Multi-source data fusion commodity recommendation method based on attribute contribution degree mask

Abstract

The application discloses a multi-source data fusion commodity recommendation method based on attribute contribution degree mask, which comprises the following steps: 1) collecting comments on commodities through a network, including pictures, texts and videos; 2) training a self-encoder; 3) extracting dimension-reduced data; 4) fusing the data; 5) training a deep model; 6) calculating weights by using a mask method; 7) obtaining the weights of each dimension data, and fine-tuning the weight of each dimension in each data; 8) retraining a scoring prediction model to obtain a final scoring prediction model; 9) predicting in the mode of a knowledge graph, and predicting new product data by using old product data; and 10) performing scoring prediction on new and old products, and recommending according to the scoring result. The application improves the accuracy of the system and the credibility of the system output.

Description

Technical Field

[0001] This invention relates to the field of machine learning, and more specifically to a product recommendation method based on multi-source data fusion. Background Technology

[0002] Traditional recommendation methods suffer from the drawback of lagging information collection and processing, making them increasingly unable to meet the needs of the ever-changing market. Furthermore, traditional methods rely heavily on manual processing, which is inefficient and requires a great deal of subjective judgment based on the experience of system personnel, making it difficult to effectively develop recommendation systems. Therefore, exploring new recommendation methods is imperative.

[0003] In the field of recommender systems, people have begun to explore the possibility of combining deep learning with recommendation methods. Currently, a large number of recommender systems have emerged, such as the general-purpose UserCF and ItemCF systems. After training, although they can predict users' possible ratings for beverages based on big data, they also contain several problems, the main three of which are as follows:

[0004] First, there is the problem of multi-source data fusion: existing recommendation algorithms mainly process some structured data, recording relevant information of users and items, without considering the impact of multi-source data on recommendation results. For example, combining structured attribute information data sources with beverage image data sources, while we often need to collect a large amount of unstructured data, such as images and videos, when collecting information. Therefore, models that cannot perform multi-source fusion often miss a lot of potential information.

[0005] Secondly, there's the issue of the contribution of attribute information: existing recommendation systems assign equal weight to product attributes, ignoring the impact of different attribute information on the recommendation results. However, in reality, different attribute information contributes differently to the recommendation results, necessitating further filtering to identify important key attribute information, thereby improving the effectiveness of the recommendation method.

[0006] Thirdly, there's the cold start problem: The beverage market is incredibly diverse. Besides everyday drinks like mineral water and cola, each new niche beverage (such as recently launched tea drinks) has a short lifespan and is rapidly updated. Therefore, when a new beverage launches or new users join, the lack of historical user ratings or preference data makes effective recommendations difficult.

[0007] Current recommendation systems cannot effectively solve all three problems mentioned above at the same time. Summary of the Invention

[0008] This invention aims to overcome the aforementioned shortcomings of existing technologies by providing a multi-source data fusion product recommendation method based on attribute contribution masks, thereby solving the above three problems and improving the effectiveness and accuracy of recommendation systems.

[0009] The multi-source data fusion product recommendation method based on attribute contribution mask of the present invention includes the following steps:

[0010] 1) Data collection: Collect reviews of beverages online, including pictures, text, and videos, and store them in dataset S0 after processing.

[0011] 2) Training the autoencoder: Since data collected from the internet is often multi-source data, including unstructured, semi-structured, and structured data, unstructured or semi-structured data needs to be structured. An autoencoder method is used to solve this problem. The autoencoder is first trained, and the specific method is as follows:

[0012] 2.1) An autoencoder is a self-supervised learning algorithm that reduces the dimensionality of data and trains a neural model by training a mapping f: x → x'. The specific method is as follows:

[0013] (I) Given an unlabeled dataset S0, store 70% of it into an empty dataset S. 01 In this dataset, 30% of the data is used as the training set, and the remaining 30% is stored as the validation set in an empty dataset S. 02 middle.

[0014] (II) Train subnetwork M using data from the training set. 11 Let's learn the mapping f1:x→z, where z is a data point with a dimension less than m, and store the obtained z into the dataset S. 13 middle.

[0015] (III) Using the obtained dataset S 13 The data z in the dataset is used as input to train the subnetwork M. 12 Learn the mapping f:z→x`, where x` is a data point with the same dimension as the original data x, and store the resulting x` in the dataset.

[0016] (IV) The two sub-networks are adjusted using backpropagation to minimize the distance metric, thereby obtaining the final autoencoder M1. The formula for calculating the distance metric is as follows:

[0017]

[0018] Where n represents the number of training data and L represents the distance metric.

[0019] 3) Extracting dimensionality-reduced data: Based on the autoencoder M1 trained in step 2), the multivariate data is fused. The data x to be fused in the existing dataset S0 is input into the autoencoder M1, and the intermediate data z is extracted. Since the intermediate data z of the autoencoder is equivalent to performing dimensionality reduction on the input data, it is equivalent to extracting the common features in the data.

[0020] 4) Data fusion: Add the data in S0 to its corresponding intermediate data z, and store the result in dataset S1 as data for training the model.

[0021] 5) Training the deep model: Train the deep model M2 using the data in dataset S1. Use backpropagation to train the model and minimize the loss function. The specific loss function is as follows:

[0022]

[0023] 6) Calculate weights: A preliminary product rating model M2 is obtained through step 5), but the weights of each dimension attribute are still the same. To improve accuracy, a masking method is used to solve this problem. The specific method is as follows:

[0024] 6.1) Input the data from S1 into the existing scoring model M2 to obtain the output result, i.e., the score f(x). Each data point corresponds to one output result f(x). i ).

[0025] 6.2) For each data score f(x) i Perform the differentiation, and the result g is obtained. ia That is, the data of dimension a in the i-th data for the output result f(x) i The degree of influence of ), where the specific formula for differentiation is as follows:

[0026]

[0027] Where i represents the i-th data, x ia The a-th dimension represents the i-th data.

[0028] 6.3) For each dimension of g ia The formula for calculating the mean is as follows:

[0029]

[0030] g ia This represents the result of taking the derivative of the i-th data point in dimension a, where n is the total number of data points.

[0031] 6.4) Solve for the mask w. The specific formula for solving for w is as follows:

[0032] w a =lg(|ga|) (5)

[0033] g a This represents the average influence level of dimension a.

[0034] 7) Adjusting weights: Obtain the weights for each dimension according to formula (5). Next, fine-tune the weights for each dimension in each data point. The specific formula for fine-tuning is as follows:

[0035] x ia =x ia *w a (6)

[0036] 8) Retrain the model: Retrain the rating prediction model M2 in the same way as step 5). The training result is the final rating prediction model M2.

[0037] 9) Predicting scores for new products: Due to the limited sample data for new products, steps 1) to 8) cannot be used for score prediction. Therefore, this invention employs a knowledge graph approach to predict scores, utilizing existing product data to predict new product data. The specific method is as follows:

[0038] 9.1) Collect relevant data on old users and old products and store them in dataset S2.

[0039] 9.2) Calculate the correlation coefficient ρ between existing users, between existing products, and between existing users and existing products. If the correlation coefficient is greater than 0.5, then establish a connection between them. The specific formula for calculating the correlation coefficient is as follows:

[0040]

[0041] 9.3) The depth-first search algorithm is used to solve for the relationship between users and products. The specific method is as follows:

[0042] (I) Set the initial point to the location of the old user.

[0043] (II) The relevance between users and products is calculated based on different paths. The value is the product of the relevance coefficients of all edges along the path. The specific formula is as follows:

[0044] ρ x,y =ρ x,y *ρ a*b (8)

[0045] Where ρ a*b This represents the correlation coefficient between two points on the path.

[0046] (III) Select the path with the highest correlation coefficient between a user and a product, and use the correlation coefficient corresponding to the path as the final correlation coefficient between the user and the product.

[0047] 9.4) Establish connections between new users, new products and the knowledge graph, calculate the correlation coefficient between new users and old users and old products, and a correlation coefficient greater than 0.5 is considered to be established.

[0048] 9.5) Calculate the ratings for new users, new products, and each node using the following formula:

[0049]

[0050] ρ i,j This represents the rating of the j-th returning user for the i-th returning product. x,i This represents the degree of similarity between a new product and the i-th old product as determined by a human evaluation, where n is the total amount of data.

[0051] 10) Formulate a recommendation scheme: The recommendation system of the present invention has now been successfully established, which can predict the ratings of new and old products and make recommendations based on the rating results.

[0052] The working principle of this invention (analysis of the reasons for the advantages of this invention):

[0053] 1. To address the problem of multi-source data fusion, this invention uses an autoencoder structure for dimensionality reduction, achieving a unified effect for data with different structures and sources.

[0054] 2. To address the issue of identical attribute weights, this invention employs gradient interpretation to achieve the effects of dimensional weight analysis and adjustment.

[0055] 3. To address the cold start problem, this invention employs a knowledge graph construction method, enabling the prediction of new product data using data from older products.

[0056] The advantages of this invention are:

[0057] 1. By using multi-source data fusion, the sources and structures of data are expanded, and core attributes are extracted through dimensionality reduction, which effectively improves the accuracy of the system.

[0058] 2. By adjusting the weights, the relationships between the various dimensions of the data were effectively adjusted, thereby improving the reliability of the system output.

[0059] 3. By using knowledge graphs, connections between old and new products are established, greatly expanding the scope of the system's application and improving the applicability of the recommendation system. Attached Figure Description

[0060] Figure 1 This is an example diagram of the self-encoder of the present invention.

[0061] Figure 2 This is a schematic diagram of the scoring process for the old product of this invention.

[0062] Figure 3 This is a schematic diagram of the new product scoring process of the present invention.

[0063] Specific Implementation Cases

[0064] The following is in conjunction with the appendix Figure 1-3 The technical solution of the present invention will be illustrated by examples.

[0065] The multi-source data fusion product recommendation method based on attribute contribution mask of the present invention includes the following steps:

[0066] 1) Data collection: Collect beverage review data about the beverage through forums, official websites, and review websites. After image processing, text processing and other operations, convert it into computer language and transform it into data containing information such as "user's past consumption records, product popularity, user's gender, age, education level, region, and season of purchase". This is the initial collected dataset S0.

[0067] 2) Training the autoencoder: Store 70% of the data from S0 into an empty dataset S. 01 S0 is used as the training set, and 30% of the data in S0 is stored in dataset S. 02 The test set is used in S. 01 The data in the code uses the backpropagation algorithm for the autoencoder M1. The specific structure of the autoencoder is as follows: Figure 1 Training is performed by minimizing the distance metric function, where the specific function of the distance metric is:

[0068]

[0069] 3) Extracting dimensionality-reduced data: Input all data in S0 into the autoencoder M1 to obtain the intermediate data z for each of them after dimensionality reduction.

[0070] 4) Data fusion: Add the data in S0 to its corresponding intermediate data z, and store the result in dataset S1 as data for training the model.

[0071] 5) Training the deep model: To train the rating prediction model M2, we use 70% of the data in S1 as the training set. We use the backpropagation algorithm to minimize the loss function, where the formula for the loss function is:

[0072]

[0073] 6) Calculate Dimension Weights: The weights for each dimension in the current model remain the same. The following steps adjust the weights for each dimension:

[0074] 6.1) Input all the data in S1 into the existing model M2, and the output is f(x) i This is the final product's score.

[0075] 6.2) Find f(x) i The derivative for each dimension of data is given by the following formula:

[0076]

[0077] 6.3) For each dimension, the derivative is averaged using the following formula:

[0078]

[0079] 6.4) Next, calculate the mask w. The calculation formula is:

[0080] w a =lg(|g a |) (5)

[0081] 7) Adjusting Weights: Having obtained the mask for the relevant data, i.e., the weights for each dimension, we will now adjust the weights for each dimension of data in S1. The calculation formula is as follows:

[0082] x ia =x ia *w a (6)

[0083] 8) Retrain the model: Retrain the model M2 using the same method as in step 5), and the result is the final product rating prediction model M2.

[0084] 9) Rating and predicting new products: Since new products lack relevant data, knowledge graphs are needed for newly developed products and new user access. The basic approach to establishing such a knowledge graph is as follows:

[0085] 9.1) Collect relevant data on old products and store them in dataset S2 as training data.

[0086] 9.2) The correlation coefficient between the data needs to be calculated. If the correlation coefficient is greater than 0.5, a relationship is established. The formula for calculating the correlation coefficient is:

[0087]

[0088] 9.3) Using a depth-first search algorithm, starting with a specific user, calculate the correlation coefficient between that user and each product, i.e., the possible ratings. The highest rating among the calculated results is taken as the user's rating for that product. The rating calculation formula is as follows:

[0089] ρ x,y =ρ x,y *ρ a*b (8) 9.4) Add the new product to the knowledge graph and calculate the correlation coefficient as in 9.2). If the correlation coefficient is greater than 0.5, a connection is established.

[0090] 9.5) The rating of a new product can be calculated using user ratings of the product. The formula is as follows:

[0091]

[0092] 10) Develop a Recommendation Plan: Having completed the above steps to predict ratings for both existing and new products, the recommendation system is now successfully established. Next, since we have obtained the possible ratings for the products, we can proceed with brand positioning. Furthermore, through attribute weighting, not only are the ratings more accurate, but we can also identify the most important dimensions among each attribute. For example, experimental results show that the key indicators influencing the sales of a certain tea beverage are "packaging, taste, and age." Based on these results, we can optimize the product, scientifically improving market satisfaction through product optimization, thus effectively creating a market recommendation system.

Claims

1. A multi-source data fusion product recommendation method based on attribute contribution masks, comprising the following steps: 1) Data collection: Collect reviews about beverages online, including pictures, text, and videos, and store them in dataset S0 after processing; 2) Training the autoencoder: Since data collected from the internet is often multi-source data, including unstructured, semi-structured, and structured data, unstructured or semi-structured data needs to be structured. An autoencoder method is used to solve this problem. The autoencoder is first trained, and the specific method is as follows: 2.1) An autoencoder is a self-supervised learning algorithm that reduces the dimensionality of data and trains a neural model by training a mapping f: x → x'. The specific method is as follows: (I) Given an unlabeled dataset S0, store 70% of it into an empty dataset S. 01 In this dataset, 30% of the data is used as the training set, and the remaining 30% is stored as the validation set in an empty dataset S. 02 middle; (II) Train subnetwork M using data from the training set. 11 Let's learn the mapping f1:x→z, where z is a data point with a dimension less than m, and store the obtained z into the dataset S. 13 middle; (III) Using the obtained dataset S 13 The data z in the dataset is used as input to train the subnetwork M. 12 Learn the mapping f:z→x`, where x` is a data point with the same dimension as the original data x, and store the resulting x` in the dataset. (IV) The two sub-networks are adjusted using backpropagation to minimize the distance metric, thereby obtaining the final autoencoder M1. The formula for calculating the distance metric is as follows: Where n represents the number of training data and L represents the distance metric; 3) Extracting dimensionality-reduced data: Use the autoencoder M1 trained in step 2) to fuse multivariate data. Input the data x to be fused from the existing dataset S0 into the autoencoder M1, and extract the intermediate data z. Since the intermediate data z of the autoencoder is equivalent to performing dimensionality reduction on the input data, it is equivalent to extracting the common features in the data. 4) Data fusion: Add the data in S0 to its corresponding intermediate data z, and store the result in dataset S1 as data for training the model; 5) Training the deep model: Train the deep model M2 using the data in dataset S1. Use backpropagation to train the model and minimize the loss function. The specific loss function is as follows: 6) Calculate weights: A preliminary product rating model M2 is obtained through step 5), but the weights of each dimension attribute are still the same. To improve accuracy, a masking method is used to solve this problem. The specific method is as follows: 6.1) Input the data from S1 into the existing scoring model M2 to obtain the output result, i.e., the score f(x). Each data point corresponds to one output result f(x). i ); 6.2) Score each data point f(x) i Perform the differentiation, and the result g is obtained. ia That is, the data in dimension a of the i-th data with respect to the output result f(x) i The degree of influence of ), where the specific formula for differentiation is as follows: Where i represents the i-th data, x ia The a-th dimension represents the i-th data point; 6.3) For each dimension of g ia The formula for calculating the mean is as follows: g ia This represents the result of taking the derivative of the i-th data point in the a-th dimension, where n is the total number of data points. 6.4) Solve for the mask w. The specific formula for solving for w is as follows: w a =lg(|g a |) (5) g a This represents the average degree of influence of dimension a; 7) Adjusting weights: Obtain the weight of each dimension according to formula (5), and then fine-tune the weight of each dimension in each data point. The specific formula for fine-tuning is as follows: x ia =x ia *w a (6) 8) Retrain the model: Retrain the rating prediction model M2 in the same way as step 5). The training result is the final rating prediction model M2. 9) Predicting scores for new products: Due to the limited sample data for new products, steps 1) through 8) cannot be used for score prediction. Instead, a knowledge graph approach is employed to predict new product data using existing product data. The specific method is as follows: 9.1) Collect relevant data on existing users and products, and store them in dataset S2; 9.2) Calculate the correlation coefficient ρ between existing users, between existing products, and between existing users and existing products. If the correlation coefficient is greater than 0.5, then establish a connection between them. The specific formula for calculating the correlation coefficient is as follows: 9.3) The depth-first search algorithm is used to solve the relationship between users and products. The specific method is as follows: (I) Set the initial point to the location of the old user; (II) Calculate the correlation coefficient between users and products based on different paths. The value is the product of the correlation coefficients of all edges along the path. The specific formula is as follows: r x,j =ρ x,i *r i,j (8) Where ρ i,j The correlation coefficient between two points in the path; (III) Select the path with the highest correlation coefficient between a user and a product, and use the correlation coefficient corresponding to the path as the final correlation coefficient between the user and the product; 9.4) Establish connections between new users, new products and the knowledge graph, calculate the correlation coefficient between new users and old users and old products, and a correlation coefficient greater than 0.5 is considered to be established; 9.5) Calculate the ratings for new users, new products, and each node using the following formula: ρ i,j This represents the rating of the j-th returning user for the i-th returning product. x,i This represents the degree of similarity between a new product and the i-th old product as determined by a human evaluation, where n is the total amount of data. 10) Develop a recommendation scheme: The recommendation system has now been successfully established and can predict the ratings of new and old products and make recommendations based on the rating results.

Images