CN116701773B

CN116701773B - Interpretable recommendation method and device for interpretable recommendation

Info

Publication number: CN116701773B
Application number: CN202310975611.7A
Authority: CN
Inventors: 马江虹; 王蓉
Original assignee: Shenzhen Graduate School Harbin Institute of Technology
Current assignee: Shenzhen Graduate School Harbin Institute of Technology
Priority date: 2023-08-04
Filing date: 2023-08-04
Publication date: 2023-11-10
Anticipated expiration: 2043-08-04
Also published as: CN116701773A

Abstract

The embodiment of the disclosure provides an interpretable recommendation method and a device for interpretable recommendation. The interpretable recommendation method comprises the following steps: acquiring actual evaluations of the item by the user, each actual evaluation comprising at least one of an actual score and an interpretable evaluation tag; training the first and second deep learning models using the actual evaluations such that the first deep learning model is capable of predicting a first preference score of a user for an item and the second deep learning model is capable of predicting a second preference score of a single user and a single item for a binary set of interpretable evaluation tags; predicting a first preference score of the target user for the item by a first deep learning model; selecting a recommended item for the target user according to the first preference score; predicting a second preference score of the target user and the binary group formed by each recommended item on the interpretable evaluation label through a second deep learning model; the interpretable rating tag is selected as an interpretation of the recommended item to the target user based on the second preference score.

Description

Interpretable recommendation method and device for interpretable recommendation

Technical Field

The embodiment of the disclosure relates to the field of computer technology, in particular to an interpretable recommendation method and a device for interpretable recommendation.

Background

With the rapid development of e-commerce and social media platforms, recommendation systems have become an integral tool for many industries, such as product suggestions on online e-commerce websites (e.g., amazon and panoramas), or playlist generators for video and music services, etc. The recommendation system predicts the preference degree of the user on the project according to the basic characteristics and the historical behaviors of the user and the attributes of the project, and further provides personalized recommendation service for the user. Users rely on recommender systems to alleviate the problem of information overload and explore the content of interest from a vast commodity ocean (such as a product, movie, news, or restaurant). Recommendation systems have become an important component of different information systems, even as a standard configuration for some information systems.

With the development of technology, users are no longer satisfied with recommendation systems that only recommend items, and users are more required to know the reason (explanation) why items are recommended. An interpretable recommendation (also referred to interchangeably as an "interpretable recommendation" in the context) helps to improve transparency, persuasion, effectiveness, credibility, and satisfaction of the recommendation system. In recent years, it has become an important topic in many network application studies such as social media, electronic commerce and content sharing websites to explain that recommendation has become a popular topic.

Disclosure of Invention

Embodiments described herein provide an interpretable recommendation method, an apparatus for interpretable recommendation, and a computer readable storage medium storing a computer program.

According to a first aspect of the present disclosure, an interpretable recommendation method is provided. The interpretable recommendation method comprises the following steps: acquiring actual evaluations of a plurality of items by a plurality of users, wherein each of the actual evaluations includes at least one of an actual score and an interpretable evaluation tag, the interpretable evaluation tag including at least one of an evaluation phrase and an emoticon; training a first deep learning model and a second deep learning model using the obtained actual evaluations to enable the first deep learning model to predict a first preference score of a user for an item and the second deep learning model to predict a second preference score of a single user and a single item for an interpretable evaluation tag; predicting a plurality of first preference scores of the target user for the plurality of items by a trained first deep learning model; selecting at least one recommended item for the target user from the plurality of items according to the plurality of first preference scores; predicting a plurality of second preference scores for the plurality of interpretable rating labels for the target user and for the doublet of each recommended item by a trained second deep learning model; and selecting at least one interpretable rating label from the plurality of interpretable rating labels as an interpretation for recommending the at least one recommended item to the target user according to the plurality of second preference scores.

In some embodiments of the present disclosure, the first deep learning model includes: the system comprises a first matrix decomposition module, a first multi-layer sensor module and a first output layer. The output of the first matrix decomposition module and the output of the first multi-layer sensor module are used as the input of the first output layer. The output of the first output layer is used as the output of the first deep learning model.

The first matrix factorization module is represented as:

，

wherein,representing the output of the first matrix factorization module, +.>A first potential feature vector representing a u-th user of the plurality of users,/for>A first potential feature vector representing an ith item of the plurality of items,/and a second potential feature vector representing an ith item of the plurality of items>A first sparse bias representing the u-th user, < >>First sparse bias representing item i, < +.>Representing the element-wise product of the vectors.

The first multi-layer sensor module is represented as:

，

wherein,，

wherein,representing the output of the first multi-layer sensor module, < >>A second potential feature vector representing the u-th user, < >>A second potential feature vector representing item i, < +.>Representing the input of the first multi-layer sensor module, a->Representing the activation function of the kth layer sensor in the first multi-layer sensor module, +.>Weight matrix representing kth layer sensor in first multi-layer sensor module, +. >A bias vector representing a kth layer sensor in the first multi-layer sensor module, +.>Representing the output of the kth layer sensor in the first multi-layer sensor module.

The first deep learning model is expressed as:

，

wherein,a first preference score of the ith item, which represents the ith user output by the first deep learning model,/for the ith item>Representing the activation function of the first output layer, +.>Representing the weight of the first output layer.

In some embodiments of the present disclosure, in the first deep learning model, the first potential feature vector of the u-th user is updated to the first enhanced potential feature vector of the u-th user, the first potential feature vector of the i-th item is updated to the first enhanced potential feature vector of the i-th item, the second potential feature vector of the u-th user is updated to the second enhanced potential feature vector of the u-th user, and the second potential feature vector of the i-th item is updated to the second enhanced potential feature vector of the i-th item.

The first enhanced potential feature vector for the u-th user is represented as:

，

the first enhanced potential feature vector of the i-th item is represented as:

，

wherein,a first enhanced potential feature vector representing a u-th user,/->A first potential feature vector representing a u-th user, < > >A first enhanced potential feature vector representing item i,/->A first potential feature vector representing item i, < +.>Representing a first enhancement weight matrix for the u-th user,>representing a first enhancement weight matrix for the ith item,/->Representing a first enhanced bias vector for the u-th user,>a first enhancement bias vector for the ith item is represented.

The second enhanced potential feature vector for the u-th user is represented as:

，

the second enhanced potential feature vector of the i-th item is represented as:

，

wherein,a second enhanced potential feature vector representing a u-th user,/->A second potential feature vector representing the u-th user, < >>A second enhanced potential feature vector representing item i,/->A second potential feature vector representing item i, < +.>The representation is for the (u)Second enhancement weight matrix of user, +.>Representing a second enhancement weight matrix for item i,/i>Representing a second enhanced bias vector for the u-th user,>representing a second enhanced bias vector for the ith item.

Wherein,representing an attribute distribution vector of an ith item, each element in the attribute distribution vector of the ith item representing a frequency of a corresponding interpretable rating label mentioned in an actual rating associated with the ith item, +_ >Representing the preference distribution vector of the u-th user.

Wherein,，/>，/>，，/>，/>，，/>，

wherein,representing a distribution matrix of attributes>An attribute distribution vector representing an ith item having an interaction with a ith user, at +.>Middle->Splicing according to columns, and adding->Explicit preference degree of the u-th user, +.>Indicating the explicit preference degree of the ith item by the ith user,/for the ith item>Indicating the implicit preference level of the u-th user, +.>Indicating the implicit preference level of the ith item by the ith user at +.>Middle->Splicing by line, in->Middle->Splicing according to rows, and (E) adding>Representing the ith item of the ith userObjective explicit polarity emotion value, +.>Explicit polarity emotion mean value representing the u-th user, < >>Implicit polarity emotion value representing the ith item of the ith user,/for>Represents the implicit polarity emotion average value of the u-th user,representing the actual score of the ith item by the ith user,/->A set of interpretable rating labels representing the ith user mentioned in the actual rating of the ith item, N represents the number of interpretable rating labels in the set, +.>Representing the relationship between the e-th interpretable rating tag and polar emotions, including positive and negative emotions.

In some embodiments of the present disclosure, the first loss function for training the first deep learning model is expressed as:

Wherein,representing a first loss function, ">Representing a set of all tuples each of the plurality of users and each of the plurality of items that can be constituted, if the ith user has an interaction with the ith item + +.>If the u-th user has no interaction with the i-th item +.>，/>Representing a first preference score for the ith item by the ith user,。

in some embodiments of the present disclosure, the second deep learning model includes: the system comprises a second matrix decomposition module, a second multi-layer sensor module and a second output layer. The output of the second matrix decomposition module and the output of the second multi-layer sensor module are used as the input of a second output layer, and the output of the second output layer is used as the output of a second deep learning model.

The second matrix factorization module is represented as:

wherein,representing the output of the second matrix factorization module, +.>A first enhanced potential feature vector representing a u-th user,/->A first enhanced potential feature vector representing item i,/->Second sparse bias representing the u-th user, < ->Second sparse bias representing item i, +.>A first potential feature matrix representing an interpretable rating label with respect to a user,a first potential feature matrix is represented that can interpret the evaluation tag with respect to the item.

The second multi-layer sensor module is represented as:

，

wherein,，

wherein,representing the output of the second multi-layer sensor module, < >>A second enhanced potential feature vector representing a u-th user,/->A second enhanced potential feature vector representing item i,/->Representing an input of a second multi-layer sensor module, < >>Representing the activation function of the kth layer sensor in the second multi-layer sensor module, +.>A weight matrix representing a kth layer sensor in the second multi-layer sensor module, +.>Representing the second multi-layer sensor moduleBias vector of the kth layer sensor, for example>Representing the output of the kth layer sensor in the second multi-layer sensor module +.>Representing a second potential feature matrix of the interpretable rating label with respect to the user,>a second potential feature matrix is represented that can interpret the evaluation tag with respect to the item.

The second deep learning model is expressed as:

wherein (1)>A plurality of second preference scores of a binary group consisting of a ith user and an ith item representing the output of the second deep learning model for a plurality of interpretable rating labels, < ->Representing an activation function of the second output layer, +.>Representing the weight of the second output layer.

In some embodiments of the present disclosure, the interpretable recommendation method further includes: and establishing a self-expression matrix of the interpretable evaluation label in the actual evaluation. Wherein, the self-expression matrix is respectively established for the evaluation phrase and the emoticon in the interpretable evaluation label.

Wherein, the self-expression matrix of the evaluation phrase in the interpretable evaluation label is calculated according to the following formula:

，

wherein, the self-expression matrix of the expression symbol in the interpretable evaluation label is obtained according to the following formula:

，

wherein,a self-expression matrix representing the evaluation phrase, C representing the self-expression matrix of the emoticon,>representing a set of all tuples each of a plurality of users and each of a plurality of items, the +_>Representing the actual preference score of the dyadic set of the ith user and the ith item for a plurality of interpretable rating labels,/for>Relation matrix representing emoticons and emotion labels in actual evaluation +.>Vector representing all elements as 1, +.>Representing elements located on the diagonal of the matrix, < >>And->Is constant.

In some embodiments of the present disclosure, the interpretable recommendation method further includes: the second preference score output by the second deep learning model is updated using the self-expression matrix of interpretable evaluation labels. Wherein, in the case where the interpretable rating tag is a rating phrase, a plurality of second preference scores of the tuple of the u-th user and the i-th item on the plurality of interpretable rating tags are updated as:

，

wherein in the case where the interpretable rating label is an emoticon, the second preference scores of the binary group of the u-th user and the i-th item for the plurality of interpretable rating labels are updated as:

，

Wherein,a second updated preference score representing a plurality of interpretable rating labels for a plurality of tuples of a u-th user and an i-th item,/->A plurality of second preference scores of a binary group consisting of a ith user and an ith item representing the output of the second deep learning model for a plurality of interpretable rating labels, < ->A self-expression matrix representing the evaluation phrase, C representing the self-expression matrix of the emoticon, ++>Representing an identity matrix>Is constant.

In some embodiments of the present disclosure, the second loss function for training the second deep learning model is expressed as:

，

wherein,representing a second loss function, ">Representing a set of all tuples each of a plurality of users and each of a plurality of items, the +_>Represents the set of all interpretable rating labels, +.>Indicating whether the actual evaluation of the ith item by the ith user includes the ith interpretable rating label,/or not>A second preference score for the e-th interpretable rating label representing a tuple of the u-th user and the i-th item,>representing a predicted preference score of a dyadic group of a ith user and an ith item on an e-th interpretable rating label based on a self-representation matrix of the interpretable rating label, < > >Is thatIn the case where the interpretable evaluation tag is an evaluation phrase, +_e->In case the interpretable rating label is an emoticon, -, a +_>，/>A self-expression matrix representing the evaluation phrase, C representing the self-expression matrix of the emoticon, ++>A plurality of second preference scores of a binary group consisting of a ith user and an ith item representing the output of the second deep learning model for a plurality of interpretable rating labels, < ->。

In some embodiments of the present disclosure, the first deep learning model and the second deep learning model are jointly trained. Wherein a joint loss function for joint training of the first deep learning model and the second deep learning model is expressed as:

，

wherein,representing a joint loss function->Representing a first loss function, ">A second loss function is indicated and is indicated,is constant.

According to a second aspect of the present disclosure, an apparatus for interpretable recommendation is provided. The apparatus includes at least one processor; and at least one memory storing a computer program. The computer program, when executed by at least one processor, causes an apparatus to perform the steps of the method according to the first aspect of the present disclosure.

According to a third aspect of the present disclosure, there is provided a computer readable storage medium storing a computer program, wherein the computer program when executed by a processor implements the steps of the method according to the first aspect of the present disclosure.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following brief description of the drawings of the embodiments will be given, it being understood that the drawings described below relate only to some embodiments of the present disclosure, not to limitations of the present disclosure, in which:

FIG. 1 is an exemplary flow chart of an interpretable recommendation method according to an embodiment of the present disclosure; and

fig. 2 is a schematic block diagram of an apparatus for interpretable recommendation in accordance with an embodiment of the present disclosure.

It is noted that the elements in the drawings are schematic and are not drawn to scale.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present disclosure more apparent, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings. It will be apparent that the described embodiments are some, but not all, of the embodiments of the present disclosure. All other embodiments, which can be made by those skilled in the art based on the described embodiments of the present disclosure without the need for creative efforts, are also within the scope of the protection of the present disclosure.

Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the presently disclosed subject matter belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the specification and relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein. As used herein, a statement that two or more parts are "connected" or "coupled" together shall mean that the parts are joined together either directly or joined through one or more intermediate parts. In addition, terms such as "first" and "second" are used merely to distinguish one component (or portion of a component) from another component (or another portion of a component).

Existing interpretable recommendation methods focus mainly on text interpretation. However, given the role of emoticons as a new form of conveying information, expressing and enhancing emotion, reducing speech ambiguity, and affecting information understanding, and the increasing popularity of emoticons in user reviews, the importance of incorporating them into interpretable recommendation systems to effectively persuade users to make purchasing decisions is highlighted. Emoticons used in product reviews have proven to have a tendency to expand the impact of active consumers, thereby enhancing the purchase intent of the user. Thus, ignoring emoticons in an interpretable recommendation system will miss opportunities to leverage them to influence consumer behavior and enhance recommendation results. By introducing the emoticons and researching the relation between the emoticons and the emotion information, the recommendation effect of the emoticons serving as the explanation can be improved.

The actual score can be generally considered as an explicit feedback of the user to the product, and the opinion or emotional tendency of the user can be explicitly expressed. However, implicit feedback from user comments is also not negligible. Implicit feedback is, for example, an interpretable rating tag, which may indirectly reflect the user's preferences and emotional tendency. By combining explicit feedback and implicit feedback of the user, the method has great help in extracting user preferences and further improving recommendation effects.

The present disclosure proposes using an item-interpretation joint decomposition framework to implement an interpretable recommendation method. The interpretable recommendation method is capable of providing a user with a plurality of project suggestions and providing an interpretation of each project (the interpretation may include both a rating phrase and an emoticon). In other words, the interpretable recommendation method is capable of performing two tasks: project recommendation tasks and interpretation recommendation tasks. The present disclosure may also enhance user and item representations by introducing interpretable rating labels and actual scores, and enhance the presentation of interpretable rating labels by utilizing high-order correlations between interpretable rating labels, thereby enhancing the effectiveness of item recommendations and interpreted recommendations. Fig. 1 shows an exemplary flowchart of an interpretable recommendation method 100, according to an embodiment of the present disclosure.

In the interpretable recommendation method 100, at block S102, actual evaluations of a plurality of items by a plurality of users are obtained. Each of the actual evaluations includes at least one of an actual score and an interpretable evaluation tag. The interpretable rating tag includes at least one of a rating phrase and an emoticon.

In some embodiments of the present disclosure, the items may include: merchandise, music, movies, travel, etc. The actual rating refers to a user-provided actual rating of the item that it used. Each user may make an actual evaluation of some or all of the plurality of items. The actual evaluations of multiple items by multiple users may be partially or fully overlapping. The user may evaluate the item using only scoring, may evaluate the item using only the scoring text (including the interpretable scoring tag), and may evaluate the item using both scoring and scoring text. The actual score in the actual evaluation is a number within a preset range. The evaluation phrase in the interpretable evaluation tag is in the form of an adjective + noun/noun phrase. The evaluation phrase may be extracted from the actual evaluation using a rule based on a dependency grammar. The emoticons in the interpretable evaluation tag may include emoticons represented by graphical images (which may also be referred to as "graphical emoticons"). The graphical emoticons may be static or dynamic. Emoticons may represent basic emotions (including anger, expectations, aversions, fear, happiness, sadness, surprise, and trust), as well as polar emotions (including positive and negative emotions).

At block S104, the first and second deep learning models are trained using the obtained actual evaluations such that the first deep learning model is capable of predicting a first preference score of the user for the item and the second deep learning model is capable of predicting a second preference score of the single user and the single item for the interpretable evaluation tag.

In some embodiments of the present disclosure, the first deep learning model may include: a first matrix decomposition (MF) module, a first multi-layer perceptron (MLP) module, and a first output layer. The first matrix factorization module is used to learn the linear relationship between the user's potential feature vectors (which may also be referred to as latent factors or latent factors) and the project's potential feature vectors. The first multi-layer perceptron module is configured to learn a non-linear relationship between potential feature vectors of the user and potential feature vectors of the item.

The first matrix factorization module models bi-directional interactions of potential factors of users and items. Each dimension of the potential space is assumed to be independent of the others and are linearly combined with the same weight. Thus, the first matrix factorization module may be considered as a linear model of a potential factor. The first matrix factorization module may learn a linear relationship between each of the plurality of users (any user in the context referred to by the u-th user) and each of the plurality of items (any item in the context referred to by the i-th item) by:

（1）

The first multi-layer perceptron module uses nonlinear functions to learn interactions between users and items, giving the first deep learning model a greater degree of flexibility.

The inputs to the first multilayer perceptron module are:

（2）

wherein,a second potential feature vector representing the u-th user, < >>A second potential feature vector representing item i, < +.>Is->And->Is a concatenation vector of (a) in the (b).

The network architecture of the first multi-layer sensor module, which follows the tower pattern with the widest bottom layer, can learn more abstract features of the data by using a small number of hidden units for higher layers.

The kth layer sensor in the first multi-layer sensor module is denoted as:

（3）

wherein,，/>representing the activation function of the kth layer sensor in the first multi-layer sensor module,representing the firstWeight matrix of kth layer sensor in multi-layer sensor module, < - >A bias vector representing a kth layer sensor in the first multi-layer sensor module, +.>Representing the output of the kth layer of sensors in the first multi-layer sensor module, m representing the number of layers of sensors in the first multi-layer sensor module.

Output of last layer sensor in first multi-layer sensor moduleAs output of the first multilayer perceptron module:

（4）

wherein,representing the output of the first multi-layer sensor module.

To provide more flexibility to the first deep learning model, the first matrix factorization module and the first multi-layer perceptron module are allowed to learn the potential feature vectors individually. The first matrix-decomposition module and the first multi-layer sensor module may be coupled together by a first output layer. The output of the first matrix decomposition module and the output of the first multi-layer sensor module are used together as the input of the first output layer. The output of the first output layer is used as the output of the first deep learning model.

The first deep learning model is expressed as:

（5）

wherein,representing a first deep learning model outputFirst preference score of the ith item for the (u) th user, <>Representing the activation function of the first output layer, +.>Representing the weight of the first output layer.

In some embodiments of the present disclosure, the interpretable rating label and the actual score may capture a user's preference for items and attribute distribution of the items from the user's opinion, thus potential feature vectors of the user and items may be enhanced by the interpretable rating label and the actual score.

In an aspect, item attributes may be learned to extract enhanced item representations (enhanced potential feature vectors of items). To establish a relationship between the interpretable evaluation tag and the item, an attribute distribution vector may be created for each item. By analyzing the frequency of user comments on particular aspects of an item, the intensity of the characteristics of the item in those aspects can be determined. To model the enhanced latent feature vector of the ith item, an attribute distribution vector of the ith item can be introduced 。/>The element (entry) of (i) represents the frequency of the corresponding interpretable rating tag mentioned in the comment associated with the ith item. />The popularity of different interpretable rating labels in user feedback can be quantified, facilitating modeling and representation of item property distributions.

In another aspect, user preferences may be learned to extract enhanced user representations (enhanced potential feature vectors of the user). User expressed preferences play a vital role in determining the importance of item attributes. To quantify this degree of preference, one can defineThe preference degree of the ith item for the ith user. />Shows the (positive/negative) emotion value of the ith user to the ith item, of the (u) th user >The average emotion value of the u-th user for all items is represented. By evaluating such a degree of preference, the degree to which a user likes or dislikes a particular item can be evaluated based on the emotional preference of the user for that particular item as compared to their overall emotional tendency. In fact, the _on>Two types of user feedback can be used to represent: explicit user emotion feedback and implicit user emotion feedback. These two types provide different views of the project by the user. Explicit user emotion feedback refers to a score provided directly by a user that explicitly expresses the opinion or emotion tendencies of the user. Implicit user emotion feedback is derived from user comments, such as interpretable rating labels, which reflect user preferences and emotion tendencies therebetween.

For emotion analysis, eight basic emotions are considered in this disclosure, namely anger, expectancy, aversion, fear, happiness, sadness, surprise and trust, and two polarity emotions, positive emotion and negative emotion.

For emoticons, the present disclosure studies the relationship of the emoticons to the eight basic emotions, and the relationship of the emoticons to the polar emotion. For the evaluation phrase, the present disclosure only investigates the relationship between the evaluation phrase and the polar emotion, because the evaluation phrase does not exhibit a series of emotions similar to emotions, and the association with the eight basic emotions is less close.

In one example, feature extraction is performed on the evaluation text in the actual evaluation using word2vec techniques to obtain a feature representation for each word (including emoticons). Then, based on an emotion dictionary disclosed by the national research council of canada (NRC), feature representations of words under various emotion tags included in the evaluation text are averaged to obtain feature representations of each emotion tag. Since emotion tags are language tags, a bias may be present when their corresponding feature representations are compared to emoticons (non-language tags), so that the feature representations of the emotion tags and emoticons need to be normalized, respectively. Specifically, an average vector of eight basic emotion tags may be calculated and subtracted from the feature representation of each emotion tag. This essentially zeroes the center of the polyhedron defined by the eight emotion tags. Similarly, the average of all the emoji vectors is subtracted from the feature representation of each emoji. To quantify the relationship between emoticons and emotion tags, cosine similarity between them can be calculated to construct a relationship matrix of emoticons and emotion tags, which is expressed in the context as 。

Similarly, the relationship between interpretable rating labels (rating phrases and emoticons) and polar emotions can be constructedRepresenting the relationship between the interpretable rating label and the polar emotion. If->If a certain element in the set is positive, the set indicates that the corresponding evaluation phrase or expression symbol represents positive emotion. If->If a certain element in the set is negative, the corresponding evaluation phrase or the expression symbol indicates negative emotion.

Assume thatFor a set of items with interaction with the u-th user, < ->Is a column vector. />Every element in (2) corresponds to the u-th user pair +.>A preference degree of each item in the list. Can use +.>Indicating explicit preference level of the u-th user (explicit user emotion feedback). Can use +.>Indicating the implicit preference level of the u-th user (implicit user emotion feedback). Use matrix +.>To represent attribute distribution->Each column of (a) represents->Attribute distribution of corresponding items in the list. Let->Is x, & gt>And->Can be expressed as:

wherein,representing a distribution matrix of attributes>An attribute distribution vector representing an ith item having an interaction with a ith user, at +.>Middle->Splicing according to columns, and adding->Explicit preference degree of the u-th user, +.>Indicating the explicit preference degree of the ith item by the ith user,/for the ith item >Indicating the implicit preference level of the u-th user, +.>Indicating the implicit preference level of the ith item by the ith user at +.>Middle->Splicing by line, in->Middle->Splicing according to rows, and (E) adding>Explicit polarity emotion value representing the ith item of the ith user, +.>Explicit polarity emotion representing a u-th userAverage value->Implicit polarity emotion value representing the ith item of the ith user,/for>Represents the implicit polarity emotion average value of the u-th user,representing the actual score of the ith item by the ith user,/->A set of interpretable rating labels representing the ith user mentioned in the actual rating of the ith item, N representing the number of interpretable rating labels in the set, +.>Representing the relationship between the e-th interpretable rating tag and polar emotions, including positive and negative emotions. />Positive numbers indicate that the evaluation phrase or emoticon in the e-interpretable evaluation tag represents a positive emotion. />Negative numbers indicate that the evaluation phrase or emoticon in the e-interpretable evaluation tag represents negative emotions.

The preference distribution of the u user can be obtained by weighted averaging the attribute distribution of the item evaluated by the u user as follows:

(13)

wherein, Indicating the l-2 norm.

The degree of preference of the u-th user and the attribute distribution of the i-th item may not be consistent with the dimensions of the potential feature vectors of the user and item in the first deep learning model. To address this discrepancy, a linear layer may be added to the first deep learning model, ensuring that their dimensions are uniform. In the first deep learning model, the first potential feature vector of the u-th user is updated to the first enhanced potential feature vector of the u-th user, the first potential feature vector of the i-th item is updated to the first enhanced potential feature vector of the i-th item, the second potential feature vector of the u-th user is updated to the second enhanced potential feature vector of the u-th user, and the second potential feature vector of the i-th item is updated to the second enhanced potential feature vector of the i-th item.

(14)

the first enhanced potential feature vector of the i-th item is represented as:

(15)

wherein,a first enhanced potential feature vector representing a u-th user,/->A first potential feature vector representing a u-th user, < >>A first enhanced potential feature vector representing item i,/->A first potential feature vector representing item i, < +. >Representation is directed toFirst enhancement weight matrix of the u-th user, < ->Representing a first enhancement weight matrix for the ith item,/->Representing a first enhanced bias vector for the u-th user,>representing a first enhanced bias vector for item i,>representing an attribute distribution vector of an ith item, each element in the attribute distribution vector of the ith item representing a frequency of a corresponding interpretable rating label mentioned in an actual rating associated with the ith item, +_>Representing the preference distribution vector of the u-th user.

(16)

(17)

wherein,a second enhanced potential feature vector representing a u-th user,/->A second potential feature vector representing the u-th user, < >>A second enhanced potential feature vector representing item i,/->A second potential feature vector representing item i, < +.>Representing a second enhancement weight matrix for the u-th user,>representing a second enhancement weight matrix for item i,/i>Representing a second enhanced bias vector for the u-th user,>representing a second enhanced bias vector for item i,>representing an attribute distribution vector of an ith item, each element in the attribute distribution vector of the ith item representing a frequency of a corresponding interpretable rating label mentioned in an actual rating associated with the ith item, +_ >Representing the preference distribution vector of the u-th user.

Obtained according to the formulas (14), (15), (16) and (17)、/>、/>And->Can be substituted into the formula (1) and the formula (2) forRespectively replace->、/>、/>And->Thereby recalculating +.>。

In some embodiments of the present disclosure, the first loss function (cross entropy loss function) for training the first deep learning model is expressed as:

(18)

wherein,representing a first loss function, ">Representing a set of all of the tuples that each of the plurality of users and each of the plurality of items can make up. If the u-th user has an interaction with the i-th item +.>. If the u-th user has no interaction with the i-th item +.>。/>Representing a first preference score of the ith user for the ith item,/for the ith user>To activate the function +.>。

By gradient descent methodMinimizing, model parameters of the first deep learning model may be obtained, including: />、/>、/>、/>、/>、/>、/>、/>、/>、/>、/>、/>、/>、/>、、/>、/>、/>。

As above, project recommendation tasks may be implemented through a first deep learning model. How the explanation recommendation task is implemented is described below.

An interpretive recommendation task is to recommend a list of interpretable rating labels for a particular user-item doublet (doublet made up of a single user and a single item). Because of the data sparsity of the user-item-interpretable evaluation tag triplet, embodiments of the present disclosure decompose the triplet into two binary relations: user-interpretable rating tags and items-interpretable rating tags. The relationship between the user-item doublet and the interpretable rating label may be obtained using the relationship between the user and the interpretable rating label, and the relationship between the item and the interpretable rating label.

Embodiments of the present disclosure use a second deep learning model to learn the relationship between the user-item doublet and the interpretable rating tag.

In some embodiments of the present disclosure, the second deep learning model includes: the system comprises a second matrix decomposition module, a second multi-layer sensor module and a second output layer.

The second matrix factorization module may learn a linear relationship between the user's preference for interpretable rating labels and the item's preference for interpretable rating labels by:

(19)

The second multi-layer perceptron module uses a nonlinear function to learn the interaction between the user's preferences for interpretable rating labels and the item's preferences for interpretable rating labels.

The inputs to the second multilayer perceptron module are:

(20)

Wherein,a second enhanced potential feature vector representing a u-th user,/->A second enhanced potential feature vector representing item i,/->Representing a second potential feature matrix of the interpretable rating label with respect to the user,>representing a second latent feature matrix of the interpretable rating label with respect to the item,>representation->And->Is a concatenation vector of (a) in the (b).

The network architecture design of the second multi-layer sensor module follows a tower pattern with the widest bottom layer, which can learn more abstract features of the data by using a small number of hidden units for higher layers.

The kth layer sensor in the second multi-layer sensor module is denoted as:

(21)/>

wherein,，/>representing the activation function of the kth layer sensor in the second multi-layer sensor module,a weight matrix representing a kth layer sensor in the second multi-layer sensor module, +.>A bias vector representing a kth layer sensor in the second multi-layer sensor module, +.>Representing the output of the kth layer of sensors in the second multi-layer sensor module, m representing the number of layers of sensors in the second multi-layer sensor module.

Output of last layer sensor in second multi-layer sensor moduleAs output of the second multilayer perceptron module:

(22)

Wherein,representing the output of the second multi-layer sensor module.

Similar to the first deep learning model, the second matrix factorization module and the second multi-layer perceptron module are allowed to learn the potential feature vectors alone in the second deep learning model. The second matrix-decomposition module and the second multi-layer sensor module may be coupled together by a second output layer. The output of the second matrix factorization module and the output of the second multi-layer perceptron module are used together as the input of the second output layer. The output of the second output layer is used as the output of the second deep learning model.

The second deep learning model is expressed as:

(23)

wherein,a plurality of second preference scores of a binary group consisting of a ith user and an ith item representing the output of the second deep learning model for a plurality of interpretable rating labels, < ->Representing an activation function of the second output layer, +.>Representing the weight of the second output layer.

In some embodiments of the present disclosure, the second loss function for training the second deep learning model may be expressed as:

(24)

wherein,representing a set of all tuples each of a plurality of users and each of a plurality of items, the +_>Represents the set of all interpretable rating labels, +. >Indicating whether the actual evaluation of the ith item by the ith user includes the ith interpretable rating label,/or not>A second preference score for the e-th interpretable rating label representing a tuple of the u-th user and the i-th item,>to activate the function +.>。

By gradient descent method to make the second loss functionMinimizing, model parameters of the second deep learning model may be obtained, including: />、/>、/>、/>、/>、/>、/>、/>、/>、/>、/>、/>、/>、、/>、/>、/>、/>、/>、/>。

Embodiments of the present disclosure also propose to help explain recommended tasks by exploiting the high-order correlation between interpretable rating labels. Interpretation recommendations can be viewed as a multi-tag learning problem. The multi-label classification performance can be successfully improved by using the knowledge related to the labels in multi-label learning. Because of the strong associative modeling capabilities of higher-order policies in terms of tag associations, this disclosure applies them to interpreting recommended tasks. To build a higher-order tag correlation model, it is assumed that each interpretable evaluation tag can be reconstructed from one or more other interpretable evaluation tags. Correlation of interpretable evaluation tags can be modeled based on co-occurrence relationships and emotional relationships.

Since users often express information of multiple aspects in their comments, the relevance of an interpretable rating tag can be modeled by studying the co-occurrence relationship of the interpretable rating tag. The greater the probability that two interpretable evaluation labels appear simultaneously, the stronger the correlation between them is indicated.

Based on this co-occurrence relationship, we can construct a correlation matrix (also referred to as a self-expression matrix) between interpretable evaluation tagsThe following are provided:

，（25）

wherein,representing each of the plurality of users and each of the plurality of items can constituteIs a set of all the tuples of +.>Vector representing all elements as 1, +.>Representing elements located on the diagonal of the matrix, < >>Is constant. />Indicating the l-2 norm. />And represents the l-1 norm for preventing overfitting. />Can maintain consistency between interpretable evaluation tag dependencies. The constraint that the diagonals are all zero avoids the trivial solution, namely the appearance of an identity matrix. Non-negative constraints may help provide visualized relationship coefficients.

It should be noted that the self-expression matrices of the evaluation phrase and the emoticon in the interpretable evaluation tag may be independent of each other. In this context,the self-expression matrix of the evaluation phrase can be represented, and the self-expression matrix of the emoticon can be represented.

The semantics of emoticons can be inferred by analyzing their associated emotion information. The closer the emotion information contained by the emoticons, the stronger the emotion relationship between them. Based on the relationship between the emoticons and the emotion information, another correlation matrix between the emoticons can be constructed The following are provided:

，（26）

wherein,relation matrix representing emoticons and emotion labels in actual evaluation +.>Vector representing all elements as 1, +.>Representing elements located on the diagonal of the matrix, < >>Is constant. />Indicating the l-2 norm. />And represents the l-1 norm for preventing overfitting. />Can maintain consistency between interpretable evaluation tag dependencies. The constraint that the diagonals are all zero avoids the trivial solution, namely the appearance of an identity matrix. Non-negative constraints may help provide visualized relationship coefficients.

By considering both co-occurrence and emotion relationships, the self-expression matrix C of the emoticon can be constructed as follows:

，（27）

wherein C represents the self-expression matrix of the emoticon,representing a set of all tuples each of a plurality of users and each of a plurality of items, the +_>Representing the actual preference score of the dyadic set of the ith user and the ith item for a plurality of interpretable rating labels,/for>Relation matrix representing emoticons and emotion labels in actual evaluation +.>Vector representing all elements as 1, +.>Representing elements located on the diagonal of the matrix, < >>And->Is constant (may be an empirical value). />Is a weight coefficient for balancing co-occurrence relationships and emotion relationships.

In some embodiments of the present disclosure, the second preference score output by the second deep learning model may be updated using a self-representative matrix of interpretable evaluation labels.

In the case where the interpretable rating tag is a rating phrase, a plurality of second preference scores of the dibasic group consisting of the u-th user and the i-th item for the plurality of interpretable rating tags are updated as:

（28）

in the case where the interpretable rating label is an emoticon, a plurality of second preference scores of the binary group constituted by the u-th user and the i-th item for the plurality of interpretable rating labels are updated as:

（29）

wherein,a second updated preference score representing a plurality of interpretable rating labels for a plurality of tuples of a u-th user and an i-th item,/->A plurality of second preference scores of a binary group consisting of a ith user and an ith item representing the output of the second deep learning model for a plurality of interpretable rating labels, < ->A self-expression matrix representing the evaluation phrase, C representing the self-expression matrix of the emoticon, ++>Representing an identity matrix>Is constant. />Is a weight that balances the interpretable evaluation of similarity between tags.

In some embodiments of the present disclosure, to take full advantage of interpretable aspect relationships obtained from self-expression learning, a new preference score function may be defined for the evaluation phrase and emoticon, respectively, that incorporates the preference scores of all other interpretable aspects.

In the case where the interpretable evaluation tag is an evaluation phrase,

（30）

in the case where the interpretable evaluation tag is an emoticon,

（31）

wherein,a self-expression matrix representing the evaluation phrase, C representing the self-expression matrix of the emoticon, ++>A plurality of second preference scores of a binary group consisting of a ith user and an ith item representing the output of the second deep learning model for a plurality of interpretable rating labels, < ->A plurality of predicted preference scores for a plurality of interpretable rating labels based on a dyadic set of a u-th user and an i-th item derived from a representation matrix of interpretable rating labels are represented.

By guiding the training of the second deep learning model in conjunction with the relationships between the respective interpretable evaluation tags, the second deep learning model can be made to learn more efficiently. Thus, the second loss function in equation (24)Can be updated as:

wherein,representing a second loss function, ">Representing a set of all tuples each of a plurality of users and each of a plurality of items, the +_>Represents the set of all interpretable rating labels, +.>Indicating whether the actual evaluation of the ith item by the ith user includes the ith interpretable rating label (if +.>The actual evaluation of the ith item by the ith user includes the ith interpretable evaluation tag, if +. >=0, then the actual rating of the ith item by the ith user does not include the ith interpretable rating label), +.>A second preference score for the e-th interpretable rating label representing a tuple of the u-th user and the i-th item,>representing a predicted preference score of a dyadic group of a ith user and an ith item on an e-th interpretable rating label based on a self-representation matrix of the interpretable rating label, < >>Is->Element e of->In order to activate the function,。/>

in some embodiments of the present disclosure, the first deep learning model and the second deep learning model may be jointly trained. Wherein a joint loss function for joint training of the first deep learning model and the second deep learning model is expressed as:

（33）

wherein the method comprises the steps of，Representing a joint loss function->Representing a first loss function, ">Representing a second loss function, ">Is constant. />For balancing weights between project recommendation tasks and interpretation recommendation tasks.

At block S106, a plurality of first preference scores for a target user of the plurality of users for the plurality of items are predicted by the trained first deep learning model. Specifically, a plurality of first preference scores for a target user of the plurality of users for the plurality of items may be predicted by equation (5).

In the formula (5), the following formula can be used for calculationAnd->：

（34）

（35）

Wherein,。

wherein,a first enhanced potential feature vector representing a u-th user,/->A first enhanced potential feature vector representing item i,/->A first sparse bias representing the u-th user, < >>First sparse bias representing item i, < +.>A second enhanced potential feature vector representing a u-th user,/->A second enhanced potential feature vector representing item i,/->Representing the activation function of the kth layer sensor in the first multi-layer sensor module, +.>Weight matrix representing kth layer sensor in first multi-layer sensor module, +.>A bias vector representing a kth layer sensor in the first multi-layer sensor module, +.>Representing the output of the kth layer of sensors in the first multi-layer sensor module, m representing the number of layers of sensors in the first multi-layer sensor module.

At block S108, at least one recommended item for the target user is selected from a plurality of items according to a plurality of first preference scores. In some embodiments of the present disclosure, the plurality of first preference scores may be ranked, and items corresponding to the highest ranked one or more first preference scores may be recommended to the target user. In other embodiments of the present disclosure, the first preference score may be compared to a first threshold. If the first preference score of the target user for any item is greater than a first threshold, the item is recommended to the target user.

At block S110, a plurality of second preference scores for the plurality of interpretable rating labels for the target user and the doublet of each recommended item are predicted by a trained second deep learning model. Specifically, in the case where the interpretable rating tag is a rating phrase, a plurality of second preference scores of the target user and the binary group constituted by each recommended item for the plurality of interpretable rating tags can be predicted by the formula (28). In the case where the interpretable rating label is an emoticon, a plurality of second preference scores of the target user and the binary group constituted by each recommended item for the plurality of interpretable rating labels can be predicted by the expression (29).

In the formulas (28) and (29),can be calculated by equation (23). In formula (23), a->Can be calculated by formula (19),>can be calculated by equation (22). />

At block S112, at least one interpretable rating label is selected from the plurality of interpretable rating labels as an interpretation for recommending at least one recommended item to the target user in accordance with the plurality of second preference scores. In some embodiments of the present disclosure, the plurality of second preference scores may be ranked, and the target user is presented with interpretable rating labels corresponding to the highest ranked one or more second preference scores. In one example, the plurality of second preference scores may be ranked for the rating phrase and the emoticon in the interpretable rating label, respectively, and then the highest ranked one or more rating phrases and highest ranked one or more emoticons are presented. In other embodiments of the present disclosure, the second preference score may be compared to a second threshold. If the second preference score of any interpretable rating label by the binary group formed by the target user and the recommended item is greater than a second threshold value, the interpretable rating label is taken as an interpretation for recommending the recommended item to the target user.

Fig. 2 shows a schematic block diagram of an apparatus 200 for interpretable recommendation, according to an embodiment of the present disclosure. As shown in fig. 2, the apparatus 200 may include a processor 210 and a memory 220 storing a computer program. The computer program, when executed by the processor 210, causes the apparatus 200 to perform the steps of the interpretable recommendation method 100 as shown in fig. 1. In one example, apparatus 200 may be a computer device or a cloud computing node. The apparatus 200 may obtain actual evaluations of a plurality of items by a plurality of users. Wherein each of the actual evaluations includes at least one of an actual score and an interpretable evaluation tag. The interpretable rating tag includes at least one of a rating phrase and an emoticon. The apparatus 200 may train the first and second deep learning models using the obtained actual evaluations to enable the first deep learning model to predict a first preference score of a user for an item and the second deep learning model to predict a second preference score of a single user and a single item for an interpretable evaluation tag. The apparatus 200 may predict a plurality of first preference scores for the target user for the plurality of items through the trained first deep learning model. The apparatus 200 may select at least one recommended item for the target user from a plurality of items according to a plurality of first preference scores. The apparatus 200 may predict a plurality of second preference scores for the plurality of interpretable evaluation tags for the target user and the doublet of each recommended item through a trained second deep learning model. The apparatus 200 may select at least one interpretable rating tag from the plurality of interpretable rating tags as an interpretation that recommends at least one recommended item to the target user according to the plurality of second preference scores.

In embodiments of the present disclosure, processor 210 may be, for example, a Central Processing Unit (CPU), a microprocessor, a Digital Signal Processor (DSP), a processor of a multi-core based processor architecture, or the like. Memory 220 may be any type of memory implemented using data storage technology including, but not limited to, random access memory, read only memory, semiconductor-based memory, flash memory, disk storage, and the like.

Furthermore, in embodiments of the present disclosure, the apparatus 200 may also include an input device 230, such as a keyboard, mouse, etc., for inputting the actual evaluation of the item by the user. Additionally, the apparatus 200 may further comprise an output device 240, such as a display or the like, for outputting at least one recommended item for the user and an interpretation of the recommendation of the at least one recommended item.

In other embodiments of the present disclosure, there is also provided a computer readable storage medium storing a computer program, wherein the computer program is capable of implementing the steps of the method as shown in fig. 1 when being executed by a processor.

In summary, an interpretable recommendation method according to embodiments of the present disclosure uses an item-interpretation joint decomposition framework to provide multiple item suggestions for a user and to provide an interpretation (which may include both a rating phrase and an emoticon form) that recommends each item. The present disclosure may also enhance user and item representations by introducing interpretable rating labels and actual scores, and enhance the presentation of interpretable rating labels by utilizing high-order correlations between interpretable rating labels, thereby enhancing the effectiveness of item recommendations and interpreted recommendations.

The flowcharts and block diagrams in the figures illustrate the architecture, functionality, and operation of possible implementations of apparatus and methods according to various embodiments of the present disclosure. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems which perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

As used herein and in the appended claims, the singular forms of words include the plural and vice versa, unless the context clearly dictates otherwise. Thus, when referring to the singular, the plural of the corresponding term is generally included. Similarly, the terms "comprising" and "including" are to be construed as being inclusive rather than exclusive. Likewise, the terms "comprising" and "or" should be interpreted as inclusive, unless such an interpretation is expressly prohibited herein. Where the term "example" is used herein, particularly when it follows a set of terms, the "example" is merely exemplary and illustrative and should not be considered exclusive or broad.

Further aspects and scope of applicability will become apparent from the description provided herein. It is to be understood that various aspects of the application may be implemented alone or in combination with one or more other aspects. It should also be understood that the description and specific examples are intended for purposes of illustration only and are not intended to limit the scope of the present disclosure.

While several embodiments of the present disclosure have been described in detail, it will be apparent to those skilled in the art that various modifications and variations can be made to the embodiments of the present disclosure without departing from the spirit and scope of the disclosure. The scope of the present disclosure is defined by the appended claims.

Claims

1. An interpretable recommendation method, the interpretable recommendation method comprising:

obtaining actual evaluations of a plurality of items by a plurality of users, wherein each of the actual evaluations comprises at least one of an actual score and an interpretable evaluation tag, the interpretable evaluation tag comprising at least one of an evaluation phrase and an emoticon;

training a first deep learning model and a second deep learning model using the obtained actual evaluations to enable the first deep learning model to predict a first preference score of a user for an item and the second deep learning model to predict a second preference score of a single user and a single item for an interpretable evaluation tag;

Predicting a plurality of first preference scores of the target user for the plurality of items by a trained first deep learning model;

selecting at least one recommended item for the target user from the plurality of items according to the plurality of first preference scores;

predicting a plurality of second preference scores for a plurality of interpretable rating labels for the target user and for the doublet of each recommended item by a trained second deep learning model; and

selecting at least one interpretable rating label from the plurality of interpretable rating labels as an interpretation for recommending the at least one recommended item to the target user according to the plurality of second preference scores;

wherein the first deep learning model comprises: a first matrix decomposition module, a first multi-layer sensor module, and a first output layer,

wherein the output of the first matrix decomposition module and the output of the first multi-layer perceptron module are used together as the input of the first output layer, and the output of the first output layer is used as the output of the first deep learning model;

the first matrix factorization module is represented as:

，

wherein,representing the output of said first matrix factorization module,/- >A first potential feature vector representing a u-th user of the plurality of users, +.>A first potential feature vector representing an ith item of the plurality of items, +.>A first sparse bias representing said u-th user,>a first sparse bias indicative of said ith item, +.>Representing an element-by-element product of the vector;

the first multi-layer perceptron module is represented as:

，

wherein,，

wherein,representing the output of said first multi-layer sensor module,/or->A second potential feature vector representing said u-th user,>a second potential feature vector representing said ith item,/i>Representation houseInput of the first multi-layer sensor module, < >>Representing an activation function of a kth layer sensor in said first multi-layer sensor module,/v>A weight matrix representing the kth layer sensor in the first multi-layer sensor module,/and/or>A bias vector representing the kth layer sensor in the first multi-layer sensor module,/v>Representing an output of the kth layer sensor in the first multi-layer sensor module;

the first deep learning model is represented as:

，

wherein,representing the first preference score of the ith item by the ith user output by the first deep learning model,/v- >An activation function representing said first output layer, < >>Representing the weight of the first output layer.

2. The interpretable recommendation method of claim 1, wherein in the first deep learning model, the first potential feature vector of the u-th user is updated to a first enhanced potential feature vector of the u-th user, the first potential feature vector of the i-th item is updated to a first enhanced potential feature vector of the i-th item, the second potential feature vector of the u-th user is updated to a second enhanced potential feature vector of the u-th user, and the second potential feature vector of the i-th item is updated to a second enhanced potential feature vector of the i-th item;

the first enhanced potential feature vector of the u-th user is represented as:

，

the first enhanced potential feature vector of the ith item is represented as:

，

wherein,representing said first enhanced potential feature vector of said u-th user,/for>Representing the first potential feature vector of the u-th user,/>The first enhanced potential feature vector representing the ith item,/v->Representing the first potential feature vector of the ith item, +. >Representing a first enhancement weight matrix for said u-th user,representing a first enhancement weight matrix for said ith item,/>Representing a first enhanced bias vector for said u-th user,>representing a first enhancement bias vector for the ith item;

the second enhanced potential feature vector of the u-th user is represented as:

，

the second enhanced potential feature vector of the ith item is represented as:

，

wherein,representing said second enhanced potential feature vector of said u-th user,/for>Representing the second potential feature vector of the u-th user,/>The second enhanced potential feature vector representing the ith item,/v->Representing the second potential feature vector of the ith item, +.>Representing a second enhancement weight matrix for said u-th user,>representing a second enhancement weight matrix for said ith item,/>Representing a second enhanced bias vector for said u-th user,>representing a second enhancement bias vector for the ith item;

wherein,representing an attribute distribution vector of the ith item, each element in the attribute distribution vector of the ith item representing a frequency of a corresponding interpretable rating label mentioned in an actual rating associated with the ith item,/a- >A preference distribution vector representing the u-th user,

wherein,，/>，/>，，/>，，/>，/>，

wherein,representing a distribution matrix of attributes>An attribute distribution vector representing an ith item having an interaction with the ith user, at +.>Middle->Splicing according to columns, and adding->Explicit preference degree of the u-th user, +.>Indicating the explicit preference level of said u-th user for said i-th item,/or->Indicating the implicit preference level of the u-th user, +.>Indicating the implicit preference level of said u-th user for said i-th item at +.>Middle->Splicing by line, in->Middle->Splicing according to rows, and (E) adding>Explicit polarity emotion value representing said ith item by said ith user,/->Explicit polarity emotion mean value representing said u-th user,>an implicit polarity emotion value representative of said ith item by said ith user,implicit polarity emotion mean value representing said u-th user,>representing the actual score of the ith item by the ith user,/for->A set of interpretable rating labels representing the ith user mentioned in the actual rating of the ith item, N representing the number of interpretable rating labels in the set,/a>Representing a relationship between the e-th interpretable rating tag and a polar emotion, the polar emotion including positive emotion and negative emotion.

3. The interpretable recommendation method of claim 2, wherein a first loss function for training the first deep learning model is represented as:

，

wherein,representing said first loss function,/->Representing a set of all tuples that each of the plurality of users and each of the plurality of items can make up if the u-th user has interaction with the i-th itemIf said u-th user has no interaction with said i-th item +.>，/>Representing said first preference score of said ith user for said ith item,/->。

4. The interpretable recommendation method of claim 3, wherein the second deep learning model includes: a second matrix decomposition module, a second multi-layer sensor module, and a second output layer,

wherein the output of the second matrix factorization module and the output of the second multi-layer perceptron module are used together as the input of the second output layer, and the output of the second output layer is used as the output of the second deep learning model;

the second matrix factorization module is represented as:

，

wherein,representing the output of said second matrix factorization module, a->Representing said first enhanced potential feature vector of said u-th user,/for >The first enhanced potential feature vector representing the ith item,/v->Second sparse bias representing said u-th user,>second sparse bias representing said ith item, +.>Representing a first potential feature matrix of an interpretable rating label with respect to a user,>representing a first potential feature matrix of an interpretable rating tag for an item;

the second multi-layer perceptron module is represented as:

，

wherein,，

wherein,representing the output of said second multi-layer sensor module,/or->Representing said second enhanced potential feature vector of said u-th user,/for>The second enhanced potential feature vector representing the ith item,/v->Representing an input of said second multi-layer sensor module,/->Representing an activation function of a kth layer sensor in said second multi-layer sensor module,/v>A weight matrix representing the kth layer sensor in the second multi-layer sensor module,/and/or>A bias vector representing the kth layer sensor in the second multi-layer sensor module,/v>Representing the output of the kth layer sensor in the second multi-layer sensor module,/->A second potential feature matrix representing an interpretable rating label with respect to the user,representing interpretable rating labels with respect to items A second latent feature matrix;

the second deep learning model is represented as:

，

wherein,a second plurality of preference scores representing the second deep learning model output for the second plurality of interpretable evaluation tags for the second set of tuples of the ith user and the ith item,>representing an activation function of the second output layer, and (2)>Representing the weight of the second output layer.

5. The interpretable recommendation method of claim 4, further comprising: establishing a self-expression matrix of the interpretable evaluation label in the actual evaluation;

wherein, a self-expression matrix is respectively established for the evaluation phrase and the emoticon in the interpretable evaluation label;

wherein the self-expression matrix of the evaluation phrase in the interpretable evaluation tag is found according to the following formula:

，

wherein the self-expression matrix of the emoticons in the interpretable evaluation tag is found according to the following formula:

，

wherein,a self-expression matrix representing the evaluation phrase, C representing the self-expression matrix of the emoticon,/->Representing a set of all tuples each of the plurality of users and each of the plurality of items being configurable, Representing actual preference scores of the ith user and the ith item for the plurality of interpretable rating labels, and->Representing a relation matrix of emoticons and emotion labels in the actual evaluation>Vector representing all elements as 1, +.>Representing elements located on the diagonal of the matrix, < >>And->Is constant.

6. The interpretable recommendation method of claim 5, further comprising: updating the second preference score output by the second deep learning model using a self-expression matrix of the interpretable evaluation tag;

wherein, in a case where the interpretable rating tag is the rating phrase, the plurality of second preference scores of the plurality of interpretable rating tags by the doublet of the u-th user and the i-th item are updated as:

，

wherein, in a case where the interpretable rating label is the emoticon, the plurality of second preference scores of the plurality of interpretable rating labels by the binary group constituted by the u-th user and the i-th item are updated as:

，

wherein,a second updated preference score representing a plurality of interpretable rating labels for the plurality of interpretable rating labels by the tuple of the u-th user and the i-th item, >A second plurality of preference scores representing the second deep learning model output for the second plurality of interpretable evaluation tags for the second set of tuples of the ith user and the ith item,>a self-expression matrix representing the evaluation phrase, C representing the self-expression matrix of the emoticon,/->Representing an identity matrix>Is constant.

7. The interpretable recommendation method of claim 5 or 6, wherein a second loss function for training the second deep learning model is expressed as:

，

wherein,representing said second loss function, +.>Representing a set of all tuples each of said plurality of users and each of said plurality of items, < +.>Represents the set of all interpretable rating labels, +.>Indicating whether the actual rating of said ith item by said ith user comprises an ith interpretable rating label,/for said ith item>Representing the second preference score of the second user and the ith item for the e-th interpretable rating tag,representing a predicted preference score of a binary group of the ith user and the ith item, derived based on a self-representation matrix of the interpretable rating label, for the ith interpretable rating label, >Is->The e element in (1)Interpretable evaluation tag is the evaluation phrase, < ->In case the interpretable rating label is the emoticon, the +_>，/>A self-expression matrix representing the evaluation phrase, C representing the self-expression matrix of the emoticon,/->A second plurality of preference scores representing the second deep learning model output for the second plurality of interpretable evaluation tags for the second set of tuples of the ith user and the ith item,>。

8. the interpretable recommendation method of claim 7, wherein the first and second deep learning models are jointly trained, wherein a joint loss function for jointly training the first and second deep learning models is represented as:

，

wherein,representing the joint loss function,/->Representing said first loss function,/->Representing said second loss function, +.>Is constant.

9. An apparatus for interpretable recommendations, the apparatus comprising:

at least one processor; and

at least one memory storing a computer program;

wherein the computer program, when executed by the at least one processor, causes the apparatus to perform the steps of the interpretable recommendation method according to any one of claims 1 to 8.