CN110851705A - Item-based collaborative storage recommendation method and recommendation device - Google Patents

Abstract

The invention discloses a collaborative storage recommendation method and device based on projects, wherein the method comprises the following steps: embedding the user information and the project information respectively to obtain similarity; adding the similarity into an attention mechanism to set a weighting function, and operating the formed weight and the project neighborhood to form neighborhood representation; setting a score function through project information embedding and neighborhood representation; setting a loss function using bayesian personalization-ranking optimization; and realizing the collaborative storage recommendation according to the score function and the loss function. The device comprises: the device comprises a similarity calculation module, a neighborhood representation module, a setting module and a recommendation module. The method effectively overcomes the limitation that only the connection between users is considered, can fully consider the distribution of potential influences between the items and the users, strengthens the relation between a specific item and the neighborhood, considers the potential relation between the items and the users, adopts a higher-order method, enhances the expandability of recommendation, and improves the accuracy and precision of information retrieval and results.

Description

Item-based collaborative storage recommendation method and recommendation device

technical field

The present invention relates to the fields of natural language processing and information retrieval, in particular to an item-based collaborative storage recommendation method and a recommendation device thereof.

Background technique

The recommendation system recommends the information and products of interest to the user according to the user's interest characteristics and historical behavior. The process of browsing a large amount of irrelevant information and products will make the user drowned in the problem of information overload. The personalized recommendation system is based on massive data mining. Based on an advanced intelligent platform, it provides users with fully personalized decision support and information services. In order to improve the accuracy and effect of recommendation, a variety of recommendation methods have emerged. The content-based recommendation method is the continuation and development of information filtering technology. It is based on the content information of the item to make recommendations, and does not need to be based on the user's evaluation of the item. The user's interest information is obtained in the case of the feature description, but it requires the feature content to have a good structure. The recommendation technology based on collaborative filtering is one of the earliest and most successful technologies in the recommendation system. It generally adopts the nearest neighbor technology, uses the user's historical preference information to calculate the distance between users, and then uses the target user's nearest neighbor user to evaluate the product. The weighted evaluation value of the evaluation is used to predict the preference degree of the target user for a specific commodity, and the system recommends the target user according to the preference degree.

In the recommendation, the use of the attention mechanism can greatly improve the sequence learning task. The attention model can perform data-weighted transformation on the data sequence, and the hard attention only pays attention to a certain position of the model sequence at each moment. The soft attention will take care of all the positions each time, and the weight of each position is different. Local attention can be regarded as a mixture of hard attention and soft attention in terms of advantages. Different from hard attention, local attention is almost everywhere differentiable and easy to train. It only pays attention to a small part of the source position at a time, while the global attention Attention needs to scan all source hidden states each time.

However, in the prior art, such as content-based recommendation methods, it requires that the content can be easily extracted into meaningful features, the feature content has a good structure, and the user's preference must be able to be expressed in the form of content features, which cannot be displayed. However, the collaborative filtering technology has sparse and scalability problems in the recommendation process, the recommendation method has poor scalability, and also faces many problems such as cold start.

SUMMARY OF THE INVENTION

The invention provides an item-based collaborative storage recommendation method and a recommendation device. The invention is used for ranking and recommending network retrieval information, which can effectively overcome the limitation of only considering the connection between users, and can fully consider items The distribution of potential impact with users, strengthen the relationship between specific items and neighborhoods, take into account the potential relationship between items and users, adopt a higher-order method, enhance the scalability of the recommendation, improve the recommendation process The accuracy and precision of information retrieval and results are described in the following description:

An item-based collaborative storage recommendation method, the method comprising:

Embed the user and item information respectively to obtain similarity;

The similarity is added to the attention mechanism to set a weighting function, and the formed weight is operated with the item neighborhood to form a neighborhood representation;

Set score function through item information embedding and neighborhood representation; set loss function using Bayesian personalization-rank optimization;

Collaborative storage recommendation is implemented according to the score function and loss function.

Further, the embedding processing is performed on the user and item information respectively, and the obtained similarity is specifically:

where t is a specific item in the item neighborhood with implicit feedback, N(u) is the set of all items that provide implicit feedback to user _u , _mu is the user memory component slice, and ei is the item memory slice.

Among them, the similarity is added to the attention mechanism to set a weighting function, and the formed weight is operated with the item neighborhood, and the neighborhood representation is specifically:

The learning process of the attention mechanism is established with an adaptive weighting function, and different weights are given to the content and items of different influences, and the unique contribution of the learning item _xiut ;

By using the weights formed by the attention mechanism, the weights are manipulated with the neighborhood to form the neighborhood representation s _iu , and the weights formed by the attention mechanism will be selectively weighted to the item and its neighborhood in different ways to get The final weighted neighborhood representation.

Wherein, the unique contribution x _iut is specifically:

Further, the neighborhood representation s _iu is specifically:

s _iu =∑ _t∈N(u) x _iut ×f _t

where _ft is a vector of the item's embedding matrix and × is a matrix-level operation.

Wherein, the setting of the scoring function through item information embedding and neighborhood information is specifically:

Capture the local structure of the neighborhood of the project and the user and the global interaction information between the project and the user, and establish the connection between the project and the user;

The latent relationship between items and users is reflected non-linearly to generate ranking scores that focus on item influence.

An item-based collaborative storage recommendation device, the device comprising:

The similarity calculation module is used to embed the user and item information respectively to obtain the similarity;

The neighborhood representation module is used to add the similarity to the attention mechanism to set the weighting function, and operate the formed weight with the item neighborhood to form the neighborhood representation;

Set module to set score function through item information embedding and neighborhood representation; set loss function using Bayesian personalization-rank optimization;

The recommendation module is used to implement collaborative storage recommendation based on the score function and loss function.

The beneficial effects of the technical scheme provided by the present invention are:

1. The present invention can effectively improve the accuracy of information acquisition in the recommendation process by considering the potential impact of items on users at the item level, better and more comprehensive learning;

2. The present invention pays attention to item information and features, which can capture the attractiveness of items to users more comprehensively and in detail, improve the accuracy of recommendation results, and improve the scalability of recommendation. The final experimental results show that this method is better than only focusing on users. The results of the recommendation algorithm based on the relationship between historical behavior and users are more accurate.

Description of drawings

1 is a flowchart of a project-based collaborative storage recommendation method;

Fig. 2 is an overall architecture diagram of an item-based collaborative storage recommendation method;

FIG. 3 is a schematic structural diagram of an item-based collaborative storage recommendation device.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the embodiments of the present invention are further described in detail below.

Example 1

In order to achieve the above object, an embodiment of the present invention proposes an item-based collaborative storage recommendation method, see FIG. 1 and FIG. 2 , and the method includes the following steps:

101: Embed the user and item information respectively to obtain similarity;

102: Add the obtained similarity to the attention mechanism to set a weighting function, and operate the formed weight with the item neighborhood to form a neighborhood representation;

103: Set a score function through item information embedding and neighborhood information;

104: Use Bayesian personalization-rank optimization to set the loss function;

105: Experiment to evaluate and verify the effect of the item-based collaborative storage recommendation method proposed by the present invention according to the score function and the loss function.

In one embodiment, step 101 performs embedding processing on the user and item information respectively, and the specific steps are as follows:

The entered user and item information is embedded in the user and item memory components, respectively, where the user's memory component is M, the item's memory component is E, the user's preferences are stored in the user memory component slice _mu , and the item i's memory component is The features are stored in the item memory slice ei, and the similarity a _iut in its given neighborhood for a specific item and a specific user is _calculated .

In one embodiment, step 102 adopts an attention mechanism (known to those skilled in the art, which will not be repeated in this embodiment of the present invention) on the basis of step 101 to set a weighting function, and the weights formed by using the attention mechanism are related to neighbors. The domain is operated to form a neighborhood representation. The specific steps are as follows:

Using adaptive weighting function to establish the learning process of attention mechanism, it can give different weights to content and items with different influences, highlighting the importance of high-impact content, after learning the unique contribution of items x _iut , by using attention The weight formed by the mechanism operates on the weight and the neighborhood to form the neighborhood representation s _iu , and the weight formed by the attention mechanism will be selectively weighted to the item and its neighborhood in different ways to obtain the final weighted The neighborhood representation of .

In one embodiment, step 103 sets a score function through item information embedding and neighborhood information on the basis of steps 101 and 102, and the specific steps are as follows:

Capture the local structure of the neighborhood of the project and the user and the global interaction information between the project and the user, and establish a deeper and wider connection between the project and the user. potential relationships to yield ranking scores focusing on project impact

In one embodiment, step 104 sets a loss function to optimize the model, and the specific steps are as follows:

Use Bayesian Personalized Rank Optimization to set up a loss function and optimize the model.

In one embodiment, step 105 conducts an experiment on an item-based collaborative storage recommendation method, and the specific steps are as follows:

The hit rate of the experiment and the cumulative gain of normalized damage are calculated to evaluate the effect of the model. In order to better balance these two indicators, eight baseline comparison experiments are used to evaluate and verify the effect of the algorithm.

Example 2

The scheme in Embodiment 1 is further introduced below in conjunction with specific calculation formulas and examples, and is described in detail below:

201: Perform embedding processing on the user and item information respectively, and the inputted user and item information is embedded in the memory components of the user and the item respectively;

where the user's memory component is M, the item's memory component is E, the user's preferences are stored in the user memory component slice _mu , and the features of item _i are stored in the item memory slice ei, to obtain specific items and The similarity a _iut of a particular user in its given neighborhood is shown in formula (1).

where t in formula (1) is a specific item in the item neighborhood with implicit feedback, and N(u) is the set of all items that provide implicit feedback to user u.

202 : Use the attention mechanism to set the weighting function. The embodiment of the present invention uses the adaptive weighting function to establish the learning process of the attention mechanism, and the unique contribution x _iut of the learning item is shown in formula (2).

Among them, a _iut in formula (2) is the similarity between a specific item and a specific user in its given neighborhood.

203: Using the weights formed by the attention mechanism, operate the weights with the neighborhood to form the neighborhood representation s _iu , as shown in formula (3).

s _iu =∑ _t∈N(u) x _iut ×f _t (3)

In formula (3), f _t is the vector of another embedding matrix of the item, and × is the matrix-level operation. According to the characteristics of the item, the weights formed by the attention mechanism will be selectively weighted to the item and its neighbors in different ways. domain to get the final weighted neighborhood.

203: Set a score function through item information embedding and neighborhood information, the specific steps are as follows:

如公式(4)、公式(5)和公式(6)所示。Capture the local structure of the neighborhood of the project and the user and the global interaction information between the project and the user, and establish a deeper and wider connection between the project and the user. potential relationship between, by linearly projecting U to generate ranking scores of attention item influence

As shown in formula (4), formula (5) and formula (6).

为存储器，为第y层的相似矩阵，v、b为要学习的参数，

为经过项目邻域更新后的邻域，

为存储器之间的非线性映射，e_t为用户t的独特贡献，

为第四层的邻域矩阵，σ(x)＝1/(1+exp(-x))为非线性激活sigmoid函数，y为存储器的层。Among them, in formula (4)(5)(6)

for memory, is the similarity matrix of the y-th layer, v and b are the parameters to be learned,

is the neighborhood after project neighborhood update,

is the nonlinear mapping between memories, e _t is the unique contribution of user t,

is the neighborhood matrix of the fourth layer, σ(x)=1/(1+exp(-x)) is the nonlinear activation sigmoid function, and y is the memory layer.

Among them, in the present invention, y is 4, and W is a weight matrix, which maps the features of the item to the latent space and combines it with the information of the previous layer. It is an element-level operation. After obtaining the elements of the item and the user After multiplication, the learned neighborhood representation and parameters are manipulated by linear projection. Using nonlinear activation ReLU function

204: Use Bayesian personalized ranking optimization to set a loss function to optimize the recommendation effect, as shown in formula (7).

σ(x)=1/(1+exp(-x)) in formula (7) is a logical sigmoid function.

205: Through the calculation of the hit rate and the normalized impairment cumulative gain, the effect of the method of the present invention can be evaluated and verified.

Example 3

An embodiment of the present invention provides an item-based collaborative storage recommendation device. Referring to FIG. 3 , the device includes:

The similarity calculation module is used to embed the user and item information respectively to obtain the similarity;

That is, item embedding matrices E, F and user embedding matrix M are obtained.

The neighborhood representation module is used to add the similarity to the attention mechanism to set the weighting function, and operate the formed weight with the item neighborhood to form the neighborhood representation;

Set module to set score function through item information embedding and neighborhood representation; set loss function using Bayesian personalization-rank optimization;

The recommendation module is used to implement collaborative storage recommendation based on the score function and loss function.

Example 4

In the experiment of the item-based collaborative storage recommendation method and its recommendation device, the final score of each item is calculated by the score function, and the ranking and recommendation are carried out by the score.

It can be seen from the experimental results that after introducing the potential impact of items on users, through the optimization of attention mechanism and loss function, the item-based collaborative storage recommendation method and its recommendation device have certain linguistic significance and excellent effects.

In the experiments of the item-based collaborative storage recommendation method and its recommendation device, eight baseline experiments are used to conduct comparative experiments, the number of negative samples in the experiment is set to 4, the weight decay of l2 The weight decay of l2 is set to 0.1, the memory embedding size d is 50, and the item information embedding matrix and user information embedding matrix are randomly and automatically generated.

The present invention evaluates the item-based collaborative storage recommendation method and the effect of the recommendation device by using two evaluation indicators: hit rate (HR) and normalized impairment cumulative gain (NDCG) value. The main purpose of the hit rate calculation is to evaluate the model to obtain The proportion of the number of items in the top-N recommendation list in the test set, and the main purpose of the normalized discount cumulative gain calculation is to evaluate the comparison between the effect of the recommendation list generated by the model and the effect of the recommendation list generated under ideal conditions. The calculation formula of the hit ratio (HR) is shown in formula (8), and the calculation formula of the normalized impairment cumulative gain (NDCG) is shown in formula (9).

The data value |GT| in formula (8) refers to all test sets, and NumberofHits@K is the sum of the number of test sets belonging to each user's top-K recommendation list. In formula (9), DCG is the average discounted cumulative gain, and ideal DCG@K is the maximum DCG value under ideal conditions. The larger the value of the hit rate and the normalized damage cumulative gain, the better the effect.

Table 1 is the experimental evaluation index table

The experimental results of the nine baseline comparison experiments are shown in Table 1. SVD++ is a hybrid model that combines neighborhood-based similarity and latent factor models. The generalized matrix factorization (GMF) model is a nonlinear extension of the latent factor model. KNN's neighborhood-based method for computing cosine item-item similarity. Bayesian Personalization Ranking (BPR) is a matrix factorization model with implicit feedback. Neural Matrix Factorization (NeuMF) is a matrix factorization model for item ranking through a multilayer perceptron model. Collaborative Denoising Autoencoder (CDAE) is an item-based deep learning model. The Factor Term Similarity Model (FISM) is a neighborhood-based model that optimizes the loss function by decomposing the item-item pair similarity matrix. CMN is a model that fuses memory components with attention mechanisms. It can be seen from the experimental results that the experimental effect of the item-based collaborative storage recommendation method and its recommendation device has the highest hit rate and normalized impairment cumulative gain, indicating that the model is effective.

Those skilled in the art can understand that the accompanying drawing is only a schematic diagram of a preferred embodiment, and the above-mentioned serial numbers of the embodiments of the present invention are only for description, and do not represent the advantages or disadvantages of the embodiments.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (7)

1. A project-based collaborative storage recommendation method, wherein the method comprises: Embed the user and item information respectively to obtain similarity; The similarity is added to the attention mechanism to set a weighting function, and the formed weight is operated with the item neighborhood to form a neighborhood representation; Set score function through item information embedding and neighborhood representation; set loss function using Bayesian personalization-rank optimization; Collaborative storage recommendation is implemented according to the score function and loss function. 2. The item-based collaborative storage recommendation method according to claim 1, wherein the embedding processing is performed on the user and item information respectively, and the obtained similarity is specifically:

where t is a specific item in the item neighborhood with implicit feedback, N(u) is the set of all items that provide implicit feedback to user _u , _mu is the user memory component slice, and ei is the item memory slice. 3. A project-based collaborative storage recommendation method according to claim 2, wherein the similarity is added to the attention mechanism to set a weighting function, and the formed weight is operated with the project neighborhood to form a neighborhood. The domain representation is specifically: The learning process of the attention mechanism is established with an adaptive weighting function, and different weights are given to the content and items of different influences, and the unique contribution of the learning item _xiut ; By using the weights formed by the attention mechanism, the weights are manipulated with the neighborhood to form the neighborhood representation s _iu , and the weights formed by the attention mechanism will be selectively weighted to the item and its neighborhood in different ways to get The final weighted neighborhood representation. 4. a kind of project-based collaborative storage recommendation method according to claim 3, is characterized in that, described unique contribution x _iut is specifically: 5. a kind of item-based collaborative storage recommendation method according to claim 3, is characterized in that, described neighborhood representation s _iu is specifically: s _iu =∑ _t∈N(u) x _iut ×f _t where _ft is a vector of the item's embedding matrix and × is a matrix-level operation. 6. a kind of item-based collaborative storage recommendation method according to claim 1, is characterized in that, described by item information embedding and neighborhood information setting score function is specifically: Capture the local structure of the neighborhood of the project and the user and the global interaction information between the project and the user, and establish the connection between the project and the user; The latent relationship between items and users is reflected non-linearly to generate ranking scores that focus on item influence. 7. An item-based collaborative storage recommendation device, wherein the device comprises: The similarity calculation module is used to embed the user and item information respectively to obtain the similarity; The neighborhood representation module is used to add the similarity to the attention mechanism to set the weighting function, and operate the formed weight with the item neighborhood to form the neighborhood representation; Set module to set score function through item information embedding and neighborhood representation; set loss function using Bayesian personalization-rank optimization; The recommendation module is used to implement collaborative storage recommendation based on the score function and loss function.

Images