

Step 7, on the basis of above-mentioned steps sequence, choose top n article and recommend user u, N thinks a certain positive integer of setting;

Through the operation of above-mentioned steps, complete the article of user u are recommended.

Beneficial effect

The present invention, by using the personalized recommendation method based on user preference and item property, solves the cold start-up problem based on new projects.By adding the attribute information of attribute, in the time not there is not preference information, make recommendation by information attribute value.Improve the recall ratio of commending system by this recommend method simultaneously.

Accompanying drawing explanation

Fig. 1 is the idiographic flow schematic diagram of personalized recommendation method involved in the present invention

Embodiment

Below by embodiment, the specific embodiment of the present invention is described in further details.

In certain website, there are 1000 of users, 5000 of films, every film has title, sells the time, 3 kinds of attributes of classification, now use the personalized distributed recommendation method based on improved similar matrix to recommend article to the 1st user in this website, as shown in Figure 1, operation steps is as follows for its idiographic flow:

According to step 1: determine project-based similar matrix;

The proper vector of definition film: item _i=(p ₁, p ₂,, p ₃), p _i(1≤i≤3) have represented the value of i feature of this project.First every film is represented to item with 3 dimensional vectors _i=(w ₁, w ₂,, w ₃), wherein w _i(1≤i≤3) represent the value of i feature of article.Then by the distance A between reckoner aspect object vector _ijrepresent item _iand item _jbetween similarity, thereby form similar matrix

Under obtain similarity by distance between project u and project v computing method employing Euclidean distance calculate.

According to step 2: determine the co-occurrence matrix based on user preference;

The list of preferences of definition user to project: prefs=(user_id, item_id, pref), the wherein scoring of pref representative of consumer to project, the scoring composition scoring list prefs of all users to project.Appear at the number of times B in same user's list of preferences by calculating every a pair of project simultaneously _ij(represent item _iand item _jappear at the number of times in identical user's list of preferences simultaneously) form co-occurrence matrix

According to step 3: determine final similar matrix;

Final similar matrix is defined as

wherein α and β are respectively 0.5.

According to step 4: compute user preferences vector;

Based on the distributed algorithm of matrix, user need to be regarded as to a vector to the preference of each project, in the data model of n project, user's preference is a n-dimensional vector.User's preference value is corresponding to the element in vector, and the preference not providing is by 0 expression.There is a user preference vector P with respect to each user like this _{(u, j)}, 1<j<n, represents the preference of user u to project j.The final user's 1 who generates preference vector is

Described preference value is the score information of user to article in website, and score value is from 1 to 5.

According to step 5: calculated candidate is recommended the recommendation of article;

Recommend the recommendation of article according to following formula calculated candidate according to the similar matrix arriving of above-mentioned steps and user preference vector;

For user u, the recommendation scores of project i is:

$P_{(u,j)}=\underset{j<n}{\mathrm{\&Sigma;}}{\mathrm{sim}}_{(j,i)}{P}_{(u,j)}$

Wherein, sim _{(j, i)}the similarity of expression project j and project i; P _{(u, j)}represent the preference value of user u to project j;

For the recommendation of user's 1 all items be

According to step 6: to user 1, calculate the recommendation scores of all items, and the project that is zero using preference value recommends article as candidate, and recommend the recommendation of article to sort according to order from big to small to candidate;

According to step 7: on the basis of above-mentioned steps sequence, choose front 50 article and recommend user 1.

Through the operation of above-mentioned steps, complete user 1 article are recommended.

Claims (1)

1. the personalized recommendation method based on user preference and item property, is characterized in that:

Step 1, definite similar matrix based on item attribute; By the distance A between the vector of computational item _ijrepresent the similarity between vector, build similar matrix:

Step 2, definite co-occurrence matrix based on user preference; The list of preferences of definition user to project: prefs=(user_id, item_id, pref), the wherein scoring of pref representative of consumer to project, the scoring composition scoring list prefs of all users to project; Appear at the number of times B in same user's list of preferences by calculating every a pair of project simultaneously _ij(represent item _iand item _jappear at the number of times in identical user's list of preferences simultaneously) form co-occurrence matrix

Step 3, determine that final similar matrix is:

wherein

with β be self-defining weight;

Step 4, compute user preferences vector: have a user preference vector P with respect to each user _{(u, j)}, 1 < j < n, represents the preference of user u to project j,

The preference vector of the final user u generating is

Step 5: calculated candidate is recommended the recommendation of article; The user preference vector that the similar matrix obtaining according to step 3 and step 4 obtain, calculated candidate is recommended the recommendation of article;

For user u, the recommendation scores of project i is:

$P_{(u,j)}=\underset{j<n}{\mathrm{\&Sigma;}}{\mathrm{sim}}_{(j,i)}{P}_{(u,j)}$

Wherein, sim _{(j, i)}the similarity of expression project j and project i; P _{(u, j)}represent the preference value of user u to project j; Recommendation for user u all items is:

Step 6, to user u, calculate the recommendation scores of all items, and the project that is zero using preference value recommends article as candidate, and recommends the recommendation of article to sort according to order from big to small to candidate;

Step 7, on the basis of above-mentioned steps sequence, choose top n article and recommend user u, N thinks a certain positive integer of setting.

Images