CN110442781B - Pair-level ranking item recommendation method based on generation countermeasure network - Google Patents

Abstract

The invention discloses a method for recommending a pair-level ranking project based on a generated confrontation network. And secondly, extracting the user interaction features and the project interaction features from the feedback data of the user through an embedding layer. And then, fusing the interactive features and the comment features of the project, generating a personalized negative sample for the user by using a generator for generating a confrontation network, updating parameters through the confrontation training, and continuously improving the recommendation performance of the generator until the recommendation result of the generator is converged finally. The method can effectively mine the comment characteristics of the items from the comments to relieve the sparsity of implicit feedback data, and overcomes the weakness of the traditional sampling mode by utilizing the personalized selection negative sample in the process of generating the countermeasures, thereby improving the accuracy of recommendation.

Description

Pair-level ranking item recommendation method based on generation countermeasure network

Technical Field

The invention belongs to the technical field of data mining, and particularly relates to a pair-level ranking item recommendation method based on a generation countermeasure network in the field of intelligent recommendation.

Background

The development of the internet brings more convenient services and more diversified information to people, and with the increasing scale of the contacted information, users inevitably suffer from the problem of information overload, so that the target information is more and more difficult to acquire quickly. The recommendation system is an important means for solving the information overload problem, and has important influence on the retention of customers and the success of Internet enterprises. The recommendation system learns the characteristics of the user from the historical feedback of the user, so that items which accord with the preference of the user are selected from a large number of item sets, and personalized recommendation is carried out on the user. In real life, users express their preferences in an implicit feedback manner, such as whether to click or not, whether to purchase or not, and the meaning of the item expression which is not observed is uncertain, so that the recommendation of items based on implicit feedback is more challenging.

The existing recommendation methods based on implicit feedback are mainly divided into two categories: point-level regression and rank ordering. The point-level regression considers that the unobserved items are not preferred by the user ([ document 1 ]), optimizes the prediction scores of the items by the user through regression, and sorts the items according to the degree of the prediction scores. The rank ordering can utilize unobserved data, and data sparsity is relieved to a certain extent. Document 2 proposes a Bayesian Personalized Ranking algorithm (BPR) based on the assumption that users prefer the observed items. And converting the item recommendation problem into a binary classification problem of whether the user prefers the item i or not in a partial order relation (i, j), and generating a recommendation list through the partial order relation between every two items. Zhao et al proposed that SBPR introduced social relationships into BPR for modeling ([ document 3 ]). Document 4 designs a ranking model including visual information, extracts visual features using a deep neural network, and generates a personalized ranking more suitable for user preferences as an aid to interactive information. However, the above methods mostly adopt a random sampling mode to select the negative sample, and the correlation between the negative sample and the user is ignored.

In recent years, with the continuous and deep learning research and the great success in the fields of image, voice, natural language processing and the like (document 5 to document 8), the introduction of deep learning into the recommendation field becomes a hot point of research, and the data sparsity and cold start problems can be effectively alleviated by utilizing the characteristic of mining auxiliary information by a deep neural network. Document 9 extracts context features of item reviews by using a stacked noise reduction auto-encoder, and tightly combines the extraction of the review features with collaborative filtering of a rating matrix, allowing bidirectional interaction between the two; document 10 constructs a word vector matrix using word vectors obtained by unsupervised learning, and classifies texts by extracting semantic features of the texts through convolution and pooling. Document 11 extracts an effective feature representation using a convolutional neural network and reconstructs a user-item score matrix by integrating probability matrix decomposition. The convolutional neural network overcomes the inherent limitation of the bag-of-words model and can successfully learn the semantic features of the text.

Document 12 proposes generation of a countermeasure network for the first time, and the basic idea is to perform an extremely small game between a generator and a discriminator, and to perform iterative training and mutual promotion. Since the generator has the ability to generate samples, it has been successfully used in many fields ([ document 13 ]). Wang et al introduced the concept of generating countermeasures into the information retrieval field (IRGAN) ([ document 14 ]), designed a generator for generating discrete data that meets the characteristics of the recommendation field, and presented a feasible optimization scheme; the item recommended by the generator is judged to conform to the preference of the user by the judger. Document 15 proposes a countermeasure pair learning recommendation system that applies the idea of generating a countermeasure network to ranking learning, generates one negative sample with a generator, provides another solution to the problem of optimization of discrete samples generated by a generator, and increases the convergence speed and improves the stability of a countermeasure model. Document 16 proposes to combine long-term score information and short-term session information to recommend a next movie to a user using antagonism training in content-aware movie recommendation. However, currently, research for fusing comment information in the generation of confrontation networks and relieving data sparsity to recommend projects is still few.

[ document 1] Pan R, zhou Y, cao B, et al, one-class collectible filtering// Proceedings of the Eighth IEEE International Conference on Data mining. IEEE, 2008.

[ document 2] Rendle S, freedenthaler C, gantner Z, et al.BPR.

[ document 3] ZHao T, mcAuley J, king I.squaring social connections to improve personal communication for a social filtering// Proceedings of the 23rd ACM international conference on information and knowledge management.ACM, 2014.

[ document 4] He R, mcAuley J.VBPR.

[ document 5] Chen Y S, wang Y C, kao M H, et al. Deep photo enhancer.

[ document 6] Nassif A B, shahin I, attili I, et al, speech registration Using Deep Neural networks

[ document 7] ZHao J, mao X, chen L.learning deep defects to recording deep CNN.IET Signal processing,2018,12 (6): 713-721.

[ document 8] Wu Y, schuster M, chen Z, et al Google's neural machine transformation system.

[ document 9] Wang H, wang N, yeung D Y.Collaborative deep learning systems// Proceedings of the 21th ACM SIGKDD international conference on knowledge display and data mining. ACM 2015 1235-1244.

[ document 10] Kim Y.capacitive neural networks for sensor classification// arXiv preprinting arXiv:1408.5882,2014.

[ document 11] Kim D, park C, oh J, et al. Capacitive matrix factorization for document context-aware registration// Proceedings of the 1th ACM Conference on recommendation systems. ACM, 2016.

[ document 12] Goodfellow I, pouget-Abadie J, mirza M, et al. Genetic additive networks// Advances in neural information processing systems 2014:2672-2680.

[ document 13] Creswell A, white T, dumoulin V, et al. Genetic additive networks.

[ document 14] Wang J, yu L, zhang W, et al, irgan.

[ document 15] Sun Z, wu B, wu Y, et al. APL

[ document 16] ZHao W, chai H, wang B, et al, leveraging Long and short-term information in content-aware movie recommendation. ArXiv preprintin arXiv:1712.09059,2017.

Disclosure of Invention

In view of the problems and deficiencies of the prior art, the present invention provides a method for recommending ranked items based on a generative confrontation network. The key point is that a generator for generating the confrontation network is used for generating an individualized negative sample, the effectiveness of parameter reasoning is improved, comment modeling is performed through the convolutional neural network to relieve data sparsity, and the recommendation performance is improved.

The technical scheme adopted by the invention is as follows: a method for recommending items in a ranking order based on a generation countermeasure network, comprising the following steps:

step 1: training data is collected, implicit feedback (such as whether clicking, purchasing, collecting and the like) given by a user and item comment information are filtered, and the filtered comment information is preprocessed.

Step 2: converting the item comment information into a word vector matrix by using a GloVe word vector model, and learning potential context characteristics of the comment information as comment characteristics of the item by using a convolutional neural network; implicit feedback given by a user is converted into user interaction characteristics and project interaction characteristics through the embedding layer.

And step 3: optimizing parameters in the training process of the confrontation;

selecting a negative sample from unobserved items in a generator in a generated countermeasure network according to the prediction score in a personalized manner, forming a partial order relationship pair with an observed positive sample, judging whether the generated partial order relationship is correct or not through a discriminator, and guiding the generator to update parameters according to the judged result; carrying out iterative confrontation training on the two generators until the recommendation result of the final generator is converged;

and 4, step 4: and outputting a converged recommendation result.

Preferably, in step 1, implicit feedback and comment information are filtered to obtain interactive data of all user IDs and item IDs which actually exist; and then preprocessing the filtered comment information, only keeping the real words, performing connection operation on the comments of the same item by all users, and supplementing the comments of all items to a uniform length.

Preferably, in the step 2, for the item comment information, a word vector of each word is obtained through GolVe word vector model mapping, a word vector matrix corresponding to the item comment information is obtained, the matrix is convolved through sharing weight, and then a vector with a specified dimension is obtained through pooling and full connection and is used as a comment feature of the item to be output;

for implicit feedback given by a user, mapping one-hot codes of sparse users or items through a potential feature matrix at an embedding layer to obtain a low-dimensional potential factor vector.

Preferably, in step 3, the interactive features and the comment features of the project are fused, and the prediction score of the project by the user is calculated through a linear inner product;

in the process of generating the countermeasure, the generator calculates the prediction score, converts the prediction score into the distribution of generated samples through a Softmax function, generates negative samples for the user according to the distribution, and forms a partial order relation with the sampling of the positive samples;

the discriminator judges whether the generated partial order relation is correct or not and whether the user prefers a positive sample or not through the prediction score; the discriminator wants to update the parameters such that the distinction between positive and negative samples is as large as possible, and therefore the parameter phi of the discriminator is updated as follows:

wherein U representsSet of users, i ⁺ ,i ^- Respectively representing positive and negative samples, p _true ,G _θ Respectively representing the distribution of the real and generated samples, D being the discriminator, D (u, i) ⁺ ,i ^- ) Judging the relation of partial order<i ⁺ ,i ^- >Whether correct, whether user u prefers positive samples;

updating the parameters of the generator firstly requires the generator to generate a partial order relation formed by a batch of negative samples and real samples, the discriminator discriminates the generated partial order relation, and the discrimination result logD (u, i) is ⁺ ,i ^- ) An update as a reward guide generator;

where I represents the set of items and theta represents the generator parameter, and the generator parameter theta is updated by minimizing the objective function.

In each iteration, a generator is fixed, parameters of the discriminator are updated, and the aim is to enable the discriminator to give a correct judgment result under the current condition; then, fixing the discriminator, training the generator to continuously update own parameters according to the discrimination result of the current discriminator, and enabling the generated negative samples to be more difficult to distinguish; the above iterative training is carried out by the two until the recommendation result of the final generator is converged.

Compared with the prior art, the invention adopting the technical scheme has the following technical effects:

1. and acquiring potential feature representation of the item comment information by using a convolutional neural network, acquiring context information of the comment by sharing weight, and integrating the context features obtained in the previous step into a specified dimension for outputting through pooling and full connection. Compared with the traditional method for acquiring the comment theme by using the bag-of-words model, the method extracts the context semantic features of the comment, and more accurately captures the attribute of the item;

2. the method integrates the interactive characteristics of the project and the comment characteristics of the project acquired from the comment information, and supplements implicit feedback data by using the comment information as auxiliary information to relieve the sparsity of the data;

3. conventional random sampling does not reflect the correlation between the sampled samples and the user. In the process of confrontation, the generator is used for capturing the preference of the user, negative samples are selected from unobserved items in a personalized mode, and the parameters of the generator are updated according to the judgment result of the judger. The defects of the traditional sampling method are overcome, and the effectiveness of parameter reasoning is improved.

By combining the three points, the method for recommending the item in the ranking order based on the generated confrontation network can learn the user preference and the item attribute more comprehensively and accurately, so that the performance of a recommendation system is improved.

Drawings

FIG. 1 is a flow chart of an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

The context characteristics of the item comment information are learned by using the convolutional neural network, the interaction characteristics of the item and the comment characteristics of the item are fused, and the sparsity of data is relieved. And finally, generating an individualized negative sample through the generator, judging the effect of the generated sample by using the discriminator, updating the parameters of the generator and the discriminator in the training process of the confrontation, and continuously improving the recommendation performance of the generator.

As shown in fig. 1, the invention discloses a method for recommending ranked items based on a generated confrontation network, which comprises the following steps:

step 1: training data is collected, implicit feedback (such as whether clicking, purchasing, collecting and the like) given by a user and item comment information are filtered, and the filtered comment information is preprocessed.

In the field of recommendations, the feedback of a user is divided into explicit feedback, which may explicitly indicate the likes and dislikes of the user, typically expressed in terms of scores or ratings, a high rating indicating that the user likes the item, and a low rating indicating that the user dislikes the item. While another feedback, such as a click, we can only consider items that the user clicked on to be likely to be liked by the user, but items that the user did not click on to be simply regarded as disliked are more likely to be not seen by the user, and this feedback is called implicit feedback.

In this embodiment, the implicit feedback and the item comment information are first filtered to obtain all the interactive data in which the user ID and the item ID actually exist. And then preprocessing the filtered comment information, only keeping real words, connecting all comments of the same item by all users, and filling all comments to the length by taking 0.85 of the longest comment length as a standard.

Step 2: converting the comment information into a word vector matrix by using a GloVe word vector model, and learning potential context characteristics of the comment information as comment characteristics of a project by using a convolutional neural network; the structured implicit feedback given by the user is converted into user interaction features and item interaction features through the submerged layer.

The item comment information is the evaluation of a user on an item represented by unstructured text, feature extraction can be carried out from two dimensions, the features of the user are extracted according to comments given by the user, the features of the item are extracted according to comments given by different users, and the comment features of the item are extracted from the comments. While implicit feedback information, i.e. user and item interactions (clicks, purchases, favorites, etc.), e.g. '1' for clicked items and '0' for non-clicked items, is a structured data, we extract user interaction features and item interaction features from the implicit feedback, respectively.

In this embodiment, for the comment information obtained through preprocessing, each real word is mapped through a word vector model Glove to obtain a corresponding word vector, so that the comment information of the item i is converted into a word vector matrix:

wherein

Indicating a concatenate operation, i.e. combining word vectors into a matrix, based on>

For a word vector, d represents the dimension of the word vector in the GloVe model, and l represents the number of words in the comment, i.e. the length of the comment.

And taking the obtained word vector matrix as the input of a convolutional neural network, extracting context characteristics in the comments through convolution operation, selecting the most representative characteristics through a pooling layer, and converting the characteristics into specified dimensions through a full connection layer for outputting. For convenience of representation, the present embodiment represents the process as:

/>

representing item review features extracted from the review information, W representing a convolution kernel, X _i The word vector matrix for item i obtained in the previous step, b is the offset vector added to prevent overfitting.

In this embodiment, for implicit feedback information given by a user, the implicit feedback information is converted into a low-dimensional implicit vector representation through latent factor matrix mapping in an embedding layer.

Wherein

Respectively represent user latent factor matrices anda project latent factor matrix. m, n respectively represent the number of users and the number of items, k is the dimension of the characteristic vector, v _u ,v _i And mapping the feature vectors of the users and the items to obtain low-dimensional potential factor vectors of the users and the items.

And step 3: the parameters are optimized during the training of the confrontation. And the generator in the generation countermeasure network selects a negative sample from the unobserved items in a personalized manner according to the prediction scores, forms a partial order relation pair with the observed positive sample, judges whether the generated partial order relation is correct or not through the discriminator, and guides the generator to update parameters according to the judged result. The two are subjected to iterative antagonistic training until the recommendation of the final generator converges.

In this embodiment, the detailed steps of step 3 are:

step 3.1: first, the interactive characteristics of the items learned from the implicit feedback data

And item review feature learned in review>

Perform fusion, and user characteristics p _u Taking a linear inner product results in a user's predictive score for an item>

Wherein, b ⁱ Indicating the bias of item i.

Step 3.2: the generator obtains the distribution of the generated samples through the prediction scores, the prediction scores represent the estimation of the user preference, and the personalized negative samples which accord with the user preference can be obtained by sampling according to the distribution. This example uses the Softmax function to convert the predicted scores into the distribution G of the samples _θ And g (i | u) is the correlation degree between the user u and the item i and is often used in the recommendation fieldThe prediction score is often used for presentation. t represents the sampling temperature, with lower sampling temperatures focusing more on higher scoring documents.

Step 3.3: the discriminator judges the positive sample i sampled from the true feedback ⁺ Partial order relationship of negative sample i-components generated by sum generator<i ⁺ ,i ^- >Whether it is correct. D (u, i) ⁺ ,i ^- ) Representing the possibility of distinguishing the true data from the generated samples.

D(u,i ⁺ ,i ^- )＝ψ(f(i ⁺ |u)-f(i ^- |u))

ψ(x)＝log(1+exp(-x))

The discriminator wants to update the parameters so that the distinction between positive and negative samples is as large as possible, and therefore, the present embodiment updates the parameters of the discriminator phi in the following manner.

/>

Where U represents a set of users, i ⁺ ,i ^- Respectively representing positive and negative samples, p _true ,G _θ Respectively representing the distribution of the real and generated samples, D being the discriminator, D (u, i) ⁺ ,i ^- ) Judging the relation of partial order<i ⁺ ,i ^- >Whether it is correct, whether user u prefers a positive sample;

step 3.4: updating the generator parameters firstly requires the generator to generate a partial order relation formed by a batch of negative samples and real samples, the discriminator discriminates the generated partial order relation, and the discrimination result is used as the update of the return guidance generator.

Where I represents the set of items and theta represents the generator parameter, and the generator parameter theta is updated by minimizing the objective function.

Step 3.5: in each iteration, the fixed generator of the embodiment updates the parameters of the discriminator, and aims to enable the discriminator to give a correct judgment result under the current condition. Then, the discriminator is fixed, and the generator is trained to continuously update own parameters according to the discrimination result of the current discriminator, so that the generated negative samples are more difficult to distinguish. The above iterative training is carried out by the two until the recommendation result of the final generator is converged.

And 4, step 4: and outputting a converged recommendation result.

It should be understood that the above examples are only for illustrating the present invention and are not intended to limit the scope of the present invention. Furthermore, it should be understood that various changes or modifications of the present invention may be made by those skilled in the art without departing from the spirit and scope of the technical solution of the present invention after reading the teaching of the present invention, and all of them should be covered in the scope of the claims of the present invention.

Claims (3)

1. A method for recommending items in a ranking order based on a generation countermeasure network, comprising the following steps:

step 1: collecting training data, filtering implicit feedback given by a user and item comment information, and preprocessing the filtered comment information;

step 2: converting the item comment information into a word vector matrix by using a GloVe word vector model, and learning potential context characteristics of the comment information as comment characteristics of the item by using a convolutional neural network; converting implicit feedback given by a user into user interaction characteristics and project interaction characteristics through an embedding layer;

and step 3: optimizing parameters in the training process of the confrontation;

selecting a negative sample from unobserved items in a generator in a generated countermeasure network according to the prediction score in a personalized manner, forming a partial order relationship pair with an observed positive sample, judging whether the generated partial order relationship is correct or not through a discriminator, and guiding the generator to update parameters according to the judged result; carrying out iterative confrontation training on the two generators until the recommendation result of the final generator is converged;

the interactive features and the comment features of the project are fused, and the prediction score of the project by the user is calculated through a linear inner product;

in the process of generating the countermeasure, the generator calculates a prediction score, converts the prediction score into the distribution of generated samples through a Softmax function, selects negative samples for user individuation according to the distribution, and forms a partial order relation with positive samples obtained by sampling from real samples;

the discriminator judges whether the generated partial order relation is correct or not and whether the user prefers a positive sample or not through the prediction score; the discriminator wants to update the parameters such that the distinction between positive and negative samples is as large as possible, and therefore the parameter phi of the discriminator is updated as follows:

where U represents a set of users, i ⁺ ,i ^- Respectively representing positive and negative samples, p _true ,G _θ Respectively representing the distribution of the real and generated samples, D being the discriminator, D (u, i) ⁺ ,i ^- ) Judging the partial order relation i ⁺ ,i ^- Whether it is correct, whether user u prefers a positive sample;

updating the parameters of the generator firstly requires the generator to generate a partial order relation formed by a batch of negative samples and real samples, the discriminator discriminates the generated partial order relation, and the discrimination result logD (u, i) is ⁺ ,i ^- ) An update as a reward guide generator;

wherein I represents a set of items, θ represents a parameter of the generator, and the generator parameter θ is updated by minimizing an objective function;

in each iteration, a generator is fixed, parameters of the discriminator are updated, and the aim is to enable the discriminator to give a correct judgment result under the current condition; then, fixing the discriminator, training the generator to continuously update own parameters according to the discrimination result of the current discriminator, and enabling the generated negative samples to be more difficult to distinguish; the two are subjected to the iterative training until the recommendation result of the final generator is converged;

and 4, step 4: and outputting a converged recommendation result.

2. The generation-based matchmaking ranking item recommendation method of claim 1 further characterized by: in the step 1, implicit feedback and comment information are filtered to obtain interactive data of all user IDs and item IDs which really exist; and then preprocessing the filtered comment information, only keeping real words, performing connection operation on all comments of the same project by users, and supplementing the comments of all the projects to a uniform length.

3. The generation-confrontation-network-based hierarchically ordered item recommendation method of claim 1, wherein: in the step 2, for the item comment information, a word vector of each word is obtained through GolVe word vector model mapping, a word vector matrix corresponding to the item comment information is obtained, the matrix is convolved through shared weight, and then a vector with a specified dimensionality is obtained through pooling and full connection and is used as a comment feature of the item to be output;

for implicit feedback given by a user, mapping one-hot codes of sparse users or items through a potential feature matrix at an embedding layer to obtain a low-dimensional potential factor vector.

Images