Deep neural network recommendation system with prediction reliability

Abstract

The deep neural network recommendation system with prediction reliability comprises the following steps: obtaining and cleaning data information from a database to obtain required information; dividing a user-item rating matrix R into corresponding several independent and only containing 0-1 binary sub-matrices according to different rating values set by the system; using a double-tower model in a deep neural network to train the obtained sub-matrices in parallel to predict the probability value of the same blank area in each sub-matrix; normalizing the obtained probability value in each sub-matrix, and taking the maximum normalized probability value as the reliability probability; finding the rating represented by the corresponding sub-matrix according to the obtained reliability probability, and taking the rating as the predicted rating of the blank area; comparing the predicted rating with the reliability threshold value set by the system, filtering the predicted rating with a reliability probability lower than the threshold value, and retaining the rating with a higher reliability probability; sorting the filtered predicted rating set, pushing the first k items with higher predicted ratings to the target user, and forming a personalized recommendation list.

Description

TECHNICAL FIELD

[0001] The present application relates to the recommendation technical field, and particularly relates to a deep neural network recommendation system with prediction reliability. BACKGROUND

[0002] In recent years, deep neural networks have shown superior performance in the field of computer vision, natural language processing and other machine learning fields. Compared with shallow networks, deep networks can better capture high-order nonlinear relationships between user and item features from data. Therefore, this technology has also been widely used in the recommendation field and has achieved good results.

[0003] However, most of the existing deep learning recommendation systems only focus on the accuracy of the recommended items, and ignore whether the recommended items have reliability. In the recommendation system, providing the corresponding reliability probability for the rating predicted by the recommendation algorithm can effectively filter out low reliability predictions, so as to realize reliable and accurate recommendation and improve user satisfaction. At the same time, since the reliability information has certain interpretability, it can also enhance the credibility of the user to the recommendation system. Therefore, it is particularly important to design a deep neural network recommendation system with prediction reliability to further improve the recommendation accuracy. SUMMARY

[0004] In view of the above problems, the present application provides a deep neural network recommendation system with prediction reliability to help the recommendation system filter out unreliable predicted ratings, comprising the following:

[0005] A data acquisition and preparation module is used to acquire and clean the data collected in the database to obtain the required information.

[0006] A rating matrix division module is used to divide the user-item rating matrix R into corresponding several independent and only containing 0-1 binary sub-matrices according to different rating values set by the system. For example, if the rating interval in the system is 1-5 points, R is divided into 5 sub-matrices, and each sub-matrix represents a rating.

[0007] A rating probability calculation module is used to use the double tower model in the deep neural network to train the obtained sub-matrices in parallel to predict the probability value of the same blank area in each sub-matrix.

[0008] A reliability acquisition module is used to normalize the probability value obtained in each sub-matrix, and take the maximum normalized probability value as the reliability probability.

[0009] A rating value acquisition module is used to find the rating represented by the corresponding sub-matrix according to the obtained reliability probability, and take the rating as the predicted rating of the blank area.

[0010] A reliability threshold filtering module is configured to filter the prediction ratings with a reliability probability lower than a threshold by comparing with a system set reliability threshold, and keep the ratings with a higher reliability probability.

[0011] A recommendation result obtaining module is configured to internally sort the kept prediction ratings with a higher reliability probability, and push the top k items with a higher prediction rating to the target user to form a personalized recommendation list.

[0012] The rating matrix dividing module is configured to divide the user-item rating matrix R into several independent and only containing 0-1 binary sub-matrices according to different rating values set by the system.

[0013] The rating probability calculating module is configured to use a two-tower model in a deep neural network to train the obtained sub-matrices in parallel to predict the probability value of the same blank area in each sub-matrix.

[0014] ① Input layer, input the feature embedding vector z of the user u (item i) u (z i ) of the user u (item i)

[0015] ② Hidden layer, use the activation function ReLU to prevent model overfitting, and obtain the output l of all hidden layers.

[0016] ReLU(z u )=max(0,z u )

[0017] l1=W1 T z u

[0018] l n =ReLU(W1 T l n-1 +b n ),n=2,...,X-1

[0019] Wherein, W n and b n respectively represent the weight matrix and bias vector of the nth hidden layer.

[0020] ③ Output layer, obtain the latent feature vector p u (q i ) of the user u (item i)

[0021]

[0022]

[0023] IV. Calculate the predicted rating probability of user u on item i as follows:

[0024]

[0025] where W out and e denote the weight matrix and element-wise product of vectors respectively; the sigmoid function σ(x) = 1 / (1+e -x ) is used to limit the output within the range (0, 1).

[0026] The reliability acquisition module of the present application is used to normalize the predicted probability obtained in the rating r represented by each sub-matrix and take the normalized maximum probability value as the reliability probability of the region.

[0027]

[0028] The rating value acquisition module of the present application is used to find the rating r represented by the corresponding sub-matrix according to the obtained reliability probability, and take the rating as the predicted rating p ui of the blank region.

[0029] The reliability threshold filtering module of the present application is used to compare the reliability probability corresponding to the predicted rating with the system set reliability threshold θ, if the reliability probability is less than the threshold, it is considered that the predicted rating is unreliable, and filtering is performed, and vice versa.

[0030] The recommendation result acquisition module of the present application is used to acquire the recommendation list of the target user u. Which includes: sorting the retained predicted ratings with higher reliability probability internally from large to small, setting the number of recommended items k, selecting the top k predicted ratings with higher reliability probability to push to the target user u, to form the personalized recommendation list of the target user u.

[0031] The deep neural network recommendation system with prediction reliability proposed in the present application only uses the interaction relationship and rating information between users and items, without other auxiliary information to obtain the reliability probability of the predicted rating. Therefore, the system has good universality and expansibility. At the same time, the system retains the predicted ratings with higher reliability by filtering the predicted ratings with lower reliability probability, which can further improve the recommendation accuracy and user satisfaction. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 is the system process schematic diagram of the present application;

[0033] Figure 2 is the system framework diagram of the present application;

[0034] Figure 3 is the HR@5 value and NDCG@5 value of the present application compared with other methods under four data sets. DETAILED DESCRIPTION

[0035] In order to make the purposes, technical solutions and advantages of the present application clearer, the following will describe in detail the recommendation system of the deep neural network with prediction reliability according to the present application with reference to the accompanying drawings. 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 by the claims of the present application.

[0036] Figure 1 is the process schematic diagram of the deep neural network recommendation system with prediction reliability of the present application. From Figure 1 It can be seen that the deep neural network recommendation system with prediction reliability proposed by the present application includes obtaining and cleaning data information from a database to obtain required information; dividing a user-item rating matrix R into corresponding several independent and only containing 0-1 binary sub-matrices according to different rating values set by the system; using a double tower model in the deep neural network to train the obtained sub-matrices in parallel to predict the probability value of the same blank area in each sub-matrix; normalizing the probability value obtained in each sub-matrix, and taking the maximum probability value after normalization as the reliability probability; finding the rating represented by the corresponding sub-matrix according to the obtained reliability probability, and taking the rating as the predicted rating of the blank area; comparing with the reliability threshold set by the system, filtering the predicted rating with a reliability probability lower than the threshold, and retaining the rating with a higher reliability probability; sorting the filtered predicted rating set, pushing the first k items with a higher predicted rating to the target user, and forming a personalized recommendation list for the target user.

[0037] Figure 2 is the framework diagram of the deep neural network recommendation system with prediction reliability of the present application. From Figure 2It can be seen that the deep neural network recommendation system with prediction reliability provided by the present application comprises a data acquisition and preparation module, a rating matrix division module, a rating probability calculation module, a reliability acquisition module, a rating value acquisition module, a reliability threshold filtering module and a recommendation result acquisition module. The data acquisition and preparation module is used for extracting data from a database and performing cleaning and preprocessing operations on the data to obtain the required information; the rating matrix division module divides the user-item rating matrix R into corresponding several independent and only containing 0-1 binary sub-matrices according to different rating values set by the system, for example, if the rating interval in the system is 1-5 points, R is divided into 5 sub-matrices, and each sub-matrix represents a rating; the rating probability calculation module uses the double tower model in the deep neural network to train the obtained sub-matrices in parallel to predict the probability value of the same blank area in each sub-matrix; the reliability acquisition module is used for normalizing the probability value obtained in each sub-matrix, and taking the maximum probability value after normalization as the reliability probability; the rating value acquisition module is used for finding the rating represented by the corresponding sub-matrix according to the obtained reliability probability, and taking the rating as the predicted rating of the blank area; the reliability threshold filtering module is used for comparing with the reliability threshold set by the system to filter the predicted rating with a reliability probability lower than the threshold, and retaining the rating with a higher reliability probability; the recommendation result acquisition module is used for acquiring the rating value acquisition module and retaining the predicted rating with a higher reliability probability by the reliability threshold filtering module, internally sorting the predicted ratings, pushing the first k items with higher predicted ratings to the target user, and forming a personalized recommendation list for the target user.

[0038] Further, the following case is taken as an example for illustration:

[0039] There are M users U = {u1, u2,..., uM} and N items I = {i1, i2,..., iN} in the MovieLens dataset, the rating domain is {1, 2, 3, 4, 5}, and the user-item rating matrix is represented as [R]M×N. M N rui represents the rating value of the user on the item i, U(i,j) represents the user set that has rated both the item i and the item j, i.e., the co-rated item; U(i) represents the user set that has rated the item i. M×N ui The rating matrix division module divides the user-item rating matrix R into corresponding several independent and only containing 0-1 binary sub-matrices according to different rating values set by the system, for example, if the rating interval in the system is 1-5 points, R is divided into 5 sub-matrices, and each sub-matrix represents a rating; the rating probability calculation module uses the double tower model in the deep neural network to train the obtained sub-matrices in parallel to predict the probability value of the same blank area in each sub-matrix; the reliability acquisition module is used for normalizing the probability value obtained in each sub-matrix, and taking the maximum probability value after normalization as the reliability probability; the rating value acquisition module is used for finding the rating represented by the corresponding sub-matrix according to the obtained reliability probability, and taking the rating as the predicted rating of the blank area; the reliability threshold filtering module is used for comparing with the reliability threshold set by the system to filter the predicted rating with a reliability probability lower than the threshold, and retaining the rating with a higher reliability probability; the recommendation result acquisition module is used for acquiring the rating value acquisition module and retaining the predicted rating with a higher reliability probability by the reliability threshold filtering module, internally sorting the predicted ratings, pushing the first k items with higher predicted ratings to the target user, and forming a personalized recommendation list for the target user. i

[0040] Firstly, the data acquisition and preparation module is used for acquiring and cleaning the system data to obtain the variables in the hypothesis. The specific implementation steps of the whole deep neural network recommendation system with prediction reliability are as follows:

[0041] ​​​S1: The data collection and preparation module obtains the required data information from the database, including user ID, project ID and rating information;

[0042] S2: The rating matrix division module divides the user-project rating matrix R into corresponding several independent and only contains 0-1 binary sub-matrices according to the different rating values set by the system;

[0043] Taking the MovieLens dataset as an example, the rating domain is {1, 2, 3, 4, 5}, and the rating matrix R can be divided into five sub-matrices corresponding to the rating values. The setting rules of the sub-matrices are as follows: if the user rating is the same as the rating represented by the sub-matrix, the value of the corresponding position in the sub-matrix is set to 1; if it is different, it is set to 0; otherwise, it is empty, that is, the position is the blank area to be predicted by the system.

[0044] S3: The rating probability calculation module uses the two-tower model in the deep neural network to train the obtained sub-matrices in parallel to predict the probability value of the same blank area in each sub-matrix;

[0045] The calculation formula of each layer of the two-tower model is as follows:

[0046] Input layer: input layer, input the feature embedding vector z of user u (project i) u (z i );

[0047] Hidden layer: use the activation function ReLU to prevent model overfitting, and obtain the output l of all hidden layers;

[0048] ReLU(z u )=max(0,z u )

[0049] l1=W1 T z u

[0050] l n =ReLU(W1 T l n-1 +b n ),n=2,...,X-1

[0051] Where W n and b n represent the weight matrix and bias vector of the nth hidden layer, respectively.

[0052] Output layer: obtain the latent feature vector p of user u (project i) u (q i ):

[0053]

[0054]

[0055] Then, the predicted rating probability of user u on item i is calculated as

[0056]

[0057] where W out and e denote the weight matrix and element-wise product of vectors respectively; the sigmoid function σ(x) = 1 / (1+e -x ) is used to limit the output within the range (0, 1).

[0058] S4: Reliability acquisition module, for normalizing the predicted probability obtained in the rating r represented by each sub-matrix and taking the normalized maximum probability value as the reliability probability of the region, the formula is as follows:

[0059]

[0060] S5: Rating value acquisition module, for finding the rating r represented by the corresponding sub-matrix according to the obtained reliability probability, and taking the rating as the predicted rating p of the blank region ui .

[0061] S6: Reliability threshold filtering module, for comparing the reliability probability corresponding to the predicted rating with the reliability threshold θ set by the system, if the reliability probability is less than the threshold, it is considered that the predicted rating is unreliable, and filtering is performed, and vice versa.

[0062] S7: Recommendation result acquisition module, for acquiring the recommendation list of target user u. Which includes: sorting the retained predicted ratings with higher reliability probability internally from large to small, setting the number of items to be recommended k, selecting the top k items with higher predicted rating to push to the target user u, to form the personalized recommendation list of the target user u.

[0063] Figure 3 The proposed method is measured in four data and two aspects of indicators, and compared with other methods. Two measurements are: hit rate (HR) and normalized discounted cumulative gain (NDCG). Their calculation methods and measurement contents are as follows:

[0064] Hit rate (HR): used to measure whether the actual recommended items in the test set are present on the top-k predicted recommendation list, which indicates the item recommendation ability of the model, the calculation method is as follows:

[0065]

[0066] where m denotes the number of users in the recommendation system, and hits(i) denotes the proportion of the top k predicted recommendation items of the i-th user in the actual recommendation item set.

[0067] Normalized Discounted Cumulative Gain (NDCG): It illustrates the position of clicks by assigning higher scores to the top-ranked clicks, which shows the item ranking recommendation quality of the model, and the calculation formula is as follows:

[0068]

[0069]

[0070]

[0071] where DCG@k and IDCG@k denote the discounted cumulative gain and ideal DCG, respectively, rel p The recommendation relevance of the item at position p is displayed, i.e., rel p = 1 if the predicted recommendation item appears in the actual recommendation list, otherwise 0. It is worth noting that we assume that the item is recommended if its predicted or actual rating exceeds the median of the rating interval.

Claims

1. A deep neural network recommendation system with predicted reliability to help the recommendation system filter out unreliable predicted ratings, comprising the following: a data collection and preparation module for obtaining and cleaning the data collected in the database to obtain the required information; a rating matrix division module for dividing the user-item rating matrix R into several independent and only containing 0-1 binary sub-matrices according to the different rating values set by the system, the setting rules of the sub-matrices are as follows: if the user rating is the same as the rating represented by the sub-matrix, the value of the corresponding position in the sub-matrix is set to 1; if not, it is set to 0; otherwise, it is empty, i.e. the position is a blank area to be predicted by the system; a rating probability calculation module for using the dual tower model in the deep neural network to train the obtained sub-matrices in parallel to predict the probability value of the same blank area in each sub-matrix; a reliability acquisition module for normalizing the probability value obtained in each sub-matrix, and taking the maximum normalized probability value as the reliability probability; a rating value acquisition module for finding the rating represented by the corresponding sub-matrix according to the obtained reliability probability, and taking the rating as the predicted rating of the blank area; a reliability threshold filtering module for filtering the predicted ratings with reliability probability lower than the threshold value by comparing with the reliability threshold set by the system, and retaining the ratings with high reliability probability; a recommendation result acquisition module for internally sorting the retained predicted ratings with high reliability probability, pushing the top k items with high predicted ratings to the target user to form the personalized recommendation list of the target user.

2. The deep neural network recommendation system with predictive reliability of claim 1, wherein, The rating probability calculation module is used to use the dual tower model in the deep neural network to train the obtained sub-matrices in parallel to predict the probability value of the same blank area in each sub-matrix, specifically comprising the following steps: ① input layer, input the feature embedding vector z of the user u u ; ②Hidden layer, using the activation function ReLU to prevent model overfitting, get the output of all hidden layers l n : ReLU(z u ) = max(0, z u ) l1 = W1 T z u where W n and b n denote the weight matrix and bias vector of the nth hidden layer, respectively. ③ output layer, obtain the potential feature vector p of the user u in the Xth layer u : Similarly, according to steps 1-3, we obtain the latent feature vector q of item i at level X i ; (4) Calculate the predicted rating probability of user u on item i: where W out and e denote the weight matrix and element-wise product of vectors, respectively; the sigmoid function σ(x) = 1 / (1 + e -x ) is used to limit the output to the range (0, 1).

3. The deep neural network recommendation system with predictive reliability of claim 1, wherein, The reliability obtaining module is configured to normalize the obtained prediction probability in the rating r represented by each sub-matrix and the normalized maximum probability value is taken as the reliability probability of the region, and the formula is as follows:

4. The deep neural network recommendation system with predictive reliability of claim 1, wherein, The rating value obtaining module is configured to find a rating r represented by the corresponding sub-matrix according to the obtained reliability probability, and take the rating as the predicted rating p of the blank area ui .

5. The deep neural network recommendation system with predictive reliability of claim 1, wherein, The reliability threshold filtering module is used to compare the reliability probability corresponding to the predicted rating with the reliability threshold θ set by the system, if the reliability probability is less than the threshold value, it is considered that the predicted rating is unreliable, and is filtered, otherwise.

6. The deep neural network recommendation system with predictive reliability of claim 1, wherein, The recommendation result acquisition module is used to obtain the recommendation list of the target user u, which includes: internally sorting the retained predicted ratings with high reliability probability from large to small, setting the number of items to be recommended k, selecting the top k items with high predicted ratings to push to the target user u to form the personalized recommendation list of the target user u.

Images