A precise service recommendation method based on analysis of economic operation uncertainty factors

Abstract

The present invention provides a precise service recommendation method based on the analysis of economic operation uncertainty factors, comprising the following steps: Step 1: raw data processing. The data is preprocessed, including missing value processing, outlier processing and standardization, and then effective features are extracted from the raw data and feature embedding is performed. Step 2: Initialize the model. According to the data features, the recommendation system model is initialized. Step 3: Train the model. The processed samples in step 1 are input into the model in step 2, and a suitable loss function is selected to train the model. Step 4: Collect difficult negative samples. Use the LIME algorithm to obtain the attribution value of each feature, and collect difficult negative samples according to the attribution value. Step 5: Train the model again and output the final model. Add difficult negative samples to the negative samples, train the model again, and improve the robustness of the model.

Description

Technical Field

[0001] The present invention relates to the recall field of recommendation systems, and in particular to an accurate service recommendation method based on analysis of economic operation uncertainty factors. Background Art

[0002] The "stay-at-home economy" and "lazy economy" have jointly promoted the development of online consumption. According to a report released by iResearch, the frequency of online consumption by users has increased significantly compared with before and after the epidemic. For example, the online consumption of fresh fruits and food and beverage categories with the highest user growth has increased by 27.6% and 17.3% respectively after the epidemic; during the 618 period, the overall transaction volume of JD.com's home appliance stores increased by 240% year-on-year.

[0003] On the one hand, the online market is expanding rapidly, and each e-commerce platform needs to provide users with accurate service recommendations and recall a batch of products that users are interested in from a large number of candidate products to facilitate users' online shopping. On the other hand, e-commerce recommendation systems should be robust. Recommendation systems capture users' consumption preferences based on historical data such as user portraits and user consumption records, so as to match users with suitable products. Existing recommendation systems overemphasize the accuracy of historical consumption data, believing that the higher the accuracy of historical data, the more likely it is to predict users' future consumption behavior. However, this may cause the recommendation system to rely too much on individual features learned from historical data, lacking robustness to cope with the impact of unstable factors that may appear in the future. In addition, in an unstable environment, user interests may change at any time, and the model should be sensitive to changes in user interests.

[0004] A robust recommendation system should not rely on individual features in the data. It should improve the ability to distinguish as much as possible, especially for those difficult-to-distinguish positive and negative samples. Usually, the negative samples in the recall phase of the recommendation system are randomly sampled. Most of the negative samples are simple negative samples, that is, the negative samples are very different from the positive samples. The recommendation system can easily make judgments based on a small number of features. In order to improve the robustness of the model, it is necessary to add difficult negative samples. These negative samples are closer to the positive samples, and the model can learn more information. Summary of the invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to improve the problems existing in the existing e-commerce recommendation system, improve the robustness of the recommendation system, and reduce the impact of economic instability factors. The present invention specifically provides an accurate service recommendation method based on the analysis of economic operation uncertainty factors, including the following steps:

[0006] Step 1: preprocess the original data to obtain data features;

[0007] Step 2: Initialize the recommendation system model based on data characteristics;

[0008] Step 3: Input the samples processed in step 1 into the recommendation system model in step 2, and train the recommendation system model;

[0009] Step 4, collect difficult negative samples;

[0010] Step 5: train the recommendation system model again, output the final recommendation system model, and use the final recommendation system model to obtain the recommendation results.

[0011] In step 1, the original data includes data describing the user, data describing the object, and interaction data between the user and the object.

[0012] In step 1, the original data is processed for missing values, outliers, and standardization. Finally, prior knowledge is used to select effective features for model training. One-hot encoding is used for feature embedding of discrete features, while numerical features remain unchanged.

[0013] In step 2, the recommendation system model includes a linear model and a deep neural network, and the linear model is expressed as:

[0014] y=w T x+b

[0015] Where x = [x1x2…x d ] T is the input sample, x i is the i-th sample feature, i ranges from 1 to d; weight w = [w1w2…w d ] T and bias b are the parameters that the linear model needs to learn; y is the output of the linear model;

[0016] The deep neural network includes a feature combination layer and a feature interaction layer. The feature combination layer is expressed as:

[0017]

[0018] Among them, V x is the set of all feature embedding vectors; x i represents the i-th feature, v i is the embedding vector of the ith feature, and there are n features in total; ⊙ represents the element-wise product vector obtained by multiplying the corresponding dimensions of two vectors. The time complexity is reduced by the following formula:

[0019]

[0020] In step 2, the feature interaction layer includes more than two hidden layers, and the first hidden layer is represented as follows:

[0021] h (1) =f(W (1) f(V x )+b (1) )+f(V x )

[0022] Among them, h (1) is the output of the first hidden layer; W (1) and b (1) are the weight and bias parameters of the first hidden layer; f(V x ) is the output of the feature combination layer; f(.) is the Relu activation function, which is expressed as follows:

[0023] f(x)=max(0,x)

[0024] The rest of the feature interaction layers are expressed as follows:

[0025] h (l+1) =f(W (l) h (l) +b (l) )+h (l)

[0026] Among them, h (l) is the output of the lth layer of the deep neural network; W (l) and b (l) are the weight and bias parameters of the lth layer respectively; f(.) is the Relu activation function.

[0027] In step 2, the output P(Y=1|x) of the recommendation system model is:

[0028]

[0029] Among them, Y represents the data label. In the click prediction task, Y = 1 means that the user clicks on the item. is the value of the last layer of neurons in the deep neural network, is the weight parameter of the last layer of neurons, σ(.) is the sigmoid function, expressed as follows:

[0030]

[0031] Among them, e is the natural base, and x is the input value of the last layer of neurons.

[0032] Step 3 includes: defining the probability p that the user's match with the positive sample is higher than the user's match with the negative sample:

[0033] p=σ(P(u,i + )-P(u,i - ))

[0034] Among them, σ is the sigmoid function; u represents the user, i + and i - Represent the items in the positive sample and the items in the negative sample respectively; σ(.) is the sigmoid function; P(.) is the output of the recommendation system model;

[0035] Combining the probability p with the cross entropy loss function, we get the Bayesian Personalized Ranking (BPR) loss function L BPR for:

[0036] L BPR =-logp

[0037] A negative sample is randomly sampled for each positive sample, and the recommendation system model is trained using the BPR loss function.

[0038] In step 4, the model-independent local interpretable (Local Interpretable Model-AgnosticExplanation, LIME) algorithm is used to collect difficult negative samples:

[0039] Let g∈G be an interpretable model, G represents a set of interpretable models, define Ω() as model complexity, the model complexity of the linear model is the non-zero weight, and the model complexity of the decision tree model is the depth of the tree; define the neighborhood of sample x as π x ,L(f,g,π x ) means in π x The error between the output value of an interpretable model g and the recommendation system model f on the sample in the defined neighborhood. The goal of the LIME algorithm is to find the error between the output value of an interpretable model g and the recommendation system model f on the sample in the neighborhood π x We can get an interpretable model ξ(x) that is closest to the recommendation system model, which is expressed as follows:

[0040] ξ(x)=arg min g∈G L(f,g,π x )+Ω(g).

[0041] Step 4 also includes the following steps:

[0042] Step 4-1, define the neighborhood π of the sample x to be explained x , in π x Samples close to x are sampled;

[0043] Step 4-2, use the recommendation system model to make predictions on samples close to x;

[0044] Step 4-3: Use samples similar to x to train an interpretable model, obtain the attribution value of each feature of sample x, and then take the average to obtain the feature x with the largest attribution value. max , let its attribute value be φ max ; For each positive sample i of user u + With a certain probability p sample , randomly sample a feature x max The sample is regarded as a difficult negative sample with probability p sample for:

[0045]

[0046] Where n is the number of features of the item; φ i represents the average attribution value of each feature on all positive samples of user u; α is a hyperparameter.

[0047] In step 5, the recommendation system model is trained again using the collected difficult negative samples, and finally a robust recommendation system model is obtained.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) The recommendation system of the present invention has high accuracy. The recommendation model consists of a memory part and a generalization part. The memory part learns the first-order interaction of features and can well remember the high-frequency patterns that appear in the historical interaction data; the generalization part displays the high-order interaction between the learning features and taps into the user's interests.

[0050] (2) The recommendation system of the present invention is highly robust. The method in game theory is used to collect difficult negative samples and retrain the model to improve the robustness of the model and reduce the disturbance caused by uncertain factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0052] Figure 1 It is a flow chart of the steps of the present invention.

[0053] Figure 2 It is the basic framework diagram of the recommendation system model of the present invention.

[0054] Figure 3 This is a diagram of the data format of a dataset published by an e-commerce platform.

[0055] Figure 4 This is a schematic diagram showing the results after the attribution values ​​of the model output are normalized by the four features of the item. DETAILED DESCRIPTION

[0056] The present invention provides a precise service recommendation method based on economic operation uncertainty factor analysis, comprising the following steps:

[0057] Step 1: Raw data processing: Preprocess the data, including missing value processing, outlier processing and standardization, then extract effective features from the raw data and perform feature embedding.

[0058] Step 2: Initialize the model. Initialize the recommendation system model based on the data characteristics.

[0059] Step 3: Train the model. Input the samples processed in step 1 into the model in step 2 and select a suitable loss function to train the model.

[0060] Step 4: Collect difficult negative samples. Use the LIME algorithm to obtain the attribution value of each feature, and collect difficult negative samples based on the attribution value.

[0061] Step 5: Train the model again and output the final model. Add difficult negative samples to the negative samples and train the model again to improve the robustness of the model.

[0062] In the present invention, first of all, there are data describing users, data describing items, and data on the interaction between users and items. The data describing users include the user's gender, age, city, etc., the data describing items include the item's category, price, label, etc., and the data on the interaction between users and items include the user's interactive behavior data on the item, such as clicking, rating, or purchasing.

[0063] Next, the original data is processed for missing values, outliers, and standardization. Missing values ​​in the data may be caused by objective storage failures, subjective user concealment, and other reasons. The processing principle is to use the most likely value to fill the missing values. You can use the same mean interpolation method, use the clustering method to predict the category to which the missing value belongs, and then fill it with the mean of the category; you can also use the approximate substitution method, use the Euclidean distance to determine the closest samples to the sample containing the missing value, and perform weighted averaging on these samples to fill the missing values. Outliers in the data refer to data that deviates from most of the data. Specifically, they can be set as data that deviates from the mean value in the data by more than three times the standard deviation. These data can be deleted directly or corrected using the method in the missing value processing. Data standardization is to convert numerical data, such as the age of the user, the price of the item, etc., into data with a mean of 0 and a variance of 1 using the z-score standardization method. The formula is:

[0064]

[0065] Among them, a represents new data, b represents original data, c represents the average value of original data, and d represents the standard deviation of original data;

[0066] Finally, we can use prior knowledge to select effective features for model training, use one-hot encoding for discrete features, and keep numerical features unchanged.

[0067] In the present invention, the recommendation system model used is divided into two parts, which realize the memory function and the generalization function respectively. The memory part of the model is used to remember the common high-frequency patterns in historical data. It is a linear model that learns the first-order effects of features and is expressed as:

[0068] y=w T x+b

[0069] Where x = [x1x2…x d ] T is the input sample, x i is the sample feature; weight w=[w1w2…w d ] T and bias b are the parameters that the linear model needs to learn; y is the output of the linear model.

[0070] The generalization part of the model is used to mine user interests from historical data. It is a deep neural network that learns high-order interactions of features. It is divided into a feature combination layer and a feature interaction layer. The feature combination layer combines features in pairs. The feature interaction layer is a multi-layer perceptron that takes the output of the feature combination layer as input. This layer automatically learns high-order interactions between features. The feature combination layer is represented as:

[0071]

[0072] Among them, V x is the set of all feature embedding vectors; x i represents the i-th feature, v i is the embedding vector of the feature, there are n features in total; ⊙ represents the element product vector obtained by multiplying the corresponding dimensions of two vectors. Obviously, the time complexity of the above formula is O(n 2 ), can be converted into the following formula, reducing the time complexity to O(n):

[0073]

[0074] The feature interaction layer contains multiple hidden layers. The first hidden layer is represented as follows:

[0075] h (1) =f(W (1) f(V x )+b (1) )+f(V x )

[0076] Among them, h(1) is the output of the first hidden layer; W (1) and b (1) are the weight and bias parameters of the first hidden layer; f(V x ) is the output of the feature combination layer; f(.) is the Relu activation function, which is expressed as follows:

[0077] f(x)=max(0,x)

[0078] The rest of the feature interaction layers are expressed as follows:

[0079] h (l+1) =f(W (k) h (k) +b (k) )+h (l)

[0080] Among them, h (l) is the output of the lth layer of the neural network; W (l) and b (l) are the weight and bias parameters of the lth layer respectively; f(.) is the Relu activation function. A residual network is used here, that is, each hidden layer learns h (l+1) -h (l) This solves the problem of worsening accuracy as the network is deepened.

[0081] Finally, the output of the recommendation system model is the logistic regression of these two parts:

[0082]

[0083] The above formula is the logarithmic probability that the model predicts that the label of sample x is Y=1, which can be regarded as the model's predicted value for sample x. T and b are the parameters of the memory part model, is the value of the last layer of neurons in the generalized model, is the weight parameter of the neurons in this layer, and σ(.) is the sigmoid function, which is expressed as follows:

[0084]

[0085] At this point, the model has been initialized. The structural diagram of the model is as follows: Figure 3 shown.

[0086] The recommendation system is mainly divided into the recall stage and the ranking stage. In the recall stage, about hundreds of items that users are interested in are selected from all items. In the ranking stage, these items are ranked again, and the items ranked in the top dozens are exposed to users. The positive samples in the recommendation system are the items that users have clicked and purchased. The ranking stage of the recommendation system often uses the samples that have been exposed but not clicked as negative samples. However, in the recall stage of the recommendation system, it is impossible for all thousands of candidate items to be exposed to users, and then select the samples that users have not clicked. Therefore, in the recall stage, negative samples are usually randomly sampled, which cannot guarantee that the negative samples obtained by random sampling are all true negative samples. The recommendation system usually completes a binary classification problem. The model predicts the items that users are interested in as 1 and the items that users are not interested in as 0. The cross entropy loss function, which is commonly used in binary classification problems, calculates the absolute accuracy of the sample, and requires the negative sample to be a true negative sample. Therefore, the cross entropy loss function is not used here, and the BPR loss function is used. This loss function models the <user, positive sample, negative sample> triple, and its goal is to ensure that the matching between the user and the positive sample is higher than the matching between the user and the negative sample. First, define the probability that the user matches the positive sample better than the negative sample:

[0087] p=σ(P(u,i + )-P(u,i - ))

[0088] Among them, u represents the user, i + and i - Represent the items in the positive sample and the items in the negative sample respectively; σ(.) is the sigmoid function; P(.) is the model output. Then combine this probability with the cross entropy loss function to get the BPR loss function:

[0089] L BPR =-logp

[0090] A negative sample is randomly sampled for each positive sample, and the model is trained using the BPR loss function. However, the negative samples obtained by random sampling are usually simple negative samples. The difference between simple negative samples and the user's positive samples is often too large, and the model can easily distinguish them. Therefore, after the model is trained, in order to enhance the robustness of the model, it is necessary to use difficult negative samples to train the model again.

[0091] Difficult negative samples are samples that are close to positive samples. The purpose is to let the model learn more information as much as possible, so as to avoid the model relying on only individual features to complete classification, thereby improving the robustness of the model to unknown data and reducing the disturbance caused by uncertain factors. Here, the LIME algorithm based on the Shapley value method in game theory is used. This algorithm can obtain the contribution of each feature to the model output value, that is, the attribution value of each feature. If the attribution value of a feature of the positive sample is large, it means that the model mainly completes the classification based on this feature. However, under the influence of uncertain factors, the user's preferences will change. Using simple negative samples to train the model will cause the model to "lazily" complete the classification based on this feature, that is, samples with this feature have a relatively high probability of being judged as positive samples, while ignoring the information in other features. Therefore, negative sampling is performed on features with large attribution values ​​to reduce the model's dependence on this feature.

[0092] If we use the linear model y=w1x1+w2x2+…+w m x m +b makes predictions, for each feature x of sample x i , its attribution value is the corresponding coefficient w i . The way the LIME algorithm obtains feature attribution values ​​is to use interpretable models such as linear models and decision tree models to explain complex models such as neural networks. The core idea is that if you want to obtain the attribution values ​​of each feature of a sample x, you can focus on a neighborhood of sample x, use the samples in the neighborhood to train an interpretable model, and use the trained interpretable model to obtain the attribution values ​​of each feature of sample x. Let g∈G be an interpretable model, G represents a set of interpretable models, define Ω(g) as the model complexity, Ω(g) of the linear model as a non-zero weight, and Ω(g) of the decision tree model as the depth of the tree. Define the neighborhood of sample x as π x , L(f,g,π x ) means in π x The error between the output value of an interpretable model g and the recommendation system model f on the sample in the defined neighborhood. The goal of the LIME algorithm is to x We can get an interpretable model that is closest to the recommendation system model, which is expressed as follows:

[0093]

[0094] The steps of the LIME algorithm are:

[0095] 1. Define the neighborhood π of the sample x to be explained x , in π x Samples close to x are sampled;

[0096] 2. Use the recommendation system model to make predictions on these samples;

[0097] 3. Use these samples to train an interpretable model, such as a linear model, to obtain the attribution values ​​of each feature of the sample x.

[0098] Use the LIME algorithm to find all positive samples i of user u + The attribution values ​​of each feature are then averaged to obtain the feature x with the largest attribution value. max , let its attribute value be φ max For each positive sample i of user u + With a certain probability, randomly sample a feature x max The sample is regarded as a difficult negative sample, and the probability is:

[0099]

[0100] Where n is the number of features of the item; φ i represents the average attribution value of each feature on all positive samples of user u; α is a hyperparameter, which can be set to

[0101] The model is trained again using the collected difficult negative samples to obtain a robust recommendation system model.

[0102] Example

[0103] This embodiment discloses a precise service recommendation method based on economic operation uncertainty factor analysis. The method flow chart is as follows: Figure 1 As shown, the following steps are included:

[0104] Step 1: Raw data processing: Preprocess the data, including missing value processing, outlier processing and standardization, then extract effective features from the raw data and perform feature embedding.

[0105] Step 2: Initialize the model. Initialize the recommendation system model based on the data characteristics.

[0106] Step 3: Train the model. Input the samples processed in step 1 into the model in step 2 and select a suitable loss function to train the model.

[0107] Step 4: Collect difficult negative samples. Use the LIME algorithm to obtain the attribution value of each feature, and collect difficult negative samples based on the attribution value.

[0108] Step 5: Train the model again and output the final model. Add difficult negative samples to the negative samples and train the model again to improve the robustness of the model.

[0109] In step 1, user data, item data, and interaction data between users and items are first obtained from the server. User data includes user ID, age, gender, city, etc. Item data includes item ID, category, brand, price, etc. User-item interaction data includes whether the user clicks on the item, the user's rating of the item, etc. Then the data is preprocessed, including missing value processing, outlier processing, and standardization. Then, according to the recommendation target, appropriate features are extracted from the original data.

[0110] Here we use a data set published by an e-commerce platform, and its data format is Figure 3 As shown;

[0111] This dataset classifies the age of users; interaction time means that the user clicked the item at that moment; the items browsed by the user every half an hour are set to the same session id. Since all users are from the same country, the feature "country" is not used here; in order to avoid overfitting of the model, the "session id" is used here to consider the time factor of the user clicking the item, and the feature "interaction time" is not used.

[0112] Finally, the features are embedded. Continuous data can be directly input into the model after being standardized, and discrete data such as user ID, city, item ID, and category are input into the model after one-hot encoding.

[0113] Wherein, in step 2, the number of neurons in each layer of the model is set according to the dimension of feature embedding. The recommendation system model used in the present invention is divided into two parts, which respectively realize the memory function and the generalization function. The memory part of the model is used to remember the common high-frequency patterns in historical data, and it is a linear model; the generalization part of the model is used to mine the user's interests from historical data, and it is a deep neural network; the output of the recommendation system model is the logistic regression of the outputs of these two parts. The number of input neurons of the linear model and the deep model is the same, which is equal to the dimension of the embedding vector. After the model is built, its parameters are randomly initialized.

[0114] Among them, in step 3, if the current task is click prediction, the positive sample items are the items that the user has clicked, and the negative sample items are randomly sampled from all candidate items. First, the user feature vector and the positive sample feature vector are spliced ​​together and input into the model to obtain the output value p; then the user feature vector and the negative sample feature vector are spliced ​​together and input into the model to obtain the output value n. The model is trained using the BPR loss function, which models the <user, positive sample, negative sample> triple. Its goal is to ensure that the matching between the user and the positive sample is higher than the matching between the user and the negative sample, that is, the above output value p is greater than the output value n.

[0115] Among them, in step 4, the LIME algorithm based on the Shapley value method in game theory is used. This algorithm can obtain the contribution of each feature to the model output value, that is, the attribution value of each feature. If it is a click prediction task, suppose user u has clicked on m items, and the user feature vector is concatenated with the feature vectors of these m items to form m positive sample feature vectors, which are input into the recommendation system model to obtain m outputs. Here, a linear model is selected to explain the recommendation system model, and the linear model is trained with the above m positive sample feature vectors and m output values. Then the feature x with the largest attribution value is obtained. max , let its attribute value be φ max Then for each positive sample i of user u + With a certain probability, randomly sample a feature x max The samples are regarded as hard negative samples.

[0116] The items in this dataset have four features, namely item id, brand, store id, and category. Here, the LIME algorithm is used for the user with user id 2399402 to obtain the attribution values ​​of the four features of the item to the model output, and then the attribution values ​​are normalized and displayed as follows: Figure 4 As shown;

[0117] The store ID has the largest attribution value, which shows that the store ID has the greatest impact on the user's click behavior. In order to avoid overfitting of the model and improve the robustness of the model, The probability of randomly collecting difficult negative samples is , and the product IDs of these negative samples are the same as the product IDs in the positive samples of the user.

[0118] Among them, in step 5, the difficult negative samples collected in step 4 are used to train the model again according to the method in step 3. Finally, the model can predict new samples, and adding difficult negative sample training improves the click prediction rate of the model.

[0119] In some embodiments, if the task is to predict the user's rating of an item, the interaction data between the user and the item needs to include the user's rating of the item. Items with ratings exceeding a certain threshold can be set as positive sample items for the user, and items with ratings below a certain threshold can be set as negative sample items for the user.

[0120] It should be noted that in steps 3 and 5, the training model needs to set an appropriate learning rate. The learning rate determines the speed of model training and can be set to 0.01; an L2 regularization term can be added to the loss function to further improve the generalization ability of the model; and the stochastic gradient descent method can be selected to update the model parameters.

[0121] It should be noted that in step 4, if the number of items clicked by the user is too small to train an interpretable model using the LIME algorithm, the users can be classified, such as clustering them according to their characteristics, and then the interaction data of all users in the same category can be used to train an interpretable model.

[0122] The present invention provides an accurate service recommendation method based on the analysis of economic operation uncertainty factors. There are many methods and ways to implement the technical solution. The above is only a preferred implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A precise service recommendation method based on economic operation uncertainty factor analysis, characterized in that: The following steps are involved: Step 1: preprocess the original data to obtain data features; Step 2: Initialize the recommendation system model based on data characteristics; Step 3: Input the samples processed in step 1 into the recommendation system model in step 2, and train the recommendation system model; Step 4: Collect difficult negative samples; Step 5: train the recommendation system model again, output the final recommendation system model, and use the final recommendation system model to obtain the recommendation results; In step 1, the original data includes data describing the user, data describing the item, and interaction data between the user and the item; In step 1, the original data is processed for missing values, outliers, and standardization. Finally, prior knowledge is used to select effective features for model training. One-hot encoding is used for feature embedding of discrete features, while numerical features remain unchanged. In step 2, the recommendation system model includes a linear model and a deep neural network, and the linear model is expressed as: y=w T x+b Where x=[x1x2...x d ] T is the input sample, x i is the i-th sample feature, i ranges from 1 to d; weight w = [w1w2...w d ] T and bias b are the parameters that the linear model needs to learn; y is the output of the linear model; The deep neural network includes a feature combination layer and a feature interaction layer. The feature combination layer is expressed as: Among them, V x is the set of all feature embedding vectors; x i represents the i-th feature, v i is the embedding vector of the ith feature, and there are n features in total; ⊙ represents the element-wise product vector obtained by multiplying the corresponding dimensions of two vectors. The time complexity is reduced by the following formula: In step 2, the feature interaction layer includes more than two hidden layers, and the first hidden layer is represented as follows: h (1) =f(W (1) f(V x )+b (1) )+f(V x ) Among them, h (1) is the output of the first hidden layer; W (1) and b (1) are the weight and bias parameters of the first hidden layer; f(V x ) is the output of the feature combination layer; f(.) is the Relu activation function, which is expressed as follows: f(x) = max(0, x) The rest of the feature interaction layers are expressed as follows: h (l+1) =f(W (l) h (l) +b (l) )+h (l) Where h(l) is the output of the lth hidden layer of the deep neural network; W(l) and b(l) are the weight and bias parameters of the lth hidden layer respectively; f(.) is the Relu activation function; In step 2, the output P(Y=1|x) of the recommendation system model is: Among them, Y represents the data label. In the click prediction task, Y = 1 means that the user clicks on the item. is the value of the last layer of neurons in the deep neural network, is the weight parameter of the last layer of neurons, σ(.) is the sigmoid function, expressed as follows: Among them, e is the natural base, and x is the input value of the last layer of neurons; Step 3 includes: defining the probability p that the user's match with the positive sample is higher than the user's match with the negative sample: p=σ(P(u,i + )-P(u,i - )) Among them, σ is the sigmoid function; u represents the user, i + and i - Represent the items in the positive sample and the items in the negative sample respectively; σ(.) is the sigmoid function; P(.) is the output of the recommendation system model; Combining the probability p with the cross entropy loss function, we get the Bayesian personalized ranking BPR loss function L BPR for: L BPR =-log p Randomly sample a negative sample for each positive sample and use the BPR loss function to train the recommendation system model; In step 4, a model-independent local interpretable algorithm is used to collect difficult negative samples: Let g∈G be an interpretable model, G represents a set of interpretable models, define Ω(g) as the model complexity, the model complexity of the linear model is the non-zero weight, and the model complexity of the decision tree model is the depth of the tree; define the neighborhood of sample x as π x , L(f, g, π x ) means in π x The error between the output value of an interpretable model g and the recommendation system model f on the sample in the defined neighborhood. The goal of the LIME algorithm is to find the error between the output value of an interpretable model g and the recommendation system model f on the sample in the neighborhood π x We can get an interpretable model ξ(x) that is closest to the recommendation system model, which is expressed as follows: Step 4 also includes the following steps: Step 4-1, define the neighborhood π of the sample x to be explained x , in π x Samples close to x are sampled; Step 4-2, use the recommendation system model to make predictions on samples close to x; Step 4-3: Use samples similar to x to train an interpretable model, obtain the attribution value of each feature of sample x, and then take the average to obtain the feature x with the largest attribution value. max , let its attribute value be φ max ; For each positive sample i of user u + With a certain probability p sample , randomly sample a feature x max The sample is regarded as a difficult negative sample with probability p sample for: Where n is the number of features of the item; φ i represents the average attribution value of each feature on all positive samples of user u; α is a hyperparameter.

2. The method according to claim 1, characterized in that In step 5, the recommendation system model is trained again using the collected difficult negative samples, and finally a robust recommendation system model is obtained.

Images