Multi-attribute neural collaborative recommendation system fusing project uncertainty

Abstract

The application relates to a multi-attribute neural collaborative recommendation system fusing project uncertainty, and belongs to the technical field of recommendation. Data information is acquired and processed from a database to obtain an initial rating matrix; based on information entropy theory, an uncertainty matrix is acquired according to the rating distribution of different users to the same project; user ID and un-rated project ID are obtained from the database and used as the input of a neural network to obtain multi-attribute predicted ratings of the user to the project; the uncertainty is introduced into the neural network framework and used as a weight factor to correct the multi-attribute predicted ratings; the corrected multi-attribute predicted ratings are used as the input of the neural network to perform result prediction and obtain the comprehensive ratings of the user to the project; the projects are sorted in descending order according to the predicted ratings; the number of recommended projects is set, the projects meeting the recommendation conditions are pushed to the target user, and a personalized recommendation list is formed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of recommendation technology, and mainly relates to a multi-attribute neural collaborative recommendation system fusing project uncertainty. BACKGROUND

[0002] With the continuous expansion of the Internet user scale and the accelerated development of commercialization process, a large amount of information begins to present in an explosive form in front of the user, so that the user is difficult to find the content of interest from it, and the information overload problem is increasingly prominent. As an efficient information filtering technology, the recommendation system successfully solves this problem. It mainly analyzes the user historical interaction information through big data technology, predicts the potential preference behavior of the user, so as to realize personalized recommendation service. This not only can save the user search cost and improve the user experience, but also can provide good commercial value for e-commerce, social media and other platforms.

[0003] At present, the collaborative filtering technology is most successful and widely used in the recommendation system, which mainly includes memory-based and model-based collaborative filtering. Compared with the memory-based method relying on the common scoring item, such as cosine similarity, the model-based method can use machine learning method or deep learning technology to capture the complex interaction relationship between the user and the project, and has better scalability and prediction effect, which can effectively solve the cold start problem. However, most of the existing deep learning-based recommendation methods only use the single scoring information of the user to the project for recommendation work. As we know, the purchase behavior of the user will be affected by many factors, because the preference degree of the user to different attributes of the project is different. Considering this problem, researchers propose a multi-attribute recommendation model to capture the user preference characteristics from multiple dimensions, so as to more comprehensively predict the user preference and improve the recommendation accuracy. However, this part of the research only uses the absolute score value of each attribute of the project, and ignores the attention degree of the user on different attributes, which will affect the personalized expression ability of the model to some extent. SUMMARY

[0004] In view of the above problems, the present application provides a multi-attribute neural collaborative recommendation system fusing project uncertainty, which comprises:

[0005] A data acquisition and preparation module is used for acquiring and processing user data to obtain the information required by the system and constructing a user-project scoring matrix R.

[0006] A multi-attribute scoring prediction module predicts the score of each attribute of the user u on the project i through a neural network.

[0007] A project uncertainty module introduces the concept of information entropy based on the acquired data, calculates the uncertainty according to the scoring distribution of different users on the same project, and obtains a project uncertainty matrix DU.

[0008] The multi-attribute score correction module corrects the multi-attribute scores predicted by the neural network through the obtained uncertainty matrix, so as to increase (or weaken) the proportion of the advantages (disadvantages) of the project i.

[0009] The result prediction module takes the corrected score data as input information of the neural network, and predicts the preference value of the target user u on the un-scored project.

[0010] The recommendation result acquisition module sorts the predicted scores obtained by the result prediction module in descending order, and finds the first k projects with the highest predicted preference values for the target user u for personalized push.

[0011] The multi-attribute neural collaborative recommendation system fusing project uncertainty, the data collection and preparation module is used to process the collected original data into a user-project score matrix R.

[0012] The multi-attribute neural collaborative recommendation system fusing project uncertainty, the multi-attribute score prediction module predicts the scores of each attribute of the project by the neural network, and the main steps are as follows:

[0013] S1: Convert the user ID and project ID into a low-dimensional feature vector, i.e. an embedding vector;

[0014] S2: Connect the two embedding vectors to form the input x of the neural network based on multi-attribute score, and the expression is as follows:

[0015]

[0016] Where, p u , q i are user embedding vector and project embedding vector respectively;

[0017] S3: Through the neural network, learn the high-order feature interaction between the user and the project, and predict the score of each attribute of the project i by the user u The expression is as follows:

[0018] h l =ReLU(W l h l-1 +b l )

[0019] Where, h l represents the output of the lth layer hidden layer of the full connection neural network, W l , b l respectively represent the weight matrix and bias vector of the lth layer hidden layer;

[0020]

[0021]

[0022] Wherein, L is the number of hidden layer of neural network, sigma is the nonlinear activation function of output layer.

[0023] The multi-attribute neural collaborative recommendation system fusing item uncertainty of the application, the item uncertainty module, by means of the concept of information entropy, introduces uncertainty to measure the uncertainty of user rating of item.The main steps are as follows:

[0024] S1: calculate the probability of each rating value r of the evaluated item i under attribute k in the rating interval RD The expression is as follows:

[0025]

[0026] S2: according to the definition of information entropy, the uncertainty of item i under each attribute k is calculated The expression is as follows:

[0027]

[0028] S3: according to the uncertainty of each attribute of the item, the uncertainty matrix is obtained, and the expression is as follows:

[0029]

[0030] The multi-attribute neural collaborative recommendation system fusing item uncertainty of the application, the multi-attribute rating correction module corrects the predicted rating of each attribute of the item through the multi-attribute uncertainty of the item.The main steps are as follows:

[0031] S1: normalize the uncertainty of each attribute of the item, and then take the inverse, to obtain the multi-attribute weight vector W of the item i i The expression is as follows:

[0032]

[0033] Wherein, n is the number of items, and c is the number of attributes.

[0034] S2: use the weight vector W i To correct the multi-attribute rating predicted by neural network To increase (or weaken) the proportion of the advantage (disadvantage) attribute of the item i.The expression is as follows:

[0035]

[0036] Wherein, is the element product operation of vector.

[0037] The result prediction module again utilizes a deep neural network to mine the potential relationship between the corrected multi-attribute prediction score and the comprehensive score, so as to obtain the final prediction score value of the user u on the item i The expression is as follows:

[0038] Z1=ReLU(W1y ui ′+b1)

[0039] Z2=ReLU(W2Z1+b2)

[0040] …

[0041] Z H =ReLU(W H Z H-1 +b H )

[0042]

[0043] Wherein, H represents the number of neural network layers; Z H , W H , b H respectively represent the output of the Hth hidden layer, the weight matrix and the bias vector; and sigma represents the nonlinear activation function of the neural network output layer.

[0044] The recommendation result acquisition module is used for acquiring the recommendation list of the target user u and performing personalized pushing.

[0045] The multi-attribute neural collaborative recommendation system fusing item uncertainty provided by the application utilizes the deep neural network method to learn the nonlinear interaction features between the user and the item and the preference features of the user on the multi-attribute labels of the item. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 It is a process schematic diagram of the recommendation system of the application;

[0047] Figure 2 It is a system framework diagram of the application;

[0048] Figure 3is the F1 value change of the present application under two data sets;

[0049] Figure 4 is the NDCG value change of the present application under two data sets. DETAILED DESCRIPTION

[0050] In order to make the purposes, technical solutions and advantages of the present application clearer, the following will combine the drawings to specifically describe the multi-attribute neural collaborative recommendation system fusing project uncertainty according to the present application. It should be understood that the specific implementation methods described herein are only used to explain the present application, and should not be used to limit the present application. Changes, modifications, additions, or replacements made by those skilled in the art within the scope of the present application should be covered in the scope of the claims of the present application.

[0051] Figure 1 is the process diagram of the multi-attribute neural collaborative recommendation system fusing project uncertainty of the present application. From Figure 1 It can be seen that the multi-attribute recommendation system proposed by the present application includes obtaining and processing data information from the database to obtain an initial rating matrix; based on the information entropy theory, the uncertainty matrix is obtained according to the rating distribution of different users to the same project; the user ID and the un-rated project ID are obtained from the database, which are used as the input of the neural network to obtain the multi-attribute predicted rating of the user to the project; the uncertainty is introduced into the neural network framework as a weight factor to correct the multi-attribute predicted rating; the corrected multi-attribute predicted rating is used as the input of the neural network for result prediction to obtain the comprehensive rating of the user to the project; the project is sorted in descending order according to the predicted rating; the number of recommended projects is set, and the projects meeting the recommendation conditions are pushed to the target user to form a personalized recommendation list.

[0052] Figure 2 is the framework diagram of the multi-attribute recommendation system of the present application. From Figure 2It can be seen that the multi-attribute recommendation system provided by the application comprises a data acquisition and preparation module, a multi-attribute score prediction module, an item uncertainty module, a multi-attribute score correction module, a result prediction module and a recommendation result acquisition module. The data acquisition and preparation module is used to extract data from a database and convert the data into a user-item score matrix; the multi-attribute score prediction module is used to mine complex interactive features of users and items by a neural network and predict scores of attributes of items by users; the item uncertainty module is used to calculate uncertainty according to score distribution of different users on the same item by means of information entropy theory; the multi-attribute score correction module is used to correct multi-attribute scores of items obtained by the multi-attribute score prediction module by taking uncertainty obtained by the item uncertainty module as a weight factor; the result prediction module is used to take the corrected multi-attribute scores as input of the neural network, mine potential relationship between the input and a comprehensive score and predict the comprehensive score of the item by the user; and the recommendation result acquisition module sorts the prediction result in descending order, sets a number of items to be recommended according to the sorted result, and recommends an item set meeting the recommendation condition and ranking at the top to the target user.

[0053] Further, the following case is taken as an example for illustration:

[0054] Suppose that there are m users I = {i1, i2, i3,..., i m} and n items J = {j1, j2, j3,..., j n}, a user-item score matrix is represented by R, and the number of item attributes is represented by c.

[0055] First, the data acquisition and preparation module is used to acquire and process system data to obtain each variable in the hypothesis. The specific implementation steps of the whole recommended system are as follows:

[0056] S1: The data acquisition and preparation module acquires required data information from a database, constructs a user-item score matrix R, and includes user ID, item ID and score information;

[0057] S2: The user ID and the item ID are converted into low-dimensional embedding vectors p u , q i ;

[0058] S3: The two embedding vectors are connected to form input x of a neural network based on multi-attribute scores, and the expression is as follows:

[0059]

[0060] S4: High-order feature interaction between users and items is learned by the neural network, and scores of attributes of an item i by a user u are predicted The expression is as follows:

[0061] h l = ReLU(W l h l-1 + b l )

[0062] where h l represents the output of the lth hidden layer of the fully connected neural network, W l and b l represent the weight matrix and bias vector of the lth hidden layer, respectively.

[0063]

[0064]

[0065] where L is the number of hidden layers of the neural network, and σ is the nonlinear activation function of the output layer.

[0066] S5: In the project uncertainty module, based on the concept of information entropy, the uncertainty is obtained according to the score distribution of different users on the same project, and the calculation process is as follows:

[0067] First, the probability of each score value r of the evaluated project i in the attribute k in the score interval RD is calculated

[0068]

[0069] Secondly, according to the definition of information entropy, the uncertainty of project i in each attribute k is calculated

[0070]

[0071] Further, the uncertainty matrix can be obtained as follows:

[0072]

[0073] S6: In the multi-attribute score correction module, the project uncertainty obtained in the project uncertainty module is taken as a weight factor to correct the multi-attribute score predicted by the neural network To increase (or weaken) the proportion of the advantages (disadvantages) of the project. Its expression is as follows:

[0074]

[0075]

[0076] where is the element product operation of the vector.

[0077] S7: In the result prediction module, the corrected multi-attribute score obtained by the multi-attribute score correction module is taken as input, and a deep neural network is used again to mine the potential relationship between the corrected multi-attribute prediction score and the comprehensive score, so as to obtain the final prediction score value of the user u on the project i The expression is as follows:

[0078] Z1=ReLU(W1y+b1) ui

[0079] Z2=ReLU(W2Z1+b2)

[0080] …

[0081] Z H =ReLU(W H Z H-1 +b H )

[0082]

[0083] Wherein, H represents the number of neural network layers; Z H , W H , b H respectively represent the output of the Hth hidden layer, the weight matrix and the bias vector; and sigma represents the nonlinear activation function of the neural network output layer.

[0084] Figures 3-4 The performance of the method of the present application on two evaluation indicators is measured respectively, and the two indicators are F1 value and normalized discounted cumulative gain (NDCG). Their calculation methods and measurement contents are as follows:

[0085] The F1 value is used to verify the recommendation accuracy of the model. The expression is as follows:

[0086]

[0087]

[0088]

[0089] Wherein, I p and I a respectively represent the predicted recommendation and the real recommendation of the project set. Recall represents recall rate, i.e. the proportion of the items that the user really likes in the actual favorite items in the recommendation list. Precision represents accuracy, i.e. the proportion of the items that the user really likes in the predicted favorite items in the recommendation list. The larger the comprehensive index F1 value is, the more accurate the recommendation result of the recommendation algorithm is.

[0090] ​NDCG is used to measure the quality of the ranking of the items to be recommended in the recommendation list. Its expression is as follows:

[0091]

[0092]

[0093] wherein DCG represents the discounted cumulative gain, IDCG represents the maximum DCG value in the ideal case; k represents the length of the recommendation list, rel j represents the relevance of the recommended item at position j.

Claims

1. A multi-attribute neural collaborative recommendation system fusing item uncertainty, characterized by comprising: a data collection and preparation module for obtaining and processing user data to obtain information required by the system and constructing a user-item rating matrix R; a multi-attribute rating prediction module for predicting ratings of each attribute of a user u on an item i through a neural network; an item uncertainty module for introducing the concept of information entropy based on the obtained data, calculating uncertainty according to the rating distribution of different users on the same item, and obtaining an uncertainty matrix DU of the item; a multi-attribute rating correction module for correcting the multi-attribute ratings predicted by the neural network through the obtained uncertainty matrix, and the specific steps are as follows: S1: normalizing the uncertainty DU of each attribute k of item i (i,k) and taking the inverse, to obtain the weight vector W of item i on attribute k (i,k) whose expression is as follows: wherein n is the number of items, c is the number of attributes; thus, the multi-attribute weight vector W of item i can be obtained i ; S2: Utilize the weight vector W i De-biasing multi-attribute scores predicted by neural networks To increase or weaken the proportion of the strength or weakness attribute of item i, the expression is as follows: where is the element-wise product of vectors y ui is the predicted rating of item i by user u. a result prediction module for predicting the preference value of a target user on an un-rated item by taking the corrected rating data as the input information of the neural network; a recommendation result acquisition module for sorting the predicted ratings obtained by the result prediction module in descending order to find the top k items with the highest predicted preference values for the target user u for personalized push.

2. The multi-attribute neuro-collaborative filtering recommendation system that fuses project uncertainty as claimed in claim 1, wherein, The data collection and preparation module is used to process the collected raw data into a user-item rating matrix R.

3. The multi-attribute neuro-collaborative filtering recommendation system that fuses project uncertainty as claimed in claim 1, wherein, The multi-attribute rating prediction module predicts ratings of each attribute of a user on an item through a neural network, and the main steps are as follows: S1: convert the user ID and item ID into a low-dimensional feature vector, i.e., an embedding vector; S2: connect the two embedding vectors to form the input x of the neural network based on multi-attribute ratings, which is expressed as follows: where p u and q i are user and item embedding vectors, respectively. S3: Predict the ratings of each attribute of item i by the user u through learning high-order feature interactions between users and items by neural networks The expression is as follows: h l = ReLU(W l h l-1 + b l ) where h l represents the output of the l-th hidden layer of the fully connected neural network, W l , b l respectively represent the weight matrix and the bias vector of the l-th hidden layer. where L is the number of hidden layers of the neural network, and σ is the nonlinear activation function of the output layer.

4. The multi-attribute neuro-collaborative filtering recommendation system that fuses project uncertainty as claimed in claim 1, wherein, The item uncertainty module introduces uncertainty to measure the uncertainty of user ratings on items by means of the concept of information entropy, and the main steps are as follows: S1: calculate the probability of each rating value r of the evaluated item i at attribute k appearing in the rating interval RD The expression is as follows: S2: According to the definition of information entropy, the uncertainty of project i in each attribute k is calculated The expression is as follows: S3: obtain the uncertainty matrix according to the uncertainty of each attribute of the item, which is expressed as follows:

5. The multi-attribute neuro-collaborative filtering recommendation system that fuses project uncertainty as claimed in claim 1, wherein, The result prediction module again uses a deep neural network to mine the potential relationship between the corrected multi-attribute prediction scores and the comprehensive scores, to obtain a final prediction score value of the user u on the item i The expression is as follows: Z1 = ReLU(W1y + b1) ui + b1) Z2=ReLU(W2Z1+b2) … Z H = ReLU(W H Z H-1 + b H ) where H denotes the number of neural network layers; Z H , W H and b H denote the output of the Hth hidden layer, the weight matrix and the bias vector, respectively; and σ denotes the non-linear activation function of the output layer of the neural network.

6. The multi-attribute neuro-collaborative filtering recommendation system that fuses project uncertainty as claimed in claim 1, wherein, The recommendation result acquisition module is used to obtain the recommendation list of the target user u and perform personalized push, and the process includes: first setting the number of items recommended to the target user k, then sorting the predicted ratings obtained in the result prediction module, and pushing the top k items with the largest predicted rating values in the result to the target user to form the personalized recommendation list of the user.

Images