CN110866145A

CN110866145A - Co-preference assisted deep single-class collaborative filtering recommendation method

Info

Publication number: CN110866145A
Application number: CN201911080135.2A
Authority: CN
Inventors: 张全贵; 胡嘉燕; 李鑫
Original assignee: Liaoning Technical University
Current assignee: Liaoning Technical University
Priority date: 2019-11-06
Filing date: 2019-11-06
Publication date: 2020-03-06
Anticipated expiration: 2039-11-06
Also published as: CN110866145B

Abstract

The invention discloses a co-preference assisted deep single-class collaborative filtering recommendation method, which comprises the following steps: data acquisition and processing; dividing a data set; constructing a co-preference auxiliary model; training models and item recommendations. The method adopts the personal depth latent factor to express that the network (DLFR) can be used for predicting the probability of user item interaction independently, and also can be used as a sub-network in a co-preference assisted deep single-class collaborative filtering recommendation model to output the coupling relation of user item interaction. The method and the system use single-class collaborative filtering (OCCF) and the deep neural network (MLP) to learn the common preference of the user neighbors to the item categories and the interaction between the users and the items, are beneficial to learning the influence of the deeper user neighbor preference on the user item interaction, and have good recommendation accuracy and interpretability.

Description

Co-preference assisted deep single-class collaborative filtering recommendation method

Technical Field

The invention belongs to the technical field of computer artificial intelligence, and particularly relates to a co-preference assisted depth single-class collaborative filtering recommendation method.

Background

The recommendation system plays an important role in helping users find items which are in line with personal preferences in a huge video-audio library. To provide useful suggestions to the user, we capture and understand the information that the user feeds back. In practical scenarios, explicit data is not well captured, while implicit data is better captured, such as page link clicks, video viewing, product purchases, and other user/item interaction history. However, the implicit data has the problems of uncertainty, quarterly sparsity, imbalance and the like, and high recommendation precision is difficult to obtain only by using implicit constraints.

Disclosure of Invention

Based on the defects of the prior art, the technical problem solved by the invention is to provide a common preference assisted deep single-class collaborative filtering recommendation method which has good recommendation accuracy and interpretability.

In order to solve the technical problems, the invention is realized by the following technical scheme:

the invention combines explicit characteristics such as demographic information and the like with potential factors, predicts the user-item interaction probability and recommendation visibility influenced by the explicit characteristics by learning the preferences of the neighbors of users to the preferences of certain type of articles, and provides a co-preference assisted deep single-type collaborative filtering recommendation method, which comprises the following steps:

step S1: data acquisition and processing;

step S2: dividing a data set, using the collected data set in a leave-one-out mode, using the latest user item interaction item as a test set of each user, using the other items as training data sets, randomly extracting items except the user interaction item set, and forming the test data set of the user together with the test items;

step S3: constructing a co-preference auxiliary model;

step S4: training models and item recommendation, wherein the training data set and the testing data set obtained in the step S2 are respectively used for training and evaluating the co-preference assisted model constructed in the step S3, and the co-preference is obtained by indirectly utilizing the explicit characteristics through learning the interaction relation between natural neighbor users having similar explicit characteristics with the current user and a certain class of articles.

Preferably, the specific step of step S3 is:

s301, constructing an input layer: the input layer comprises four input contents, namely a user natural neighbor vector, a project category vector and a user/project id;

s302, constructing a co-preference auxiliary single-type collaborative filtering model;

and S303, constructing an output layer.

The specific steps of step S302 are as follows:

s30201, constructing a personal depth potential factor representation network: predicting probability a of user-item interaction through network learning by inputting one-hot vectors of user identification u and item identification v_L，a_LIs represented by a highly abstract common preference representation a 'of user neighbors to item categories output by the network'_LThe dimensions of the data are consistent;

s30202, constructing a public preference representation network based on the neighbor: highly abstract common preference representation a 'of user neighbors to item categories is obtained through learning through full-connection layer user natural neighbors NB and item categories VF'_L；

S30203, constructing a co-preference assisted deep learning network: output a 'of step S30202'_LAnd output a of step S30201_LAnd combining the vectors into a vector connected in series, processing the vector by a full connection layer, outputting a final interactive vector, and finally outputting the interactive vector activated by a Sigmoid function.

Further, the specific steps of constructing the personal depth latent factor representation network in step S30201 are as follows:

s3020101, constructing a vector mapping layer: inputting a single-hot vector full-connection layer of a user identifier u and a project identifier v, converting the single-hot vector full-connection layer into vectors p and q of specified dimensions, wherein the p and the q respectively represent potential vectors of the user u and the project v;

s3020102, capturing interaction among user items: on top of the embedding layer, the embedding vectors p and q are input into a fusion layer that multiplies the corresponding elements of p and q, and then outputs a lineThe sexual vector r represents a linear user-item interaction, represented as:

s3020103, user item nonlinear interaction: through the processing of the full connection layer, the non-linear interaction between the user and the article can be converted into a representation;

and S3020104, outputting the user item interaction probability or the user item probability or the coupling relation of the user item interaction.

Optionally, the specific steps of constructing the co-preference assisted deep learning network in step S30203 are as follows:

s3020301, selecting a natural neighbor method: calculating a user similarity matrix of a data set according to the user explicit characteristics, and then taking the first K similar users as natural neighbors NB of the users, and respectively using two similarity indexes and a Hamming distance;

s3020302, constructing a vector mapping layer: inputting natural neighbors and item categories of layer activity users;

s3020303, embedding the sparse vector into the dense vector of the item category by using a full connection layer on the item category input layer, and then outputting the user neighbor and the item category embedded full connection layer as a vector E_cAnd inputting the parameters into a multilayer fully-connected neural network to learn the highly abstract common preference of the neighbors to the item categories.

Optionally, the specific steps of training the model and recommending the item in step S4 are:

s401, constructing a loss function: calculating a loss value of the model by using a cross entropy loss function;

s402, constructing an optimization function: using Adam as an optimization function for the model;

s403, item recommendation: and sending the information of one user and the item information into a co-preference auxiliary single-type collaborative filtering model, outputting and predicting the interaction probability of the user to the item, and judging whether the item is recommended to the user.

Therefore, the co-preference assisted deep single-class collaborative filtering recommendation method has at least the following beneficial effects:

(1) the invention provides a deep neural network model combining preference representation and implicit feedback based on explicit characteristics and personal potential factors so as to improve the performance of single-class collaborative filtering of OCCF.

(2) The invention does not directly use the explicit characteristics, but indirectly learns the high-level relation between the explicit characteristics of the user and the object types by learning the interaction between the natural neighbors of the active user and a certain class of objects so as to fully play the role of the explicit characteristics.

(3) The individual depth latent factor representation network (DLFR) can be used for predicting the probability of user item interaction independently, and can also be used as a sub-network in a co-preference assisted depth single-class collaborative filtering recommendation model to output the coupling relation of user item interaction.

(4) The method and the system use single-class collaborative filtering (OCCF) and a deep neural network (MLP) to learn the common preference of the user neighbors on the item categories and the interaction between the users and the items, and are beneficial to learning the influence of the deeper user neighbor preference on the user item interaction.

The foregoing description is only an overview of the technical solutions of the present invention, and in order to make the technical means of the present invention more clearly understood, the present invention may be implemented in accordance with the content of the description, and in order to make the above and other objects, features, and advantages of the present invention more clearly understood, the following detailed description is given in conjunction with the preferred embodiments, together with the accompanying drawings.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings of the embodiments will be briefly described below.

FIG. 1 is a block flow diagram of a co-preference assisted deep single-class collaborative filtering recommendation method of the present invention;

FIG. 2 is a block flow diagram of the present invention for constructing a co-preference assisted model;

FIG. 3 is a block diagram of a process for constructing a co-preference assisted single-class collaborative filtering model according to the present invention;

FIG. 4 is a block flow diagram of the present invention for constructing a personal depth latent factor representation network (DLFR);

FIG. 5 is a block diagram of a process for constructing a co-preference assisted deep learning network (NB-OCCF) according to the present invention;

FIG. 6 is a block flow diagram of a training model and item recommendation of the present invention;

FIG. 7 is a block diagram illustrating a structure of a collaborative filtering recommendation method for constructing a co-preference auxiliary single category according to the present invention.

Detailed Description

Other aspects, features and advantages of the present invention will become apparent from the following detailed description, taken in conjunction with the accompanying drawings, which form a part of this specification, and which illustrate, by way of example, the principles of the invention. In the referenced drawings, the same or similar components in different drawings are denoted by the same reference numerals.

As shown in fig. 1 to 7, the method for recommending deep single-class collaborative filtering assisted by common preference according to the present invention includes the following steps:

s1, data acquisition and processing: downloading MoieLens 100K and MovieLens 1M data sets in a Grouplens website, and using the original scoring matrix

The conversion is to a binarization preference matrix R e {0,1}, where each element is represented as 0 or 1, indicating whether there is an interaction between the user-item pair.

The "UserID", "MovieID", "Rating", "Timestamp" datasets in the scoring matrix after transformation are as follows:

UserID	MovieID	Rating	Timestamp
				1	1993	1	978300760

the characteristic data of the project include "MovieID", "Title", "Genres", as follows:

MovieID	Title	Genres
			1	Toy Story(1995)	Animation\|Children's\|Comedy

demographic information data of the user includes "user id", "genter", "Age", "occupancy", "Zip-code", as follows:

UserID	Gender	Age	Occupation	Zip-code
					1	F	1	10	48067

s2, dividing the data set: and (3) adopting a leave-one-out method for the collected data set, taking the latest user item interaction items (the scores of the user on the items) as a test set of each user, taking the other items as a training set, randomly extracting 99 items except the user interaction item set, and forming the test data set of the user together with the test items. The training set data is used for training the model and learning and predicting the probability of user-item interaction; the test data set data is used to evaluate the generalization ability of the model.

S3, constructing a co-preference assisted model: personal depth latent factor representation network (DLFR) and neighbor-based co-preference representation network (NB-CPR); a co-preference assisted deep learning network (NB-OCCF).

S4, training the model and recommending the item: the training data set and the testing data set obtained in the step S2 are respectively used for training and evaluating the co-preference assisted model constructed in the step S3, and the co-preference is obtained indirectly by using the explicit feature by learning the interaction relationship between the natural neighbor user having the similar explicit feature to the current user and a certain kind of articles, so that the collaborative filtering recommendation method is assisted to improve the recommendation accuracy.

As shown in fig. 2, the specific steps of establishing the model in step S3 are as follows:

s301, constructing an input layer: the input layer includes four input contents, a user natural neighbor vector, a project category vector, and a user/project id. The input layers are in the form of (user _ vec _ input, item _ attr _ input, user _ id _ input, item _ id _ input).

For example, NB1 with user _ vec _ input having a UserID of 1 is [ 2346119522855247530222022162215521372133 ].

The MovieID of item _ attr _ input is [ 000000010000000000 ] 1.

In Keras, the input layer can be implemented with the following code:

user_vector_input＝Input(shape＝(K，)，dtype＝'float32'，name＝'user_vector_input')item_attr_input＝Input(shape＝(18，)，dtype＝'float32'，name＝'item_attr_input')user_id_input＝Input(shape＝(1，)，dtype＝'float32'，name＝'user_id_input')item_id_input＝Input(shape＝(1，)，dtype＝'float32'，name＝'item_id_input')

wherein K is the number of natural neighbors of the user.

S302, constructing a co-preference auxiliary single-type collaborative filtering model: personal depth latent factor representation network (DLFR); neighbor-based co-preference representation networks (NB-CPR) and co-preference assisted deep learning networks (NB-OCCF).

S303, constructing an output layer: user-item interaction probabilities affected by the explicit features are predicted.

In Keras, the output layer can be implemented with the following code:

topLayer＝Dense(1，activation＝'sigmoid'，init＝'lecun_uniform'，name＝'topLayer')(dense_1)

as shown in fig. 3, the step S302 is to construct a personal depth latent factor representation network (DLFR); the specific steps of the neighbor-based common preference representation network (NB-CPR) and the common preference assisted deep learning network (NB-OCCF) are as follows:

s30201, constructing a personal depth latent factor representation network (DLFR): predicting probability a of user-item interaction through DLFR network learning by inputting one-hot vectors (one-hot) of user identification u and item identification v_L。a_LAnd a highly abstract common preference representation of user neighbor-to-item categories output by a neighbor-based common preference representation network (NB-CPR)a′_LAre consistent in dimension.

S30202, constructing a neighbor-based public preference representation network (NB-CPR): highly abstract common preference representation a 'of user neighbors to item categories is obtained through learning through full-connection layer user natural neighbors NB and item categories VF'_L。

S30203, constructing a co-preference assisted deep learning network (NB-OCCF): output a 'of NB-CPR'_LAnd output a of DLFR_LAnd combining the vectors into a vector (denoted as r) connected in series, and then processing the vector by a full connection layer to output a final interaction vector. The final output of NB-OCCF is activated by a Sigmoid function and is thus compressed to the range 0,1]。

Thus, the training data set for NB-OCCF may be represented as

y⁽ⁱ⁾The target of > - < i ∈ {1, 2., | D | } and NB-OCCF can be expressed as a predicted probability:

P_Θ(y＝1|u,v,NB，v_c)

of NB-OCCF

The final output expression is:

the NB-OCCF model was trained using the same loss function as DLFR.

In Keras, the output layer implementation code is as follows:

merge_attr_id_embedding＝merge([attr_2，id_2]，mode＝'concat')

dense_1＝Dense(256)(merge_attr_id_embedding)

dense_1＝Activation('relu')(dense_1)

as shown in fig. 4, the specific steps of constructing the personal depth latent factor representation network (DLFR) in step S30201 are as follows:

s3020101, constructing a vector mapping layer:converting a one-hot full-connection layer of input user identifications u and item identifications v into vectors p and q of specified dimensions, wherein p and q respectively represent potential vectors of the user u and the item v and are formulated as

Wherein the weight matrix W_uòR^k×|U|And W_vòR^k×|v|The same is true in the input layer and the embedding layer.

In Keras, the code that constructs the vector mapping layer of the personal depth latent factor representation network is as follows:

user_id_input＝Input(shape＝(1，)，dtype＝'float32'，name＝'user_id_input')user_id_Embedding＝Embedding(input_dim＝num_users，output_dim＝64，name＝'user_id_Embedding'，embeddings_initializer＝RandomNormal(mean＝0.0，stddev＝0.01，seed＝None)，W_regularizer＝l2(0)，input_length＝1)user_id_Embedding＝Flatten()(user_id_Embedding(user_id_input))item_id_inputInput(shape＝(1，)，dtype＝'float32'，name＝'item_id_input')item_id_Embedding＝Embedding(input_dim＝num_items，output_dim＝64，name＝'item_id_Embedding'，embeddings_initializer＝RandomNormalmean＝0.0，stddev＝0.01，seed＝None)，W_regularizer＝l2(0)，input_length＝1)item_id_EmbeddingFlatten()(item_id_Embedding(item_id_input))

s3020102, capturing interaction among user items: above the embedding layer, the embedding vectors p and q are input into a fusion layer that multiplies the corresponding elements of p and q, then the output linear vector r represents a linear user-item interaction, represented as:

unlike the basic matrix silver decomposition model, in which the user-item interactions are modeled by inner products, in the multilayer fully-connected neural network, deep abstract user-item interactions are learned, Wu and Wv are trained to represent potential factors of all users and items, and since the codes of users and items are expressed as unique hot codes, the codes of Wu and WvEach column represents a certain user or project latent factor p and q, respectively. For a given item v, each dimension of v measures the extent to which the item has these factors. For a given user u, each dimension of u measures the user's level of interest in the item's corresponding factor. Thus, the output vector r of the element layer captures the linear interaction between the user and the item.

In Keras, the code building the personal depth latent factor to represent the interaction between the captured user items of the network is as follows:

merge_id_embedding＝merge([user_id_Embedding，item_id_Embedding]，mode＝'mul')

s3020103, user item nonlinear interaction: through the processing of the fully connected layer, it can be converted into a representation of the non-linear interaction between the user and the article, represented as:

...

wherein, W₁,W₂,...,W_LAnd b₁,b₂,...,b_LWeight matrix and bias representing layers, a₁,a₂,...,a_LThe output of each layer activated by the ReLU activation function is represented.

In Keras, the code that constructs the transformation of user item interactions into non-linear interactions of the personal depth latent factor representation network is as follows:

id_1＝Dense(64)(merge_id_embedding)

id_1＝Activation('relu')(id_1)

id_2＝Dense(32)(id_1)

id_2＝Activation('relu')(id_2)

s3020104, outputting the user item interaction probability or the user item probability or the coupling relation of the user item interaction: the purpose of DLFR as a single network is to predict the probability of user-item interaction, the invention uses the logistic regression sigmoid function to compress the output of the model to the interval [0, 1]And interpreting the target problem as a probability p_Θ(y 1| u, v |), using

As the prediction output:

wherein W₀Is the weight matrix of the last layer, a_LIs the output of the previous layer, b₀Is the bias weight vector of the last layer. For the OCCF problem, there are no negative samples, so a negative sample strategy is needed to extract some negative samples. In the present invention, a uniform negative sampling strategy is employed, and some counter-examples are drawn from the interactions not observed in matrix Y. E.g., one example in a given training dataset, D { < (u) } {⁽ⁱ⁾,v⁽ⁱ⁾),y⁽ⁱ⁾> and corresponding prediction output

(i represents the ith example and i ∈ {1,2, 3., | D | }. if DLFR is used as the sub-network of NB-OCCF, the final output is a_LThat is, the non-linear interaction of the user item is the coupling relationship of the user item interaction.

In Keras, the code that constructs the probability that a person-deep latent factor represents a user-item interaction with a network is as follows:

topLayer＝Dense(1，activation＝'sigmoid'，init＝'lecun_uniform'，name＝'topLayer')(id_2)

as shown in fig. 5, the specific steps of constructing the co-preference assisted deep learning network (NB-OCCF) in step S30203 are as follows:

s3020301, selecting a natural neighbor method: and calculating a user similarity matrix of a data set according to the explicit characteristics of the users, and then taking the first K similar users as natural neighbors NB of the users, and respectively using two similarity indexes COS and Hamming distance. For COS, it is a coupled attribute similarity measure for classification objects that considers coupling between attributes. Hamming distance is a simple and common method of computing similarity using classification attributes. This setup was to observe the impact of different similarity measures on the OCCF performance.

For example: the implementation code for obtaining 10, 20.30, 40, 50, 100 natural neighbors of a user using hamming distance is as follows:

mat＝loadCrossDistanceMat()

argsortmat＝argSortMatrix(mat)

for i in[10，20，30，40，50，100]:

getNeighbors(argsortmat，i)

s3020302, constructing a vector mapping layer: the natural neighbors and item categories of the layer active users are entered. For example, training example (u, v) in DLFR, neighbor users u and v will be entered into the entry of the NB-CPR category. Representing the respective natural neighbors of all users as T vectors NB₁,NB₂,...,NB_T. T is the number of neighbors, each neighbor NB_i(i e {1, 2.,. T }) is a one-hot code that contains | U | natural neighbor identifications. Thus, for each training example, each active user has | U | natural neighbors. The embedding vector of each natural neighbor is learned by the embedding layer. Thus, all user neighbors are passed through the embedding layer to obtain user neighbor E_NB，E_NB(i)Representing the embedding of each neighbor.

When in use

The combining weights are assigned different values depending on the particular application. In the present invention, for each

The same values are assigned. The more hypothesisThe effect of being similar neighbors on hybrid embedding is not absolutely high. Thus, the mean of all neighborhood embeddings is calculated as a hybrid embeddings.

Another input to NB-CPR is the category of the item in the DLFR model. The item category is obtained by a category attribute of the item (e.g., movie type) or a clustering method (e.g., K-Means algorithm). In the present invention, item type attributes are treated as item categories and encoded as multiple heat vectors v_c＝(v_c1,v_c2,...,v_|c|)，c_i∈{c₁,c₂,...,c_|c|Denoted as a set of categories for all items. Each one of which is

May be 1 or 0 depending on whether the item has c or not_iThe category (2). This representation method can maintain the versatility of encoding the categories of items belonging to multiple categories.

The implementation code is as follows:

user_vector_input＝Input(shape＝(20，)，dtype＝'float32'，name＝'user_vector_input')user_vector_embedding＝Embedding(input_dim＝6041，output_dim＝64，name＝'user_vector_embedding'，embeddings_initializer＝RandomNormal(mean＝0.0，stddev＝0.01，seed＝None)，W_regularizer＝l2(0)，input_length＝10)user_vector_embedding＝user_vector_embedding(user_vector_input)user_vector_embedding＝Lambda(lambda x:K.mean(x，axis＝1))(user_vector_embedding)item_attr_input＝Input(shape＝(18，)，dtype＝'float32'，name＝'item_attr_input')item_attr_embedding＝Dense(18，activation＝'relu'，name＝'item_attr_embedding')(item_attr_input)

s3020303, embedding the sparse vector into the dense vector of the item category by using a full connection layer on the item category input layer. Then embedding user neighbors and item categories into a full connection layer and outputting the vector E_cAnd inputting the parameters into a multilayer fully-connected neural network to learn the highly abstract common preference of the neighbors to the item categories. It is expressed as:

...

wherein, W'₁,W′₂,...,W′_LAnd b'₁,b′₂,...,b′_LRepresenting weight matrix and offset, a 'of each layer'₁,a′₂,...,a′_LRepresenting the output of the layers activated by the ReLU activation function.

The implementation code is as follows:

merge_attr_embedding＝merge([user_vector_embedding，item_attr_embedding]，mode＝'concat')

attr_1＝Dense(32)(merge_attr_embedding)

attr_1＝Activation('relu')(attr_1)

attr_2＝Dense(16)(attr_1)

attr_2＝Activation('relu')(attr_2)

as shown in fig. 6, the specific steps of training the model and recommending the item in step S4 are as follows:

s401, constructing a loss function: the loss value of the model is calculated using a cross entropy loss function, which is formulated as follows:

the parameters Θ of the loss function that can then be learned:

in Keras, the cross entropy loss function implementation code is as follows:

model.compile(optimizer＝Adam(lr＝learning_rate)，

loss＝'binary_crossentropy'，metrics＝['accuracy'，'mae'])

s402, constructing an optimization function: using Adam as an optimization function for the model; the learning rate parameter is set to 0.001, and the default values of the keras are used for other parameters.

S403, item recommendation: and (3) sending information of one user and item information (user id and item id) into a common preference auxiliary single-type collaborative filtering model, outputting and predicting the interaction probability of the user to the item, and judging whether the item is recommended to the user.

In summary, the co-preference assisted depth-single-class collaborative Filtering recommendation method (NB-OCCF, One-class collaborative Filtering) of the present invention includes the following steps: s1, data acquisition; s2, dividing the data set; s3, constructing a model: (1) neighbor-based common preference representation network (NB-CPR), neighbor-based common prediction: calculating a user similarity matrix according to the explicit characteristics of the user through similarity index (cosine similarity COS and Hamming) distance to obtain a natural neighbor of the user; the natural neighbors of the user and the categories of the predicted items are input into the fully connected neural network, and highly abstract common preference expression of the user neighbors on the categories of the items is learned. (2) Personal depth potential factor representation networks (DLFR), Deep versus factors representation: respectively inputting the ids of the user and the item to obtain tensors with the same dimensionality, then learning the interaction between the user and the item through MLP, and predicting the probability of user-item interaction; (3) common preference assisted deep learning network (NB-OCCF): the probability of interacting with the user-project is input to a full-connection layer through the highly abstract common preference representation of the user neighbors on the project categories, and a final interaction vector is output; s4, training the model and recommending the item: the five-level scores of the user items are used as the input of the model, and the model continuously updates parameters in the training process, so that the relation between the common preference of the user on natural neighbors and predicted item categories and potential factors of user-item interaction is better learned.

While the foregoing is directed to the preferred embodiment of the present invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof, and the scope thereof is determined by the claims that follow.

Claims

1. A co-preference assisted deep single-class collaborative filtering recommendation method is characterized by comprising the following steps:

step S1: data acquisition and processing;

step S3: constructing a co-preference auxiliary model;

2. The co-preference assisted deep single-class collaborative filtering recommendation method according to claim 1, wherein the specific steps of the step S3 are:

and S303, constructing an output layer.

3. The method for deep single-class collaborative filtering recommendation with assistance of common preference according to claim 2, wherein the specific steps of the step S302 are as follows:

4. The co-preference assisted deep single-class collaborative filtering recommendation method according to claim 3, wherein the specific steps of constructing the personal deep latent factor representation network in the step S30201 are as follows:

5. The co-preference assisted deep single-class collaborative filtering recommendation method according to claim 3, wherein the step S30203 of constructing the co-preference assisted deep learning network comprises the following specific steps:

6. The co-preference assisted deep single-class collaborative filtering recommendation method according to claim 1, wherein the training models and item recommendations in step S4 are specifically performed by: